Prognosis & Health Management System (PHMS): a machine learning framework to support decision-making in predictive maintenance in a production system

Abstract

The search for the effective use of production assets has been constant, mainly in industries with evolving mechanization. In this way, maintenance management gains visibility as it ensures asset availability. Predictive maintenance (PDM) is one of the main maintenance management strategies. Allows early detection of failures, avoiding unscheduled downtime and unnecessary costs. As technologies have advanced, predictive maintenance improves Prognosis and Health Management (PHM). It provides the means to recognize patterns, understand anomalies and estimate the equipment’s remaining useful life (RUL). At the same time, technologies such as the internet of things (IoT), machine learning (ML), and cloud computing enable the digitization of assets, providing intelligent manufacturing. However, this scenario makes PDM a complex and expensive task when applied to systems with equipment connected in series. On the one hand, data is abundantly generated, collected, and stored. On the other hand, it is difficult to convert data into useful information to support PDM and PHM. Given the gaps related to PDM and reliability, we suggest the Prognosis and Health Management System (PHMS) in this thesis, which is supported by an analytical framework that uses a set of techniques and ML. First, we performed a case study to evaluate the proposition with real data from the process industry. In developing the framework, we used semi-supervised ML with Autoencoder (AE) to build the operational threshold and identify anomalies. For the Feature Identification step, we applied XGBoost and the SHAP method. Next, we test different deep learning architectures to predict the RUL of the system. In the RUL prediction, we present different deep learning architectures. In this sense, we highlight the N-BEATS deep learning architecture as an essential alternative to traditional architectures such as Recurrent Neural Networks (RNN). Through the framework applied to the case study, it was possible to identify an anomaly and the behavior of the most relevant variables for the failure and predict the RUL of the equipment with R2 greater than 90% with N-BEATS. In this way, according to the results presented, the operation and maintenance teams can carry out preventive actions, avoiding unscheduled stops of the production system. In this sense, the development of the framework contributes to the adoption of emerging technologies in real processes. In addition to the benefits presented, we highlight the development of PDM studies on real data unknown in the academic environment. We draw attention to this point, as most reliability studies are based on widely known and treated data.