Real-Time Defect Identification in Automotive Brake Calipers Using PCA-Optimized Feature Extraction and Machine Learning

TL;DR

This paper introduces a non-contact system using impact-acoustic measurements and machine learning to detect defects in automotive brake calipers in real time.

Contribution

The novel contribution is a PCA-optimized feature extraction method combined with machine learning for real-time defect detection in brake calipers.

Findings

PCA identified Shannon Entropy as the most discriminative feature for defect classification.

Machine learning models achieved over 95% accuracy in classifying normal and defective calipers.

A GUI-based software was developed for real-time defect identification and visualization.

Abstract

This study aims to develop a non-contact automated impact-acoustic measurement system (AIAMS) for real-time detection of manufacturing defects in automotive brake calipers, a key component of the Electric Parking Brake (EPB) system. Calipers hold brake pads in contact with discs, and defects caused by repeated loads and friction can lead to reduced braking performance and abnormal vibration and noise. To address this issue, an automated impact hammer and a microphone-based measurement system were designed and implemented. Feature extraction was performed using Fast Fourier Transform (FFT) and Principal Component Analysis (PCA), followed by defect classification through machine learning algorithms including Support Vector Machine (SVM), k-Nearest Neighbor (KNN), and Decision Tree (DT). Experiments were conducted on five normal and six defective caliper specimens, each subjected to 200 repeated measurements, yielding a total of 2200 datasets. Twelve statistical and spectral features were extracted, and PCA revealed that Shannon Entropy (SE) was the most discriminative feature. Based on SE-centric feature combinations, the SVM, KNN, and DT models achieved classification accuracies of at least 99.2%/97.5%, 98.8%/98.0%, and 99.2%/96.5% for normal and defective specimens, respectively. Furthermore, GUI-based software (version 1.0.0) was implemented to enable real-time defect identification and visualization. Field tests also demonstrated an average defect classification accuracy of over 95%, demonstrating its applicability as a real-time quality control system.

Figures