关键词: 统计过程控制, 偏移幅度, 支持向量回归, 多核函数, 粒子群优化

Abstract:

In order to estimate the mean shift size in manufacturing process more accurately, a multi-kernel SVR was proposed. The multiple kernel function was convex combination of three basic single kernel functions, i.e. the linear kernel, the polynomial kernel and the radial basis kernel. For constructing the multi-kernel function, PSO algorithm was adopted to optimize the kernel parameters, the weight coefficient and the punishment coefficient C of SVR jointly. And the five-fold cross validation method was applied to compute the fitness value of each particle, which was the average of determination coefficient values of SVR on each of the five groups of samples. As a result, the final multi-kernel SVR obtains good generalization ability. This multi-kernel SVR was integrated with CUSUM chart to monitor and quantify mean shifts in process. The simulation results indicate that it improves a lot on estimating mean shift size compared with the ANN method, and also is superior to SVR with single radial basis kernel. An application examination on actual gear machining process further verified its validity and practicability.

Key words: statistic process control(SPC);shift size;support vector regression(SVR), multiple kernel function, particle swarm optimization(PSO)