In the field of fault diagnosis, data imbalance is a common issue, which leads to the degradation of the recognition rate of faults classification models. The synthetic minority oversampling technique (SMOTE) can mitigate the impact of data imbalance by generating new fault samples. In this research, an adaptive geometric SMOTE is proposed for bearing fault diagnosis under imbalanced data. First, different sample attention levels are assigned according to the sample distribution, and the edge samples are given higher attention levels to enhance the classification boundary. Second, the sample generation area is adaptively adjusted according to its attention level, and the generation area for boundary samples is set relatively conservatively to avoid the sample overlapping issue. Finally, the multiple classification models are used for classification testing. Several related experiments show the effectiveness and superiority of the proposed method in handling imbalanced classification tasks by compared with tradition methods.
{"title":"A New Adaptive Geometric SMOTE for Bearing Imbalanced Fault Diagnosis","authors":"Chengbin Wei;Chenyu Tian;Yutao Chen;Bo Zhao;Long Wen","doi":"10.1109/JSEN.2026.3661375","DOIUrl":"https://doi.org/10.1109/JSEN.2026.3661375","url":null,"abstract":"In the field of fault diagnosis, data imbalance is a common issue, which leads to the degradation of the recognition rate of faults classification models. The synthetic minority oversampling technique (SMOTE) can mitigate the impact of data imbalance by generating new fault samples. In this research, an adaptive geometric SMOTE is proposed for bearing fault diagnosis under imbalanced data. First, different sample attention levels are assigned according to the sample distribution, and the edge samples are given higher attention levels to enhance the classification boundary. Second, the sample generation area is adaptively adjusted according to its attention level, and the generation area for boundary samples is set relatively conservatively to avoid the sample overlapping issue. Finally, the multiple classification models are used for classification testing. Several related experiments show the effectiveness and superiority of the proposed method in handling imbalanced classification tasks by compared with tradition methods.","PeriodicalId":447,"journal":{"name":"IEEE Sensors Journal","volume":"26 6","pages":"9375-9386"},"PeriodicalIF":4.3,"publicationDate":"2026-02-10","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147440611","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 : 2026-02-09DOI: 10.1109/JSEN.2026.3660748
Wenyan Xiong;Fagui Liu;Jun Jiang;Xi Yao;C. L. Philip Chen
Improving energy efficiency in wireless sensor networks (WSNs) is critical to ensuring effective transmission and prolonging network lifetime. However, most existing clustering methods are designed for relatively static scenarios and thus fail to cope with dynamic scenarios such as node mobility, topology variations, and uneven energy consumption, leading to poor adaptability, suboptimal real-time performance, and low energy efficiency, and significantly hinder the applicability in real-world WSNs. To address these challenges, we propose a structural lightweight fuzzy broad learning-based competitive clustering algorithm (FBLSC) for efficient cluster head (CH) selection and adaptive clustering in dynamic WSNs. Specifically, a fuzzy logic system (FLS) is employed to extract and evaluate node features under uncertainty, generating compete value (CV), which are then used to train a broad learning system (BLS) for fast and accurate CH prediction. Compared with traditional fuzzy systems and particle swarm optimization (PSO)-based models, the proposed hybrid structure improves adaptability, learning speed, and decision accuracy while maintaining low computational complexity. The experimental results show that, averaged over six representative network scenarios, FBLSC reduces average energy consumption by 61.8%, achieves +107.2% longer Last Node Dies (LND) network lifetime, and improves throughput by 97.7% compared with the same baseline. Due to its ability to enhance energy efficiency and prolong network lifetime, FBLSC is well-suited for deployment in dynamic, energy-constrained application scenarios such as environmental monitoring, smart agriculture, emergency response systems, and others.
{"title":"FBLSC: Fuzzy BLS-Based Competitive Clustering Optimization Algorithm for Energy Efficient in Wireless Sensor Networks","authors":"Wenyan Xiong;Fagui Liu;Jun Jiang;Xi Yao;C. L. Philip Chen","doi":"10.1109/JSEN.2026.3660748","DOIUrl":"https://doi.org/10.1109/JSEN.2026.3660748","url":null,"abstract":"Improving energy efficiency in wireless sensor networks (WSNs) is critical to ensuring effective transmission and prolonging network lifetime. However, most existing clustering methods are designed for relatively static scenarios and thus fail to cope with dynamic scenarios such as node mobility, topology variations, and uneven energy consumption, leading to poor adaptability, suboptimal real-time performance, and low energy efficiency, and significantly hinder the applicability in real-world WSNs. To address these challenges, we propose a structural lightweight fuzzy broad learning-based competitive clustering algorithm (FBLSC) for efficient cluster head (CH) selection and adaptive clustering in dynamic WSNs. Specifically, a fuzzy logic system (FLS) is employed to extract and evaluate node features under uncertainty, generating compete value (CV), which are then used to train a broad learning system (BLS) for fast and accurate CH prediction. Compared with traditional fuzzy systems and particle swarm optimization (PSO)-based models, the proposed hybrid structure improves adaptability, learning speed, and decision accuracy while maintaining low computational complexity. The experimental results show that, averaged over six representative network scenarios, FBLSC reduces average energy consumption by 61.8%, achieves +107.2% longer Last Node Dies (LND) network lifetime, and improves throughput by 97.7% compared with the same baseline. Due to its ability to enhance energy efficiency and prolong network lifetime, FBLSC is well-suited for deployment in dynamic, energy-constrained application scenarios such as environmental monitoring, smart agriculture, emergency response systems, and others.","PeriodicalId":447,"journal":{"name":"IEEE Sensors Journal","volume":"26 6","pages":"9300-9316"},"PeriodicalIF":4.3,"publicationDate":"2026-02-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147440598","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 : 2026-02-06DOI: 10.1109/JSEN.2026.3652315
Dong Li;Xiaohua Wang;Jiacheng Qi;Wenjie Wang
Visual simultaneous localization and mapping (SLAM) underpins many robotic applications; yet, both traditional dense SLAM and neural point cloud-based approaches still struggle to balance real-time tracking with high-fidelity dense reconstruction in cluttered indoor scenes. To enhance the quality of dense mapping reconstruction for robots in complex indoor environments, this article proposes an improved neural point cloud-based dense SLAM method joint encoding and adjustable neural point cloud-based RGB-D SLAM (JA-SLAM). First, JA-SLAM employs a dual-multilayer perceptron (MLP) architecture consisting of a geometric MLP and a color MLP: the geometric MLP is used to predict occupancy probabilities of neural point clouds, while the color MLP predicts RGB values of neural point clouds. Specifically, the geometric MLP employs a hybrid encoding approach called high-frequency and multiscale collaborative encoding (HFMSCE), which effectively leverages both the high-frequency and multiscale spatial information of point clouds. Second, a region-justable neural point cloud densification strategy that performs adjustable optimization based on the scene information density is designed to optimize point cloud distribution according to the scene information density. Third, we regularize the mapping objective with a weighted L2 term to balance reconstruction accuracy and robustness. Experimental results show that JA-SLAM achieves significant performance improvements in complex scenarios; on the Replica, TUM RGB-D, and ScanNet datasets, it outperforms state-of-the-art neural point cloud-based methods in terms of mapping fidelity while maintaining competitive tracking performance, achieving an average 2.3-dB improvement in PSNR and a 25% reduction in the number of point clouds.
视觉同步定位和地图（SLAM）是许多机器人应用的基础；然而，传统的密集SLAM和基于神经点云的方法仍然难以在杂乱的室内场景中实现实时跟踪和高保真密集重建的平衡。为了提高机器人在复杂室内环境下密集映射重建的质量，本文提出了一种改进的基于神经点云的密集SLAM方法联合编码和基于可调神经点云的RGB-D SLAM （JA-SLAM）。首先，JA-SLAM采用由几何MLP和颜色MLP组成的双多层感知器（MLP）架构：几何MLP用于预测神经点云的占用概率，而颜色MLP用于预测神经点云的RGB值。具体而言，几何MLP采用了一种称为高频多尺度协同编码（HFMSCE）的混合编码方法，有效地利用了点云的高频和多尺度空间信息。其次，设计基于场景信息密度可调优化的区域可调神经点云密度策略，根据场景信息密度优化点云分布；第三，我们用加权L2项正则化映射目标，以平衡重建精度和鲁棒性。实验结果表明，在复杂场景下，JA-SLAM算法的性能得到了显著提高；在Replica、TUM RGB-D和ScanNet数据集上，它在映射保真度方面优于最先进的基于神经点云的方法，同时保持有竞争力的跟踪性能，实现了平均2.3 db的PSNR改进，点云数量减少了25%。
{"title":"JA-SLAM: Joint Encoding and Adjustable Neural Point Cloud-Based RGB-D SLAM","authors":"Dong Li;Xiaohua Wang;Jiacheng Qi;Wenjie Wang","doi":"10.1109/JSEN.2026.3652315","DOIUrl":"https://doi.org/10.1109/JSEN.2026.3652315","url":null,"abstract":"Visual simultaneous localization and mapping (SLAM) underpins many robotic applications; yet, both traditional dense SLAM and neural point cloud-based approaches still struggle to balance real-time tracking with high-fidelity dense reconstruction in cluttered indoor scenes. To enhance the quality of dense mapping reconstruction for robots in complex indoor environments, this article proposes an improved neural point cloud-based dense SLAM method joint encoding and adjustable neural point cloud-based RGB-D SLAM (JA-SLAM). First, JA-SLAM employs a dual-multilayer perceptron (MLP) architecture consisting of a geometric MLP and a color MLP: the geometric MLP is used to predict occupancy probabilities of neural point clouds, while the color MLP predicts RGB values of neural point clouds. Specifically, the geometric MLP employs a hybrid encoding approach called high-frequency and multiscale collaborative encoding (HFMSCE), which effectively leverages both the high-frequency and multiscale spatial information of point clouds. Second, a region-justable neural point cloud densification strategy that performs adjustable optimization based on the scene information density is designed to optimize point cloud distribution according to the scene information density. Third, we regularize the mapping objective with a weighted L2 term to balance reconstruction accuracy and robustness. Experimental results show that JA-SLAM achieves significant performance improvements in complex scenarios; on the Replica, TUM RGB-D, and ScanNet datasets, it outperforms state-of-the-art neural point cloud-based methods in terms of mapping fidelity while maintaining competitive tracking performance, achieving an average 2.3-dB improvement in PSNR and a 25% reduction in the number of point clouds.","PeriodicalId":447,"journal":{"name":"IEEE Sensors Journal","volume":"26 6","pages":"9330-9338"},"PeriodicalIF":4.3,"publicationDate":"2026-02-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147440608","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 : 2026-02-03DOI: 10.1109/JSEN.2026.3658771
Yipeng Chen;Ziwei Zhang;Jun Liu
The robustness of perception systems in adverse weather is critical for the safety of autonomous vehicles, with millimeter-wave (mmWave) radar being an indispensable sensor. However, current radar-based segmentation models are trained offline on static datasets and suffer from catastrophic forgetting when encountering unseen object classes in dynamic real-world environments. To address this limitation, we introduce a class-incremental continual learning (CIL) framework specifically designed for automotive radar point cloud semantic segmentation. Our approach employs a model-agnostic student–teacher architecture, where a frozen model from a previous task provides supervisory signals to the current model via knowledge distillation (KD). This is combined with a focal loss to handle the inherent class imbalance of radar data. Our framework is comprehensively evaluated on the RadarScenes dataset across several state-of-the-art segmentation architectures, including both point- and transformer-based models, to demonstrate its general applicability. Our experiments demonstrate that the proposed strategy effectively mitigates catastrophic forgetting. This work establishes a benchmark for continual learning on radar point clouds, paving the way for more adaptive and long-term autonomous perception systems.
{"title":"Continual Learning for Automotive Radar Semantic Segmentation","authors":"Yipeng Chen;Ziwei Zhang;Jun Liu","doi":"10.1109/JSEN.2026.3658771","DOIUrl":"https://doi.org/10.1109/JSEN.2026.3658771","url":null,"abstract":"The robustness of perception systems in adverse weather is critical for the safety of autonomous vehicles, with millimeter-wave (mmWave) radar being an indispensable sensor. However, current radar-based segmentation models are trained offline on static datasets and suffer from catastrophic forgetting when encountering unseen object classes in dynamic real-world environments. To address this limitation, we introduce a class-incremental continual learning (CIL) framework specifically designed for automotive radar point cloud semantic segmentation. Our approach employs a model-agnostic student–teacher architecture, where a frozen model from a previous task provides supervisory signals to the current model via knowledge distillation (KD). This is combined with a focal loss to handle the inherent class imbalance of radar data. Our framework is comprehensively evaluated on the RadarScenes dataset across several state-of-the-art segmentation architectures, including both point- and transformer-based models, to demonstrate its general applicability. Our experiments demonstrate that the proposed strategy effectively mitigates catastrophic forgetting. This work establishes a benchmark for continual learning on radar point clouds, paving the way for more adaptive and long-term autonomous perception systems.","PeriodicalId":447,"journal":{"name":"IEEE Sensors Journal","volume":"26 6","pages":"9364-9374"},"PeriodicalIF":4.3,"publicationDate":"2026-02-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147440629","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 : 2026-02-02DOI: 10.1109/JSEN.2026.3658197
Soufiane Ben Othman;Chinmay Chakraborty;Saranjit Singh;Mohamed Amine Frikha
Sensor devices and Internet of Things (IoT) devices face a critical, fundamental challenge: deploying robust security while operating under severe constraints on energy, processing power, and memory. This article presents biologically inspired entropy security (BioEnS), a novel, closed-loop framework designed to overcome the inherent security–privacy–efficiency trilemma by achieving paretooptimal adaptive security. BioEnS models adaptive defense as a real-time, constrained multiobjective optimization problem, dynamically resolving the trade-off between security assurance ($Phi$ ) and resource consumption ($Psi$ ) based on current context. The framework core relies on a hardware root-of-trust entropy source (HRTES), which provides a quantifiable PUF-derived min-entropy rate ($E_{text{rate}}$ ) for nondeterministic key derivation, feeding into an Adaptive Security Manager (ASM). This mechanism rigorously enforces context-dependent security requirements ($Phi_{text{req}}$ ) through a dominant $lambda$ -penalty term, enabling ultralow latency policy decisions. Experimental validation on an ARM Cortex-M platform demonstrates exceptional performance: BioEnS maintains a near-zero security violation rate (SVR) (0.02%) while simultaneously yielding a superior lifetime extension ratio (LER) of $0.69 times$ relative to the high-security baseline (HSB), confirming the validity of the guaranteed policy enforcement.
传感器设备和物联网（IoT）设备面临着一个关键的、根本性的挑战：在能源、处理能力和内存受到严格限制的情况下，部署强大的安全性。本文介绍了生物启发熵安全（BioEnS），这是一种新颖的闭环框架，旨在通过实现paretooptimal自适应安全来克服固有的安全-隐私-效率三难困境。BioEnS将自适应防御建模为一个实时的、有约束的多目标优化问题，根据当前环境动态解决安全保障（$Phi$）和资源消耗（$Psi$）之间的权衡。框架核心依赖于硬件信任根熵源（HRTES），它为非确定性密钥派生提供可量化的puf派生的最小熵率（$E_{text{rate}}$），并将其输入自适应安全管理器（ASM）。该机制通过一个占主导地位的$lambda$惩罚项严格执行与上下文相关的安全需求（$Phi_{text{req}}$），从而支持超低延迟策略决策。在ARM Cortex-M平台上的实验验证显示了卓越的性能：BioEnS保持了接近零的安全违规率（SVR） （0.02）%) while simultaneously yielding a superior lifetime extension ratio (LER) of $0.69 times$ relative to the high-security baseline (HSB), confirming the validity of the guaranteed policy enforcement.
{"title":"A Neuroentropy-Driven Nature-Inspired Framework for Adaptive Privacy and Lightweight Security in Sensor Devices","authors":"Soufiane Ben Othman;Chinmay Chakraborty;Saranjit Singh;Mohamed Amine Frikha","doi":"10.1109/JSEN.2026.3658197","DOIUrl":"https://doi.org/10.1109/JSEN.2026.3658197","url":null,"abstract":"Sensor devices and Internet of Things (IoT) devices face a critical, fundamental challenge: deploying robust security while operating under severe constraints on energy, processing power, and memory. This article presents biologically inspired entropy security (BioEnS), a novel, closed-loop framework designed to overcome the inherent security–privacy–efficiency trilemma by achieving paretooptimal adaptive security. BioEnS models adaptive defense as a real-time, constrained multiobjective optimization problem, dynamically resolving the trade-off between security assurance (<inline-formula> <tex-math>$Phi$ </tex-math></inline-formula>) and resource consumption (<inline-formula> <tex-math>$Psi$ </tex-math></inline-formula>) based on current context. The framework core relies on a hardware root-of-trust entropy source (HRTES), which provides a quantifiable PUF-derived min-entropy rate (<inline-formula> <tex-math>$E_{text{rate}}$ </tex-math></inline-formula>) for nondeterministic key derivation, feeding into an Adaptive Security Manager (ASM). This mechanism rigorously enforces context-dependent security requirements (<inline-formula> <tex-math>$Phi_{text{req}}$ </tex-math></inline-formula>) through a dominant <inline-formula> <tex-math>$lambda$ </tex-math></inline-formula>-penalty term, enabling ultralow latency policy decisions. Experimental validation on an ARM Cortex-M platform demonstrates exceptional performance: BioEnS maintains a near-zero security violation rate (SVR) (0.02%) while simultaneously yielding a superior lifetime extension ratio (LER) of <inline-formula> <tex-math>$0.69 times$ </tex-math></inline-formula> relative to the high-security baseline (HSB), confirming the validity of the guaranteed policy enforcement.","PeriodicalId":447,"journal":{"name":"IEEE Sensors Journal","volume":"26 6","pages":"9356-9363"},"PeriodicalIF":4.3,"publicationDate":"2026-02-02","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147440503","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}
Reliable air data is essential for safe and stable flight operations. However, physical air data sensors are susceptible to failure due to environmental disturbances, especially in ultralight manned aircraft, where hardware redundancy is often impractical due to strict size, weight, and power constraints. Although model-based synthetic air data systems (SADS) have been proposed to reduce sensor reliance, they require precise aerodynamic coefficients and are sensitive to modeling errors. To address these limitations, this study proposes a lightweight, data-driven SADS framework based on a hybrid deep learning model that combines temporal and trend-based features. An unscented Kalman filter (UKF) is applied as a postprocessing step to enhance robustness against noise and anomalous inputs. The system is trained and validated on real-world flight data and demonstrates improved accuracy and stability over conventional deep learning baselines. These results suggest that the proposed method offers a robust and complementary alternative to model-based SADS, particularly in resource-constrained flight environments.
{"title":"Robust Synthetic Air Data Estimation via Kalman-Aided Deep Learning Approach for Analytical Redundancy","authors":"Hyuntae Bang;Angelo Lerro;Wonkeun Youn;Hyojung Ahn","doi":"10.1109/JSEN.2026.3658151","DOIUrl":"https://doi.org/10.1109/JSEN.2026.3658151","url":null,"abstract":"Reliable air data is essential for safe and stable flight operations. However, physical air data sensors are susceptible to failure due to environmental disturbances, especially in ultralight manned aircraft, where hardware redundancy is often impractical due to strict size, weight, and power constraints. Although model-based synthetic air data systems (SADS) have been proposed to reduce sensor reliance, they require precise aerodynamic coefficients and are sensitive to modeling errors. To address these limitations, this study proposes a lightweight, data-driven SADS framework based on a hybrid deep learning model that combines temporal and trend-based features. An unscented Kalman filter (UKF) is applied as a postprocessing step to enhance robustness against noise and anomalous inputs. The system is trained and validated on real-world flight data and demonstrates improved accuracy and stability over conventional deep learning baselines. These results suggest that the proposed method offers a robust and complementary alternative to model-based SADS, particularly in resource-constrained flight environments.","PeriodicalId":447,"journal":{"name":"IEEE Sensors Journal","volume":"26 6","pages":"9347-9355"},"PeriodicalIF":4.3,"publicationDate":"2026-02-02","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147440544","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：