Gearbox fault diagnosis and feature visualization based on multiscale convolution network with LSTM and attention mechanism

Abstract

Aiming at the problems that traditional gearbox fault diagnosis methods rely on manual feature extraction, have poor adaptability to non-stationary vibration signals, and are difficult to effectively capture multi-scale fault features and key time periods, this paper proposes an end-to-end fault diagnosis model combining Multi-Scale Convolutional Neural Network (MSCNN), Long Short-Term Memory (LSTM) network and attention mechanism. The model uses dual-branch multi-scale convolution to extract local and global fault features, employs bidirectional LSTM to model temporal dependencies, and introduces an attention mechanism to adaptively assign weights to critical fault segments, so as to improve diagnostic accuracy and model interpretability. Experiments are carried out on the Southeast University gearbox fault dataset, with a traditional fully connected network based on handcrafted time-frequency features as the baseline. The results show that proposed model achieves 99.42% classification accuracy for five gear states which is 9.12% higher than the baseline. The t-SNE visualization and clustering metrics verify that the learned features have better intra-class aggregation and inter-class separation. This model effectively overcomes the limitations of traditional methods and provides a high-precision and interpretable scheme for gearbox fault diagnosis.