Application of Neural Networks for Real-Time Decision Support in Virtual Approval of Brake Components

Abstract

This study aims to present a virtual numerical validation procedure for durability in brake system components, using artificial neural networks and based on experimental bench tests. The study focus was concentrated on the drum brake spider component, responsible for mechanically connecting the brake system subassemblies. To develop the validation procedure, engineering software such as ABAQUS, Fe-Safe, Minitab, and MATLAB was used. These were crucial for carrying out stress analyses, statistical data validation, and construction of an Artificial Neural Network (ANN) capable of predicting finite element responses, fatigue life, and supporting real-time decision-making for structural validation of mechanical components. The results obtained from these tools allowed the calibration of a numerical virtual model using the Finite Element Method (FEM) based on mechanical theories and results obtained in bench tests with the brake system, thus, a finite element database was generated for the application of the ANN, containing 130 data from a total of 4,800 possible combinations. The training, validation, and testing of the ANN were determined using a performance analysis algorithm. Finally, the results obtained with the artificial neural network were compared with the results of finite elements and computational fatigue life. The efficiency of the real-time response prediction method was measured using the Mean Squared Error (MSE). With the use of ANN, it was possible to obtain an average error of 0.85% for predicting maximum principal stress and an error of 10.33% for predicting fatigue life. For the classification of fatigue life results, the ANN presented an accuracy of 100%, enabling decision-making in real-time.