AI-ENABLED PREDICTIVE MAINTENANCE AND FAULT DIAGNOSIS OF INDUSTRIAL MECHANICAL SYSTEMS USING MACHINE LEARNING

Abstract

Industrial mechanical systems generate various types of vibration, temperature, current, torque, rotation and pressure data streams, all of which can be harnessed to detect incipient faults and help with maintenance planning. The use of artificial intelligence (AI) for a fault diagnosis and remaining useful life estimation (RUL) framework for predicting failure of rolling bearings is proposed in this study. In the proposed framework, the sensor preprocessing, frequency domain transformation, feature selection, ensemble machine learning and temporal deep learning are included. The numerical experiment is not based on any proprietary industrial data, but rather is a controlled-benchmark simulation with the goal of creating as realistic a simulation of class imbalance and sensor noise and variation of operating conditions as possible. A synthetic benchmark of 12 000 sensor windows on 6 operating states was designed, with 70 % of the windows used for training, 15 % for the validation and 15 % for the test partition. Random Forest, XGBoost, Support Vector Machine, and CNN-LSTM models along with a proposed Hybrid Ensemble were chosen to be evaluated and compared along the lines of accuracy, precision, recall, F1-score, ROC-AUC, false-alarm rate and RUL error. An accuracy of 98.1%, a F1-score of 97.8% and an ROC-AUC of 0.992 were obtained for the proposed hybrid model, whereas for the CNN-LSTM the results were 97.2% accuracy and 96.7% F1-score. The hybrid system was able to predict the RUL with a mean absolute error of 1.02 h over a four-hour prediction horizon and an illustrative false alarm rate of 1.6%. The most significant diagnostic variables based on feature analysis are the RMS vibration, the kurtosis, the temperature and the torque. The obtained results indicate that a physically meaningful signal description with a nonlinear ensemble learning and temporal representation may be a good architecture for mechanical fault diagnosis in industry. The manuscript also concentrates on the deployment of the model, with explainability, uncertainty estimation and readiness for digital twin.