Development and comparative evaluation of various fault detection algorithms for a drum brake using artificial neural networks and a physics-based model.

Drum brakes are among the essential safety components in automobiles; hence, it is critical to detect any fault in the system that might hinder its performance. A fault detection framework based on artificial intelligence and a rigid-body model of a system (as a digital experiment) is proposed to detect four of the most common faults in automotive drum brakes. First, a contact mechanics-based nonlinear model of a drum brake (incorporating conformal contact between the brake shoes and the brake drum) is developed to estimate the dynamic response accurately. The predictions from this model are compared with the data available in the literature. The responses estimated from the model with a known level of faults form the input data set for various fault quantification algorithms. Sensitivity-based pruning technique in conjunction with an artificial neural network is used to select the optimal set of features for which the physical interpretations are deduced. The developed fault detection methodologies are quantitatively compared, and the most applicable among them is determined based on the detection accuracy and the computational cost. Finally, the robustness of the selected methodology is tested under the presence of white Gaussian noise in the measured signals. The proposed framework and the selected methodology will help design an accurate health monitoring tool to diagnose faults in automotive drum brakes. Though the current study focuses on drum brakes, the possible generalization of its framework to other systems is discussed. • Formulated a novel conformal contact applicable to similar contact-based systems. • Developed a conformal contact-based mathematical model for a drum brake. • Developed a fault detection framework for a drum brake to quantify four major faults. • Compared multiple fault detection algorithms and selected the optimum algorithm. • Analyzed the robustness of the optimum algorithm using noisy signals. [ABSTRACT FROM AUTHOR]