Conference: Defining Degradation States for Diagnosis Classification Models in Real Systems based on Monitoring Data
Published in 31st European Safety and Reliability Conference (ESREL 2021), 2021
Recommended citation: Sergio Cofre-Martel, Camila Correa-Jullian, Enrique López Droguett, Katrina M. Groth, and Mohammad Modarres Modarres. Defining Degradation States for Diagnosis Classification Models in Real Systems Based on Monitoring Data. In Proceedings of the 31st European Safety and Reliability Conference (ESREL 2021), pages 1286–1293, Singapore. Research Publishing Services. doi:10.3850/978-981-18-2016-8_303-cd
Abstract
As the complexity of modern engineering systems increases, data-driven approaches have become valuable tools to aid maintenance decision-making. However, raw data collected from monitoring sensors require a comprehensive and systematic preprocessing to separate healthy from faulty states before their use in data-driven models. Frequently, anomaly detection models implemented for this purpose are based on statistical relationships or rule-based thresholds, rather than on information provided from maintenance logs related to the internal operation of the system. In this work, we propose a framework to establish a link between the recorded sensor data behavior and the system’s degradation processes. In particular, this framework aims to obtain a labeled degradation dataset from raw monitoring data to train a diagnosis classifier for a system with multiple failure modes. A dataset obtained from two years of sensor monitoring and reported failure logs of a copper mining process line is used to exemplify the framework. Different machine learning classifiers are presented for each failure mode, individually and combined. Results show that the degradation labelling procedure is effective, and classifiers obtain up to 95% accuracy for the detection task for a two-class problem. Cross-comparison of the classifiers per failure mode allows the identification of problematic classes, showing the benefits of addressing each failure mode individually rather than for the entire system simultaneously.
Keywords: Diagnostics, Classification, Machine learning, Machinery data processing, Degradation detection.
Recommended citation: Sergio Cofre-Martel, Camila Correa-Jullian, Enrique López Droguett, Katrina M. Groth, and Mohammad Modarres Modarres. Defining Degradation States for Diagnosis Classification Models in Real Systems Based on Monitoring Data. In Proceedings of the 31st European Safety and Reliability Conference (ESREL 2021), pages 1286–1293, Singapore. Research Publishing Services. doi:10.3850/978-981-18-2016-8_303-cd