Hydraulic brake fault detection from acoustic signals using frequency domain filter bank integrated deep learning framework

Brake failure is a critical safety issue in the automotive industry, often resulting from mechanical wear, hydraulic malfunctions, or electronic failure. Common causes include brake fluid leakage, worn-out brake pads, and damaged brake lines, exacerbated by poor maintenance and ignored warning signs, such as squealing. Effective fault detection is essential to ensure road safety. This study is motivated by the critical need to enhance the accuracy of brake fault detection and classification, which is a key factor in improving road safety and reducing the risk of accidents caused by brake system failures. This paper presents an automated approach for detecting hydraulic brake faults by analyzing five types of brake sound anomalies, Type I to Type V, covering various disk pad wear scenarios. These time-domain sound signals are converted to frequency-domain image representation using a mel spectrogram, and a custom 14 layer Deep Sequential Convolutional Neural Network (CNN) model classifies these images. The model achieved 97.7% accuracy in binary classification and 97.8% in 5-class classification, with perfect results at 20 epochs during hold-out cross-validation. This study significantly enhances the safety of hydraulic brake systems by automating fault detection and classification using a deep sequential CNN framework.