Robust energy centric machine learning assisted fault detection & diagnosis: an extrusion case

Abstract

Condition monitoring for fault detection and diagnosis in industrial systems typically relies on vibration, temperature, or pressure signals, with machine learning playing an increasingly crucial role. While traditionally reserved for environmental objectives, energy monitoring is now gaining significant attention for predictive maintenance. Yet, its potential remains under-explored, with past decade works focusing mainly on fault detection. In this study, we propose an improved energy-centric condition monitoring framework for fault detection and isolation. Since energy is a single aggregated system-wide variable, we emphasise the importance of feature engineering to extract discriminative fault signatures. Our approach was evaluated on a plastic extrusion system under seven fault conditions and only system-wide energy measurements and timestamps were used for fault detection, isolation and robustness analysis. We use deep unsupervised models, specifically autoencoders and variational autoencoders, trained on engineered features to detect deviations from normal behaviour and isolate fault types. Our models achieved up to 98% F1-scores in fault detection and enabled the isolation of all seven fault modes (compared to four with classical models). Moreover, the models showed robust performance under measurement noise arising from adjacent machinery or unmonitored operational variables, maintaining high detection accuracy even with 20% additive noise, where conventional models dropped below 60%. Robustness to sensor drift was also evaluated. Isolation scores remained high when detection was achieved with our models. These findings demonstrate that energy-based monitoring, when properly engineered, can be a cost-effective, robust alternative for industrial condition monitoring.