Reliability-driven constrained Pareto optimization for knowledge-informed predictive maintenance decisions under uncertainty

Purpose – A novel hybrid machine learning approach based on a multi-objective zero-shot optimization algorithm considering multiple uncertainties is proposed in this study. To identify damaged regions, multi-domain processing functions for multi-directional vibration signals are adapted to form a productive transfer learning method. This technique also selects domain-invariant features for diagnosing dynamic fault severities with various uncertainties. Stochastic gradient descent (SGD)-based neural networks (NNs) minimize the loss functions based on the detection error and knowledge-oriented constrained functions with virtual, semi-physics and physics-guided learning approaches to improve identification speed and enhance feature engineering on source and target domains. Design/methodology/approach – The purpose of this paper is to develop an industrial knowledge-oriented predictive maintenance decision-making approach by utilizing a real-world dataset under multiple uncertainties, considering variable operational-environmental conditions and imbalanced failure propagation based on the constrained Pareto optimization for the reliability assessment and online health monitoring of the real-plant rotating machinery via the hybrid machine learning method by applying multi-domain signal processing functions through a fused feature bank by exploiting an SGD-based NN that is adaptable with various defects and robust to the variation of damage severities. Findings – As two major comparative achievements of two types of uncertainty including single uncertainty (time-varying operational conditions) and multiple uncertainty (time-varying operational conditions and imbalanced failure propagation), it can be concluded that the second layout learns a more adaptable feature space with real industrial plants than the configuration of the first framework due to a more domain-independent Pareto optimization-oriented learning approach. Another finding is that the possibility of failure risks of the technique regarding the 1st layout is much greater than that of the second framework due to more complex uncertainty. In conclusion, depending on the machine knowledge and uncertainty quantification, both strategies efficiently establish an intelligent predictive maintenance decision-making system. Practical implications – By exploiting the smart decision support metric as the zero-shot reliability metric, it is proved that the proposed method performs efficiently under real-world processes to predict different phases of rotating machinery damages with unseen data. Originality/value – This paper proposes a multi-objective decision-making model capable of overcoming the two main challenges related to condition-based maintenance: real-time health indicators concerning industrial machines are restricted under time-varying operational conditions with a few-sample strategy and imbalanced failure propagation with a zero-sample strategy.