Toward Resilient Predictive Maintenance in Complex Industrial Processes

Industrial technological advancements and the surge of data they generate, which typically presents noise, drifts, and missingness, can bear tangled problems in predictive maintenance (PdM) practices by adversely impacting the Remaining Useful Life (RUL) modeling and its subsequent forecasting. Utilizing the N-CMAPSS datasets, this work focuses on leveraging the impact of irregular data into the RUL modeling process, aiming to continuously discern noise and anomalous drifts from the system’s expected behavior, enhancing the estimation of remaining life through an innovative diagnostic-based strategy. The results, which have been obtained by benchmarking a vast set of machine learning algorithms and then comparing the proposed dynamic approach with a reliability-centred linear degradation model, present a general improvement in RUL’s prognostic metrics. In detail, utilizing the best model from the benchmark (BiLSTM) and the innovative approach detailed in this study, a better model alignment is outlined (3% increase of R2) with conservative predictions near failures (17.5% reduction of S-score), paving a promising path toward Resilient Predictive Maintenance (ResPdM).