Remaining useful life prediction based on a modified SVR method with health stage division, cluster sampling and similarity matching

Abstract

Remaining useful life (RUL) prediction is an advanced methodology of prognostics and health management (PHM), and is in advantage of life cycles management of equipment and maintenance cost reduction. Among the data driven methods, support vector regression (SVR) is one of the most suitable methods when there are limited failure history data for analysis. However, many uncertain factors such as individual variation and time varying operating conditions, will lead that the failure time of all equipment statistically dispersive, and with the increase of sample set, such dispersity may inevitably increase and further reflects in the model training. In consequent, the dispersity may cause two drawbacks. On the one hand, the linearity of SVR model will increase with the increase of sample set, and overfitting or underfitting tends to occur. And on the other hand, a single model only performs well in its generalization and robustness, but may lost its effectiveness that it may fail to work well for a new on-service equipment. In order to deal with the two drawbacks, this paper proposes a modified SVR method with health stage division, cluster sampling and similarity matching. Through health stage division and cluster sampling, each state of the whole degradation process can obtain the optimal parameters, then the irrelevant linearity can be reduced. In addition, since that similar input results similar output, the optimal parameters of the most similar testing sample are also suitable for the on-service equipment, and through similarity matching the most similar testing sample can be obtained, thus the drawback of a single model can be avoided. Finally, the effectiveness of the proposed method is verified systematically by a simulated dataset of fatigue crack growth and a real-world degradation dataset of GaAs-based lasers.