A Proposal of Metric for Improving Remaining Useful Life Prediction in Industrial Systems

Abstract

The deterioration of engineering systems due to wear and working conditions impact directly on their performance, requiring more efficient maintenance programs to prevent unexpected stops and increase production quality. Neural networks have shown significant results in predicting the remaining useful life (RUL) of systems. A neural network for prognostic is generally trained to minimize the mean square error (MSE) between the RUL prediction and its true value. This metric gives equal importance to the error at the beginning and at the end of a system’s useful life. However, the prediction of the RUL is more critical as a system approaches the end of its useful life. Therefore, making an accurate evaluation of prognostic models requires to take this into account. In this paper, a new performance metric for the evaluation of prognostic models is proposed with the objective of establishing a direct relation between RUL prediction and maintenance planning. In addition, a procedure to use this metric for training a multilayer perceptron (MLP) network is proposed to improve the network’s capacity to learn degradation patterns near the end of the useful life. The procedure is applied to NASA Commercial Modular Aero-Propulsion System Simulation (C-MAPSS) dataset, improving the prediction results significantly.