Research On Control Chart Mixed Pattern Recognition For Variable Window Size

Nowadays,statistical process control methods are widely used in the production process of enterprises.On the one hand,the goal is to improve production capacity and reduce costs;on the other hand,the quality of the product can be guaranteed.Statistical process control uses control charts as an analytical tool to determine if the production process is under control.Control chart patterns can not only indicate whether the process is under control but also help diagnose abnormal causes when abnormal patterns occur.By applying machine learning and deep learning algorithms in pattern recognition,quality problems with the production process can be detected promptly,improving accuracy and efficiency of control charts.However,most studies on control chart pattern recognition assumed a fixed window size of considering real-world scenarios.This work aims to study blending of control chart patterns of varying window sizes.The research content includes:(1)This study addresses the issue of recognizing mixed patterns in control charts with varying window sizes.To generate simulation data,the Monte Carlo method is utilized to create 10 different window sizes that are then converted into images.By optimizing the network structure and parameters of Le Net-5,a convolutional neural network is trained to automatically extract features from control chart images and identify 13 mixed patterns.The results indicate that this approach outperforms support vector machine and one-dimensional convolutional neural networks using time series as input for control chart mixed pattern recognition with variable window size.The proposed method achieves a recognition rate of 98.41%.(2)To address the issue of recognizing mixed patterns in control charts with random changes in window length,we propose a model based on a convolutional neural network and whale algorithm to optimize support vector machines.First,the Monte Carlo method is applied to generate simulation data,which is then transformed into grayscale images.Since training a convolutional neural network requires ample samples and time,we fuse it with support vector machines.The former extracts image features while the latter classifies 13 mixed patterns of control graphs.Next,we compare and analyze how different kernel functions affect recognition rates to determine an optimal one before optimizing kernel parameters and penalty parameters using the whale optimization algorithm.Finally,our proposed model achieves higher recognition rates than four other tested methods with an average rate of 96.86% for all13 control chart mixed patterns.(3)Recognition of control chart patterns in real production processes using models trained on simulation data.Experiments show that the proposed method can effectively identify anomalous patterns in actual production and has certain application value.