Diagnosing Faults in Suspension System Using Machine Learning and Feature Fusion Strategy

Abstract

Comfort and safety in automobiles can be enhanced with predictive maintenance by means of early problem detection and isolation, collectively referred to as fault diagnosis. To maintain a lead position in automotive industry, fleet managers are turning to predictive analysis. In this study, an attempt was made involving feature fusion to determine most significant features required to determine the suspension faults using vibration signals and machine learning approach. Three different features extraction techniques such as statistical, histogram and autoregressive moving average model (ARMA) were extracted from the acquired vibration signals for different fault conditions at the three different loads by means of a specially fabricated experimental setup. Feature selection was done for individual features using J48 decision tree algorithm. The performance of tree-based classifiers was assessed on the chosen individual features. Additionally, each individual feature was paired with others in four distinct combinations: statistical-histogram, statistical-ARMA, ARMA-histogram and ARMA-histogram-statistical, across three different loads. These combined features were then input into tree-based algorithms to identify the optimal classification algorithm, regardless of the load conditions. The results obtained in this study indicate that the combination of the ARMA-histogram-statistical feature with a random forest classifier yields optimal classification accuracy, regardless of load conditions, with values of 98.125%, 99.375% and 96.250%, respectively, and an average computational time of 0.10 s.