Automatic detection of brake faults using a 1D CNN applied to NVH signals

Abstract The reliability of braking systems is one of the foundational pillars on which vehicle safety rests. However, conventional diagnostic approaches rely on expert inspection of NVH phenomena and suffer from well-known limitations, including high costs and the inherent subjectivity related to the complexity of these systems. To this extent, this article suggests a method to diagnose brake faults through a 1D-CNN using the NVH measurements performed on the test bench FIVE@ECL. The contribution of the current research lies in a data-driven 1D-CNN approach for extracting discriminatory temporal features from these signals for multi-state classification. In this work, three different working conditions are taken into consideration: a healthy system, an intermediate fault, and a severe failure. After a structured pre-processing phase including filtering and normalization, the synchronized multichannel NVH temporal observations were subsequently used for training the proposed model, which achieved an accuracy of 98.66%, with F1 scores higher than 0.97 and AUC values close to 1.00. This testifies that the adopted 1D-CNN architecture is able to extract time features that are discriminant for each operating condition of the braking system and, therefore, to safely tell apart the studied conditions. This provides a basis for the development of continuous monitoring algorithm for the braking systems, opening wide perspectives in predictive maintenance embedded systems using deep learning approaches. Future work will focus on increasing the dataset size and compressing the network model to allow for its implementation in real-time.