Human Activity Recognition (HAR) based on inertial sensors plays a pivotal role in consumer electronics applications such as health monitoring, fitness tracking, and human-computer interaction. However, achieving robust HAR performance remains challenging due to significant variations in activity duration and rhythms, restricted views in attention modeling, and information loss caused by excessive dimensionality compression. In this paper, we propose MSNCVGA, a multi-scale network enhanced by cross-view gated attention for inertial sensor-based HAR tasks. This method explicitly models sequence temporal scale diversity through multi-scale convolutions and a cross-scale feature fusion branch, while utilizing the cross-view gated attention mechanism to synergistically focus on complementary and dependent cross-view global key features among channel, sequence, and sample dimensions, effectively avoiding excessive information loss. Finally, a recurrent neural network is employed to integrate contextual features from different scale branches. Experiments on five public datasets demonstrate that MSNCVGA achieves better performance than the baseline methods, reaching F1-scores of 0.848, 0.989, 0.985, 0.642, and 0.979 on PAMAP2, UCI HAR, mHealth, RealWorld, and KU-HAR, respectively. Notably, the proposed model has only 0.54M parameters, providing a high-accuracy, low-latency solution suitable for deployment on resource-constrained consumer electronic devices.

Accurate indoor localization is fundamental to a wide range of modern engineering applications, from smart environments to autonomous robotics. Information fusion of ultra-wideband (UWB) and inertial measurement unit (IMU) signals shows exceptional potential, as UWB provides high-accuracy ranging while IMU supplies continuous motion tracking, making their fusion particularly effective for complex indoor environments. However, accurate localization and trajectory tracking in dynamic indoor scenarios remain a major challenge. Changes in obstacle types and positions impose stringent requirements on the robustness and adaptability of localization models. To address these challenges, this paper proposed a novel adaptive fusion localization framework, termed the spatiotemporal graph-based adaptive Kalman filter (STG-AKF), which combines a spatiotemporal graph attention network (STGAT) with an adaptive error state Kalman filter (AESKF). Specifically, we pioneered modeling the localization problem as a dynamic graph using STGAT to generate quantified bias and uncertainty estimates. Building on this, we proposed an AESKF fusion algorithm to adaptively adjust the update weights of the error state Kalman filter according to the bias and uncertainty estimation. Furthermore, an asynchronous batch-based module was designed to ensure real-time inference. Extensive experiments on the publicly available dataset demonstrate the effectiveness of our proposed STG-AKF framework. Compared to the baseline, STG-AKF reduces the average absolute trajectory error (ATE) by 60.53% under cross-scenario validation and 52.06% under cross-trajectory validation. The results highlight that the proposed STG-AKF significantly improves positioning accuracy and robustness in unseen non-line-of-sight (NLOS) scenarios, while maintaining real-time performance.

Accurate objective physical fatigue (OPF) assessment is critical for enhancing safety in labor-intensive industries, optimizing athletic performance, and supporting personalized health management. With the advancement of wearable sensing technologies, wearable OPF assessment methods based on multi-source information fusion is emerging is gradually becoming an critical approach for sports and health monitoring. However, wearable OPF assessment methods face the following challenges: spatial modeling deficiency, ineffective multimodal fusion and limited long-range dependency capture. To address these limitations, we propose DBMNet, a dual-branch multi-modal network with spatial-temporal fusion for fatigue level classification. DBMNet captures temporal dynamics and spatial patterns separately from raw 12-lead electrocardiogram (ECG) and multi-site inertial measurement unit (IMU) signals. A novel Convolutional Additive Adaptive Cross-refinement Fusion (CAACF) module is introduced to enable adaptive, efficient, and interpretable fusion of heterogeneous modalities. We collected synchronized ECG and IMU data from 65 subjects during treadmill exercise following a modified Bruce protocol, annotated with both coarse- and fine-grained fatigue levels. Extensive experiments demonstrate that DBMNet achieves significant improvements over baseline and state-of-the-art methods, with up to 7% increase in classification accuracy and superior performance across multiple evaluation metrics. Moreover, DBMNet maintains a lightweight architecture suitable for deployment on mobile and wearable devices. This work provides an effective and scalable framework for real-time, objective fatigue monitoring using multi-source physiological signals.

Objective.Due to the growing demand for personal health monitoring in extreme environments, continuous monitoring of core temperature has become increasingly important. Traditional monitoring methods, such as mercury thermometers and infrared thermometers, may have limitations in tracking real-time fluctuations in core temperature, especially in special application scenarios such as firefighting, military, and aerospace. This study aims to develop a non-invasive, continuous core temperature prediction model based on machine learning, addressing the limitations of traditional methods in extreme environments.Approach.This study develops a novel machine learning-based non-invasive continuous body core temperature monitoring model. A wearable dual temperature sensing device is designed to collect skin and environment temperature, six machine learning algorithms are trained utilizing data from 62 subjects.Main results.Performance evaluations on a test set of 10 subjects reveal outstanding results, achieving a mean absolute error of 0.15 °C ± 0.04 °C, a root mean square error of 0.17 °C ± 0.05 °C, and a mean absolute percentage error of 0.40% ± 0.12%. Statistical analysis further confirms the model's superior predictive capability compared to traditional methods.Significance.The developed temperature monitoring model not only provides enhanced accuracy in various conditions but also serves as a robust tool for individual health monitoring. This innovation is particularly significant in scenarios requiring continuous and precise temperature tracking, and offering entirely new insights for improved health management strategies and outcomes.

Running is a popular sport, but abnormal running postures can lead to injuries. With the growing popularity of wearable electronics, real-time detection of abnormal running postures has become increasingly feasible. However, the variability of real-world running scenarios poses significant challenges to the robustness of deep learning-based detection methods. Models trained on fixed scenarios often experience a substantial decline in performance when applied to different environments. To address these issues, this article proposes a cross-scenario transfer learning (CSTL) framework to detect abnormal running postures by wearable inertial sensors, addressing the challenges of high time complexity and limited adaptability across various running scenarios. In the pre-training stage, we propose a Light-Norm Transformer (LNT) model to improve recognition accuracy using data collected from a single running scenario. In the fine-tuning stage, the pre-trained LNT model is transferred to adapt to five running scenarios with a contribution-guided freezing and dynamic adapter fine-tuning strategy, and the leave-one-out cross-validation method is adopted to evaluate the cross-scenario classification performance on each running scenario in the test module. The proposed CSTL method significantly reduces training time and enhances recognition accuracy, achieving an accuracy of 87.0% on the five running scenarios with only 10 epochs of fine-tuning, outperforming the non-transfer learning approach and achieving the optimal performance-efficiency balance. The proposed CSTL method enhances the adaptability and robustness of abnormal running posture detection systems in practical, real-world applications.

Human activity recognition using wearable sensors has become a critical task in ubiquitous computing, with applications in healthcare, fitness monitoring, and smart environments. However, sensor-based HAR often involves high-dimensional, multi-channel time series data, where redundant or irrelevant features may degrade both classification performance and model interpretability. In this study, we propose a lightweight feature screening framework guided by a multi-head attention mechanism to address this challenge. The model first applies channel-wise linear transformations to extract localized representations from each sensor axis, and then employs a multi-head attention module to dynamically assess the importance of each feature across all channels. This design enables the model to emphasize the most informative components while suppressing noise and redundancy. Experiments on the KU-HAR dataset demonstrate that the proposed method achieves 96.0% classification accuracy using only 60 selected features. In addition, the selected features provide valuable references for future research in feature selection, model simplification and multimodal sensor fusion.

Achieving prompt hemostasis after femoral artery access is critical for patient comfort and complication reduction in percutaneous coronary interventions. Most of the existing hemostasis methods either rely on manual control or are difficult to adapt to personality feelings. In this paper, we introduce an advanced framework for hemostatic robot control based on Reinforcement Learning from Human Feedback (RLHF). A Reinforcement Learning (RL) algorithm is developed to learn a control strategy to guide the robot to adaptively adjust direction based on real-time pressure sensor data and accurately track the puncture site for effective hemostasis. Virtual environment simulations demonstrate the algorithm's efficiency and accuracy, surpassing traditional reinforcement learning methods in convergence speed and performance, highlighting the potential of human feedback to humanize robotic operations.

Cross-domain human activity recognition using wearable inertial sensors remains a challenging task, especially in scenarios where no labeled data is available in the target domain. To address this, our team (SIAT-BIT) propose a semi-supervised learning framework based on MixMatch for locomotion and transportation mode recognition. Our approach leverages labeled data from multiple public HAR datasets and unlabeled data from the Sussex-Huawei Locomotion-Transportation Recognition Challenge Task 2 (Kyutech IMU) dataset. The framework integrates pseudo-label generation, data augmentation, soft label sharpening, and cross-sample mixing to mitigate domain shift and label scarcity. Experimental results demonstrate that the proposed method achieves competitive performance in Task 2 with an accuracy of 76.8% and F1 score of 76.5%, confirming its effectiveness in modeling real-world cross-domain activity scenarios without relying on target domain labels.

Inertial measurement unit (IMU)-based gait biometrics have attracted increasing attention for unobtrusive identity recognition. While recent studies often fuse signals from multiple sensor positions and time–frequency features, the actual contribution of each sensor location and signal modality remains insufficiently explored. In this work, we present a comprehensive quantitative analysis of the role of different IMU placements and feature domains in gait-based identity recognition. IMU data were collected from three body positions (shank, waist, and wrist) and processed to extract both time-domain and frequency-domain features. An attention-gated fusion network was employed to weight each signal branch adaptively, enabling interpretable assessment of their discriminative power. Experimental results show that shank IMU dominates recognition accuracy, while waist and wrist sensors primarily provide auxiliary information. Similarly, the contribution of time-domain features to classification performance is the greatest, while frequency-domain features offer complementary robustness. These findings illustrate the importance of sensor and feature selection in designing efficient, scalable IMU-based identity recognition systems for wearable applications.

Human Activity Recognition has been widely applied in mobile analysis, mobile health, and intelligent sensing. However, the existing HAR algorithms still face challenges in sensor modality dropout, varying device placement, and limited model generalization. To address these issues, the Sussex-Huawei Locomotion (SHL) recognition challenge provides a complex real-world dataset for algorithm development. In this study, our team (SIAT-BIT) proposes a classification framework based on handcrafted features, extracting 156 statistical, time-domain, and frequency-domain features from a total of 12 channels, including the tri-axial signals and their amplitudes of accelerometers, gyroscopes, and magnetometers. The experimental results show that our approach achieves strong recognition performance across multiple device placements, achieving an accuracy rate of 71.4% and an F1 score of 71.8% on the verification dataset which confirms the effectiveness and efficiency of our method.

Indoor location and navigation technologies are crucial for healthcare, security and other location-based services. Wi-Fi and inertial sensors have become mainstream indoor localization technologies for wearable device platforms due to simple deployment and low cost. This study proposes an extended Kalman filtering (EKF)-based multimodal sensor fusion algorithm for indoor localization, combining Wi-Fi fingerprint and inertial measurement unit (IMU) data to provide accurate and continuous pedestrian localization. The main contributions of this work are threefold. First, a Wi-Fi fingerprint data augmentation method based on access point (AP) location sorting is proposed and a regression network model with a convolutional denoising autoencoder for WiFi-based indoor localization (CDAELoc) is designed to improve the robustness. Second, a dual-branch deep inertial odometry (DbDIO) network model for IMU-based indoor localization is introduced, consisting of two branches with various convolutional kernel sizes to extract features at different scales. Finally, an EKF-based Wi-Fi and inertial odometry (WIO-EKF) fusion localization system is presented, utilizing the predicted results from the proposed CDAELoc and DbDIO models as the system observations and mitigating the initial heading error of DbDIO. The proposed models are applied to the UJIIndoorLoc, RoNIN public data sets and self-collected data set. Experimental results prove that the proposed CDAELoc model outperforms other Wi-Fi localization models, reducing the average positioning error (APE) by 12.5%. The proposed DbDIO model achieves higher accuracy and requires fewer model parameters than any other deep inertial odometry model. Finally, the APE of WIO-EKF is lower than those of CDAELoc and DbDIO by 34% and 42%.

Gait analysis based on sensors integrated into wearable devices has become an emerging research field with the development of miniaturized sensors and the wide application of wearable devices. Accurate gait event detection is essential for clinical gait analysis. It is also the basis of gait phase division, which is of great importance for gait abnormality diagnosis, rehabilitation evaluation, professional running guidance, etc. This study aimed to propose a lightweight inertial sensor-based real-time gait event detection method. First, an improved method to enhance the plantar pressure-based gait event detection’s robustness was proposed, providing a reference standard for gait event detection. Then, significant signal features corresponding to each gait event were found on the angular velocity signals and were studied by mapping the specific gait events on the angular velocity signals. Four gait events could be accurately detected by the proposed significant feature detection method. The method was evaluated on a dataset collected from 15 participants (including 9 healthy individuals and 6 recovering patients). The results demonstrated high accuracy, with mean absolute error (MAE) of 12.1, 19.6, 18.7, and 11.7 ms for the heel-strike, foot-flat, heel-off, and toe-off events, respectively. These findings suggested that the proposed method was effective for detecting gait events.

Remote health monitoring is one kind of E-health service, which transfer the users’ physiological data to the medical data center for analysis or diagnosis. Wireless body area network (WBAN) is a promising technology to achieve physiological information acquiring and delivering and thus has been widely adopted in remote health-monitoring applications. For WBAN, energy consumption is the major concern which has been addressed in many researches. Different from existing works, this work studies a joint scheduling and admission control problem with objective of optimizing the energy efficiency of both intra- and beyond-WBAN link. The problem is formulated as constrained Markov decision processes, and the relative value iteration and Lagrange multiplier approach are used to derive the optimal intelligent algorithm. Simulation results show the proposed algorithm is capable of, in comparison with greedy scheme, achieving nearly 100% throughput improvement in various power consumption budgets. Moreover, the proposed algorithm can achieve up to 5.5× power consumption saving for sensor node in comparison with other scheduling algorithms.

Gait recognition offers non-contact, non-intrusive, easily sensed, and hard-to-hide features, demonstrating significant research value and application potential. Although prevailing gait recognition techniques have produced certain results, the formulation and actualization of a functional gait recognition system continue to encounter numerous obstacles, such as the resources of the wearable devices, user experience and so on. This paper proposes a gait recognition method based on the fusion of information from multiple sensors, aiming to overcome the limitations of traditional identification technologies and enhance the accuracy and user experience in practical applications. A novel lightweight network model was designed in this study, integrating an attention mechanism to reduce the number of parameters and improve recognition accuracy. The most individual-differentiating sensor locations were identified by comparing the gait recognition performance of sensors in different body parts. Furthermore, decision-level and data-level fusion methods are studied and found that the data-level fusion strategy outperforms the decision-level one. Combinations of sensors in various body parts are investigated and the optimal combination is proposed. Finally, the impact of the scale of identified subjects on model performance was analyzed, demonstrating efficient data processing capabilities and robustness.

Human activity recognition (HAR) has developed rapidly in recent years due to its widespread applications in motion analysis, mobile health monitoring, security, and rehabilitation. However, due to missing sensor data, complex application scenarios, poor model robustness, existing HAR algorithms still cannot meet application requirements. In this context, the Sussex-Huawei Locomotion (SHL) recognition challenge provides a dataset for improving HAR algorithms. In this study, our team (SIAT-BIT) proposes a three-branch convolutional neural network framework for SHL recognition challenge. Firstly, the data is preprocessed for feature extraction, and then three classifiers are trained in parallel using three cross-entropy loss functions. The experimental results show that the proposed model achieves the best performance with the least model parameters. In addition, we further improved the performance through post-smoothing. Finally, we get an average accuracy of 0.9274 on the validation dataset.

Measuring core body temperature has significant implications in various fields, including healthcare, sports medicine, and military operations. Existing core body temperature measurement, such as rectal and esophageal measurements, are invasive, costly, and may not provide continuous monitoring, highlighting the need for non-invasive and reliable alternatives. With the development of wearable and computer technology, predicting core body temperature based on continuous measurement of skin temperature, ambient temperature, and other physiological parameters has become a trend for future development. In this study, we proposed a novel non-invasive method for predicting core body temperature using wearable sensors. A stacking fusion model was proposed to improve the core temperature prediction accuracy. The skin temperature and ambient temperature were collected by a wearable device and processed to remove outliers. Simultaneously, the core body temperature collected by a capsule temperature sensor was used as the gold standard. Then, we used the preprocessed data to train a machine learning algorithm, which was validated on collecting human data. The results showed that the algorithm based on stacking model fusion performed best, with an RMSE of 0.0448, MAE of 0.0214, and MAPE of 0.0582%, indicating high precision in predicting core body temperature, which can be used for continuous monitoring and accurate diagnosis, effective medical treatment.

Accurate assessment of physical fatigue is crucial to preventing physical injury caused by excessive exercise, overtraining during daily exercise and professional sports training. However, as a subjective feeling of an individual, physical fatigue is difficult for others to objectively evaluate. Heart rate variability (HRV), which is derived from electrocardiograms (ECG) and controlled by the autonomic nervous system, has been demonstrated to be a promising indicator for physical fatigue estimation. In this paper, we propose a novel method for the automatic and objective classification of physical fatigue based on HRV. First, a total of 24 HRV features were calculated. Then, a feature selection method was proposed to remove useless features that have a low correlation with physical fatigue and redundant features that have a high correlation with the selected features. After feature selection, the best 11 features were selected and were finally used for physical fatigue classifying. Four machine learning algorithms were trained to classify fatigue using the selected features. The experimental results indicate that the model trained using the selected 11 features could classify physical fatigue with high accuracy. More importantly, these selected features could provide important information regarding the identification of physical fatigue.

Walking locomotion needs the cooperation of nervous, muscular and skeletal system. The posture of walking, which is also named gait, has been verified to be an effective index of human health states. Accurate gait quality evaluation is an important index of the individual’s health states. Traditional gait evaluation methods usually need to be carried out in the laboratory equipped with special instruments (e.g., high speed cameras, pressure carpet, etc.). Although these methods have high accuracy, they are not suitable for long-term and continuous gait monitoring due to the high cost and complex operation. With the popularization of the Inertial Measurement Units (IMU) in various kinds of wearable devices, wearable and quantitative gait quality evaluation has become a new research hotspot. In this paper, we proposed a human wearable gait evaluation method which can extract gait characteristics and quantitatively analyze gait quality. A zero-velocity updating (ZUPT) algorithm was proposed to reduce the cumulative error; The gait phase portrait was calculated based on the angle and angular velocity of the ankles; the Elliptic Fourier Analysis (EFA) method was used on the gait phase portrait to quantitatively evaluate gait complexity; The gait symmetry was then evaluated based on the characteristics extracted from the gait phase. Finally, the proposed methods were tested on the dataset collected form 8 subjects and the experimental results verified the effectiveness of the proposed methods.

Wireless Body Sensor Network (WBAN) is a promising technology which can provide pervasive E-health services. However, the power consumption of wireless communications in WBAN greatly restricts its applications. Recently Ambient Backscatter Communication (AmBC) technology has been introduced to realize ultra-low power passive wireless communication, thus it is highly desirable for WBAN. In this paper we consider that AmBC, powered by a multi-antenna Base Station (BS), is incorporated in WBAN to achieve ultra-low power intra-WABN communication. To fully exploit the potential of the AmBC in WBAN, the transmit beamforming at BS, the AmBC information symbol period as well as the transmission time allocation are jointly designed, taking into account Quality of Service (QoS) requirement of WBAN and cellular users, and BS power budget. The formulated optimization problem is highly nonconvex and very challenging to solve. To tackle this issue, we employ Inner Approximate (IA) framework to reformulate the original problem and solve it in an iterative manner. Finally, numerical results are provided to show the effectiveness of the proposed algorithm under various of system parameters settings.