FCHG: Fuzzy Cognitive Hypergraph for interpretable fault detection

Abstract

While deep learning-based fault detection methods for rolling body mechanical components have achieved good results, their application in real-world scenarios is limited, and their reliability is not satisfactory due to the non-interpretability of deep learning. Meanwhile, automated component vibration signals are usually affected by irregularities, which make it challenging to obtain advanced features by conventional feature extraction methods. To solve the above problems, we develop a new fuzzy cognitive graph model that combines the interpretability of fuzzy cognitive graphs with the high-precision categorization of hypergraphs, called fuzzy cognitive hypergraph (FCHG). First, frequency domain analysis is combined with graph analysis to help the network extract high-level features. Second, causal learning is used to construct the FCHG weight matrix to make the FCHG interpretable and enhance the reliability of fault diagnosis. Finally, the FCHG network is built using spectrogram theory and Chebyshev polynomial expansion to capture multiple feature relationships and improve fault diagnosis accuracy. The practicality and effectiveness of the FCHG model for monitoring the health and stability status of rolling body mechanical components are verified on three datasets.