Pub Date : 2025-11-10DOI: 10.1109/JSEN.2025.3628663
Yuanzhe Li;Steffen Müller
Pedestrian crossing intention prediction is crucial for autonomous vehicles (AVs), enabling timely reactions to prevent potential accidents, especially in urban areas. The prediction task is challenging because the pedestrian’s behavior is highly diverse and influenced by various environmental and social factors. Although various networks have shown the potential to exploit complementary cues through multimodal fusion in this task, certain issues remain unresolved. First, critical contextual information, such as geometric depth and its associated modalities, has not been adequately explored. Second, the effective multimodal fusion strategies—particularly in terms of fusion scales and fusion order—remain underexplored. To address these limitations, a multimodal Transformer with cross-modality guided attention (MTC) is proposed. MTC fuses seven visual and motion modality features extracted from multiple Transformer-based encoding modules, incorporating depth maps (DMs) as a new modality to supplement the model’s understanding of scene geometry and pedestrian-centric distance information. MTC follows a multimodal fusion strategy in the spatial–modality–temporal order. Specifically, a novel cross-modality guided attention (CMGA) mechanism is designed to capture complementary feature maps through comprehensive interactions between coregistered visual modalities. Additionally, intermodal attention (IMA) and Transformer-based temporal feature fusion (TFF) are designed to effectively facilitate cross-modal interaction and capture temporal dependencies. Extensive evaluations on the JAAD dataset validate the proposed network’s effectiveness, outperforming the state-of-the-art (SOTA) methods.
{"title":"MTC: Multimodal Transformer With Cross-Modality Guided Attention for Pedestrian Crossing Intention Prediction","authors":"Yuanzhe Li;Steffen Müller","doi":"10.1109/JSEN.2025.3628663","DOIUrl":"https://doi.org/10.1109/JSEN.2025.3628663","url":null,"abstract":"Pedestrian crossing intention prediction is crucial for autonomous vehicles (AVs), enabling timely reactions to prevent potential accidents, especially in urban areas. The prediction task is challenging because the pedestrian’s behavior is highly diverse and influenced by various environmental and social factors. Although various networks have shown the potential to exploit complementary cues through multimodal fusion in this task, certain issues remain unresolved. First, critical contextual information, such as geometric depth and its associated modalities, has not been adequately explored. Second, the effective multimodal fusion strategies—particularly in terms of fusion scales and fusion order—remain underexplored. To address these limitations, a multimodal Transformer with cross-modality guided attention (MTC) is proposed. MTC fuses seven visual and motion modality features extracted from multiple Transformer-based encoding modules, incorporating depth maps (DMs) as a new modality to supplement the model’s understanding of scene geometry and pedestrian-centric distance information. MTC follows a multimodal fusion strategy in the spatial–modality–temporal order. Specifically, a novel cross-modality guided attention (CMGA) mechanism is designed to capture complementary feature maps through comprehensive interactions between coregistered visual modalities. Additionally, intermodal attention (IMA) and Transformer-based temporal feature fusion (TFF) are designed to effectively facilitate cross-modal interaction and capture temporal dependencies. Extensive evaluations on the JAAD dataset validate the proposed network’s effectiveness, outperforming the state-of-the-art (SOTA) methods.","PeriodicalId":447,"journal":{"name":"IEEE Sensors Journal","volume":"25 24","pages":"44929-44939"},"PeriodicalIF":4.3,"publicationDate":"2025-11-10","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145729451","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"综合性期刊","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-11-10DOI: 10.1109/JSEN.2025.3628713
Bin Chen;Jinlong Zhang;Junhai Yang;Bohao Pan
The volumetric proportions of ice crystals, water, and air within snowpack are highly susceptible to environmental disturbances, leading to multistate phase transitions, such as dry snow, wet snow, and slush. This study introduces a new method for runway snow identification using planar electrode impedance detection. Based on dielectric polarization theory, the effects of water content (0%–30% by volume) and density (100–600 kg/m3) on the complex permittivity of snow are analyzed. A multidimensional identification space is established using the sensitive excitation bands identified at 20 and 100 kHz to accurately classify snow types. A multidimensional identification space is defined to accurately classify snow types. Electrode design is optimized for runway conditions, and a calibration method is applied to mitigate impedance drift caused by interference. Field tests show the developed contact sensor achieves 85% identification accuracy. This work provides a new technique for real-time, automated runway snow condition monitoring, aligning with global reporting format (GRF) standards.
{"title":"Runway Snow State Identification Method Based on Impedance Characteristic Differences","authors":"Bin Chen;Jinlong Zhang;Junhai Yang;Bohao Pan","doi":"10.1109/JSEN.2025.3628713","DOIUrl":"https://doi.org/10.1109/JSEN.2025.3628713","url":null,"abstract":"The volumetric proportions of ice crystals, water, and air within snowpack are highly susceptible to environmental disturbances, leading to multistate phase transitions, such as dry snow, wet snow, and slush. This study introduces a new method for runway snow identification using planar electrode impedance detection. Based on dielectric polarization theory, the effects of water content (0%–30% by volume) and density (100–600 kg/m3) on the complex permittivity of snow are analyzed. A multidimensional identification space is established using the sensitive excitation bands identified at 20 and 100 kHz to accurately classify snow types. A multidimensional identification space is defined to accurately classify snow types. Electrode design is optimized for runway conditions, and a calibration method is applied to mitigate impedance drift caused by interference. Field tests show the developed contact sensor achieves 85% identification accuracy. This work provides a new technique for real-time, automated runway snow condition monitoring, aligning with global reporting format (GRF) standards.","PeriodicalId":447,"journal":{"name":"IEEE Sensors Journal","volume":"25 24","pages":"44940-44950"},"PeriodicalIF":4.3,"publicationDate":"2025-11-10","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145729396","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"综合性期刊","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
This article proposes a real-time errorcompensated multisensor acquisition system for a self-weight multiphysics cone penetration apparatus that performs marine geotechnical investigation. Conventional methods such as standard penetration test (SPT) and cone penetration test (CPT) provide reliable, high-resolution data but require dedicated offshore vessels, which are expensive to operate. To address these limitations, the apparatus with the proposed acquisition system has been developed for a lightweight and cost-effective solution. The proposed acquisition system drives hydro-compensated dual pressure transducers, strain gauges with Wheatstone bridges, and an inertial measurement unit (IMU) to obtain accurate geotechnical parameters as well as determine soil strength and stiffness properties during dynamic penetration. Additionally, the acquisition system uses an RS-485 communication protocol to transmit data over long distances up to 1.2 km at a data rate up to 100 kb/s. A 10.7 V lithium-ion (Li-ion) battery powers the proposed system, generating supply voltages of 9, 5, and 2 V through onboard voltage regulators to drive analog and digital subsystems. The proposed apparatus was verified to acquire reliable geotechnical parameters through field tests, providing a viable solution for offshore wind power development and submarine cable installations.
{"title":"A Real-Time Error-Compensated Multisensor Acquisition System for Marine Geotechnical Investigation","authors":"Seung-Beom Ku;Hyungjin Jung;Hyungjin Cho;Jiseok Oh;Jang-Un Kim;JunA Lee;Sungjun Cho;Jongmuk Won;Junghee Park;Hyunwook Choo;Hyung-Min Lee","doi":"10.1109/JSEN.2025.3628740","DOIUrl":"https://doi.org/10.1109/JSEN.2025.3628740","url":null,"abstract":"This article proposes a real-time errorcompensated multisensor acquisition system for a self-weight multiphysics cone penetration apparatus that performs marine geotechnical investigation. Conventional methods such as standard penetration test (SPT) and cone penetration test (CPT) provide reliable, high-resolution data but require dedicated offshore vessels, which are expensive to operate. To address these limitations, the apparatus with the proposed acquisition system has been developed for a lightweight and cost-effective solution. The proposed acquisition system drives hydro-compensated dual pressure transducers, strain gauges with Wheatstone bridges, and an inertial measurement unit (IMU) to obtain accurate geotechnical parameters as well as determine soil strength and stiffness properties during dynamic penetration. Additionally, the acquisition system uses an RS-485 communication protocol to transmit data over long distances up to 1.2 km at a data rate up to 100 kb/s. A 10.7 V lithium-ion (Li-ion) battery powers the proposed system, generating supply voltages of 9, 5, and 2 V through onboard voltage regulators to drive analog and digital subsystems. The proposed apparatus was verified to acquire reliable geotechnical parameters through field tests, providing a viable solution for offshore wind power development and submarine cable installations.","PeriodicalId":447,"journal":{"name":"IEEE Sensors Journal","volume":"25 24","pages":"44951-44961"},"PeriodicalIF":4.3,"publicationDate":"2025-11-10","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145729432","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"综合性期刊","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Locating and tracking targets in indoor environments is a challenging field of research. The complexity and variability of the environment limit the suitability of many technologies for this application. In this context, mmWave frequency modulated continuous wave (FMCW) radars can prove to be valuable sensors when combined with deep learning (DL) techniques, in order to extend performance in target locating and tracking. This article presents an original approach to locate and track moving targets in indoor environments, based on a YOLOv3 DL network that can be applied to radar data. To quantify the performance of the proposed method, here named mmTracking, tests were designed in accordance with the ISO/IEC 18305:2016 reference standard. The results show a mean error in localization of 0.39 m with a variance of 0.01 m2, and a root mean square error (RMSE) in the tracking of 0.40 m.
{"title":"mmTracking: A DL-Based mmWave RADAR Data Processing Algorithm for Indoor People Tracking","authors":"Michela Raimondi;Gianluca Ciattaglia;Antonio Nocera;Maria Gardano;Linda Senigagliesi;Susanna Spinsante;Ennio Gambi","doi":"10.1109/JSEN.2025.3628185","DOIUrl":"https://doi.org/10.1109/JSEN.2025.3628185","url":null,"abstract":"Locating and tracking targets in indoor environments is a challenging field of research. The complexity and variability of the environment limit the suitability of many technologies for this application. In this context, mmWave frequency modulated continuous wave (FMCW) radars can prove to be valuable sensors when combined with deep learning (DL) techniques, in order to extend performance in target locating and tracking. This article presents an original approach to locate and track moving targets in indoor environments, based on a YOLOv3 DL network that can be applied to radar data. To quantify the performance of the proposed method, here named mmTracking, tests were designed in accordance with the ISO/IEC 18305:2016 reference standard. The results show a mean error in localization of 0.39 m with a variance of 0.01 m2, and a root mean square error (RMSE) in the tracking of 0.40 m.","PeriodicalId":447,"journal":{"name":"IEEE Sensors Journal","volume":"25 24","pages":"45071-45083"},"PeriodicalIF":4.3,"publicationDate":"2025-11-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145729287","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"综合性期刊","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-11-07DOI: 10.1109/JSEN.2025.3628211
Vincenzo Saroli;Emiliano Schena;Carlo Massaroni
In recent years, additive manufacturing techniques, particularly 3-D printing methods like fused deposition modeling (FDM), have been increasingly explored for the development of systems for physiological monitoring, such as respiratory activity and joint kinematics, while retaining advantages such as rapid prototyping, low costs, and high customizability. This study presents the design, fabrication, and metrological characterization of single-layer strain bare sensor (BS) produced via FDM, with a thickness of only 0.15 mm, composed of a thermoplastic polyurethane (TPU) matrix filled with carbon black (CB) particles. In addition, the work investigates the impact of integrating the BS into flexible substrates—specifically kinesiology tape-integrated sensor (TS) and silicone-integrated sensor (SS)—to enhance mechanical robustness, a factor often neglected in existing literature. Electromechanical characterization was performed through quasi-static and cyclic tensile tests up to 5% strain. The resistance response exhibited nonlinear behavior, with maximum relative resistance changes of 40%, 38%, and 30% for the BS, TS, and SS configurations, respectively. The highest gauge factor (GF) of -14.7 was observed for the TS at 1% strain. During cyclic loading/unloading tests, all configurations demonstrated low hysteresis errors (~4%), even at high frequencies (90 cycles/min), despite the intrinsic piezoresistive nature of the sensors. In hygrothermal characterization, while substrate integration did not significantly mitigate the effect of temperature, silicone encapsulation proved effective in reducing humidity sensitivity, with the SS configuration showing only a 4% variation compared to ~13% for BS and TS. Finally, pilot tests conducted on a healthy volunteer demonstrated the feasibility of using the developed sensors for respiratory monitoring and joint kinematics assessment.
{"title":"Fabrication and Metrological Characterization of Bare and Integrated 3-D-Printed Single-Layer CB-TPU Strain Sensors","authors":"Vincenzo Saroli;Emiliano Schena;Carlo Massaroni","doi":"10.1109/JSEN.2025.3628211","DOIUrl":"https://doi.org/10.1109/JSEN.2025.3628211","url":null,"abstract":"In recent years, additive manufacturing techniques, particularly 3-D printing methods like fused deposition modeling (FDM), have been increasingly explored for the development of systems for physiological monitoring, such as respiratory activity and joint kinematics, while retaining advantages such as rapid prototyping, low costs, and high customizability. This study presents the design, fabrication, and metrological characterization of single-layer strain bare sensor (BS) produced via FDM, with a thickness of only 0.15 mm, composed of a thermoplastic polyurethane (TPU) matrix filled with carbon black (CB) particles. In addition, the work investigates the impact of integrating the BS into flexible substrates—specifically kinesiology tape-integrated sensor (TS) and silicone-integrated sensor (SS)—to enhance mechanical robustness, a factor often neglected in existing literature. Electromechanical characterization was performed through quasi-static and cyclic tensile tests up to 5% strain. The resistance response exhibited nonlinear behavior, with maximum relative resistance changes of 40%, 38%, and 30% for the BS, TS, and SS configurations, respectively. The highest gauge factor (GF) of -14.7 was observed for the TS at 1% strain. During cyclic loading/unloading tests, all configurations demonstrated low hysteresis errors (~4%), even at high frequencies (90 cycles/min), despite the intrinsic piezoresistive nature of the sensors. In hygrothermal characterization, while substrate integration did not significantly mitigate the effect of temperature, silicone encapsulation proved effective in reducing humidity sensitivity, with the SS configuration showing only a 4% variation compared to ~13% for BS and TS. Finally, pilot tests conducted on a healthy volunteer demonstrated the feasibility of using the developed sensors for respiratory monitoring and joint kinematics assessment.","PeriodicalId":447,"journal":{"name":"IEEE Sensors Journal","volume":"25 24","pages":"44919-44928"},"PeriodicalIF":4.3,"publicationDate":"2025-11-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145729522","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"综合性期刊","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-11-07DOI: 10.1109/JSEN.2025.3627958
Biswajit Haldar;Boby George;M. Arul Muthiah;M. A. Atmanand
The high cost and power requirements of the acoustic Doppler velocimeter (ADV) restrict its use. This type of current meter is also susceptible to biofouling. A recently reported innovative approach where the wide range of ocean current speed is estimated from the buoy measurement data, such as load cell, GPS, anemometer, and wave sensor, using the advanced machine learning (ML) technique, is a viable option for ocean current speed measurement with advantages such as lower power requirements, lower cost, and resistance to biofouling. However, the reported method is limited to the measurement of current speed alone. Although the speed of ocean currents has been widely studied, the direction of ocean currents is equally significant for various scientific, economic, and environmental applications. In this article, an attempt is made to estimate both the speed and direction of the surface ocean current from buoy sensor data using ML. The performance of the ML models is evaluated and validated using buoy data collected from the northern Bay of Bengal for the duration of December 2019 to February 2021. This study compares four different ML models, ultimately identifying the random forest (RF) as the best-performing model for the estimation of current speed and direction. The study shows a correlation value of 0.94 and a root mean square error (RMSE) of 0.065 m/s between the observed and estimated current speed for the entire range of measurements (0–1.56 m/s). On the other hand, the correlation between the estimated and observed current direction is found to be 0.98 with an RMSE value of 13.320 for the measurement range of 0.4–1.56 m/s. The result shows that the model is capable of reliably estimating the current speed and direction with significant accuracy. However, the accuracy of the speed estimation is good for the full range of current, whereas the estimation of the current direction is good for the current above a threshold value of 0.4 m/s.
{"title":"Performance Evaluation of ML Models for Ocean Current Speed and Direction Estimation From Buoy Sensor Data","authors":"Biswajit Haldar;Boby George;M. Arul Muthiah;M. A. Atmanand","doi":"10.1109/JSEN.2025.3627958","DOIUrl":"https://doi.org/10.1109/JSEN.2025.3627958","url":null,"abstract":"The high cost and power requirements of the acoustic Doppler velocimeter (ADV) restrict its use. This type of current meter is also susceptible to biofouling. A recently reported innovative approach where the wide range of ocean current speed is estimated from the buoy measurement data, such as load cell, GPS, anemometer, and wave sensor, using the advanced machine learning (ML) technique, is a viable option for ocean current speed measurement with advantages such as lower power requirements, lower cost, and resistance to biofouling. However, the reported method is limited to the measurement of current speed alone. Although the speed of ocean currents has been widely studied, the direction of ocean currents is equally significant for various scientific, economic, and environmental applications. In this article, an attempt is made to estimate both the speed and direction of the surface ocean current from buoy sensor data using ML. The performance of the ML models is evaluated and validated using buoy data collected from the northern Bay of Bengal for the duration of December 2019 to February 2021. This study compares four different ML models, ultimately identifying the random forest (RF) as the best-performing model for the estimation of current speed and direction. The study shows a correlation value of 0.94 and a root mean square error (RMSE) of 0.065 m/s between the observed and estimated current speed for the entire range of measurements (0–1.56 m/s). On the other hand, the correlation between the estimated and observed current direction is found to be 0.98 with an RMSE value of 13.320 for the measurement range of 0.4–1.56 m/s. The result shows that the model is capable of reliably estimating the current speed and direction with significant accuracy. However, the accuracy of the speed estimation is good for the full range of current, whereas the estimation of the current direction is good for the current above a threshold value of 0.4 m/s.","PeriodicalId":447,"journal":{"name":"IEEE Sensors Journal","volume":"25 24","pages":"44910-44918"},"PeriodicalIF":4.3,"publicationDate":"2025-11-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145729414","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"综合性期刊","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-11-04DOI: 10.1109/JSEN.2025.3626889
Zihua Chen;Xueze Zhang;Yangjie Luo;Haoran Wang;Lihua Zhang;Xiaoyang Kang
With the development of deep learning (DL) technology, there is a great possibility of decoding surface electromyography (sEMG) for human窶田omputer interaction (HCI) applications such as robot control. The sEMG signals have been used to complete movement classification tasks using machine learning (ML) and DL measures. However, the high-density sEMG (HD-sEMG) may not be suitable for application due to the electrode displacement. Here, we proposed a novel network architecture to decode sEMG signals acquired from low-cost armbands. We accomplished extensive experiments to validate our methods on both public dataset Ninapro DB5 and self-collected data. Adopting the sliding window strategy, our method got an average accuracy of 92.16%, 89.44%, 81.92%, and 73.41% corresponding to window sizes 1500, 1000, 500, and 200 ms. For the self-collected data, we classified seven types of movements (including rest) using a window size of 200 ms and attained an average accuracy of 95.57%, demonstrating the generalizability of the proposed architecture. To comprehensively evaluate the architecture, we also conducted experiments with different channel numbers (8 and 16 channels). Furthermore, we carried out ablation experiments to validate the effectiveness of the proposed network. All the precision rates declined after removing the multiscale attention (MSCA) module with a significant difference, which indicates that the proposed module is of great benefit to the movement classification. The overall experiment results show that our architecture has great potential for low-cost EMG movement recognition.
{"title":"A Multiscale Attention Network for sEMG Gesture Recognition Using a Portable Armband","authors":"Zihua Chen;Xueze Zhang;Yangjie Luo;Haoran Wang;Lihua Zhang;Xiaoyang Kang","doi":"10.1109/JSEN.2025.3626889","DOIUrl":"https://doi.org/10.1109/JSEN.2025.3626889","url":null,"abstract":"With the development of deep learning (DL) technology, there is a great possibility of decoding surface electromyography (sEMG) for human窶田omputer interaction (HCI) applications such as robot control. The sEMG signals have been used to complete movement classification tasks using machine learning (ML) and DL measures. However, the high-density sEMG (HD-sEMG) may not be suitable for application due to the electrode displacement. Here, we proposed a novel network architecture to decode sEMG signals acquired from low-cost armbands. We accomplished extensive experiments to validate our methods on both public dataset Ninapro DB5 and self-collected data. Adopting the sliding window strategy, our method got an average accuracy of 92.16%, 89.44%, 81.92%, and 73.41% corresponding to window sizes 1500, 1000, 500, and 200 ms. For the self-collected data, we classified seven types of movements (including rest) using a window size of 200 ms and attained an average accuracy of 95.57%, demonstrating the generalizability of the proposed architecture. To comprehensively evaluate the architecture, we also conducted experiments with different channel numbers (8 and 16 channels). Furthermore, we carried out ablation experiments to validate the effectiveness of the proposed network. All the precision rates declined after removing the multiscale attention (MSCA) module with a significant difference, which indicates that the proposed module is of great benefit to the movement classification. The overall experiment results show that our architecture has great potential for low-cost EMG movement recognition.","PeriodicalId":447,"journal":{"name":"IEEE Sensors Journal","volume":"25 24","pages":"45041-45049"},"PeriodicalIF":4.3,"publicationDate":"2025-11-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145729431","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"综合性期刊","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-11-03DOI: 10.1109/JSEN.2025.3626533
Haikang Zhu;Lubing Wang;Xufeng Zhao
Rolling bearings fault diagnosis serves as an essential tool to save costs and ensure safety in manufacturing systems. The inability to identify early stage damage of bearings may trigger abrupt equipment failures. However, current diagnostic methods are not only constrained by large amounts of data and costly computational resources but also rarely account for small-sample scenarios. This study investigates the practical problem of limited data by proposing CWT-MSAnet. MSAnet is a novel multisensory fusion framework integrating multistream attention (MSA) and convolutional block attention module (CBAM) module. The proposed MSA module achieves cross-stream feature enhancement through self-calibrated attention weights derived from parallel sensor streams, simultaneously expanding contextual receptive field and prioritizing informationrich data streams. First, each raw signal is segmented into samples and converted into images by CWT. Second, MSAnet is constructed by incorporating a hybrid CNN that integrates the CBAM with the proposed MSA. Finally, a series of experimental evaluations was systematically performed to demonstrate the efficacy of CWT-MSAnet. Experimental validation demonstrates that the performance of CWT-MSAnet is superior to other deep learning models under dataconstrained conditions. Moreover, CWT-MSAnet shows better robustness in data imbalance scenarios, noisy working conditions, and new categories.
{"title":"A Novel Small-Sample and Multisensory Fusion Fault Diagnosis Method via Continuous Wavelet Transform and Attention Mechanism","authors":"Haikang Zhu;Lubing Wang;Xufeng Zhao","doi":"10.1109/JSEN.2025.3626533","DOIUrl":"https://doi.org/10.1109/JSEN.2025.3626533","url":null,"abstract":"Rolling bearings fault diagnosis serves as an essential tool to save costs and ensure safety in manufacturing systems. The inability to identify early stage damage of bearings may trigger abrupt equipment failures. However, current diagnostic methods are not only constrained by large amounts of data and costly computational resources but also rarely account for small-sample scenarios. This study investigates the practical problem of limited data by proposing CWT-MSAnet. MSAnet is a novel multisensory fusion framework integrating multistream attention (MSA) and convolutional block attention module (CBAM) module. The proposed MSA module achieves cross-stream feature enhancement through self-calibrated attention weights derived from parallel sensor streams, simultaneously expanding contextual receptive field and prioritizing informationrich data streams. First, each raw signal is segmented into samples and converted into images by CWT. Second, MSAnet is constructed by incorporating a hybrid CNN that integrates the CBAM with the proposed MSA. Finally, a series of experimental evaluations was systematically performed to demonstrate the efficacy of CWT-MSAnet. Experimental validation demonstrates that the performance of CWT-MSAnet is superior to other deep learning models under dataconstrained conditions. Moreover, CWT-MSAnet shows better robustness in data imbalance scenarios, noisy working conditions, and new categories.","PeriodicalId":447,"journal":{"name":"IEEE Sensors Journal","volume":"25 24","pages":"45061-45070"},"PeriodicalIF":4.3,"publicationDate":"2025-11-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145729391","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"综合性期刊","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-11-03DOI: 10.1109/JSEN.2025.3626282
Jihoon Sung;Yeunwoong Kyung
Multihop wireless networks (MWNs) are critical for supporting diverse mobile services, including Internet and Internet-of-Things (IoT) applications. Their deployment flexibility and cost-effectiveness make them well-suited for industrial environments. However, achieving high throughput and low delay in such networks remains a significant challenge, particularly in the presence of network holes, areas lacking active nodes necessary for packet forwarding. In this context, we address the joint routing and scheduling problem in MWNs, specifically focusing on network holes that are often caused by irregular node deployment, which significantly degrades network performance. This article revisits potential-field routing as a foundational model for addressing network holes. Through extensive theoretical analysis, we explore its suitability for resolving network hole challenges and introduce an enhanced version of potential-field routing that incorporates topology awareness. We propose a new joint routing and scheduling solution that not only aims to reduce delays but also maintains throughput optimality in MWNs with network holes. This solution, an enhanced version of the back-pressure algorithm, leverages the potential-field routing metric to improve delay performance, particularly in lightly loaded regions, which are often problematic in existing models. It uniquely addresses the challenges posed by network holes, an area that has seen limited exploration in previous research. Simulation results demonstrate that our proposed algorithm significantly outperforms baseline models in mitigating end-to-end delays, a notable limitation of traditional back-pressure (TBP) algorithms, thus establishing it as a superior alternative.
{"title":"Exploring Delay Challenges With Integrated Potential-Field Routing and Back-Pressure Algorithm","authors":"Jihoon Sung;Yeunwoong Kyung","doi":"10.1109/JSEN.2025.3626282","DOIUrl":"https://doi.org/10.1109/JSEN.2025.3626282","url":null,"abstract":"Multihop wireless networks (MWNs) are critical for supporting diverse mobile services, including Internet and Internet-of-Things (IoT) applications. Their deployment flexibility and cost-effectiveness make them well-suited for industrial environments. However, achieving high throughput and low delay in such networks remains a significant challenge, particularly in the presence of network holes, areas lacking active nodes necessary for packet forwarding. In this context, we address the joint routing and scheduling problem in MWNs, specifically focusing on network holes that are often caused by irregular node deployment, which significantly degrades network performance. This article revisits potential-field routing as a foundational model for addressing network holes. Through extensive theoretical analysis, we explore its suitability for resolving network hole challenges and introduce an enhanced version of potential-field routing that incorporates topology awareness. We propose a new joint routing and scheduling solution that not only aims to reduce delays but also maintains throughput optimality in MWNs with network holes. This solution, an enhanced version of the back-pressure algorithm, leverages the potential-field routing metric to improve delay performance, particularly in lightly loaded regions, which are often problematic in existing models. It uniquely addresses the challenges posed by network holes, an area that has seen limited exploration in previous research. Simulation results demonstrate that our proposed algorithm significantly outperforms baseline models in mitigating end-to-end delays, a notable limitation of traditional back-pressure (TBP) algorithms, thus establishing it as a superior alternative.","PeriodicalId":447,"journal":{"name":"IEEE Sensors Journal","volume":"25 24","pages":"45009-45024"},"PeriodicalIF":4.3,"publicationDate":"2025-11-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145729421","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"综合性期刊","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-11-03DOI: 10.1109/JSEN.2025.3626674
Yan Yu;Shaojuan Ma;Chenghui Wang;Xiaona Wu;Changlin Xu
Accurate temperature estimation of environmental sensors is crucial in industrial monitoring and control systems. However, electromagnetic interference, vibration noise, and multisource signal coupling in complex industrial environments can introduce significant random errors and systematic biases, posing a major challenge to precise temperature estimation. This article proposes a temperature state estimation method based on deep learning and the unscented Kalman filter (UKF). First, the temporal convolutional network (TCN)-gated recurrent unit (GRU)-Attention framework is constructed to extract spatiotemporal features through the dilated convolutional structure of TCN to model temporal dependencies using GRU, and introduce the attention module to highlight the impact of key environmental features. Subsequently, to further enhance the robustness of the model, the predictions of the deep learning model are used as observation inputs to the UKF, constructing a hybrid deep state estimation model that adaptively suppresses environmental noise. Experimental results show that the performance of TCN-GRU-Attention is substantially improved compared to traditional deep learning models. After integration with the UKF, compared with the TCN-GRU-Attention model, both mean absolute error (MAE) and root mean square error (RMSE) decrease by approximately 20%, and maximum absolute error (MaxAE) decreases by about 30%, verifying the superior generalization performance and stability of the proposed method.
{"title":"State Estimation of Environmental Temperature Based on Deep Learning and Unscented Kalman Filtering","authors":"Yan Yu;Shaojuan Ma;Chenghui Wang;Xiaona Wu;Changlin Xu","doi":"10.1109/JSEN.2025.3626674","DOIUrl":"https://doi.org/10.1109/JSEN.2025.3626674","url":null,"abstract":"Accurate temperature estimation of environmental sensors is crucial in industrial monitoring and control systems. However, electromagnetic interference, vibration noise, and multisource signal coupling in complex industrial environments can introduce significant random errors and systematic biases, posing a major challenge to precise temperature estimation. This article proposes a temperature state estimation method based on deep learning and the unscented Kalman filter (UKF). First, the temporal convolutional network (TCN)-gated recurrent unit (GRU)-Attention framework is constructed to extract spatiotemporal features through the dilated convolutional structure of TCN to model temporal dependencies using GRU, and introduce the attention module to highlight the impact of key environmental features. Subsequently, to further enhance the robustness of the model, the predictions of the deep learning model are used as observation inputs to the UKF, constructing a hybrid deep state estimation model that adaptively suppresses environmental noise. Experimental results show that the performance of TCN-GRU-Attention is substantially improved compared to traditional deep learning models. After integration with the UKF, compared with the TCN-GRU-Attention model, both mean absolute error (MAE) and root mean square error (RMSE) decrease by approximately 20%, and maximum absolute error (MaxAE) decreases by about 30%, verifying the superior generalization performance and stability of the proposed method.","PeriodicalId":447,"journal":{"name":"IEEE Sensors Journal","volume":"25 24","pages":"45025-45040"},"PeriodicalIF":4.3,"publicationDate":"2025-11-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145729458","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"综合性期刊","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
京公网安备 11010802042870号
京ICP备2023020795号-1
扫码关注我们
求助内容：
应助结果提醒方式：