Hybrid Deep Learning and Fractional Brownian Motion Approach for Probabilistic RUL Prediction in Industrial Equipment

Abstract

Accurate prediction of the remaining useful life (RUL) constitutes the core task of predictive maintenance, and the selection of an appropriate degradation model is pivotal to enhancing the accuracy of RUL prediction. However, deep learning models cannot characterize RUL uncertainty, and traditional stochastic process degradation models are challenging in depicting the long-range dependence (LRD) of the degradation process, adversely impacting the accuracy and credibility of RUL prediction. To address these challenges, this article unveils a synergistic temporal convolutional network Kolmogorov–Arnold network fractional Brownian motion (TCN-KAN-FBM) prediction framework using TCNs and KAN for robust device RUL prognostication. The TCN-KAN module is designed to realize RUL prediction. The TCN-KAN module captures temporal features of degraded data and adaptively learns degradation knowledge for point estimation prediction. Complementing this, the FBM module, then, masterfully constructs the probability distribution of the prediction results based on its LRD and self-similarity, thus realizing RUL prediction’s uncertainty quantification (UQ). The effectiveness of the proposed method is confirmed by practical examples of rolling bearings and two sets of servo tool holder power head systems under different operating conditions.