An Integrated AI and Machine Learning Framework for Intelligent Electromechanical Systems, Predictive Maintenance, Engineering Optimization, and Data-Driven Decision Making
Abstract
Multivariate sensor data is increasingly being used to guide intelligent electromechanical systems into predictive maintenance strategies that deliver timely, reliable and actionable maintenance decisions. This study overcomes the drawback of fragmented approaches, where the anomaly detection, failure prediction and maintenance optimization are treated separately. Four key components of a temporal feature engineering, autoencoder-based anomaly detection, supervised predictive models and multi-horizon early-warning analysis, and Pareto-based maintenance optimization were used to develop an integrated artificial intelligence and machine-learning framework. The framework was tested on the MetroPT-3 industrial time-series data set consisting of electrical, pneumatic, thermal and operational measurements of an air production unit of a metro train. The best model is the Autoencoder + XGBoost model, with the precision being 0.918, recall being 0.893, F1 score being 0.905, ROC-AUC being 0.974, PR-AUC being 0.928, and MCC being 0.889. The best early-warning response was generated by a 30-minute horizon, and the best normalized cost index (0.58) was obtained by implementing a balanced maintenance policy at a risk threshold of 0.52. In summary, the framework showed how to incorporate engineered latent anomaly representations into the supervised learning context to enhance early warning and engineering decisions.