循环神经网络用作加速度计或磁传感器噪声抑制

The accelerometer signal is capable of providing crucial information regarding object motion, posture, and vibration. Therefore, it is of great significance to investigate noise suppression techniques for accelerometer signals. However, in complex industrial environments and under various equipment conditions, the noise generated exhibits a diverse nature, leading to suboptimal noise reduction outcomes when employing conventional denoising algorithms. This paper proposes a neural network model (AgentUNet) for denoising accelerometer signals. The network consists of a combination of UNet and Agent Attention, enhancing its ability to learn contextual information and achieve superior denoising performance. Additionally, instead of using clean signals as labeled data, we train the neural network using pairs of noisy data, thereby alleviating the challenge of collecting clean signals in industrial environments. Finally, we conduct experiments on experiment dataset. The experimental results indicate that AgentUNet achieves superior denoising performance compared to other baseline models. Furthermore, the denoising effectiveness of the neural network trained without using clean signals as labeled data is similar to that of the network trained using supervised learning methods.

Accelerometer signals play a critical role in many fields, for example navigation and vehicle safety. However, uncontrollable factors such as defective equipment and harsh environments make the signals recorded by sensors contain a large amount of noise, which poses a great challenge. Most existing methods are based on traditional signal processing and often face problems of incomplete noise reduction or signal distortion after denoising. In this paper, we propose a data-driven denoising method for accelerometer signals (ADNet) based on the Wave_U_Net network architecture, incorporating Multi-Head Attention mechanism and Spatial Attention mechanism. This enhances the network’s feature extraction capability and accelerates its convergence speed. Meanwhile, the attention mechanism focuses the model’s attention on the clean signal’s feature information, improving the network’s fitting capability to the signal distribution. Therefore, ADNet can achieve more thorough denoising while reducing signal distortion after denoising. Finally, we conduct experiments on Walking speed dataset and field experiment dataset to verify the effectiveness of ADNet. The experimental results show that ADNet outperforms other baseline models in terms of Mean Square Error, Mean Absolute Error, Root Mean Square Error, and Signal-to-Noise Ratio.

No abstract available

Vibration signals have been increasingly utilized in various engineering fields for analysis and monitoring purposes, including structural health monitoring, fault diagnosis and damage detection, where vibration signals can provide valuable information about the condition and integrity of structures. In recent years, there has been a growing trend towards the use of vibration signals in the field of bioengineering. Activity-induced structural vibrations, particularly footstep-induced signals, are useful for analyzing the movement of biological systems such as the human body and animals, providing valuable information regarding an individual’s gait, body mass, and posture, making them an attractive tool for health monitoring, security, and human-computer interaction. However, the presence of various types of noise can compromise the accuracy of footstep-induced signal analysis. In this paper, we propose a novel ensemble model that leverages both the ensemble of multiple signals and of recurrent and convolutional neural network predictions. The proposed model consists of three stages: preprocessing, hybrid modeling, and ensemble. In the preprocessing stage, features are extracted using the Fast Fourier Transform and wavelet transform to capture the underlying physics-governed dynamics of the system and extract spatial and temporal features. In the hybrid modeling stage, a bi-directional LSTM is used to denoise the noisy signal concatenated with FFT results, and a CNN is used to obtain a condensed feature representation of the signal. In the ensemble stage, three layers of a fully-connected neural network are used to produce the final denoised signal. The proposed model addresses the challenges associated with structural vibration signals, which outperforms the prevailing algorithms for a wide range of noise levels, evaluated using PSNR, SNR, and WMAPE.

No abstract available

Modern navigation solutions are largely dependent on the performances of the standalone inertial sensors, especially at times when no external sources are available. During these outages, the inertial navigation solution is likely to degrade over time due to instrumental noises sources, particularly when using consumer low-cost inertial sensors. Conventionally, model-based estimation algorithms are employed to reduce noise levels and enhance meaningful information, thus improving the navigation solution directly. However, guaranteeing their optimality often proves to be challenging as sensors performance differ in manufacturing quality, process noise modeling, and calibration precision. In the literature, most inertial denoising models are model-based when recently several data-driven approaches were suggested primarily for gyroscope measurements denoising. Data-driven approaches for accelerometer denoising task are more challenging due to the unknown gravity projection on the accelerometer axes. To fill this gap, we propose several learning-based approaches and compare their performances with prominent denoising algorithms, in terms of pure noise removal, followed by stationary coarse alignment procedure. Based on the benchmarking results, obtained in field experiments, we show that: (i) learning-based models perform better than traditional signal processing filtering; (ii) non-parametric kNN algorithm outperforms all state of the art deep learning models examined in this study; (iii) denoising can be fruitful for pure inertial signal reconstruction, but moreover for navigation-related tasks, as both errors are shown to be reduced up to one order of magnitude.

Human activity recognition (HAR) has received more and more attention, which is able to play an important role in many fields, such as healthcare and intelligent home. Thus, we have discussed an application of activity recognition in the healthcare field in this paper. Essential tremor (ET) is a common neurological disorder that can make people with this disease rise involuntary tremor. Nowadays, the disease is easy to be misdiagnosed as other diseases. We have combined the essential tremor and activity recognition to recognize ET patients’ activities and evaluate the degree of ET for providing an auxiliary analysis toward disease diagnosis by utilizing stacked denoising autoencoder (SDAE) model. Meanwhile, it is difficult for model to learn enough useful features due to the small behavior dataset from ET patients. Thus, resampling techniques are proposed to alleviate small sample size and imbalanced samples problems. In our experiment, 20 patients with ET and 5 healthy people have been chosen to collect their acceleration data for activity recognition. The experimental results show the significant result on ET patients activity recognition and the SDAE model has achieved an overall accuracy of 93.33%. What’s more, this model is also used to evaluate the degree of ET and has achieved the accuracy of 95.74%. According to a set of experiments, the model we used is able to acquire significant performance on ET patients activity recognition and degree of tremor assessment.

Human activity recognition (HAR) and localization are green research areas of the modern era that are being propped up by smart devices. But the data acquired from the sensors embedded in smart devices, contain plenty of noise that makes it indispensable to design robust systems for HAR and localization. In this article, a system is presented endowed with multiple algorithms that make it impervious to signal noise and efficient to recognize human activities and their respective locations. The system begins by denoising the input signal using a Chebyshev type-I filter and then performs windowing. Then, working in parallel branches, respective features are extracted for the performed activity and human’s location. The Boruta algorithm is then implemented to select the most informative features among the extracted ones. The data is optimized using a particle swarm optimization (PSO) algorithm, and two recurrent neural networks (RNN) are trained in parallel, one for HAR and other for localization. The system is comprehensively evaluated using two publicly available benchmark datasets i.e., the Extrasensory dataset and the Sussex Huawei locomotion (SHL) dataset. The evaluation results advocate the system’s exceptional performance as it outperformed the state-of-the-art methods by scoring respective accuracies of 89.25% and 90.50% over the former dataset and 95.75% and 91.50% over the later one for HAR and localization.

Applications like smart homes, security surveillance, healthcare, and workplace safety now depend heavily on Human Activity Recognition (HAR) and monitoring. Automated systems can respond quickly, stop dangerous situations, and provide individualized services when they can precisely identify and categorize human activity. The intricate temporal dependencies present in human motion are frequently difficult to capture by conventional HAR techniques that rely on hand-crafted features and shallow learning models [1]. This paper investigates the effectiveness of a pure Recurrent Neural Network (RNN) architecture for activity recognition using smartphone sensor data, despite the fact that deep learning models like CNNs and LSTMs have dominated recent HAR research. Without the need for manually created features, we suggest a simplified RNN-based framework that can identify temporal dependencies in unprocessed accelerometer and gyroscope data. Pure RNNs can attain competitive accuracy, as shown by experiments on the UCI HAR and WISDM datasets, which yielded results of 92.4% and 90.7%, respectively. The results imply that, when properly optimized, RNNs despite their simplicity remain feasible for HAR tasks.  

Human activity recognition (HAR) is a rapidly growing field of study with important practical implications in areas as diverse as medicine, sports performance analysis, and the care of the elderly. There has been a lot of interest in studying HAR because of the possibility it offers to improve human performance and quality of life through the use of ubiquitous sensors. Due to their ability to model sequential data and capture temporal dependencies in time-series data, Recurrent Neural Networks (RNNs) have become a useful tool for HAR. In this research, many recurrent neural network (RNN) architectures and hyperparameters are tested on accelerometer data from wearable sensors to determine which is most suited for HAR. To be more specific, we analyze the relationship between the number of hidden units and the accuracy of the models for three distinct RNN architectures (Simple RNN, LSTM, and GRU). The HARUS dataset is used, which contains accelerometer data from 30 people engaged in 6 distinct activities. In comparison to the other two architectures, our results demonstrate that the LSTM architecture performs best on the HARUS dataset, with an accuracy of 95.0%. Furthermore, we find that increasing the number of hidden units generally improves accuracy, with the best results being achieved with 128 hidden units. Accuracy improves alongside sequence length, while exceeding a specific value can cause overfitting. In a separate study, researchers used accelerometer data from a wearable sensor to train an RNN model for activity recognition, and found that it achieved an overall accuracy of 99.54 percent on the test set. The model did fairly well for more complex activities like walking up and down stairs, but it excelled at the simpler tasks of walking and jogging, standing, and sitting. Our research shows that LSTM is an effective architecture for HAR applications, and that hyperparameters like the number of hidden units and sequence length are particularly important. Our results add to the existing HAR literature by disclosing the best architecture and hyperparameters for recognising human activities from wearable sensor data obtained via accelerometers.

MEMS accelerometers are significantly impacted by temperature and noise, leading to a considerable compromise in their accuracy. In response to this challenge, we propose a parallel denoising and temperature compensation fusion algorithm for MEMS accelerometers based on RLMD-SE-TFPF and GRU-attention. Firstly, we utilize robust local mean decomposition (RLMD) to decompose the output signal of the accelerometer into a series of product function (PF) signals and a residual signal. Secondly, we employ sample entropy (SE) to classify the decomposed signals, categorizing them into noise segments, mixed segments, and temperature drift segments. Next, we utilize the time-frequency peak filtering (TFPF) algorithm with varying window lengths to separately denoise the noise and mixed signal segments, enabling subsequent signal reconstruction and training. Considering the strong inertia of the temperature signal, we innovatively introduce the accelerometer’s output time series as the model input when training the temperature compensation model. We incorporate gated recurrent unit (GRU) and attention modules, proposing a novel GRU-MLP-attention model (GMAN) architecture. Simulation experiments demonstrate the effectiveness of our proposed fusion algorithm. After processing the accelerometer output signal through the RLMD-SE-TFPF denoising algorithm and the GMAN temperature drift compensation model, the acceleration random walk is reduced by 96.11%, with values of 0.23032 g/h/Hz for the original accelerometer output signal and 0.00895695 g/h/Hz for the processed signal.

Cardiac vibration signal analysis emerges as a remarkable tool for the diagnosis of heart conditions. Our recent study shows the feasibility of the longitudinal monitoring of chronic heart diseases, particularly heart failure, using a gastric fundus implant. However, cardiac vibration data, captured from the implant, positioned at the gastric fundus, can be highly affected by different noises and artefacts. This study introduces a novel methodology for addressing denoising challenges in the longitudinal monitoring of chronic heart diseases, using gastric fundus implants. More precisely, a novel method is designed, by repurposing pre-trained convolutional neural network models, originally designed for classification tasks, with adequately chosen convolution filters. The proposed approach efficiently tackles noise and artefacts reduction in the acquired accelerometer signals. Moreover, the integration of additional Hilbert and Homomorphic envelopes enhances the implant's ability to better segment heart sounds, namely S1 and S2. The quality assessment of this denoising strategy is performed, in the lack of ground truth, by rather evaluating its impact on a classification stage that is introduced to the proposed pipeline. Compared to standard denoising matrix factorization and tensor decomposition-based methods, results on a real 3D accelerometer dataset acquired from a set of pigs, with and without heart failure, demonstrate the efficacy of such a proposed optimized CNN-based approach with the best balance between enhancing the segmentation accuracy and preserving a maximum usable record.

Cardiac vibration signals provide insights into the mechanical function of the heart, making them potentially useful for the diagnosis and follow-up of heart failure. These signals can be captured using implantable cardiac devices incorporating a 3D accelerometer, as recently proposed in our group, enabling continuous longitudinal monitoring. However, the quality of these signals is highly affected by the presence of noise and artefacts originating from several physiological and non-physiological sources, thereby reducing the clinical effectiveness. To address this issue, a graph-based approach to improve the quality of cardiac vibration signals captured through 3D accelerometer recordings from an implantable device located in the gastric fundus, is proposed in this paper. By harnessing the inherent repetitive nature of these heart vibration signals across recorded heart cycles, the denoising problem is reformulated as the fact of inferring a target signal matrix incorporating both the smoothness on graph and low-rank constraints. The effectiveness of the proposed approach is shown through experiments conducted on real recordings acquired from a group of pigs, both with and without heart failure. The proposed approach shows superior performance compared to traditional de noising techniques.

This article focuses on the challenge of accuracy degradation in high-G accelerometer shock calibration due to noise interference in the signal. An efficient signal-denoising method is proposed to address this issue. The method employs adaptive frequency segmentation based on complementary ensemble empirical mode decomposition (CEEMD), effectively eliminating high-frequency noise while accurately preserving the peak information of the shock response. Additionally, a time-domain zeroing (TZ) strategy is integrated into the proposed denoising method, significantly reducing noise and correcting the frequency response amplitude. Simulation results reveal that the method exhibits remarkable performance in noise reduction, sensitivity calibration accuracy, and amplitude-frequency characteristic calibration accuracy, surpassing the other methods. Furthermore, experimental results indicate the ability of this method to enhance the stability of real sensitivity and amplitude-frequency characteristic calibration, thereby providing robust technical support for high-precision calibration measurements.

Cardiac vibration signals provide critical insights into the heart's mechanical function, making them essential for diagnosing and monitoring heart failure. An implantable cardiac device equipped with a 3D accelerometer has been recently developed in our research group to capture these signals. It enables continuous long-term monitoring of heart conditions. However, noise and artifacts can significantly degrade the quality of these signals, obscuring important cardiac events and diminishing the clinical utility of the implant. To improve signal quality, a graph-based method is introduced in this paper to denoise cardiac vibration signals captured by the 3D accelerometer which is embedded in an implantable device located in the gastric fundus. The proposed method leverages the intrinsic pseudo-periodicity of heart vibration signals across cardiac cycles. More precisely, it reformulates the denoising problem as the inference of a low-rank matrix along with target signals being assumed smooth on a graph structure which is optimally learned from the observed data. The efficacy of the proposed approach compared to standard denoising methods is confirmed using real recordings from a cohort of pigs, encompassing both those with and without heart failure.

This research presents an Enhanced Long Short-Term Memory (LSTM) deep learning model for robust noise reduction in automotive wheel speed sensors. While wheel speed sensors are pivotal to vehicle stability, high-intensity or non-stationary noise often degrades their performance. Traditional filtering methods, including adaptive approaches and basic digital signal processing, frequently underperform under complex conditions. The proposed model addresses these limitations by incorporating an attention mechanism that selectively emphasizes transient high-noise frames, preserving essential rotational information. Comprehensive experiments, supported by Variational Mode Decomposition (VMD) and the Hilbert-Huang Transform (HHT), demonstrate that the Enhanced LSTM surpasses conventional techniques and baseline LSTM architectures in suppressing interference. T results yield significantly improved metrics across varying noise intensities, confirming both efficacy and stability. Although factors such as computational cost and the need for extensive labeled data remain, the Enhanced LSTM shows strong potential for real-time applications in wheel speed sensing. This work offers valuable insights into advanced noise mitigation and serves as a foundation for future deep learning research in complex automotive signal processing tasks.

The effective extraction of key features in non-stationary signals measurement is crucial in numerous engineering fields, including fault diagnosis, geological exploration, and state detection. To accomplish a more accurate and efficient extraction of key feature information from non-stationary signals, we design a novel approach based on variational mode decomposition (VMD) optimization by northern goshawk optimization (NGO) algorithm, convolutional neural network (CNN), and long short-term memory network (LSTM). First, NGO is used to optimize multiple intrinsic mode functions of VMD and reconstruct the signal according to the linear correlation method. Subsequently, the features of moving root mean square, moving kurtosis, and upper envelope are calculated, thereby constructing the feature matrix. Finally, the CNN-LSTM model is established with the chosen optimal hyperparameters prior to the training phase. The experimental results demonstrate that the proposed NGO-VMD-CNN-LSTM method, with a high accuracy reaching 98.22%, can more accurately extract the key information of typical non-stationary signals.

An ultra-high precision magnetic field measurement is of great scientific and economic significance. An atomic magnetometer that operates in a spin-exchange relaxation-free (SERF) regime has superior sensitivity and, thus, is of great significance for the ultra-high precision magnetic field measurement. In order to reduce the noise of a SERF atomic magnetometer and further improve its sensitivity, a method for noise reduction based on the multigene genetic programming (MGGP) is presented. Different from the existing methods, in this method, the model of magnetometer noise is established based on the experimental data. Namely, the noise of the SERF atomic magnetometer is first modeled by the MGGP algorithm and then reduced by the obtained model. Besides, in this way, the sensitivity is improved. The experimental results indicate that our MGGP model can adequately reflect the characteristics of the SERF magnetometer noise. Moreover, after applying the proposed method, the sensitivity of the SERF magnetometer is improved about 13 times at 1 Hz, and there is also a significant sensitivity increase at the frequencies less than 10 Hz. Therefore, the proposed method can effectively reduce the noise and improve the sensitivity of the SERF magnetometer in the low-frequency band.

Micro-electro-mechanical system inertial measurement unit (MEMS-IMU), a core component in many navigation systems, directly determines the accuracy of inertial navigation system; however, MEMS-IMU system is often affected by various factors such as environmental noise, electronic noise, mechanical noise and manufacturing error. These can seriously affect the application of MEMS-IMU used in different fields. Focus has been on MEMS gyro since it is an essential and, yet, complex sensor in MEMS-IMU which is very sensitive to noises and errors from the random sources. In this study, recurrent neural networks are hybridized in four different ways for noise reduction and accuracy improvement in MEMS gyro. These are two-layer homogenous recurrent networks built on long short term memory (LSTM-LSTM) and gated recurrent unit (GRU-GRU), respectively; and another two-layer but heterogeneous deep networks built on long short term memory-gated recurrent unit (LSTM-GRU) and a gated recurrent unit-long short term memory (GRU-LSTM). Practical implementation with static and dynamic experiments was carried out for a custom MEMS-IMU to validate the proposed networks, and the results show that GRU-LSTM seems to be overfitting large amount data testing for three-dimensional axis gyro in the static test. However, for X-axis and Y-axis gyro, LSTM-GRU had the best noise reduction effect with over 90% improvement in the three axes. For Z-axis gyroscope, LSTM-GRU performed better than LSTM-LSTM and GRU-GRU in quantization noise and angular random walk, while LSTM-LSTM shows better improvement than both GRU-GRU and LSTM-GRU networks in terms of zero bias stability. In the dynamic experiments, the Hilbert spectrum carried out revealed that time-frequency energy of the LSTM-LSTM, GRU-GRU, and GRU-LSTM denoising are higher compared to LSTM-GRU in terms of the whole frequency domain. Similarly, Allan variance analysis also shows that LSTM-GRU has a better denoising effect than the other networks in the dynamic experiments. Overall, the experimental results demonstrate the effectiveness of deep learning algorithms in MEMS gyro noise reduction, among which LSTM-GRU network shows the best noise reduction effect and great potential for application in the MEMS gyroscope area.

Autonomous vehicles require the development of sensing technologies and intelligent control of localization capabilities to guide the vehicle in unknown areas. A reliable localization system of accurate positioning and heading information is one of the critical requirements of highly challenging autonomous vehicle technology. This paper proposes an accurate estimation scheme using a Recurrent Neural Network (RNN) architecture for mobile robots in indoor environments via an Inertial Measurement Unit (IMU). The main objective is to assess the potential performance of LSTM or GRU network architecture to obtain estimation values using only low-cost IMU sensor data to create accurate heading angles. The preprocessing stage is carried out to be able to reduce or even eliminate the bad impact of noise generated on each data. The test shows that the model generated from the LSTM network architecture with 32-16 cells of neurons layer can provide heading estimates with MSE value of 0.02 and an accuracy 94.65%.

In practical applications of MIMU, due to the influence of measurement noise and external environment, the raw data contains many outliers and noise signals, which reduces its system performance. In view of the above problems, this paper combines the traditional AKF algorithm with the LSTM network to obtain a joint noise reduction algorithm. Based on the completion of noise modeling and model parameter update, the construction principle of the joint noise reduction algorithm is briefly analyzed. This method makes the advantages of AKF and LSTM complement each other, and overcomes the shortcomings of traditional AKF's poor stability. The experimental results show that the method can effectively suppress the measurement noise interference and improve the accuracy of pure inertial navigation.

Microelectromechanical System (MEMS) Inertial Measurement Unit (IMU) is popular in the community for constructing a navigation system, due to its small size and low power consumption. However, limited by the manufacturing technology, MEMS IMU experiences more complicated noises and errors. Thus, noise modeling and suppression is important for improving accuracy of the navigation system based on MEMS IMU. Motivated by this problem, in this paper, a deep learning method was introduced to MEMS gyroscope de-noising. Specifically, a recently popular Recurrent Neural Networks (RNN) variant Simple Recurrent Unit (SRU-RNN) was employed in MEMS gyroscope raw signals de-noising. A MEMS IMU MSI3200 from MT Microsystem Company was employed in the experiments for evaluating the proposed method. Following two problems were furtherly discussed and investigated: (1) the employed SRU with different training data length were compared to explore whether there was trade-off between the training data length and prediction performance; (2) Allan Variance was the most popular MEMS gyroscope analyzing method, and five basic parameters were employed to describe the performance of different grade MEMS gyroscope; among them, quantization noise, angle random walk, and bias instability were the major factors influencing the MEMS gyroscope accuracy, the compensation results of the three parameters for gyroscope were presented and compared. The results supported the following conclusions: (1) considering the computation brought from training dataset, the values of 500, 3000, and 3000 were individually sufficient for the three-axis gyroscopes to obtain a reliable and stable prediction performance; (2) among the parameters, the quantization noise, angle random walk, and bias instability performed 0.6%, 6.8%, and 12.5% improvement for X-axis gyroscope, 60.5%, 17.3%, and 34.1% improvement for Y-axis gyroscope, 11.3%, 22.7%, and 35.7% improvement for Z-axis gyroscope, and the corresponding attitude errors decreased by 19.2%, 82.1%, and 69.4%. The results surely demonstrated the effectiveness of the employed SRU in this application.

The nine-axis inertial and measurement unit (IMU)-based three-dimensional (3D) orientation estimation is a fundamental part of inertial motion capture. Recently, owing to the successful utilization of deep learning in various applications, orientation estimation neural networks (NNs) trained on large datasets, including nine-axis IMU signals and reference orientation data, have been developed. During the training process, the limited amount of training data is a critical issue in the development of powerful networks. Data augmentation, which increases the amount of training data, is a key approach for addressing the data shortage problem and thus for improving the estimation performance. However, to the best of our knowledge, no studies have been conducted to analyze the effects of data augmentation techniques on estimation performance in orientation estimation networks using IMU sensors. This paper selects three data augmentation techniques for IMU-based orientation estimation NNs, i.e., augmentation by virtual rotation, bias addition, and noise addition (which are hereafter referred to as rotation, bias, and noise, respectively). Then, this paper analyzes the effects of these augmentation techniques on estimation accuracy in recurrent neural networks, for a total of seven combinations (i.e., rotation only, bias only, noise only, rotation and bias, rotation and noise, and rotation and bias and noise). The evaluation results show that, among a total of seven augmentation cases, four cases including ‘rotation’ (i.e., rotation only, rotation and bias, rotation and noise, and rotation and bias and noise) occupy the top four. Therefore, it may be concluded that the augmentation effect of rotation is overwhelming compared to those of bias and noise. By applying rotation augmentation, the performance of the NN can be significantly improved. The analysis of the effect of the data augmentation techniques presented in this paper may provide insights for developing robust IMU-based orientation estimation networks.

The joint angle during gait is an important indicator, such as injury risk index, rehabilitation status evaluation, etc. To analyze gait, inertial measurement unit (IMU) sensors have been used in studies and continuously developed; however, they are difficult to utilize in daily life because of the inconvenience of having to attach multiple sensors together and the difficulty of long-term use due to the battery consumption required for high data sampling rates. To overcome these problems, this study propose a multi-joint angle estimation method based on a long short-term memory (LSTM) recurrent neural network with a single low-frequency (23 Hz) IMU sensor. IMU sensor data attached to the lateral shank were measured during overground walking at a self-selected speed for 30 healthy young persons. The results show a comparatively good accuracy level, similar to previous studies using high-frequency IMU sensors. Compared to the reference results obtained from the motion capture system, the estimated angle coefficient of determination (R2) is greater than 0.74, and the root mean square error and normalized root mean square error (NRMSE) are less than 7° and 9.87%, respectively. The knee joint showed the best estimation performance in terms of the NRMSE and R2 among the hip, knee, and ankle joints.

In this work, we propose a novel method for performing inertial aided navigation, by using deep neural net-works (DNNs). To date, most DNN inertial navigation methods focus on the task of inertial odometry, by taking gyroscope and accelerometer readings as input and regressing for integrated IMU poses (i.e., position and orientation). While this design has been successfully applied on a number of applications, it is not of theoretical performance guarantee unless patterned motion is involved. This inevitably leads to significantly reduced accuracy and robustness in certain use cases. To solve this problem, we design a framework to compute observable IMU integration terms using DNNs, followed by the numerical pose integration and sensor fusion to achieve the performance gain. Specifically, we perform detailed analysis on the motion terms in IMU kinematic equations, propose a dedicated network design, loss functions, and training strategies for the IMU data processing, and conduct extensive experiments. The results show that our method is generally applicable and outperforms both traditional and DNN methods by wide margins.

With the rapid advancements in artificial intelligence (AI), 5G technology, and robotics, multi-sensor fusion technologies have emerged as a critical solution for achieving high-precision localization in mobile robots operating within dynamic and unstructured environments. This study proposes a novel hybrid fusion framework that combines the Extended Kalman Filter (EKF) and Recurrent Neural Network (RNN) to address challenges such as sensor frequency asynchrony, drift accumulation, and measurement noise. The EKF provides real-time statistical estimation for initial data fusion, while the RNN effectively models temporal dependencies, further reducing errors and enhancing data accuracy. A complementary fusion mechanism integrating LiDAR (Light Detection and Ranging) data ensures robustness against noise and disturbances. The algorithm is validated through comprehensive simulations on the Gazebo platform, demonstrating a localization error within 8 cm under various noise levels and dynamic disturbances. The method also outperforms state-of-the-art algorithms, including Particle Filter (PF) and Graph SLAM, in both accuracy and computational efficiency, achieving an average runtime of 30.1 ms per frame, suitable for real-time applications. These results highlight the efficacy of the proposed EKF-RNN framework, which balances accuracy, robustness, and computational efficiency, offering significant contributions to autonomous robotic navigation.

Recurrent neural network (RNN), as a kind of neural network with outstanding computing capability, improvability, and hardware realizability, has been widely used in various fields, especially in robotics. In this paper, an RNN with noise rejection is deliberately constructed to remedy the issue of joint-angle drift frequently occurring during the cyclic motion generation (CMG) of a manipulator in a noisy environment. Different from general RNNs, the proposed RNN possesses inherent noise immunity, especially for time-varying polynomial noises. Besides, proofs on the convergence of the proposed RNN in the absence and presence of noises are given. Furthermore, we carry out simulations on manipulators PUMA 560 and UR5 to demonstrate the reliability of the proposed RNN in remedying joint-angle drift, and comparison simulations under different noisy conditions further verify its superiority. In addition, experiments are conducted on manipulator FRANKA Panda to elucidate the realizability of the proposed RNN.

This study delves into the utilization of wearable devices equipped with inertial measurement units (IMUs) to gather data for human activity recognition. It explores the application of deep neural networks for automatically identifying gym workouts using IMU sensors. Our methodology involves developing a framework that integrates convolutional neural networks (CNNs) to extract features from raw sensor data, followed by using long short-term memory (LSTM) recurrent neural networks to classify sequences. IMU data from accelerometers and gyroscopes are collected from 10 individuals performing 30 standard gym routines. The CNN-LSTM pipeline is supervised on a comprehensive dataset comprising multiple sensors and subjects to distinguish between different workouts accurately. During the evaluation, the CNN-LSTM model achieved an accuracy of 93.81 % in categorizing 30 workout categories based solely on accelerometer data. Through augmentation, this accuracy is further improved to 95.75 %. This solution outperforms independently utilized CNN and LSTM models and traditional machine learning approaches. Detailed assessments offer valuable insights into the benefits of combining diverse sensor types and model architectures for robust exercise classification. This research raises a precise and dependable wearable system for identifying gym exercises using deep neural networks. The findings suggest promising avenues for future exploration in human activity recognition, mainly focusing on utilizing on-body sensors and deep learning to analyze more complex human movements.

As an important component of inertial guidance and navigation, micro-electromechanical-system (MEMS) gyroscope is widely used in many fields. However, the accumulation of noise errors limits the long-term accuracy and further application of MEMS gyroscope. In this article, a novel denoising method for MEMS gyroscope based on interpolated complementary ensemble local mean decomposition with adaptive noise (ICELMDAN) and gated recurrent unit-unscented Kalman filter (GRU-UKF) has been proposed. First, the original signal of MEMS gyroscope is decomposed into multiple product functions (PFs) by ICELMDAN. Second, the PFs are classified into useful component, mixed component, and noise component based on the corresponding sample entropies (SEs). Finally, the mixed component is filtered by GRU-UKF and combined with the useful component to reconstruct the denoised signal. Experiments are carried out to validate the proposed method. The bias instability of MEMS gyroscope is reduced from 0.375°/h to 0.016°/h, and the standard deviation suppression rate reaches 89.28%, demonstrating the effectiveness and superiority of the proposed method.

In the realm of inertial sensor-based gait recognition, data loss during collection, arising from device-related issues or interruptions in the transmission process, introduces complexities for processing. Moreover, extracting identity-distinguishing features for identity recognition encounters challenges due to diverse sensor modalities. To solve those two issues, we propose a method named INSENGA for inertial sensor gait recognition incorporating multigated recurrent unit (multi-GRU) variational autoencoder (VAE) data imputation and channel attention weight redistribution (CAWR). INSENGA consists of two modules: multi-GRU VAE and CAWR gait recognition network (CAWR-GRN). Multi-GRU VAE is introduced to tackle the data loss, in which inertial sensor data of diverse modality are variationally autoencoded through separate hybrid convolutional neural network (CNN) and GRU networks, and GRU networks are served as decoders to capture temporal features. CAWR-GRN is featured by the CAWR mechanism, which uses the expected channel damage matrix (ECDM) to score channel weights, and employs channel attention for weight redistribution to improve gait recognition performance. CAWR-GRN also integrates a hybrid CNN and bidirectional long short-term memory (BLSTM) network to capture temporal features and a pretraining network transferring weights to CNNs. To evaluate the performance of INSENGA, experiments are conducted on one self-built dataset and two widely used public datasets. The experimental results show that INSENGA achieved an accuracy of 90.06% on the ID-Sensor dataset, 97.26% on the whuGait #1 dataset, 99.52% on the whuGait #2 dataset, and 93.49% on the osaka university-intelligent sensor and information robotics (OU-ISIR) dataset, indicating that INSENGA has reached state-of-the-art gait recognition performance.

This study introduces an adaptive sensor fusion system for AR tracking in art exhibitions, addressing environmental variability and erratic motion. A GRU-based mechanism dynamically adjusts visual-inertial data fusion by evaluating lighting conditions and movement intensity. The framework combines RGB-D and IMU inputs via a modified Kalman filter, automatically favoring visual data in stable settings and inertial data during rapid motion or poor lighting. Trained on synthetic exhibition sequences with a dual-objective loss function, the system maintains stable AR overlays despite lighting changes or sudden movements. Compatible with existing AR platforms like Unity MARS, it outperforms fixed fusion methods in robustness tests across synthetic and real-world datasets. This advancement enhances adaptive sensor fusion for AR in culturally sensitive environments where precise tracking is crucial for immersion.

Inertial Measurement Unit (IMU) plays an important role in inertial aided navigation on robots. However, raw IMU data could be noisy, especially for low-cost IMUs, and thus requires efficient pre-processing or denoising before applying further navigation algorithms. Conventional IMU denoising approaches are mostly hand-crafted and may face concerns such as sensor modelling errors and generalization issues. Several recent works leverage deep neural networks (DNNs) to tackle this problem and achieve promising results. However, currently reported deep learning methods are based on supervised learning, requiring sufficient and accurate annotations. While in real-world applications, such annotations can be expensive or unavailable, making these methods not practical. To address the above research gap, we propose incorporating self-supervised learning and future-aware inference for IMU denoising. The end-to-end navigation evaluation results on EuRoC and TUM-VI datasets are promising. The code will be publicly available at https://github.com/KleinYuan/IMUDB.

To enhance system reliability and mitigate the vulnerabilities of the Global Navigation Satellite Systems (GNSS), it is common to fuse the Inertial Measurement Unit (IMU) and visual sensors with the GNSS receiver in the navigation system design, effectively enabling compensations with absolute positions and reducing data gaps. To address the shortcomings of a traditional Kalman Filter (KF), such as sensor errors, an imperfect non-linear system model, and KF estimation errors, a GRU-aided ESKF architecture is proposed to enhance the positioning performance. This study conducts Failure Mode and Effect Analysis (FMEA) to prioritize and identify the potential faults in the urban environment, facilitating the design of improved fault-tolerant system architecture. The identified primary fault events are data association errors and navigation environment errors during fault conditions of feature mismatch, especially in the presence of multiple failure modes. A hybrid federated navigation system architecture is employed using a Gated Recurrent Unit (GRU) to predict state increments for updating the state vector in the Error Estate Kalman Filter (ESKF) measurement step. The proposed algorithm’s performance is evaluated in a simulation environment in MATLAB under multiple visually degraded conditions. Comparative results provide evidence that the GRU-aided ESKF outperforms standard ESKF and state-of-the-art solutions like VINS-Mono, End-to-End VIO, and Self-Supervised VIO, exhibiting accuracy improvement in complex environments in terms of root mean square errors (RMSEs) and maximum errors.

Recently, remarkable successes have been achieved in the quality assurance of automotive software systems (ASSs) through the utilization of real-time hardware-in-the-loop (HIL) simulation. Based on the HIL platform, safe, flexible and reliable realistic simulation during the system development process can be enabled. However, notwithstanding the test automation capability, large amounts of recordings data are generated as a result of HIL test executions. Expert knowledge-based approaches to analyze the generated recordings, with the aim of detecting and identifying the faults, are costly in terms of time, effort and difficulty. Therefore, in this study, a novel deep learning-based methodology is proposed so that the faults of automotive sensor signals can be efficiently and automatically detected and identified without human intervention. Concretely, a hybrid GRU-based denoising autoencoder (GRU-based DAE) model with the k-means algorithm is developed for the fault-detection and clustering problem in sequential data. By doing so, based on the real-time historical data, not only individual faults but also unknown simultaneous faults under noisy conditions can be accurately detected and clustered. The applicability and advantages of the proposed method for the HIL testing process are demonstrated by two automotive case studies. To be specific, a high-fidelity gasoline engine and vehicle dynamic system along with an entire vehicle model are considered to verify the performance of the proposed model. The superiority of the proposed architecture compared to other autoencoder variants is presented in the results in terms of reconstruction error under several noise levels. The validation results indicate that the proposed model can perform high detection and clustering accuracy of unknown faults compared to stand-alone techniques.

Inertial sensor errors are an important error source in strapdown inertial navigation systems (SINS). These errors are divided into two groups: systematic and random. Systematic errors include offset and scale factor errors, which are estimated using special calibration methods. Random errors, however, can lead to instability of the offset or scale factor over time, which is a key component leading to the increase in SINS errors over time. In this paper we consider a method for reducing random noise in an inertial gyroscope signal using a hybrid deep recurrent neural network consisting of subnets: long short-term memory (LSTM) and gated recurrent units (GRU). One of the most important tasks of constructing an adequate deep neural network structure is finding optimal network hyperparameters, which is solved in this work using the Bayesian optimization method. The obtained results are of great importance for solving the problem of the initial SINS alignment.

Human activity recognition (HAR) has a wide range of applications in medical care, elderly care and so on. Due to the development of wearable sensor technology, HAR based on wearable sensors has gradually become a hot research topic. In terms of feature extraction of raw data, traditional methods usually use manually extracted features, which has many limitations. The method based on deep learning has gradually become the mainstream. In this paper, a deep bidirectional GRU network for human activity recognition is proposed, which uses few data preprocessing to classify and recognize daily activities. An accuracy rate of 95.69% is achieved on the UCL-HAR dataset which is an increase of 2.8% compared to the bidirectional LSTM network. In addition, the proposed method in this paper reduces the training time by 11.8 %, which will help the neural network to be deployed in real-time applications.

Visualinertial odometry (VIO) has been extensively studied for navigation in GNSS-denied environments, but its performance can be heavily impacted by the complexity of the navigation environments such as weather conditions, illumination variation, flight dynamics, and environmental structure. Hybrid fusion approaches integrating neural networks (NNs), especially gated recurrent units (GRUs) with the Kalman filters (KFs), such as error-state KF (ESKF) have shown promising results mitigating system nonlinearities due to challenging environmental conditions data issues, there is a lack of systematic studies quantitively analyzing and comparing performance differences unhand. To address this gap and enable robust navigation in complex conditions, this study proposes and systematically analyses the performance of three hybrid fusion schemes for VIO-based navigation of autonomous aerial vehicles (AAVs). These three hybrid VIO schemes include visual odometry (VO) error compensation using NN, KF error compensation using NN, and prediction of Kalman gain using NN. The comparative analysis is performed using data generated in MATLAB incorporating the unreal engine involving diverse challenging environmental conditions: fog, rain, illumination level variability, and variability in the number of features available for extraction during the AAV flight in the urban environment. The results demonstrate the performance improvement achieved by hybrid VIO fusion schemes compared to ESKF-based traditional fusion methods in the presence of multiple visual failure modes. Comparative analysis reveals notable improvement achieved by method 1 with enhancements of 93% in sunny, 91% in foggy, and 90% in rainy conditions than the other two hybrid VIO architectures.

Robust localization of autonomous robots remains a critical challenge in environments where LiDAR performance is degraded, such as nuclear facilities or underground tunnels. Conventional Strapdown Inertial Navigation Systems (SINS) accumulate errors over time due to sensor bias and noise, a problem that is further exacerbated in environments with sparse or ambiguous LiDAR observations. In this study, we propose a GRU-based learning motion prediction model to replace the SINS-based inertial state prediction within a UKF-based LiDAR-Inertial SLAM framework. By directly learning the robot's motion dynamics and time-varying sensor biases from raw IMU data, the proposed model mitigates the cumulative errors that arise when integrating discrete IMU measurements and provides more accurate motion priors. The effectiveness of the proposed approach was validated in both simulation and real-world environments. In simulation experiments, UKF+GRU achieved equal or lower Absolute Trajectory Error (ATE) compared to conventional SLAM methods, while also significantly reducing computational cost. In real-world environments, the proposed model provided stable pose estimates, although further validation in more diverse environments and against high-precision ground-truth measurements is necessary.

This paper aims to propose a gyroscope denoising method based on the real data obtained from the inertial measurement unit to acquire the robot’s attitude. Experiments show that, compared with existing algorithms, this network structure and loss function can achieve better accuracy. This method is based on dilated convolution, which overcomes the limitations of traditional recurrent neural networks and improves the training speed. Innovatively, the authors transform the time-domain task into a frequency-domain task. Such a change can reduce the limitations of time series data in terms of noise sensitivity and small data sets. Specifically, the module the authors propose captures the interaction between long- and short-term information by using the inertial convolutional block (ICB) and dilated convolution and enhances the feature representation through Fourier analysis. In addition, the authors also adopt the idea of channel attention to capture more detailed temporal information in the sensor data. The adaptive threshold method selectively filters out high-frequency noise information, enabling us to better predict the required attitude. The authors also reconstruct a loss function, which takes into account both the incremental directions of each small range and the global incremental errors. The authors have compared this method with the widely used European Robotics Challenge data set and the publicly available TUM Stereo Visual-Inertial Event data set. Finally, the authors conclude that this method outperforms the current algorithms on both data sets. The value of this paper is that to overcome the limitations of time-domain analysis, the authors designed a Fourier adaptive attention module. Meanwhile, to handle non-continuous information, the authors introduced the ICB module. The authors reconstructed a loss function, which takes into account both the incremental directions of each small range and the global incremental errors.

This work demonstrates an airflow inertial based odometry system with multi-sensor data fusion, including thermal anemometer, IMU, ESC, and barometer. This goal is challenging because low-cost IMUs and barometers have significant bias, and anemometer measurements are very susceptible to interference from spinning propellers and ground effects. We employ a GRU-based deep neural network to estimate relative air speed from noisy and disturbed anemometer measurements, and an observer with bias model to fuse the sensor data and thus estimate the state of aerial vehicle. A complete flight data, including takeoff and landing on the ground, shows that the approach is able to decouple the downwash induced wind speed caused by propellers and the ground effect, and accurately estimate the flight speed in a wind-free indoor environment. IMU, and barometer bias are effectively estimated, which significantly reduces the position integration drift, which is only 5.7m for 203s manual random flight. The open source is available on https://github.com/SyRoCo-ISIR/Flight-Speed-Estimation-Airflow.

Future UAV (unmanned aerial vehicle) operations in urban environments demand a PNT (position, navigation, and timing) solution that is both robust and resilient. While a GNSS (global navigation satellite system) can provide an accurate position under open-sky assumptions, the complexity of urban operations leads to NLOS (non-line-of-sight) and multipath effects, which in turn impact the accuracy of the PNT data. A key research question within the research community pertains to determining the appropriate hybrid fusion architecture that can ensure the resilience and continuity of UAV operations in urban environments, minimizing significant degradations of PNT data. In this context, we present a novel federated fusion architecture that integrates data from the GNSS, the IMU (inertial measurement unit), a monocular camera, and a barometer to cope with the GNSS multipath and positioning performance degradation. Within the federated fusion architecture, local filters are implemented using EKFs (extended Kalman filters), while a master filter is used in the form of a GRU (gated recurrent unit) block. Data collection is performed by setting up a virtual environment in AirSim for the visual odometry aid and barometer data, while Spirent GSS7000 hardware is used to collect the GNSS and IMU data. The hybrid fusion architecture is compared to a classic federated architecture (formed only by EKFs) and tested under different light and weather conditions to assess its resilience, including multipath and GNSS outages. The proposed solution demonstrates improved resilience and robustness in a range of degraded conditions while maintaining a good level of positioning performance with a 95th percentile error of 0.54 m for the square scenario and 1.72 m for the survey scenario.

This paper aims to design and implement an adaptive super-twisting observer aided by an optimized hybrid GRU-LSTM sequential model (OHGLSM) for an accurate and reliable integrated navigation system, even in an environment with missing global navigation satellite system (GNSS) signals. The method addresses strap-down inertial navigation system (SINS) perturbations and accounts for unknown inertial measurement errors by integrating extra states into the nominal model. The proposed approach uses the position and velocity components of the GNSS to estimate the states. In this respect, a super-twisting observer is utilized to estimate perturbations along with the state of the system. An adaptive solution is also introduced to adjust the estimator coefficients according to the vehicle's dynamic maneuvers. Innovatively, the OHGLSM, in combination with the air-data sensors, is considered in the observer design to provide reliable vehicle position information during GNSS signal blockages. In this respect, a pitot tube and a barometric pressure sensor are used to define vehicle velocity according to the non-holonomic constraints in addition to altitude data. The proposed method undergoes mathematical analysis and experimental evaluation using real-world vehicle tests in various scenarios to demonstrate its high accuracy and reliability under various conditions. Comparative results with different estimation methods highlight the superior accuracy of the proposed method in constructing a precise navigation system even during GNSS outages.

This paper addresses the issues of significant noise and noticeable bias in the signals of low-cost inertial sensor integrated into smartphones. It proposes a method that combines Convolutional Neural Network (CNN) denoising techniques with Pedestrian Dead Reckoning (PDR) to enhance the accuracy of gait estimation and step length detection. To further resolve the problem of error accumulation in long-term PDR usage, an Extended Kalman Filter (EKF) is introduced to fuse the denoised PDR results with WiFi positioning. This approach effectively mitigates the error accumulation in PDR and reduces the instability of WiFi positioning, providing more accurate localization services in complex indoor environments. After applying the CNN denoising technique, the PDR positioning accuracy improved from 2.28 meters to 1.70 meters, a 25.44% improvement compared to the original accuracy. The positioning accuracy of the PDR and WiFi fusion method improved from 1.13 meters to 1.02 meters, representing a 9.73% improvement.

Wearable inertial motion capture (MoCap) provides a portable, occlusion-free, and privacy-preserving alternative to camera-based systems, but its accuracy depends on tightly attached sensors - an intrusive and uncomfortable requirement for daily use. Embedding IMUs into loose-fitting garments is a desirable alternative, yet sensor-body displacement introduces severe, structured, and location-dependent corruption that breaks standard inertial pipelines. We propose GID (Garment Inertial Denoiser), a lightweight, plug-and-play Transformer that factorizes loose-wear MoCap into three stages: (i) location-specific denoising, (ii) adaptive cross-wear fusion, and (iii) general pose prediction. GID uses a location-aware expert architecture, where a shared spatio-temporal backbone models global motion while per-IMU expert heads specialize in local garment dynamics, and a lightweight fusion module ensures cross-part consistency. This inductive bias enables stable training and effective learning from limited paired loose-tight IMU data. We also introduce GarMoCap, a combined public and newly collected dataset covering diverse users, motions, and garments. Experiments show that GID enables accurate, real-time denoising from single-user training and generalizes across unseen users, motions, and garment types, consistently improving state-of-the-art inertial MoCap methods when used as a drop-in module.

With the rise of artificial intelligence, sensor-based human activity recognition (S-HAR) is increasingly being employed in healthcare monitoring for the elderly, fitness tracking, and patient rehabilitation using smart devices. Inertial sensors have been commonly used for S-HAR, but wearable devices have been demanding more comfort and flexibility in recent years. Consequently, there has been an effort to incorporate stretch sensors into S-HAR with the advancement of flexible electronics technology. This paper presents a deep learning network model, utilizing aggregation residual transformation, that can efficiently extract spatial–temporal features and perform activity classification. The efficacy of the suggested model was assessed using the w-HAR dataset, which included both inertial and stretch sensor data. This dataset was used to train and test five fundamental deep learning models (CNN, LSTM, BiLSTM, GRU, and BiGRU), along with the proposed model. The primary objective of the w-HAR investigations was to determine the feasibility of utilizing stretch sensors for recognizing human actions. Additionally, this study aimed to explore the effectiveness of combining data from both inertial and stretch sensors in S-HAR. The results clearly demonstrate the effectiveness of the proposed approach in enhancing HAR using inertial and stretch sensors. The deep learning model we presented achieved an impressive accuracy of 97.68%. Notably, our method outperformed existing approaches and demonstrated excellent generalization capabilities.

Motion intention prediction is an essential challenge in improving the efficacy of exoskeleton human-machine interactions. Traditional gait phase detection and motion pattern recognition methods can only identify the current gait state, lacking the ability to adjust motion parameters based on step speed dynamically, and real-time performance is difficult to guarantee. In this paper, we proposed a deep learning model TCN-LSTM, which integrated Time Convolutional Network (TCN) and Long Short-Term Memory Network (LSTM), to predict human ankle joint angles under different walking speeds based on data from Inertial Measurement Unit (IMU) and Goniometer (GON). We selected 5 subjects in a public dataset for experimental validation. Experimental results demonstrated that the proposed model could accurately and rapidly predict ankle joint angles. The MAE of the individual-specific model was 1.191°, outperforming the 3 baseline models: LSTM (1.750°), Gated Recurrent Unit (GRU) (1.653°), and TCN (1.340°). The Mean Absolute Error (MAE) of the general model is reduced by 14.3% compared to the 3 baseline models. The prediction error of TCN-LSTM with GON is reduced by 13.69%. Furthermore, the average prediction time of our model is only 3.7ms. Therefore, the accuracy and real-time performance of the model can meet the continuous prediction of human ankle joint angle.

Fall accidents in the construction industry have been studied over several decades and identified as a common hazard and the leading cause of fatalities. Inertial sensors have recently been used to detect accidents of workers in construction sites, such as falls or trips. IMU-based systems for detecting fall-related accidents have been developed and have yielded satisfactory accuracy in laboratory settings. Nevertheless, the existing systems fail to uphold consistent accuracy and produce a significant number of false alarms when deployed in real-world settings, primarily due to the intricate nature of the working environments and the behaviors of the workers. In this research, the authors redesign the aforementioned laboratory experiment to target situations that are prone to false alarms based on the feedback obtained from workers in real construction sites. In addition, a new algorithm based on recurrent neural networks was developed to reduce the frequencies of various types of false alarms. The proposed model outperforms the existing benchmark model (i.e., hierarchical threshold model) with higher sensitivities and fewer false alarms in detecting stumble (100% sensitivity vs. 40%) and fall (95% sensitivity vs. 65%) events. However, the model did not outperform the hierarchical model in detecting coma events in terms of sensitivity (70% vs. 100%), but it did generate fewer false alarms (5 false alarms vs. 13).

Construction of learning model under computational and energy constraints, particularly in highly limited training time requirement is a critical as well as unique necessity of many practical IoT applications that use time series sensor signal analytics for edge devices. Yet, majority of the state-of-the-art algorithms and solutions attempt to achieve high performance objective (like test accuracy) irrespective of the computational constraints of real-life applications. In this paper, we propose Instant Adaptive Learning that characterizes the intrinsic signal processing properties of time series sensor signals using linear adaptive filtering and derivative spectrum to efficiently construct a low-cost learning model followed by standard classification algorithms. Our empirical studies on a number of time series sensor signals from publicly available time series database (UCR) demonstrate that with slight trade-off in performance, the proposed method achieves very fast learning capability.

Koopman operator theory provides a framework for nonlinear dynamical system analysis and time-series forecasting by mapping dynamics to a space of real-valued measurement functions, enabling a linear operator representation. Despite the advantage of linearity, the operator is generally infinite-dimensional. Therefore, the objective is to learn measurement functions that yield a tractable finite-dimensional Koopman operator approximation. In this work, we establish a connection between Koopman operator approximation and linear Recurrent Neural Networks (RNNs), which have recently demonstrated remarkable success in sequence modeling. We show that by considering an extended state consisting of lagged observations, we can establish an equivalence between a structured Koopman operator and linear RNN updates. Building on this connection, we present SKOLR, which integrates a learnable spectral decomposition of the input signal with a multilayer perceptron (MLP) as the measurement functions and implements a structured Koopman operator via a highly parallel linear RNN stack. Numerical experiments on various forecasting benchmarks and dynamical systems show that this streamlined, Koopman-theory-based design delivers exceptional performance.

Technological advances in industry have made it possible to install many connected sensors, generating a great amount of observations at high rate. The advent of Industry 4.0 requires analysis capabilities of heterogeneous data in form of related multivariate time series. However, missing data can degrade processing and lead to bias and misunderstandings or even wrong decision-making. In this paper, a recurrent neural network-based denoising autoencoder is proposed for gap imputation in related multivariate time series, i.e., series that exhibit spatio-temporal correlations. The denoising autoencoder (DAE) is able to reproduce input missing data by learning to remove intentionally added gaps, while the recurrent neural network (RNN) captures temporal patterns and relationships among variables. For that reason, different unidirectional (simple RNN, GRU, LSTM) and bidirectional (BiSRNN, BiGRU, BiLSTM) architectures are compared with each other and to state-of-the-art methods using three different datasets in the experiments. The implementation with BiGRU layers outperforms the others, effectively filling gaps with a low reconstruction error. The use of this approach is appropriate for complex scenarios where several variables contain long gaps. However, extreme scenarios with very short gaps in one variable or no available data should be avoided.

Geospatial data have been transformative for the monitoring of the Earth, yet, as in the case of (geo) physical monitoring, the measurements can have variable spatial and temporal sampling and may be associated with a significant level of perturbations degrading the signal quality. Denoising geospatial data is, therefore, essential, yet often challenging because the observations may comprise noise coming from different sources, including both environmental signals and instrumental artifacts, which can be spatially and temporally correlated, thus hard to disentangle. This study addresses the denoising of multivariate time series acquired by irregularly distributed networks of sensors, requiring specific methods to handle the spatiotemporal correlation of the noise and the signal of interest. Specifically, our method focuses on the denoising of geodetic position time series, used to monitor ground displacement worldwide with centimeter-to-millimeter precision. Among the signals affecting global navigation satellite system (GNSS) data, slow slip events (SSEs) are of interest to seismologists. These are transients of deformation that are weakly emerging compared to other signals. Here, we design SSEdenoiser, a multistation spatiotemporal graph-based attentive denoiser that learns latent characteristics of GNSS noise to reveal SSE-related displacement with submillimeter precision. It is based on the key combination of graph recurrent networks and spatiotemporal Transformers. The proposed method is applied to the Cascadia subduction zone, where SSEs occur along with bursts of tectonic tremors, a seismic rumbling identified from independent seismic recordings. The extracted events match the spatiotemporal evolution of tremors. This good space–time correlation of the denoised GNSS signals with the tremors validates the proposed denoising procedure.

Soft sensor modeling for dynamic processes has become a trending topic and a pending challenge in industrial data analysis, especially in limited labeled data scenarios. Alternatively, data augmentation strategies provide a way to address the deficiency of samples. However, current time-series data augmentation methods do not consider the spatiotemporal dependencies among samples during the data generation procedure. To address the issue, a time-series denoising diffusion probabilistic model (TimeDDPM) is proposed to construct a soft sensor for finite time-series samples. First, the long short-term memory (LSTM) units and 1-D convolutional neural networks are implemented in the noise prediction network of TimeDDPM to mine both temporal and spatial properties of samples. Then, virtual samples are reconstructed step by step in the reverse process to enlarge the sample space of insufficient data. Finally, based on the augmented samples, the LSTM network is constructed as a base model to evaluate the quality of new training data. Two cases are employed to demonstrate the superiorities of the proposed method in comparison to several cutting-edge methods.

In the existing rolling bearing performance degradation assessment methods, the input signal is usually mixed with a large amount of noise and is easily disturbed by the transfer path. The time information is usually ignored when the model processes the input signal, which affects the effect of bearing performance degradation assessment. To solve the above problems, an end-to-end performance degradation assessment model of railway axle box bearing based on a deep residual shrinkage network and a deep long short-term memory network (DRSN-LSTM) is proposed. The proposed model uses DRSN to extract local abstract features from the signal and denoises the signal to obtain the denoised feature vector, then uses deep LSTM to extract the time-series information of the signal. The healthy time-series signal of the rolling bearing is input into the DRSN-LSTM reconstruction model for training. Time-domain, frequency-domain, and time–frequency-domain features are extracted from the signal both before and after reconstruction to form a multi-domain features vector. The mean square error of the two feature vectors is used as the degradation indicator to implement the performance degradation assessment. Artificially induced defects and rolling bearings life accelerated fatigue test data verify that the proposed model is more sensitive to early failures than mathematical models, shallow networks or other deep learning models. The result is similar to the development trend of bearing failures.

Abstract Detecting anomalies in time series data is becoming mainstream in a wide variety of industrial applications in which sensors monitor expensive machinery. The complexity of this task increases when multiple heterogeneous sensors provide information of different nature, scales and frequencies from the same machine. Traditionally, machine learning techniques require a separate data pre-processing before training, which tends to be very time-consuming and often requires domain knowledge. Recent deep learning approaches have shown to perform well on raw time series data, eliminating the need for pre-processing. In this work, we propose a deep learning based approach for supervised multi-time series anomaly detection that combines a Convolutional Neural Network (CNN) and a Recurrent Neural Network (RNN) in different ways. Unlike other approaches, we use independent CNNs, so-called convolutional heads, to deal with anomaly detection in multi-sensor systems. We address each sensor individually avoiding the need for data pre-processing and allowing for a more tailored architecture for each type of sensor. We refer to this architecture as Multi-head CNN–RNN. The proposed architecture is assessed against a real industrial case study, provided by an industrial partner, where a service elevator is monitored. Within this case study, three type of anomalies are considered: point, context-specific, and collective.The experimental results show that the proposed architecture is suitable for multi-time series anomaly detection as it obtained promising results on the real industrial scenario.

Time series data naturally exist in many domains including medical data analysis, infrastructure sensor monitoring, and motion tracking. However, a very small portion of anomalous time series can be observed, comparing to the whole data. Most existing approaches are based on the supervised classification model requiring representative labels for anomaly class(es), which is challenging in real-world problems. So can we learn how to detect anomalous time ticks in an effective yet efficient way, given mostly normal time series data? Therefore, we propose an unsupervised reconstruction model named BeatGAN which learns to detect anomalies based on normal data, or data which majority of samples are normal. BeatGAN provides a framework to adversarially learn to reconstruct, which can cooperate with both 1-d CNN and RNN. Rarely observed anomalies can result in larger reconstruction errors, which are then detected based on extreme value theory. Moreover, data augmentation with dynamic time warping regularizes reconstruction and provides robustness. In the experiments, effectiveness and sensitivity are studied in both synthetic data and various real-world time series. BeatGAN achieves better accuracy and fast inference.

Controlled-Source Audio Frequency Magnetotellurics (CSAMT) is an artificial-source electromagnetic technique that partially mitigates the limitations of weak natural field signals. However, practical field surveys inevitably encounter strong interference, severely affecting signal quality. Traditional methods like Fourier transformation, which directly computes apparent resistivity from frequency-domain information, are inadequate for this context, so we need alternative denoising approaches. However, research on CSAMT denoising is currently limited. Given the excellent performance of Long Short-Term Memory (LSTM) neural networks in the processing of Magnetotelluric (MT) data, as demonstrated by previous studies, this paper proposes the use of LSTM to denoise CSAMT signals in the time domain. Unlike traditional denoising methods, we aim to directly extract the target frequency signal from the time series data for denoising. For MT data, target frequency signals and noise are all mixed together, so noise suppression can only be achieved by identifying noise characteristics in the time series. However, unlike MT data, CSAMT data has an artificial transmitting source, and the frequency of the valid signal is fixed within a time interval. This allows for the direct extraction of target frequency signals without considering the complex characteristics of noise. In this study, we developed a neural network based on bidirectional LSTM to accomplish the task of noise suppression. After conducting both simulated and measured data tests, this method was able to, on average, improve the signal-to-noise ratio (SNR) of CSAMT data by approximately 20dB and partially address the challenge of denoising when the data's SNR falls below 0dB.

Multi-dimensional force sensors are of great significance to improve the perception of robots. It's very important to remove the drift and noise of the multi-dimensional force sensor signal caused by environmental changes. Recurrent Neural Network based on Long-Short Term Memory (LSTM-RNN) is proposed for real-time signal processing of multi-dimensional force sensors. Firstly, Adaptive Empirical Mode Decomposition (AEMD) is verified to be effective in removing drift and noise from multi-dimensional force sensor signals. Then, AEMD is utilized to process the force sensor signal and LSTM-RNN is trained by the processed signal. In the force test experiment, the errors of different signals processed by LSTM-RNN are very small and smaller than those of RNN signal processing, which proves that the trained LSTM-RNN can effectively process multi-dimensional force sensor signals in real time.

The anomaly detection in Internet of Things (IoT) sensor data has become an important research area because of the possibility of noise and unavailability of labels in the sensors readings. The conventional machine learning algorithms cannot detect the anomalies when there is high correlation between the data points of the sensor data. Further, the volume and velocity of the data generated by the sensors in the IoT also a reason that the conventional statistical and machine learning algorithms fails to detect the anomalies. In recent years, the Deep Learning (DL) is gaining significant attention in the anomaly detection research due to the property of unsupervised learning of the high volume data and high detection accuracy of abnormalities. To this end, this paper proposed to study three DL models such as Autoencoders, Long Short Term Memory (LSTM) Autoencoder, and LSTM Recurrent Neural Networks (LSTM-RNN) for detecting anomalies in time series IoT sensor data. Simulations have been conducted using the Intel Berkeley Research Labs (IBRL) Sensor data to evaluate the performance. The results reveal which method performed better in terms of detection accuracy and training time.

Deep learning has achieved significant success on intelligent medical treatments, such as automatic diagnosis and analysis of medical data. To train an automatic diagnosis system with high accuracy and strong robustness in healthcare, sufficient training data are required when using deep learning-based methods. However, given that the data collected by sensors that are embedded in medical or mobile devices are inadequate, it is challenging to train an effective and efficient classification model with state-of-the-art performance. Inspired by generative adversarial networks (GANs), we propose TS-GAN, a Time-series GAN architecture based on long short-term memory (LSTM) networks for sensor-based health data augmentation, thereby improving the performance of deep learning-based classification models. TS-GAN aims to learn a generative model that creates time-series data with the same space and time dependence as the real data. Specifically, we design an LSTM-based generator for creating realistic data and an LSTM-based discriminator for determining how similar the generated data are to real data. In particular, we design a sequential-squeeze-and-excitation module in the LSTM-based discriminator to better understand space dependence of real data, and apply the gradient penalty originated from Wasserstein GANs in the training process to stabilize the optimization. We conduct comparative experiments to evaluate the performance of TS-GAN with TimeGAN, C-RNN-GAN and Conditional Wasserstein GANs through discriminator loss, maximum mean discrepancy, visualization methods and classification accuracy on health datasets of ECG_200, NonInvasiveFatalECG_Thorax1, and mHealth, respectively. The experimental results show that TS-GAN exceeds other state-of-the-art time-series GANs in almost all the evaluation metrics, and the classifier trained on synthetic datasets generated by TS-GAN achieves the highest classification accuracy of 97.50% on ECG_200, 94.12% on NonInvasiveFatalECG_Thorax1, and 98.12% on mHealth, respectively.

Some multimedia data from real life can be collected as multivariate time series data, such as community-contributed social data or sensor data. Many methods have been proposed for multivariate time series forecasting. In light of its importance in wide applications including traffic or electric power forecasting, appearance of the Transformer model has rapidly revolutionized various architectural design efforts. In Transformer, self-attention is used to achieve state-of-the-art prediction, and further studied for time series modeling in the frequency recently. These related works prove that self-attention mechanisms can reach a satisfied performance whether in time or frequency domain, but we used recurrent neural network (RNN) to verify that these are not critical and necessary. The correlation structure of RNN has time series specific inductive bias, but there are still some shortcomings in long multivariate time series forecasting. To break the forecasting bottleneck of traditional RNN architectures, we introduced RNNGAN, a novel and competitive RNN-based architecture combining the generation capability of Generative Adversarial Network (GAN) with the forecasting power of RNN. Differentiated from the Transformer, RNNGAN uses long short-term memory (LSTM) instead of the self-attention layers to model long-range dependencies. The experiment shows that, compared with the state-of-the-art models, RNNGAN can obtain competitive scores in many benchmark tests when training on multivariate time series datasets in many different fields.

Two-dimensional (2D) convolutional neural networks (CNNs) are implemented for machinery fault diagnosis owing to CNN’s capability of feature extraction from input data. One-dimensional (1D) signals are converted to two-dimensional (2D) data because 2D image data is a much more powerful information representation. However, the computation complexity of 2D CNN is high due to the 2D operation on many stages, and it requires more data to achieve higher training performance. 1D CNNs are implemented to utilize 1D signal directly, which reduces the computational complexity and enables real-time classification. In this paper, we compare the performance of the 1D CNN and 2D CNN for the multi-class classification of time-series sensor data. The architecture of each CNN considered in this work has a single convolutional layer and one fully connected layer. Real measured data are used for the training and testing of the models. 1D signal extracted from sensor signal is converted into 2D data array for the input of the 2D CNN. Classification accuracy and time complexity of both CNNs are evaluated for a given dataset. Via simulations, we show that both CNNs classify the time series data with high accuracy, achieve approximately the same classification performance.

Medical time-series analysis is crucial for the diagnosis and treatment of various diseases. As modern medical sensors evolve, the complexity and dimensionality of medical signals have increased, posing challenges for data labeling and classification. Self-supervised learning has emerged as a promising solution by automatically extracting meaningful feature representations from raw data without manually annotated labels. However, existing research has often focused on the time-point or instance level input and overlook the intricate relationships among different channels, which are crucial for precise data representation. In this study, we propose CLHi-MTS, a novel Contrastive Learning-Based Hierarchical Framework for Masked Medical Time-Series Modeling that leverages three distinct levels of proximity: channel interdependencies, temporal fluctuations, and instance-level correlations, to guide the reconstruction of masked segments enhanced with contrastive learning. We also introduce a new augmentation strategy for generating positive examples in contrastive learning, further enhancing the model’s representation learning ability. Evaluated on five datasets for both in-domain and cross-domain medical time-series classification tasks, our framework outperforms six state-of-the-art self-supervised methods.

Deep time series anomaly detection (TSAD) essentially relies on learning data "normality". Current approaches leverage various neural network architectures, including RNNs, CNNs, Transformers, and graph neural networks, effectively modeling temporal and inter-variable dependencies within time series data. However, these approaches often overlook the equally crucial asynchronous inter-variable dependencies, potentially missing anomalous patterns that arise from intricate, temporally offset interactions among variables. Therefore, we propose a novel time series anomaly detection method by perceiving comprehensive intrinsic dependencies of time series data. Specifically, we propose a Comprehensive Dependency-aware Attention (CDA) that computes attention between each element and its criss-cross counterparts (i.e., intra-variable values across the whole temporal dimension and synchronous values of different variables) and captures asynchronous inter-variable dependency by repeating this step. Additionally, to facilitate more effective downstream dependency learning, we propose a Frequency Domain Rectifier (FDR) to capture the signal’s underlying patterns and reduce noise in the frequency domain. We replace the self-attention in the classic Transformer with CDA, and introduce the FDR module to construct a new method for TSAD, i.e., Frequency-enhanced Comprehensive Dependency Attention-based Time Series Anomaly Detection (FCDATA). Extensive experiments show that: (i) our detection method significantly outperforms state-of-the-art competitors by 15% in the F1-score; and (ii) by incorporating the proposed network architecture, existing anomaly detectors achieve over 20% performance gain compared to current temporal network structures.

No abstract available

This paper presents a novel approach for bearing fault diagnosis in induction motor utilizing an improved hybrid Continuous Wavelet Transform-Deep Convolutional Neural Network-Long Short-Term Memory (CWT-DCNN-LSTM) model. The vibration data, recorded using an low-cost ADXL355 accelerometer, was preprocessed by converting the one-dimensional (1D) signals into two-dimensional (2D) images using Continuous Wavelet Transform (CWT). The dataset, comprising 13 classes with varying fault conditions, was segmented and shuffled before model training. Three datasets, corresponding to different load conditions (100W, 200W, and 300W), were used to evaluate the model’s performance. Experimental results demonstrated high training accuracy of 100% and validation accuracies of 96.43%, 97.47%, and 95.06% for the 100W, 200W, and 300W load conditions, respectively. Validation losses were recorded at 12.33%, 9.81%, and 20.33% for the respective loads. Furthermore, performance results using accuracy, sensitivity, specificity, balanced accuracy and geometric mean were computed for all three load conditions. The results indicate the robustness and effectiveness of the proposed CWT-DCNN-LSTM model for bearing fault diagnosis of induction motor using low-cost ADXL335 accelerometer, highlighting its potential for real-world industrial applications.

No abstract available

This study presents a contactless vibration measurement system leveraging a hybrid computer vision and neural network approach. The proposed framework combines Lucas-Kanade (L-K) Optical Flow with LSTM regression to predict acceleration metrics from high-speed video sequences. Image data captured via Raspberry Pi 5 and Pi Camera are processed into spatiotemporal pixel vectors, which serve as input to the LSTM network for regression-based acceleration prediction. For ground truth validation, synchronized measurements are obtained from a direct-connect accelerometer paired with a controlled vibration motor, utilizing multiprocessing techniques to ensure efficient dataset curation. Through systematic evaluation of architectural configurations, the LSTM-64-32 model with 20-step input sequences demonstrated optimal performance, achieving a training loss of 0.425 mm2/s4 squared differences compared to physical sensor readings. While results validate the feasibility of optical flow for vibration analysis, limitations in dataset diversity were identified as a contributor to model overfitting. This work advances non-contact vibration sensing by integrating low-cost hardware with deep learning, offering a scalable alternative to traditional accelerometer-based methods in structural health monitoring and industrial applications.

Mobile technologies, particularly smartphones, have become integral to human life, influencing nearly every aspect of daily activities. Their ubiquity and advanced capabilities have led to them being a solution to a wide range of problems. Recent research has focused on leveraging the existing sensors in smartphones to supplant more costly and less accessible dedicated hardware. In this vein, we introduce MobiScale, a novel system designed to measure the weight of light objects using the built-in accelerometer and vibration motor of smartphones. This system utilizes advanced feature extraction techniques to effectively capture the temporal and spectral characteristics of the weight data. The core of MobiScale is its CNN-LSTM architecture, which processes these features to estimate weights accurately. The implementation of various regularization techniques has been vital in enhancing the system's ability to generalize, thereby improving its performance. Tested across multiple devices and under various conditions, MobiScale has shown consistent and precise performance, achieving a mean square error of just 12 grams in weight estimation, underscoring its potential as a versatile and reliable tool for weight measurement using smartphone technology.

Underground high-voltage transmission cables, especially high-pressure fluid-filled (HPFF) pipe-type cable systems, are critical components of urban power networks. These systems consist of insulated conductor cables housed within steel pipes filled with pressurized fluids that provide essential insulation and cooling. Despite their reliability, HPFF cables experience faults caused by insulation degradation, thermal expansion, and environmental stressors, which, due to their subtle and gradual nature, complicate incipient fault detection and subsequent fault localization. This study presents a novel, proactive, and retrofit-friendly predictive condition monitoring method. It leverages distributed accelerometer sensors non-intrusively mounted on the HPFF steel pipe within existing manholes to continuously monitor vibration signals in real time. A physics-enhanced convolutional neural network–long short-term memory (CNN–LSTM) deep learning architecture analyzes these signals to detect incipient faults before they evolve into critical failures. The CNN–LSTM model captures temporal dependencies in acoustic data streams, applying time-series analysis techniques tailored for the predictive condition monitoring of HPFF cables. Experimental validation uses vibration data from a scaled-down HPFF laboratory test setup, comparing normal operation to incipient fault events. The model reliably identifies subtle changes in sequential acoustic patterns indicative of incipient faults. Laboratory experimental results demonstrate a high accuracy of the physics-enhanced CNN–LSTM architecture for incipient fault detection with effective data feature extraction. This approach aims to support enhanced operational resilience and faster response times without intrusive infrastructure modifications, facilitating early intervention to mitigate service disruptions.

Elevators play an essential role for safe and efficient vertical transportation in modern buildings. However, traditional maintenance strategies based on scheduled inspections often fail to detect faults that appear between inspection intervals, this leads to unplanned downtime. This letter presents an Internet of Things-based elevator health monitoring system that integrates vibration sensing, wireless communication, and machine learning to enable real-time condition monitoring. The system utilizes an accelerometer sensor interfaced with an microcontroller to capture elevator vibration and motion data. The collected data are transmitted to a cloud server, where a long short-term memory-based autoencoder is used for anomaly detection by calculating threshold based on reconstruction error. The proposed system detects real-time fault data with 98.33% accuracy and 100% recall, validated through four-month experimental study.

The integration of the use of Artificial Intelligence (AI) in industrial environments often faces the lack of labeled data, as well as historical records. This lack of information becomes a problem when implementing predictive maintenance solutions, particularly in monitoring the condition of industrial machines and automatic fault detection. This work addresses this issue in an industrial scenario, through the analysis of vibrations in a spindle motor of an ornamental stone cutting machine. Unsupervised learning techniques are explored for anomaly detection through vibration data, using the training and implementation of an LSTM (Long Short-Term Memory) Autoencoder model. Datasets consist only of unlabeled accelerometer signals acquired during normal machine operation. An analysis based on the extraction of statistical features from the signal is adopted to use them as inputs of the Machine Learning algorithm, to learn the normal behavior of the machine and detect deviations that may correspond to potential anomalies. The experimental results show that even in the absence of labeled data, it is possible to extract meaningful insights from the machine state and establish a practical pipeline for anomaly detection in industrial machines through vibration analysis.

Nowadays, most smartphones are equipped with accelerometer sensor, which can be used to record acceleration caused by movements or vibration. It opens the opportunity to use them as a personal earthquake early warning system where people can use their smartphone to detect the incoming earthquake. However, to avoid false alarm, we must be able to recognize the source of sensor acceleration. One of the most common sources of vibration recorded by smartphone accelerometer is human activities. In this work, we used machine learning to distinguish the source of movement by observing the acceleration value caused by human activities and earthquakes. RNN-LSTM neural network is trained with labeled time series acceleration data and used to recognize the movement source. Our model show potential to differentiate between human activity and earthquake movement with training accuracy value of 97% and test loss value equal to 0.3.

Structural health monitoring (SHM) of long-span bridges is frequently challenged by non-stationary behavior, superimposed multi-scale features, and significant noise interference. Traditional physics-based methods often struggle to accurately capture the complex data patterns to predict structural dynamic responses. To address these limitations, this study proposes a deep learning framework integrating adaptive signal decomposition with intelligent optimization for dynamic response prediction of long-span bridges. Specifically, an improved method based on variational modal decomposition and long short-term memory is developed to predict the structural responses of a steel-concrete composite rib-arch bridge. The SHM data were continuously collected from the arch bridge over a of 31 d period, with wind speed, equivalent vehicle load, and structural temperature designed as input variables to model key response indicators, including crown displacement, deflection, strain, cable force, and vibration acceleration. Adaptive signal decomposition facilitates multi-scale features and noise suppression, while intelligent optimization enhances hyperparameter tuning for time-series modeling. Numerical results demonstrate reasonable prediction accuracy for the key responses. The coefficient of determination (R2) for the predictions of strain, crown displacement and deflection reached 0.986, 0.968, and 0.975, respectively. Notably, for highly non-stationary vibration acceleration signals, the proposed framework significantly outperformed mainstream baseline models, reducing multiple error metrics to minimal levels. The proposed method has advantages in terms of prediction accuracy, stability, and generalization, which provides robust support for condition assessment and early safety warnings of long-span bridges.

Weak light intensity positions induced by interference fading adversely affect the sensing performance of phase-sensitive optical time-domain reflectometry (Φ-OTDR). Most effective fading suppression methods rely on frequency or phase modulation of the light source, which requires complex hardware modifications. To solve the above issue, this paper proposes a novel multi-channel data synthesizing method based on deep neural network (MDS-DNN) to reduce the impact of interference fading on the signal-to-noise ratio (SNR) of Φ-OTDR. The proposed algorithm can work efficiently without any modification of the conventional Φ-OTDR setup. The spatial sampling rate of the Φ-OTDR systems is typically much higher than the spatial resolution. This means that neighboring sampling points carry the same external vibration signal, providing redundant information. Therefore, it is possible to perform comprehensive analysis on these multi-channel data to improve the suppression capability of interference fading noise. This work designs a long short-term memory (LSTM) network-based framework and an end-to-end training strategy to automatically learn the correlation between these multi-channel data and the ideal sensing signal. Simulation and experimental results show that the MDS-DNN algorithm can effectively suppress phase noise and improve the SNR at fading positions. Experiments using the data collected from the actual Φ-OTDR system demonstrate that the output SNR can reach 49.88 dB, which is 19.65 dB higher than the average level of the input channels. Moreover, the MDS-DNN method reduces the false alarm rate caused by interference fading by one order.

Quadcopters are widely used in a variety of military and civilian mission scenarios. Real-time online detection of the abnormal state of the quadcopter is vital to the safety of aircraft. Existing data-driven fault detection methods generally usually require numerous sensors to collect data. However, quadcopter airframe space is limited. A large number of sensors cannot be loaded, meaning that it is difficult to use additional sensors to capture fault signals for quadcopters. In this paper, without additional sensors, a Fault Detection and Identification (FDI) method for quadcopter blades based on airframe vibration signals is proposed using the airborne acceleration sensor. This method integrates multi-axis data information and effectively detects and identifies quadcopter blade faults through Long and Short-Term Memory (LSTM) network models. Through flight experiments, the quadcopter triaxial accelerometer data are collected for airframe vibration signals at first. Then, the wavelet packet decomposition method is employed to extract data features, and the standard deviations of the wavelet packet coefficients are employed to form the feature vector. Finally, the LSTM-based FDI model is constructed for quadcopter blade FDI. The results show that the method can effectively detect and identify quadcopter blade faults with a better FDI performance and a higher model accuracy compared with the Back Propagation (BP) neural network-based FDI model.

No abstract available

Maintaining bridges that support road infrastructure is critical to the economy and human life. Structural health monitoring of bridges using vibration includes direct monitoring and drive-by monitoring. Drive-by monitoring uses a vehicle equipped with accelerometers to drive over bridges and estimates the bridge’s health from the vehicle vibration obtained. In this study, we attempt to identify the driving segments on bridges in the vehicle vibration data for the practical application of drive-by monitoring. We developed an in-vehicle sensor system that can measure three-dimensional behavior, and we propose a new problem of identifying the driving segment of vehicle vibration on a bridge from data measured in a field experiment. The “on a bridge” label was assigned based on the peaks in the vehicle vibration when running at joints. A supervised binary classification model using C-LSTM (Convolution—Long-Term Short Memory) networks was constructed and applied to data measured, and the model was successfully constructed with high accuracy. The challenge is to build a model that can be applied to bridges where joints do not exist. Therefore, future work is needed to propose a running label on bridges based on bridge vibration and extend the model to a multi-class model.

Artificial intelligence technologies are widely applied in the field of power equipment condition monitoring, with convolutional neural networks (CNNs) demonstrating strong adaptability in extracting features from non-stationary vibration signals. This study aims to enhance the accuracy and real-time performance of transformer fault diagnosis by constructing an intelligent diagnostic method based on CNNs. Methodologically, raw vibration signals under multiple operating conditions were first collected using triaxial accelerometers. Short-Time Fourier Transform (STFT) was applied to construct a three-channel two-dimensional time-frequency map. Subsequently, a multi-layer CNN architecture was designed, integrating residual connections and an adaptive learning rate decay strategy to enhance the model's feature modeling capability in noisy environments. A real-world dataset from a 220kV main transformer was employed for training and validation. Results demonstrate that the constructed model achieves accuracy ranges between 90% and 96% across five operating conditions, with a maximum F1 score of 0.95, significantly outperforming comparison models including 1D-CNN, SVM, and LSTM. Findings indicate that this approach effectively enhances fault identification accuracy and stability, providing a foundational model and strategic support for subsequent online monitoring system deployment.

The switch and crossing (S&C) is one of the most important parts of the railway infrastructure network due to its significant influence on traffic delays and maintenance costs. Two central questions were investigated in this paper: (I) the first question is related to the feasibility of exploring the vibration data for wear size estimation of railway S&C and (II) the second one is how to take advantage of the Artificial Intelligence (AI)-based framework to design an effective early-warning system at early stage of S&C wear development. The aim of the study was to predict the amount of wear in the entire S&C, using medium-range accelerometer sensors. Vibration data were collected, processed, and used for developing accurate data-driven models. Within this study, AI-based methods and signal-processing techniques were applied and tested in a full-scale S&C test rig at Lulea University of Technology to investigate the effectiveness of the proposed method. A real-scale railway wagon bogie was used to study different relevant types of wear on the switchblades, support rail, middle rail, and crossing part. All the sensors were housed inside the point machine as an optimal location for protection of the data acquisition system from harsh weather conditions such as ice and snow and from the ballast. The vibration data resulting from the measurements were used to feed two different deep-learning architectures, to make it possible to achieve an acceptable correlation between the measured vibration data and the actual amount of wear. The first model is based on the ResNet architecture where the input data are converted to spectrograms. The second model was based on a long short-term memory (LSTM) architecture. The proposed model was tested in terms of its accuracy in wear severity classification. The results show that this machine learning method accurately estimates the amount of wear in different locations in the S&C.

Bearing failures cause machinery breakdowns, resulting in financial losses due to production downtimes. To address this, accurate bearing condition monitoring is essential. This paper introduces a cross-domain approach to fault diagnosis using a combination of convolutional neural networks (CNNs) and long short-term memory (LSTM) models, applied to the Case Western Reserve University (CWRU) dataset and the University of Ottawa Rolling-element Dataset- Vibration and Acoustic Faults under Constant Load and Speed conditions (UORED-VAFCLS), which contain both artificial and naturally developed bearing faults. The proposed experimental framework assesses the estimators, training and testing them with raw time-domain data from both acoustic and accelerometer signals, enhancing fault detection across various operating conditions. Results demonstrate that the CNN-LSTM model, when combined with statistical preprocessing, outperforms advanced models in both performance, computational time, and stability, particularly when fusing data from multiple sources. This approach shows promise for practical implementations in industrial predictive maintenance, offering a more reliable solution for reducing downtime and improving operational efficiency. Future work will focus on further optimization of the model and minimizing the data required for effective condition monitoring.

This paper establishes an integrated framework combining self-induced vibration measurements with deep learning for vibration-based remaining useful life (RUL) prediction of mechanical frame structures in mobile robots. The main innovations comprise (1) a self-induced vibration excitation system that utilizes the robot’s drive wheels to generate controlled mechanical oscillations, using a five-sensor micro-electro-mechanical system (MEMS) accelerometer array to capture non-uniform vibration mode shapes across the robot’s structure, and (2) a processing pipeline for RUL prediction using accelerometer data and early feature fusion in two machine-learning models (long short-term memory (LSTM) and a convolutional neural network (CNN)). Our research methodology includes (i) modal analysis to identify the robot’s natural frequencies, (ii) verification platform evaluation, comparing low-cost MEMS accelerometers against a reference integrated electronic piezoelectric (IEPE) accelerometer, demonstrating industrial-grade measurement quality (coherence > 98%, uncertainty 4.79–7.21%), and (iii) data-driven validation using real data from the mechanical frame, showing that the LSTM model outperforms the CNN with a 2.61× root-mean-square error (RMSE) improvement (R2 = 0.99). Our solution demonstrates that early feature fusion provides sufficient information to model degradation and detect faults early at a lower cost, offering a feasible alternative to classical maintenance procedures through combined hardware validation and lightweight software suitable for Industrial Internet-of-Things (IIoT) deployment.

No abstract available

Bearing clearance is a common issue in mechanical systems due to unavoidable assembly errors, leading to weak fault features that are challenging to detect. This study introduces a novel diagnostic technique for detecting bearing clearance faults using the Elman Neural Network (ENN) based Long Short-Term Memory (LSTM). The raw vibration data from an accelerometer is processed using the Fast Fourier Transform (FFT) to extract frequency-domain features. ENN is employed to identify clearance faults under various operating conditions, while LSTM captures temporal dependencies in the data. This hybrid ENN-LSTM approach eliminates the need for manual feature extraction, reducing the risk of errors associated with expert-driven methods. The proposed method demonstrates robust generalization performance and achieves an average fault identification accuracy of 99.16% across different operating conditions. This research offers valuable insights for improving fault diagnostics in rotor-bearing systems.

In practical industrial applications, obtaining a sufficient number fault samples for specific types of equipment fault can be challenging. As a result, there are frequently significantly fewer defect samples obtained than healthy samples, and the data samples that are obtained typically have a high noise level. To overcome these issues, this paper introduces a novel approach termed the improved hybrid dilated convolution network (HDCN) to address these limitations and enhance classification accuracy. The proposed method involves transforming the time domain vibration signal into a time-frequency domain image using short time fourier transform (STFT), enabling simultaneous extraction of frequency domain and time domain features. A multi-scale hybrid dilated convolution network is constructed to extract multiple scale fault features and identify characteristic information. Subsequently, an adaptive weight long short-term memory (LSTM) unit is designed to perform weighted fusion of multi-scale features. It can be amplifying the contribution of important features and minimizing the influence of non-relevant features. The scaled exponential linear unit (SELU) is utilized to mitigate the significant suppression of the activation function on a few class samples. Finally, the network model is simulated using the focal loss function to make it more suitable for the case where the fault samples are small and confusing. To assess the effectiveness of the suggested approach, extensive tests are carried out on simulated datasets as well as a public dataset.

Faults in an Induction Motor (IM) can lead to unexpected downtime, resulting in considerable economic and productivity losses. From existing literature, conventional fault diagnosis approaches in an IM struggle to reliably identify fault patterns at different speeds, particularly under variable speed and changing load conditions. To resolve this issue, this paper presents a unique hybrid Convolutional Neural Network (CNN) along with the Long Short Term Memory (LSTM) topology for diagnosing faulty patterns in an IM under loaded and unloaded variable speed settings. The proposed method can identify faults such as rotor imbalances, misalignment, stator winding issues, voltage imbalances, broken rotor bars, and broken bearings. Experiments performed using the University of Ottawa Electric Motor Dataset – Vibration and Acoustic Faults under Constant and Variable Speed Conditions (UOEMD-VAFCVS) dataset reveals that all three accelerometers are 99.93% accurate at constant speed and 99.96% at variable speed under both loaded and unloaded conditions. In terms of fault diagnostic accuracy in an IM operating at different speeds and load conditions, this methodology outperforms cutting-edge methodologies in the literature. Moreover, using the publicly available CWRU dataset, this study validates the robustness of the proposed methodology in terms of operational issues in an IM. Finally, the proposed method achieves incredible results at varying speeds, stressing the need to improve industrial equipment reliability and maintenance methods.

Predictive maintenance in industrial settings has seen significant advancements through the integration of Internet of Things (IoT) technologies and data analytics. This study outlines a comprehensive methodology for implementing IoT-Driven Predictive Maintenance, focusing on data collection, machine learning model selection, real-time data analysis, and continuous improvement. Data Collection and Sensor Deployment details the deployment of sensors on machinery, including temperature, vibration, pressure, humidity, and accelerometer sensors. Real-time data collection, strategic sensor placement, and data preprocessing ensure the acquisition of high-quality data for analysis. Machine Learning Model Selection highlights the selection of Random Forest and LSTM models for predictive maintenance. These models are trained using historical data, cross-validated, and fine-tuned to optimize accuracy. Real-Time Data Analysis and Integration emphasizes the importance of real-time analysis using streaming analytics tools. Alerts are generated when sensor data deviates from predefined thresholds, and the integration with maintenance workflows ensures timely responses. Performance Evaluation and Continuous Improvement presents performance metrics, including precision, recall, and F1-score, demonstrating the effectiveness of the predictive maintenance system. Reduced downtime and decreased maintenance costs showcase its impact. Continuous improvement involves regular updates to models and data preprocessing techniques, with a focus on adapting to emerging technologies.

Structural Health Monitoring (SHM) is essential in the aetiology and maintenance of the safety and durability of civil infrastructures. The current manuscript outlines an Artificial Intelligence (AI) powered computational framework for real-time SHM using sensing Internet of Things (IoT) systems. The proposed architecture has multiple sensor modalities incorporated into edge-based ai models, namely strain gauges, accelerometers, and vibration sensors, to enable the concept of detecting anomalies in their early stages and locating the damage. A combination deep learning approach of Convolutional Neural Networks (CNNs) and Long ShortTerm Memory (LSTM) networks is used in combination to process heterogeneous sensor time series data in real time. Experimental evaluations performed with simulated datasets from bridges and buildings show that the model achieves an accuracy of more than 95 per cent in categorizing faults, and the mean squared error (MSE) reduction of 18 per cent with regard to the conventional baseline techniques. This study highlights the game-changing power of AI-IoT combination in terms of predictive maintenance, which can lead to cost savings, improved safety and resilience of critical infrastructure.

Global Navigation Satellite System (GNSS) time series are widely used for structural health monitoring (SHM) and deformation analysis, but real-world recordings frequently contain short and long contiguous gaps that degrade downstream interpretation. This study addresses the challenge of accurate imputation for GNSS displacement series by proposing a self-supervised learning framework that trains models directly on real, unlabelled data using contiguous-span masking. We evaluate four neural architectures (ANN, CNN, GRU, LSTM) under a unified pipeline comprising signal denoising (moving-average, Kalman smoothing, Haar wavelet), sliding-window segmentation, Z-score normalization, and middle-region masking. Experiments use a year-long 10-minute-sampled dataset from the Can Tho cable-stayed bridge (sensor can519501, x/y/z components) and assess reconstruction quality via R², MAE, and MSE on withheld masked segments. Results indicate that recurrent architectures, particularly LSTM, produce the most faithful reconstructions: LSTM attains the highest validation R² (≈0.948) and the lowest MAE (≈0.137) and MSE (≈0.052) among tested models, while GRU offers competitive performance and CNN/ANN show substantially weaker recovery. These findings demonstrate that masking-based self-supervision is an effective strategy for GNSS gap recovery and that LSTM-like sequence models are well suited to capture the long-range temporal dependencies in bridge displacement data. The proposed approach enhances the reliability and continuity of GNSS-derived time series for structural monitoring and can inform future multi-sensor fusion and uncertainty quantification work

To overcome the limitations of traditional periodic power equipment maintenance—characterized by insufficient scientific basis, low reliability, and poor economic efficiency—this study proposes a big data-driven fault early warning system. Integrating Internet of Things (IoT), Artificial Intelligence (AI), and digital twin technologies, the system enables a shift from reactive maintenance to proactive prevention. Its three-layer architecture includes: (1) a Smart Perception Layer collecting real-time operational data via multi-sensor networks; (2) an Edge Processing Layer performing lightweight preprocessing (filtering/denoising/compression); and (3) an Application Service Layer deploying Recurrent Neural Network (RNN)-based prediction models with dynamic threshold algorithms to forecast fault probability and time-to-failure (TTF) windows. Supported by a hybrid 4G/5G-fiber-ZigBee communication architecture and modular design for secure, adaptable operation, the system demonstrably enhances fault response efficiency, reduces maintenance costs, and strengthens grid resilience.

No abstract available

In an industrial IoT setting, ensuring the quality of sensor data is a must when data-driven algorithms operate on the upper layers of the control system. Unfortunately, the common place in industrial facilities is to find sensor time series heavily corrupted by noise and outliers. In this work, a purely data-driven self-supervised learning-based approach based on a blind denoising autoencoder is proposed for real time denoising of industrial sensor data. The term \textit{blind} stresses that no prior knowledge about the noise is required for denoising, in contrast to typical denoising autoencoders where prior knowledge about the noise is required. Blind denoising is achieved by using a noise contrastive estimation (NCE) regularization on the latent space of the autoencoder, which not only helps to denoise but also induces a meaningful and smooth latent space. Experimental evaluation in both a simulated system and a real industrial process shows that the proposed technique outperforms classical denoising methods.

Inertial measurement unit (IMU) sensors are widely used in motion tracking for various applications, e.g., virtual physical therapy and fitness training. Traditional IMU-based motion tracking systems use 9-axis IMU sensors that include an accelerometer, gyroscope, and magnetometer. The magnetometer is essential to correct the yaw drift in orientation estimation. However, its magnetic field measurement is often disturbed by the ferromagnetic materials in the environment and requires frequent calibration. Moreover, most IMU-based systems require multiple IMU sensors to track the body motion and are not convenient for use. In this paper, we propose a novel approach that uses a single 6-axis IMU sensor of a consumer smartwatch without any magnetometer to track the user's 3D arm motion in real time. We use a recurrent neural network (RNN) model to estimate the 3D positions of both the wrist and the elbow from the noisy IMU data. Compared with the state-of-the-art approaches that use either the 9-axis IMU sensor or the combination of a 6-axis IMU and an extra device, our proposed approach significantly improves the usability and potential for pervasiveness by not requiring a magnetometer or any extra device, while achieving comparable results.

Monitoring the status of the facilities and detecting any faults are considered an important technology in a smart factory. Although the faults of machine can be analyzed in real time using collected data, it requires a large amount of computing resources to handle the massive data. A cloud server can be used to analyze the collected data, but it is more efficient to adopt the edge computing concept that employs edge devices located close to the facilities. Edge devices can improve data processing and analysis speed and reduce network costs. In this paper, an edge device capable of collecting, processing, storing and analyzing data is constructed by using a single-board computer and a sensor. And, a fault detection model for machine is developed based on the long short-term memory (LSTM) recurrent neural networks. The proposed system called LiReD was implemented for an industrial robot manipulator and the LSTM-based fault detection model showed the best performance among six fault detection models.

No abstract available

No abstract available

Sensors measuring real-life physical processes are ubiquitous in today's interconnected world. These sensors inherently bear noise that often adversely affects the performance and reliability of the systems they support. Classic filtering approaches introduce strong assumption on the time or frequency characteristics of sensory measurements, while learning-based denoising approaches typically rely on using ground truth clean data to train a denoising model, which is often challenging or prohibitive to obtain for many real-world applications. We observe that in many scenarios, the relationships between different sensor measurements (e.g., location and acceleration) are analytically described by laws of physics (e.g., second-order differential equation). By incorporating such physics constraints, we can guide the denoising process to improve performance even in the absence of ground truth data. In light of this, we design a physics-informed denoising model that leverages the inherent algebraic relationships between different measurements governed by the underlying physics. By obviating the need for ground truth clean data, our method offers a practical denoising solution for real-world applications. We conducted experiments in various domains, including inertial navigation, CO2 monitoring, and HVAC control, and achieved state-of-the-art performance compared with existing denoising methods. Our method can denoise data in real time (4ms for a sequence of 1s) for low-cost noisy sensors and produces results that closely align with those from high-precision, high-cost alternatives, leading to an efficient, cost-effective approach for more accurate sensor-based systems.

No abstract available