Time-frequency Hypergraph Neural Network for Rotating Machinery Fault Diagnosis with Limited Data

Abstract

Due to the scarcity of fault samples and the weakness of processing higher-order interactive information, the most existing intelligence methods fail to achieve the optimal effect in fault diagnosis. To address these problems, a time-frequency hypergraph neural network-based fault diagnosis method is proposed. In the proposed network, the limited data is initially segmented using the sliding window mechanism to obtain a set of time-domain signal instances. Additionally, the Fast Fourier Transform (FFT) is applied to each signal instance to extract corresponding frequency-domain signals, so as to capture more fault-sensitive features. Subsequently, a two-layer convolutional neural network is used to extract fault-attention features from both the time and frequency domain signals. Also, in order to reduce computational complexity, the time-frequency domain features are adaptively stacked based on a self-attention mechanism. Furthermore, a feature similarity graph is constructed for the time-frequency domain features using a k-nearest neighbor algorithm. This graph is then input into the hypergraph neural network (HGNN) to obtain the final diagnosis results. One comparative experiment shows that the proposed method not only mitigates the performance degradation caused by limited samples and noisy environments, but also effectively leverages the higher-order interaction information among nodes in the hypergraph.