Optimization-enhanced machine failure classification using critical sensor features and hybrid learning models with advanced optimization techniques.

Industrial machine failures pose a significant challenge in modern manufacturing, leading to unexpected interruptions, increased maintenance costs, and reduced equipment reliability. Accurate machine failure prediction enables proactive maintenance strategies that minimize unexpected failures and improve operational efficiency. This study proposes an optimization-driven hybrid machine learning framework for predicting machine failure using a milling-machine condition dataset derived from operational sensor data collected in a milling-machine environment comprising 9,976 observations and 14 machine-condition attributes. Three predictive models, including Light Gradient Boosting (LGBC), Random Forest Classification (RFC), XGBoost Classification (XGBC), and a hybrid voting ensemble, were developed and subsequently optimized using the Tasmanian Devil Optimization (TDO) algorithm to identify optimal hyperparameter configurations. The models' performance was evaluated by using several evaluation metrics. Experimental results demonstrated that optimization significantly improved forecasting performance. Among the baseline models, LGBC achieved a test accuracy of 0.965, while RFC and the voting ensemble achieved 0.947 and 0.959, respectively. Following optimization, the TDO-enhanced LGBC (LGTD) achieved the highest predictive accuracy of 0.985, outperforming all competing approaches. The optimized hybrid ensemble (LG-RF-TD) also demonstrated strong predictive capability with a test accuracy of 0.979 and balanced classification metrics under both normal and fault-inducing operating conditions. Sensitivity analysis further revealed that torque and process temperature were the most influential variables in predicting machine-related failures.