Multi-scale Attention-based xLSTM for Rolling Bearing Fault diagnosis

Abstract In modern industrial production, the health status of rolling bearings is critical to the safe and stable operation of mechanical equipment. Deep learning methods are able to automatically capture the deep abstract features contained in rolling bearing vibration signals in an end-to-end manner, which displays prominent and unparalleled advantages over traditional signal processing and feature selection algorithms. Recently, xLSTM (Extended Long Short-Term Memory) realizes a major innovation to the classic time series model LSTM. In this research, a novel intelligent fault diagnosis model, multiscale attention-based xLSTM (MAXLSTM), is proposed to enhance the capability of capturing long-term dependent features in nonlinear bearing signals with diverse signal sources under complex operating conditions. First, a multiscale convolution module (MSC) is developed to extract multiscale features considering the strong coupling of fault features; moreover, a hierarchical xLSTM module (HXLSTM) is proposed to accurately model diverse, yet more explicit, long-term dependencies of signals. In particular, multi-head self-attention (MHSA) enables the model to learn various aspects of the hierarchical features in parallel, thereby further improving the recognizability of bearing fault features. The comparative experiments on three different fault diagnosis tasks demonstrate the superiority of MAXLSTM over five state-of-the-art diagnostic methods. The noise immunity tests at different signal-to-noise ratios suggest that MAXLSTM maintains high diagnostic accuracy and generalization even in high-intensity noise cases. The ablation experiments quantify the performance contributions of the core components of the model, especially for the effectiveness of the synergy between MSC and MHSA.