Research On Improved Support Vector Machine Algorithm And Its Application In PHM Technology

In recent years, modern equipment and devices become more and more large-scale and complex, which have increasing functions and increasingly complex structures, their safety and reliability are drawing more attentions by all sides. Prognostics and health management(PHM) technology provides a reliable guarantee for equipment systems. As a small sample model of machine learning methods, the support vector machine (SVM) is very practical for the limited observational data system. This paper studies the SVM algorithm used in limited samples of PHM. It improves the way of the selection of the SVM's kernel parameters and the penalty factor. And it improves the poor generalization ability of least squares support vector machines (LSSVM).Firstly,this article substitutes the variable penalty factor (VPF) into support vector classification machine (SVCM), in order to reflects importance of each point. Under the situation that is approximately linearly separable and the situation that is linear inseparable,the SVCM dual problem is deduced based on VPF, namely VPF-SVCM. To solve the VPF-SVCM problem, the formula based on the least squares method is deduced. The genetic algorithms (GA) is then utilized to search the optimal kernel parameters of the VPF-LSSVCM to construct the GAVPF-LSSVCM algorithm. What is more, binary tree(BT) structure is used to improve the GAVPF-LSSVCM algorithm to apply to fault diagnosis of the liquid rocket engine. Experiment results show that BTGAVPF-LSSVCM performs better than SVCM and LSSVCM in terms of the training time and the correct rate.This paper proposes IEC-LSSVR algorithm based on iterated error compensation (IEC),which is an improved version of LSSVR. Then through the regression examples the effectiveness and superiority of IEC-LSSVR are demonstrated. For prognostics experiment,this paper applies IEC-LSSVR to forecast the liquid rocket engine thrust and the failure rates of Boeing's airliner. Experiment results show that although the training time of the IEC-LSSVR fault prediction method is longer than that of LSSVR, its generalization ability is better than LSSVR.