A FAULT DIAGNOSIS METHOD OF ACTIVE BRAKING SYSTEMS BASED ON 1D-CNN AND MHAM

To overcome insufficient feature representation and limited status quantification in active braking systems (ABS), an end-to-end fault diagnosis method of ABS is constructed to combine a one-dimensional convolutional neural network (1D-CNN) and a multi-head attention mechanism (MHAM) under conditions of stronger noise interference and multi-source nonlinear coupling. Based on a 1D-CNN front-end encoder to directly decouple local high-frequency transient features from the original time-series waveform, the long-distance dependency topology within the sequence is deeply reconstructed to quantify the contribution weights of features at different time steps to the evolution of the system's health status. according to the joint constraints of the Adam adaptive optimization algorithm and regularization penalty term, the model effectively avoids the overfitting risk of deep networks and significantly enhances its generalization robustness under unknown and complex conditions. Empirical results strongly demonstrate the superiority of this method. The regression prediction determination coefficient R2 for the degree of failure reaches 0.9472, and the root mean square error (RMSE) is reduced by 1.56%, achieving high-precision quantitative perception of the system's health status.