Abstract: Condition-based maintenance (CbM) is a successful cost-cutting approach that is used to reduce production losses by avoiding breakdowns via effective continuous monitoring of the production equipment. Machine Learning (ML) algorithms can be used for this purpose in terms of predicting defects and diagnostics due to the huge amount of data generated by the incorporation of advance analog and digital technologies in manufacturing industries. To improve the prediction accuracy of the CbM, this study proposed a new framework using various ML algorithms. The new framework has two prediction stages: one to classify whether the status of the machine is on an operation mode or a failure mode (binary classification), and a second one to classify the type of the failure (multi-class classification). A comparative study using a public datasets is used to evaluate the proposed ML algorithms. To address the problem of data unbalance, a new modified data was introduced using the RandomOverSampler method. The proposed ML models were implemented on both the original public dataset and its modified version to perform binary and multi-class classification tasks. The hybrid XGBoost-DT prediction model achieves the best and most robust classification and prediction accuracy.

Key words: condition-based maintenance, machine learning, artificial intelligence, manufacturing industries, prediction accuracy, classification