A feature identification method to explain anomalies in condition monitoring

Abstract Through the Internet of Things (IoT), the generation of data, Cyber-Physical Systems (CPS) has shown a steady increase. The search for approaches in order to take advantage of generated data is a recurring theme on several managers’ agenda. To this end, data mining techniques, combined with asset health management, contribute to Industry 4.0 relevance in production systems. In this context, the continuous process industry has a high maturity due to high-level instrumentation and control. This type of manufacturing has usually divided sensors into process control and equipment monitoring. Traditional reliability studies on development of models use sensor data to monitor the health of the equipment without counting on valuable process information. In contrast, this study proposes an approach that seeks to increase the reliability of a productive system, using data from operational control and health monitoring of equipment, which provides a more robust model of reliability. To this end, we used a semi-supervised Machine Learning (ML) with convolutional neural networks (CNN), Autoencoder (AE), and a bagged decision tree to identify which variables are responsible for an abnormal condition. With this approach, it is possible to detect which variables are most important in the occurrence of failures, enabling preventive actions that increase the reliability of the system. The main contribution of this approach is the integration of analytical techniques, in order to improve the reliability of the system. To test and validate the method, we performed a case study with real data in a styrene petrochemical plant. As a result, it was possible to identify the beginning of an anomaly that failed the vacuum system. With the proposed approach, mitigation actions could be taken and, consequently, avoid unnecessary downtime.