Real-Time Predictive Maintenance System of Industrial Equipment without Historical Failure Data

Abstract

Predictive maintenance (PdM) appears to predict faults before they occur because unexpected industrial equipment failures directly affect workers' safety, cost, and work continuity. In this context, developing a PdM system needs historical data for failures, but those data are not available in our case, which is a sewage centrifugal pump, where failures had not been recorded before. Therefore, this research aims to develop a system for PdM that works efficiently in real-time and does not need historical data for failures; it can also predict failures at different periods and conditions. The "Data-Driven" method is a suitable methodology to apply to conditions data; also, time series forecasting and anomaly detection (AD) are the most applicable models for the studied case. The Main Bearing Temperature was chosen in the models because it is the most applicable parameter. The models' performance was evaluated in two ways: accuracy and resource consumption (execution time and RAM). After that, the most important accuracy metrics are a root mean square error (RMSE) for forecasting models and excess Mass and Mass Volume for AD models. The experimental results presented "TBATS" as the best forecast model and QuantileAD from the "ADTK" model as the best AD method.