Multi-scale attention-based xLSTM for rolling bearing fault diagnosis

Abstract In modern industrial manufacturing, the condition monitoring of rolling bearings is vital for ensuring the reliable operation of mechanical systems. However, existing deep learning approaches often struggle to effectively capture the complex characteristics of bearing vibration signals, which exhibit long-term dependencies, strong feature coupling, multi-periodic patterns, and nonlinear behaviors. To address these challenges, this paper proposes a novel intelligent fault diagnosis framework called multi-scale attention-based xLSTM. The proposed model integrates three key components: a multi-scale convolution (MSC) module, a hierarchical extended long short-term memory (HXLSTM) module, and a multi-head self-attention (MHSA) module. Specifically, the MSC module employs a multi-branch parallel architecture to extract fault features across multiple scales, enabling effective multi-scale representation of complex signal patterns. The core HXLSTM module further facilitates the modeling of long-term dependencies and nonlinearities of fault features through a hierarchical xLSTM-based structure. Additionally, the MHSA module captures critical features and temporal dependencies across different timescales while suppressing noise and redundant information. The superior performance of the proposed model is verified on three public bearing fault datasets. The experimental results show that the proposed model achieves an average diagnostic accuracy of 98.89%, and outperforms other models in terms of accuracy, sample dependency, generalization, and diagnostic stability under complex conditions.