An Evolutionary Optimized Residual CNN with Attention-Enhanced LSTM for Intelligent Bearing Fault Diagnosis

Abstract

Rolling bearings are a critical component for the reliability of rotating machinery, but traditional fault diagnosis methods rely on manually designed features, resulting in insufficient robustness under varying operating conditions. To address this issue, this paper proposes an evolutionary optimization framework that integrates a Residual Convolutional Neural Network (ResCNN) and an Attention-Enhanced Long Short-Term Memory Network (A-LSTM) to achieve intelligent fault diagnosis of bearings. This model autonomously extracts spatial features and temporal dependencies from raw vibration signals and utilizes an attention mechanism to highlight important fault modes. To further improve performance, three innovative improvements are introduced: an Adaptive Gated Feature Fusion (AGFF) mechanism for learning the fusion ratio dimension by dimension; a Multi-Scale Temporal Convolution (MSTC) module that uses parallel dilated convolutions to capture multi-scale temporal information; and a Structured Genetic Programming (GP) hyperparameter optimization strategy. Extensive experiments were conducted on the CWRU and HUST bearing datasets, including cross-load validation, noise robustness analysis, and cross-domain transfer testing. The results show that the proposed method achieves superior diagnostic accuracy and robustness compared to baseline CNNs, LSTMs, traditional hybrid models, and current state-of-the-art methods. This framework provides an efficient solution for predictive maintenance applications.