A Cross-Condition Fault Disentangling Matching Network for Industrial Early Fault Enhancement

Industrial early faults suffer from strong fault information coupling in weak signal strength, which further exacerbates the insufficiency of industrial early faults. A widely recognized approach is to enhance early fault information with historical fault information from other working conditions, but it suffers from low accuracy and weak generalization. To address this issue, a cross-condition fault disentangling matching network (CFDM-Net) is proposed, which consists of the fault disentangling subnet and fault matching subnet. Specifically, 1) the fault disentangling subnet disentangles multiple condition faults into private features that can reflect strong stylization of the working condition and shared features that can transfer fault information across conditions, so as to alleviate the strong coupling and provide the feature basis for the early fault enhancement; 2) the fault matching subnet adaptively bias matches early faults and historical faults, so as to enhance the number and types of early faults. The CFDM-Net ensures the accuracy and diversity of the enhanced early faults, which supports the reliable modeling of industrial early fault diagnosis. The real-world application experiments of the alternating current (AC) asynchronous motor early fault diagnosis and the magnetic flux leakage (MFL) based defect early detection are conducted to evaluate the proposed method. Experiment results show the CFDM-Net outperforms the existing methods for modeling under industrial early faults, which achieves the 16.14% improvement for the AC asynchronous motor early fault diagnosis and the 1.476mm error reduction for the MFL-based defect depth inversion.