A data-driven approach to fault modeling and diagnosis of brake-by-wire system

Abstract

E-Booster plays an important role in braking safety as the pressure source for vehicle hydraulic systems. According to the design feature of its series controller, the signals on key nodes, which can be directly measured, can be used as the data source for fault diagnosis. In this paper, a deep learning technique of residual 1-D CNN equipped with parallel structure is proposed for fault recognition and classification in a dynamic process. With the help of wavelet transform and probabilistic heat map, it is found that the phase current fault feature is distinct in the frequency domain. In contrast, the uniform demagnetization fault and pressure recession fault are more obvious in the time domain. Therefore, the parallel network structure with a wide and narrow convolutional kernel is used, which can handle multiple complex fault features simultaneously. Simulation models based on both data-driven and mathematical formulations are also established to better fit the actual nonlinear conditions. Finally, the proposed network structure can reach 92.1% classification accuracy compared with the commonly used lightweight 2-D CNN model. It can be concluded that 1-D CNN achieves similar classification results as 2-D CNN with less computational resource consumption.