Research And Development Of Quality Control Chart Pattern Recognition Based On LSTM

Statistical process control(SPC)is an important part of total quality management and a key means for enterprises to implement product or control service quality.Its application effect determines the company's core competitiveness in the product or service market.As the core tool of SPC,traditional control charts are widely used because they can detect abnormalities in the production process.The control chart pattern recognition technology combined with machine learning methods further strengthens and expands the abnormality detection and recognition capabilities of control charts.It has opened a new stage in the development of control charts and has become a research hotspot in the field of quality control in recent years.In this thesis,a deep learning method—Long-Short Time Memory Network(LSTM)was used for control chart pattern recognition.The recognition model was constructed,and verified by simulation experiments.An application prototype system was designed and developed.The specific work is as follows:The research and application of traditional machine learning methods in the field of control chart pattern recognition was summarized in this thesis,as well as its shortcomings and deficiencies.The advantages of deep learning compared to traditional machine learning was analyzed,meanwhile,how to combine deep learning methods with control chart pattern recognition was studied,and on the basis of in-depth study of the LSTM classification principle,proposed a control chart pattern recognition method based on LSTM,and through Monte Carlo simulation method created a simulation data set for simulation experiment verification.Experiments show that,compared with traditional machine learning methods such as SVM and multi-layer perceptron,LSTM shows superiority in the offline recognition efficiency or accuracy of the eight basic patterns of control charts.A three-stage model for online recognition of control chart patterns based on LSTM was studied.First,the eight basic control chart modes were classified into fouraccording to their geometric characteristics,which were defined as controlled mode,up mode,down mode and complex mode.The LSTM classifiers for these four modes were trained and the threshold adjustment method was used to balance the first and second errors,then calculate the model's in control average running chain length(ARL0)and out of control average running chain length(ARL),and charge the first stage of abnormal monitoring,and the Bi-LSTM coarse classifier that uses the delay method to train the abnormal patterns was responsible for the coarse classification of the last three modes in the second stage.Finally,the LSTM fine classifiers were trained for the latter three coarse classification abnormal patterns,and the third Mode breakdown of the stage.For this model,the normal and abnormal quality characteristic data streams were simulated by Monte Carlo simulation method to conduct online recognition experiments.The test results show that the average online recognition accuracy of abnormal patterns reaches 0.948.Based on the constructed control chart pattern recognition model,combined with the actual needs of production process quality monitoring,the application prototype system was analyzed,designed,and developed in accordance with the software engineering process.The system adopts a six-layer hierarchical architecture design and uses related technologies developed by the popular Java Web and Python Web to achieve start-stop control of production lines,quality data collection,online identification of control chart patterns,storage of difficult identification samples,monitoring model management,etc.The core function proves that it has certain applicability through simulation data operation.