Machine Learning (ML) Based Repair-Count and Periodic Maintenance Policy for Multipurpose CNC Machinery

This study deals with developing a maintenance policy optimization framework using a machine learning approach for multipurpose CNC machinery in an automotive part manufacturer. Throughout the 59958 minutes of observation, the mean time between failure (MTBF) indicates improper maintenance policy, which is 1100.81 minutes, which causes high maintenance costs. Several censored data were utilized to predict failure times using machine learning algorithms, such as process temperature, air temperature, rotational speed, and torque. It is found that extreme gradient boosting (XGBoost) outperforms other classification algorithms with about 96% accuracy in validation and test data. The best model from the XGBoost classifier is used to predict future failure times using the additional data from 10080 minutes of observation. The expected failure times then become the input for obtaining optimal maintenance policy by considering its statistical distribution. This study observes the repair count (strategy 1) and the periodic maintenance (strategy 2) to produce a minimum maintenance cost. The component replacement in strategy 1 is performed if only the repair frequency reaches the threshold number (i.e., k). Strategy 2 is based on a predetermined preventive maintenance time (i.e., T). Subsequently, strategy 1 produces a lower cost with the expected cost rate of IDR 335.48 and an optimal failure threshold of k = 5. Meanwhile, strategy 2 has an expected cost rate of IDR 465.58 at the optimal periodic time of T = 15597.12 minutes. Finally, this study contributes to a novel framework for establishing a data-driven maintenance strategy that satisfies the manufacturer’s interest.