A machine-learning based data-oriented pipeline for Prognosis and Health Management Systems

The search for effective asset utilization has been constant, especially in industries with evolving mechanization. In this context, maintenance management gains visibility because it is responsible for ensuring the availability of assets. Predictive maintenance (PDM) is one of the main maintenance management strategies. It allows early detection of failures, avoiding unscheduled downtime and unnecessary costs. As technologies have advanced, PDM has evolved into Prognosis and Health Management (PHM), which provides the means to recognize patterns, understand anomalies, and estimate equipment’s Remaining Useful Life (RUL). In parallel, technologies such as the Internet of Things (IoT), Machine Learning (ML), and cloud computing enable the digitalization of assets, creating smart manufacturing. However, this scenario makes PDM a complex and costly task when applied to systems with interconnected equipment. On the one hand, data is abundantly generated and collected. On the other hand, there is difficulty in converting the data into useful information to support PDM and PHM. In this regard, we propose an analytical pipeline using ML with raw data from equipment and operation. As a result, we suggest the Prognosis and Health Management System (PHMS). Therefore, we used semi-supervised ML with Autoencoder (AE), XGBoost, and the SHAP method. Furthermore, we tested different Deep Learning (DL) architectures for RUL prediction. In order to evaluate the approach, we conducted a case study with real data from the process industry. Consequently, it was possible to identify an anomaly, the behavior of the Features most relevant to failure, and to predict the RUL with significant accuracy. Mitigation actions can be taken through the proposed approach. Thus, avoiding production system downtime and contributing to adopting emerging technologies in real processes.