Stochastic fractal search-optimized multi-support vector regression for remaining useful life prediction of bearings

The remaining useful life (RUL) prediction of rolling bearings is of great significance in engineering industries. Support vector regression (SVR) is a widely used machine learning algorithm for RUL prediction which shows effectiveness in small sample cases. However, the prediction accuracy of SVR is largely dependent on the initial parameters, and the overfitting problem reduces the accuracy of the prediction results. In this paper, a multi-SVR method for bearing RUL prediction based on stochastic fractal search (SFS) is proposed. The time-domain features are extracted to describe the degeneration process of bearings. The Butterworth filter is applied for de-noising and the principal component analysis is introduced for dimensional reduction. To improve the effectiveness of SVR, the SFS algorithms are used to achieve the appropriate SVR parameters. With optimized parameters, multiple SVR models are trained by bearing datasets and the weight of each model is determined with crossover performance tests. The proposed method is validated on IMS experimental bearing datasets and the performance is compared with three novel RUL prediction methods. The results show that the predicted RUL trend of proposed method is in good agreement with actual value and the accuracy and convergence are satisfactory compared with other methods.

Automated summary

In this paper, a multi-SVR method based on SFS for bearing RUL prediction is proposed, which achieves satisfactory accuracy and convergence by extracting time-domain features and training multiple models with optimized parameters.