Extensive Feature Extraction Approach For Fault Diagnosis In Predictive Maintenance Enabled By Machine Learning

Abstract

With the advent of autonomous vehicles, there is a growing need for reliable and efficient classifiers in complex vehicular environments to detect faults. The diagnosis of faults is a key step to detect operating conditions that are abnormal early and in a way that will help in effective predictive maintenance. Machine-learning-based approaches offer potential solutions for fault diagnosis of vehicles based on the data collected from the vehicles, and with the development of wireless and intelligent vehicular technologies, it is now possible to use this observed data to complete this diagnosis. Four feature extraction and dimensionality-reduction methods, Principal Component Analysis (PCA), t-distributed Stochastic Neighbor Embedding (t-SNE), Non-Negative Matrix Factorization (NNMF), and Linear Discriminant Analysis (LDA), are tested and compared with a Support Vector Machine (SVM) for vehicle fault diagnosis. The extracted feature representations are fed to the SVM classifier, which classifies the vehicle conditions being observed. The results of the experiments reveal that the accuracy of PCA-SVM, t-SNE-SVM and NNMF-SVM is 99.14%, while that of LDA-SVM is 88.03%. In the set of the evaluated methods, t-SNE-SVM has the highest recall on 79.05% and F-measure of 79.02%, keeping an accuracy of 80% as PCA-SVM and NNMF-SVM methods. The results show the effect of feature representation on the detection of faults and that the representation t-SNE gives a relatively good representation of the data under evaluation. The proposed method can be used for automating vehicle fault diagnosis and be used as a framework for a predictive-maintenance vehicle application for the observed operating conditions of the vehicle.