Improved time-based maintenance in aeronautics with regressive support vector machines

Abstract

In modern preventive maintenance, time-based management is still the mainstream approach. This strategy continues to be the preferred choice to manage the risk of equipment failure when other alternatives, such as condition-based management, are technically or economically unfeasible. In this paper we propose a novel approach to time-based maintenance based on (linear) regressive Support Vector Machines (SVM). In the proposed modeling approach, expected lifetime is estimated based on the equipment past failure times combined with the maintenance history of similar components. Time series analysis combined with outlier detection techniques and concepts from technical analysis, such as resistance and support levels, are used to establish the SVM model prediction bounds. The proposed SVM model is compared with the traditional approach to time-based maintenance - life usage modeling - and the autoregressive moving average (ARMA) forecasting method. Results are shown on an industrial case study of data describing the maintenance life-cycle of a critical component of the aircraft bleed air system. Results suggest that the SVM model can outperform the other tested approaches both in regards to the squared, percentage and absolute errors. Also, the proposed model showed homoscedasticity of residuals, small over-prediction error as well as residual variance and bias across distinct aircraft.