WCBLM: an improved convolutional neural network for intelligent diagnosis of power-shift system faults in hydro-mechanical transmission tractors with limited sample length

Abstract To enhance the shift reliability of agricultural hydro-mechanical transmission tractors and alleviate controller area network congestion due to the transmission of extensive hydraulic pressure data, an intelligent diagnostic method for power-shift system faults, constrained by limited sample length, was investigated. Firstly, the types of faults were introduced, and sample data were obtained through experiments; Secondly, the target sample length was determined to be 6 using the traditional convolutional neural network (CNN); Thirdly, an enhanced CNN model, designated as WCBLM, was proposed. This model integrates wide convolutions to broaden the receptive field and a convolutional block attention module to enhance both global contextual and local discriminative features. These refined representations are then processed through an LSTM-based temporal modeling framework to facilitate precise fault classification. Finally, the robustness of the model was tested using Gaussian white noise and real environmental noise. The results indicate that the proposed algorithm attained an accuracy of 97.61% and an F1-score of 97.59%, surpassing traditional models. Even with real environmental noise, the model maintained an accuracy of 95.33%, outperforming support vector machine, multilayer perceptron, ResNet50, CNN-LSTM, and the improved echo state network (Attention_ConvESN) by 3% to 13%.