Abstract: Due to the new round of industrial revolution,digital technology has further promoted the combination of quality management tools with mathematical modeling and simulation techniques. Quality prediction with data-driven methods has become an important trend in the modern manufacturing industry. Among them, building a quality model with high accuracy and good robustness plays a vital role in implementing quality prediction. As one of the popular quality models, the Gaussian process model is widely used due to its superior performance in practical applications. However, due to the uncertainty in selecting the type of kernel functions, the hyper-parameter estimation in the likelihood function may be not obtained precisely, which leads to unreliable quality prediction results. To deal with this issue, a model construction strategy based on a selective ensemble kernel learning algorithm is proposed in the framework of Gaussian process models. First, the Bootstrap method is used to repeatedly extract the training samples, and each training sample is used to obtain the approximate values of hyper-parameters under different kernel functions. Then, a multi-dimensional Gaussian process model can be constructed based on the above information. Second, the predictive performance of the Gaussian process model under different kernel scenarios is analyzed, and then the kernel functions can be determined by the quality tool Pareto plot. Meanwhile, the ensemble parameters are incorporated into the construction of an improved likelihood function. The hyper-parameters of the ensemble kernel Gaussian process model can be obtained by maximizing the improved likelihood function. Finally, the effectiveness of the proposed method is verified by simulation tests and case studies. The comparison results show that the Gaussian process model based on the selective ensemble kernel not only provides a feasible method for the determination of kernel functions, but also improves the accuracy and precision of the quality prediction. Then the proposed model can be used for the following quality optimization or Bayesian optimization.

Key words: quality prediction; Gaussian process model; ensemble kernel learning; parameter optimization