Abstract

Abstract Various data-driven methods based on acoustic emission (AE) signals have been proposed to monitor and accurately identify the wear stages of the grinding wheel. However, extracting effectively generalized and discriminative features from AE signals remains a challenging task. This paper proposes a new lightweight hybrid deep learning model that combines enhanced convolution with enhanced vision transformer to effectively address the above challenges. Specifically, the key contributions of this paper are three-fold: (1) A two-stage signal preprocessing mechanism based on variational mode decomposition and continuous wavelet transform is proposed to improve the signal-to-noise ratio and feature representation of the AE signals. (2) To prevent model overfitting, a new regularization strategy based on stabilizing sparse convolutional weights and a weight penalty mechanism is designed. This approach improves the hybrid mode’s capacity to extract generalized features. (3) To concentrate on capturing multi-scale discriminative features between different wear conditions, a parameter-efficient residual convolution module based on the dropout depthwise convolution is designed, which is utilized to reconstruct the encoder of the vision transformer. In particular, to improve the training efficiency of the model, a lightweight mechanism using a stage-stride decreasing strategy is used to compress the spatial dimensions of the feature maps in the attention mechanism. The ablation experiment demonstrates the rationality of the proposed model structure. Comparative experiments show that the proposed method achieves a diagnostic accuracy of 99.6% on the test set and outperforms other state-of-the-art deep learning methods.