基于GRU神经网络的自相关过程残差控制图

GRU Neural Network-based Residual Control Chart for Autocorrelated Processes

为提升自相关过程监控的效率,提出基于门控循环单元(gated recurrent unit,GRU)神经网络的自相关过程残差控制图。采用受控下的自相关过程数据对GRU网络进行离线训练与测试,对预测误差进行监控,形成控制用残差控制图。采用训练好的GRU网络预测当前过程波动,利用控制用残差控制图判定当前过程是否失控。运用蒙特卡洛仿真法,与基于一阶自回归模型、BP神经网络以及支持向量回归构建的残差控制图进行性能对比。研究表明,过程受控时,所提残差控制图与其他3种的稳态平均运行链长相差不大,即4者的性能表现相当;而在均值偏移异常过程中,所提残差控制图的平均运行链长远小于其他3种,对自相关过程均值偏移具有较好的监控性能。

In order to further improve the efficiency of autocorrelation process monitoring,the residual control chart for autocorrelation process using gated recurrent unit(GRU)neural network is proposed.The GRU network is off-line trained and tested with the autocorrelation process data in control to monitor the prediction error and form the residual control chart for control.The trained GRU network is used to predict the current process variation and the residual control chart is used to determine whether the current process is out of control.Monte Carlo simulation method is used to compare the performance with the residual control chart based on first-order autoregressive model,BP neural network and support vector regression.The experiment results indicate that the difference of ARL between the proposed residual control chart and the other three kinds of control charts is small,that is,the performance of the four kinds of control charts is equivalent when the process is in control;while in the process of abnormal mean shift,the ARL of the proposed residual control chart is smaller than the other three kinds of control charts.The monitoring efficiency of the residual chart presented in this study has a remarkable improvement for mean shift of autocorrelation process.