Multi-Scale Feature Fusion Fault Diagnosis Method Based on Attention Mechanism

Abstract

As a key component of electromechanical equipment in the intelligent manufacturing process, rolling bearings play an important role in securing a stable operation. Deep learning techniques facilitate the extraction of features with datasets; however, when dealing with data from multiple working conditions, conventional deep learning approaches tend to inaccurately represent the features of any given condition, consequently affecting the fault diagnosis accuracy of the model. Methods that pre-distinguish operating conditions and subsequently model them separately cannot guarantee real-time fault diagnosis, and the limited amount of labelled data for each condition further hinders the real-time nature of fault diagnosis. Consequently, designing a multi-condition feature extraction method becomes imperative. This study aims to propose a multi-scale feature fusion approach based on the attention mechanism, which addresses the issue of insufficient information filtering in traditional multi-scale feature fusion methods under multi-condition, high-noise scenarios, ultimately leading to subpar model fault diagnosis accuracy. The proposed method leverages multiple networks to extract features from both single-condition and mixed-condition data. By utilizing the attention mechanism, features with distinguishable working conditions are selectively identified, thereby enhancing the effectiveness of information fusion and ultimately improving the accuracy of multi-condition fault diagnosis. Experimental validation was conducted using the bearing dataset from Case Western Reserve University. The results demonstrate that, under multi-condition and high-noise scenarios, the proposed method exhibits higher diagnostic accuracy compared to other multi-scale learning approaches.