Online robust parameter design using sequential support vector regression based Bayesian optimization

Abstract

Traditional robust parameter design (RPD) is an effective offline quality improvement method, but it cannot update the optimal settings in a timely manner. To resolve the drawback of traditional one-time modeling methods, online RPD comes into being. In online RPD, when the setting for controllable factor is unreasonable, it can be regarded as a new sample to update the optimal setting, which improves the efficiency of constructing response surface. Now, one of the mainstream approaches is to use support vector regression (SVR) as a response surface for online RPD. However, in the process of adding samples and updating the response surface, the parameters of the model are not updated. If the parameters do not change after adding samples, the accuracy of the model may decrease, which is obviously not desirable. To address this problem, this paper proposes Bayesian optimization sequential SVR based on online RPD (BOSSVR-RPD) strategy, which can simultaneously update the parameters and response surface in real time. In the proposed strategy, we can use the optimal setting of previously controllable factors, the currently observed noise factor and the corresponding response as new samples to add to the response surface to improve the accuracy of the response surface, and at the same time use the Bayesian optimization method to optimize the parameters of the model. By repeating this process, we can find the desirable optimal setting for controllable factor. This paper verifies the method through four cases, proving that the proposed BOSSVR-RPD strategy can find the optimal setting for controllable factors. Compared with the existing method without parameter optimization, the model proposed in this paper is more accurate, so the optimal setting for controllable factors can be found faster and the controllable factors obtained are more reliable.