Study On Autocorrelation Process Control Based On Neural Network Approach

For the property of auto-correlation process, the observations can not meet the requirement for traditional statistical process control (SPC)– Observations are independent, fault alarm happens very often and sensitivity to detect abnomal is low. In this article, the neural network will be tried to control auto-correlation process. As a result, it performs very well.The model of auto-correlation process will be set up by time series model. In the previous research, step shift of process mean were studied a lot. But there is very little focusing on increasing shift. The model of increasing shift was set up in this article.There is some study of BP neural network on auto-correlated process control aboadly before. But most of the result is similar to that of residual charts. To improve the performance of BP neural network, the training data has to be optimized. There are some priciples to select training data in this article, through which the neural networks were trained very well.The trained neural network was applied to control auto-correlation process. The result of normal status, step shift and increasing shift were analyed. From the result, the BP neural network performs better than residual charts.To decrease the sample size, the number of neuron in input layer was studied. The proposal was raised to select a proper number of neuron in input layer by diffent shift altitude and correlation matrix.The number of neuron in hidden layer was studied to improve the identifying ratio of neural network. This number varies from different the number of neuron in input layer.A comparision study was conducted of Shewhart chart, residual chart and BP neural network on the control of dipping process and ceramic filter manufacturing. It's affirmed that the performance of BP neural network is better than the two others.