Abnormal Process Quality Recognition In Bivariate Autocorrelated Process Based On Random Forest

The traditional multivariate control charts are available to distinguish whether the process state is in-control or not.However,one of limitations of the traditional control charts is that could not provide details directly to identify the specific abnormal variate.Moreover,to some extent,it leads to autocorrelated existing in samples due to the automated sampling in the production process.This goes against one of basic assumptions for traditional multivariate control charts in which have focused on data that follow different points of time are independent and identically distributed(I.I.D).The great classification performance of multivariate control charts would collapse with the autocorrelated effects soon,such outcomes reflect immediately upon soar of probabilities in false alarms and missed alarms.Therefore,some scholars attempt to introduce machine learning stem from Artificial Intelligence into process quality control of multivariated autocorrelated.Machine learning has been applied in detecting and recognizing the abnormal patterns.At present,this approach is research hotspot.Compared with statistical methods,machine learning based methods escape from I.I.D assumption and do not need human involvement in practice.Therefore,it is an effective way to achieve the automation and intelligence in the process quality control.On the basis of previous studies,randome forest has many virtues,such as much fewer parameters to be determined,fast training speed,high fault-tolerance,hard to overfit etc.In this thesis,we introduced the random forest and combines the requirements of sampling and recognition of abnormal patterns in bivariate autocorrelated process to conduct the study as follows.1)The vector autoregression model commonly used for descirbing multivariate autocorrelated process.The AR(1)model is given to model a bivariate autocorrelated process.According to the characteristics of up-shift pattern,the process is divided into four patterns.The Monte-Carlo simulation method is applied as an effective way to generate data sets of four patterns.2)On the basis of ensemble learning and the classification theory of random forest,a new model with optimized random forest and multi-feature extractin is suggested to recognize the abnormal patterns in bivariate autocorrelated process.Based on the appropriate span of two key parameters of random forest,the particle swarm optimization is introduced to optimize the two key parameters with one objective,which is the average recognition accuracy of four patterns.Then a PSO-RF model is contruced.Experimental results shows that the PSO-RF classifier performs well on classification due to multi-feature extraction and fusion,and the recognition accuracy rises from 91.25% to 98.33%.Moreover,the PSO-RF model is better than BPNN model(92.3%)and SVM model(97.16%)in this interest.3)Based on the PSO-RF model,according to the software engineering ideas and process,a prototype system is suggested to collect the datas and recognize the abnormal patterns in multivariate autocorrelated process.The synthetic data is applied to validate the application effectiveness of the prototype system.Based on the above findings,the proposed model(PSO-RF model)is an effective and applicable approach to recognize the abnormal patterns in bivaraitae autocorrelated process.