Abstract

AbstractGiven that physics-based models can be difficult to derive, data-driven models have been widely used for remaining useful life (RUL) prediction, which is a key element for predictive maintenance. In industrial applications, although the models have to be trained in a short time with limited computational resources, recent research using back propagation neural networks (BPNNs) has focused only on minimizing the RUL prediction error, without considering the time needed for training. Driven by this motivation, here we consider a simple and fast neural network, named extreme learning machine (ELM), and we optimize it for the specific case of RUL prediction. In particular, we propose to apply both single-objective and multi-objective optimization to search for the best ELM architectures in terms of a trade-off between RUL prediction error and training time, the latter being determined by the number of trainable parameters. We perform a comparative analysis on a recent benchmark dataset, the N-CMAPSS, in which we compare the proposed methods with other algorithms based on BPNNs. The results show that while the optimized ELMs perform slightly worse than the BPNNs in terms of RUL prediction error, they require a significantly shorter (up to 2 orders of magnitude) training time.KeywordsEvolutionary algorithmMulti-objective optimizationExtreme learning machineRemaining useful lifeN-CMAPSS

Full Text
Published version (Free)

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call