Predicting Post-Fatigue Life of Electric Power Steering Components Using Machine Learning and Reliability Models

Abstract

This study investigates the effectiveness of machine learning and statistical methods for predicting the life expectancy of Electric Power Steering (EPS) components post-fatigue testing. Using a synthetic dataset of 800 samples with five failure modes across three stress levels, we compare ARIMA, SARIMA, Decision Tree, Random Forest, Gradient Boosting, Weibull Analysis, Kaplan-Meier estimation, and Cox Proportional Hazard models. Weibull analysis reveals that high-stress components exhibit an increasing failure rate (β = 2.18) with characteristic life of 92,807 cycles, while low-stress components show approximately constant failure rates (β = 1.02, η = 192,682 cycles). The Cox Proportional Hazard model achieves a concordance index of 0.914, and log-rank tests confirm statistically significant survival differences between all stress level pairs (p < 0.001). ARIMA and SARIMA achieve MAE between 33,512 and 181,189 cycles. Among classifiers, Random Forest achieves the highest test accuracy (29.4%) across five failure modes. Results demonstrate that integrating statistical reliability methods with machine learning provides complementary insights for predictive maintenance.