Multivariate Process Monitoring And Fault Identification Method Based On Decision Tree Learning Techniques

Modern manufacturing process often involves several variables that are highly correlated. Process monitoring for a manufacturing or service process with correlated variables is referred to as multivariate quality control(MQC) or multivariate statistical process control(MSPC). Hotelling's T2 control chart is the first proposed MSPC approach. It was tested that the T2 control chart is insensitive to minor shifts in processes. Thus, Exponentially Weighted Moving Average (EWMA) and Cumulative Sum (CUSUM) control charts were extended to multivariate process monitoring scenarios. However, they cannot provide information on which variable or subset of variables causes the out-of-control signals.Decomposition methods decompose the MSPC statistics into different components and identify the most influential component responsible for the out-of-control signals. Major decomposition techniques include the Principal Component Analysis (PCA), the eigenspace comparison method, the MTY method, the step-down method, and the multiway kernel PCA method. However, the implementation of these techniques often involves advanced and complex statistical procedures. As the use of computer integrated manufacturing (CIM) and automation of SPC implementation becomes more and more popular in manufacturing processes, machine learning methods have been widely used as an automated process monitoring technique. Both Artificial Neural Network (ANN) and Decision Tree (DT) algorithms have been proposed in MSPC. In order to make up for the shortage of traditional multivariate control chart, a multivariate process monitoring and fault identification model based on decision tree algorithm is presented in this paper. Its efficiency, accuracy and feasibility is testified by a series of Monte Carlo simulations. To optimize the performance of the decision tree, some measures are taken: proper preprocessing and composition optimization for the input data decrease the input dimensions; Using different decision trees for process monitoring and fault identification lessen the number of output categories; a sampling mode based on Mahalanobis distance contours is proposed to reduce the requirement for training samples. Numerical experimentations for bivariate and ternary process are made then. The results show that this model can achieve good performance with a small amount of training samples. Also, it is clear that sample size has close relationship with the performance of this model, but correlation coefficients can only influence the fault identification part rather than the process monitoring part.