An Ensemble Learning-Based Alert Trigger System for Predictive Maintenance of Assets with In-Situ Sensors

Abstract

Failure of assets, such as machines, engines, or equipment, can cause significant loss to an enterprise. Real-time monitoring of the asset operation and early detection of failure could guide mechanics or engineers to check the assets in time and reduce the chance of a breakdown. With the availability of Internet of Things (IoT) technologies and the development of machine learning, predictive maintenance has become an effective way to monitor asset performance and detect anomalies using sensor data and historical information. However, many use cases need more information to evaluate the results. Lack of validation can undermine confidence in predictive results. To solve this problem, a novel alert trigger schema that integrates sensor fusion, feature extraction, machine learning, ensemble strategy, and an alert format is proposed. The method offers a comprehensive and reliable approach to detecting anomalies based on unlabeled data, utilizing sensor and decision-level fusion techniques. Meanwhile, eight anomaly detection techniques were investigated, including K-Means Clustering, Gaussian Mixture Model, Autoencoder, and Isolation Forest. Various algorithms generate results with differing degrees of confidence. These results are consolidated into a single indicator representing the alert level. This amalgamation of data ensures the provision of robust and reliable predictions. Instead of simply combing the alert information, confidence in different algorithms is reflected in adding different weights in the ensemble process. In addition, while other existing frameworks focus on evaluating the algorithm’s accuracy, more effort was put into demonstrating a levelbased alert system, showing diagnosis changing with time. In that way, mechanics or engineers can get information to check the asset’s status. The proposed framework was validated using a synthetic dataset based on recordings from a rotating fault simulator that generates multi-modal data, including accelerometer, acoustic, and tachometer data, representing the run state of the rotating components. The alert system showed different levels of warning for predictive maintenance. The framework is designed with high flexibility and scalability. Therefore, this framework can be generalized to other in situsensor data from various assets.