Vibration-based fault diagnosis of automotive suspension systems using voting-based ensemble learning

• Novel voting-based approach to diagnose suspension faults using vibration signals. • Statistical and histogram features were extracted with J48 for feature selection. • Eight standalone classifiers were assessed for three load conditions. • Different classifier combinations (two, three, four and five) were considered. • Voting improved classification by 1.20 % (statistical) and 1.51 % (histogram). This study presents a voting-based ensemble learning approach for the fault diagnosis of vehicle suspension systems using vibration signals captured under three load conditions: no-load (0 psi), half-load (200 psi), and full-load (400 psi). The objective is to accurately classify eight suspension conditions (seven faulty and one healthy) using two types of time-domain features: statistical and histogram-based. The J48 algorithm was employed for initial feature selection. Eight machine learning classifiers namely, Support Vector Machines (SVM), Logistic Model Tree (LMT), Naïve Bayes (NB), Logistic Regression (LR), Multilayer Perceptron (MLP), K Nearest Neighbour (KNN), J48 and Random Forest (RF) were first evaluated individually to identify three top-performing models under each load condition. Subsequently, the best-performing classifiers were combined using voting ensembles across multiple configurations (two to three classifiers) and five voting schemes. Results showed that the highest classification accuracy for statistical features was achieved by LMT (91.25 %) under no load, RF (81.25 %) under half load, and J48 (85.00 %) under full load. For histogram features, LMT yielded 89.37 % (no load) and 83.75 % (half load) while Naïve Bayes achieved 75.00 % (full load). The ensemble voting approach significantly enhanced performance: (i) Statistical features – LMT+RF (92.50 %, no load), RF+J48+NB (82.50 %, half load), J48+NB (85.62 %, full load); (ii) Histogram features – LMT+SVM (90.00 %, no load), LMT+SVM (84.37 %, half load), NB+MLP+RF (77.50 %, full load). These results confirm that the voting-based ensemble method consistently improves classification accuracy over individual models and demonstrates robust performance across varying operational conditions.