Data-Driven Fault Detection for Dynamic Systems With Performance Degradation: A Unified Transfer Learning Framework
Hongtian Chen, Zheng Chai, Bin Jiang, Biao Huang
Abstract
Continuous operations can result in performance degradation of industrial systems, which naturally increases complexity in fault detection (FD). In this study, a transfer learning method is proposed for detecting sensor faults in dynamic systems with consideration of actuator-performance degradation, whose structure is a federated neural network. By retaining and reusing previous knowledge stored in neural networks, the proposed scheme can adaptively adjust system models (or parameters) based on which sensor FD design objectives can be achieved. One of the superior strengths of the proposed method to the existing transfer learning-based methods is its ability in dealing with any type of data regardless of their probabilistic distributions. Also, the proposed method avoids a trial-and-error endeavor when establishing the structure and training the proposed federated neural networks. Furthermore, the proposed structure can be regarded as a new unified framework for data-driven parameter identification with adaptive model calibration, based on which the developed FD methods become efficient for handling industrial systems with any type of parameter changes. A study on an electrical drive system demonstrates the effectiveness and feasibility of the proposed method.