Quality Diagnosis Of Product Manufacturing Process Based On Intelligence Methods

The monitoring of the quality state of the manufacturing process is the start point of continuous quality improvement, and quality diagnosis can provide direction for continuous quality improvement of manufacturing process. The abnormalities of process can be found and corrected by diagnosing the abnormal state of the manufacturing process, and then the process can restore and maintain the controlled state. With the advance of the modernization and complexity of manufacturing process, higher demands are brought out for process control and quality diagnosis of manufacturing process. It is very difficult to fulfill these demands only using the traditional technologies of diagnosing the process quality. It is the important development direction of process quality diagnosis to introduce and use different latest technologies of other fields of technology, these technologies usually include computer technology, artificial intelligence technology, and so on. This dissertation primarily researches and investigates these problems. The main contents of this papae are as follows:Firstly, this dissertation studies the problem of control chart pattern recognition based on probabilistic neural network. Traditional uni-variable control charts are important tools to diagnose abnormalities of process, but they are helpless to diagnose the abnormal control chart patterns. To solve this problem, this dissertation proposes to recognize the control chart patterns using probabilistic neural network (PNN) technology. The structure and the parameters of probabilistic neural network are investigated and designed, and the performance of this method is evaluated using simulation experiment. The proposed PNN method can solve the problem of difficulties of designing neural network structure, and it can also solve the problem of lower accuracy rate of pattern recognition when other kinds of artificial neural network are used to recognize the control chart patterns.Secondly, this dissertation studies the problem of control chart pattern recognition based on least squares support vector machine. Traditional SPC methods of diagnosing process abnormalities are valid only in large sample conditions, and the neural network methods of diagnosing process abnormalities need a large number of training samples, these two methods can not work well when samples are limited. This dissertation proposes to recognize the control chart patterns using least squares support vector machine (LS-SVM), and the performance of the proposed method is evaluated. Meanwhile, to improve the performance of the proposed method, this dissertation proposes to optimize the parameters of the least squares support vector machine using particle swarm optimization algorithm or genetic algorithm. By the proposed methods, the control chart patterns can be recognized efficiently under the condition of limited sample size.Thirdly, this dissertation studies the problem of detecting the anticipated abnormal signal of manufacturing process using Cuscore statistics. According to the accumulated prior knowledge of process, some abnormal signals of process are expectable. Cuscore statistics can detect anticipated abnormal signals of manufacturing process very well, and can effectively use the accumulated priori knowledge of process abnormality. This dissertation investigates and evaluates the performance of Cuscore statistics technique of detecting the nonlinear two order anticipated signal. In order to solve the mismatch problem of standard Cuscore statistics technique and improve the detection performance of Cuscore statistics technique, this dissertation proposes to detect the change-point using moving window and LS-SVM pattern recognition methods.Finally, this dissertation studies the problem of quality diagnosis of multivariate process and the identification of abnormal variables. Most of the actual manufacturing processes are multivariate processes, and the variables are auto-correlated. The present multivariate statistical process control technologies can only diagnose the status of the multivariate process as a whole, and can not locate and identify the abnormal variables (or the subset of them). This dissertation combines traditional multivariate statistical process control method with artificial intelligence, and build two intelligent diagnosis models for mean vector and covariance matrix of multivariate process respectively. The two diagnosis models can format the problem of identifying the abnormal variables into the problem of pattern recognition of mean vector and covariance matrix respectively. LS-SVM pattern recognizers are designed and the performances of the two models are evaluated. The proposed two models can not only diagnose the status of the multivariate as a whole, but also can effectively identify the abnormal variables. Cases of bivariate process are given to illustrate the application of the two proposed models.According to the different characters of the processes and the different demands of process quality diagnosis, this dissertation studies the previous process diagnosis techniques while introduce artificial intelligence technology to diagnose the process quality. The proposed methods can solve the problems of the previous process diagnose technologies and strengthen the ability of process diagnosis. This dissertation provides some new research ideas and realization technology for process diagnosis.