Lifelong Learning for Fault Prognostics in Predictive Maintenance with Bayesian Neural Networks

Fault prognostics is one of the key enablers for the realisation of predictive maintenance. In today's era of digital transformation, deep learning (DL) has proven to be a promising data-driven solution for the task of fault prognostics with the ability to accurately predict the remaining useful life of industrial assets based on their historical condition-monitoring data. However, the deployment of DL-based fault prognostics models in practice still faces a number of critical challenges, especially in application scenarios with dynamic or evolving contexts suffering from data distribution shifts. Jointly training DL models using data from all contexts at once is typically impossible due to practical requirements regarding privacy constraints and resource limitations. Moreover, fine-tuning or training DL models in a classical sequential manner has been observed to typically suffer from catastrophic forgetting where adapting to a new context leads to drastically forgetting what has been learned previously. To address this problem, we proposed a novel lifelong learning framework with Bayesian neural networks for fault prognostics. Our proposed Bayesian lifelong learning method focuses not only on preserving the old knowledge learned so far but also allows to reserve the model flexibility for learning new knowledge from upcoming contexts. Experimental results on the benchmark C-MAPSS dataset of turbofan engine degradation data show the superiority of our proposed framework over other relevant lifelong learning methods. On average, we achieve a performance improvement of 10.2% and 75.3% in terms of final prediction accuracy and forgetting measure, respectively.