TY - GEN
T1 - Mechanism-Aware Low-Rank Sparse Neural Unrolling for Incipient Bearing Anomaly Detection
AU - Liu, Yunpeng
AU - Jiang, Hongkai
AU - Jiang, Wenxin
AU - Li, Zhenning
AU - Zio, Enrico
N1 - Publisher Copyright:
© 2026 IEEE.
PY - 2026
Y1 - 2026
N2 - Intelligent anomaly detection with deep learning is promising for mechanical systems. However, it is struggling due to its 'black-box' issue and opacity. Model-driven sparse representation provides a clearer physical meaning, whereas its parameters and threshold operators are usually tuned manually and are sensitive to industrial noise. In this paper, a mechanismaware low-rank sparse unrolling network (LRSU-Net) is proposed for incipient bearing anomaly detection. Firstly, vibration signals are formulated with low-rank and sparse components, where stationary operating backgrounds and transient fault impulses are considered separately. Based on this formulation, the iterative optimization process is unrolled into a feed-forward network via the half quadratic splitting algorithm. Secondly, a low-rank estimation module is designed to suppress periodic operational components and provide a cleaner residual space for weak impulse extraction. Thirdly, an adaptive convex proximal network is constructed to learn the sparse estimation operator. It preserves the adaptive ability of non-convex sparse regularization while keeping the convexity of proximal mapping. The proposed method is validated on the NASA IMS bearing run-to-failure dataset. Experimental results indicate that LRSU-Net improves incipient anomaly detection under strong interference. Moreover, the reconstructed sparse components are consistent with fault characteristic frequencies.
AB - Intelligent anomaly detection with deep learning is promising for mechanical systems. However, it is struggling due to its 'black-box' issue and opacity. Model-driven sparse representation provides a clearer physical meaning, whereas its parameters and threshold operators are usually tuned manually and are sensitive to industrial noise. In this paper, a mechanismaware low-rank sparse unrolling network (LRSU-Net) is proposed for incipient bearing anomaly detection. Firstly, vibration signals are formulated with low-rank and sparse components, where stationary operating backgrounds and transient fault impulses are considered separately. Based on this formulation, the iterative optimization process is unrolled into a feed-forward network via the half quadratic splitting algorithm. Secondly, a low-rank estimation module is designed to suppress periodic operational components and provide a cleaner residual space for weak impulse extraction. Thirdly, an adaptive convex proximal network is constructed to learn the sparse estimation operator. It preserves the adaptive ability of non-convex sparse regularization while keeping the convexity of proximal mapping. The proposed method is validated on the NASA IMS bearing run-to-failure dataset. Experimental results indicate that LRSU-Net improves incipient anomaly detection under strong interference. Moreover, the reconstructed sparse components are consistent with fault characteristic frequencies.
KW - algorithm unrolling
KW - incipient bearing anomaly detection
KW - low-rank sparse unrolling network
KW - sparse representation
UR - https://www.scopus.com/pages/publications/105053006980
U2 - 10.1109/SRSE70977.2026.11694471
DO - 10.1109/SRSE70977.2026.11694471
M3 - 会议稿件
AN - SCOPUS:105053006980
T3 - 2026 the 8th International Conference on System Reliability and Safety Engineering, SRSE 2026
SP - 214
EP - 220
BT - 2026 the 8th International Conference on System Reliability and Safety Engineering, SRSE 2026
PB - Institute of Electrical and Electronics Engineers Inc.
T2 - 8th International Conference on System Reliability and Safety Engineering, SRSE 2026
Y2 - 7 August 2026 through 10 August 2026
ER -