Fault Diagnosis And Quality Monitoring Method Based On Partial Least Squares

The scale of product and process flow has sharply risen with the development of modern industry,and the quality and safety of production attract increasing attention.In fact,some process measure variants contain the underlying information which related to the final production' quality in industry process,thus,the relationship between process variant and quality variant contributes to process modeling and quality monitoring.The PLS-based faults diagnose and process monitoring focus on the information contained in process data,to achieve fault's detection,separation,evaluation and decision-making for a complex process,which provides a guideline for operation and reduces the cost of error brings.In this paper,some types of quality monitoring and prediction methods have been proposed for the typical chemistry and semiconductor manufacturing process.To begin with,considering to the shortages of contribution plot method,such as the poor indication of deviation degree when the normal state has shifted to the fault,and the fault identify result often affected by the variant that has a large mean value but tiny variance.Then,a novel CPLS-base contribution plot method is carried out.In this method,it locates the source of fault by the relative contribution value,which improves the interpretation and accuracy for fault identification.Secondly,a new prediction method for multi-stage industrial processes is studied.The quality prediction model is established from the perspective of local mixture and date clustering,which overcomes the shortcomings of the traditional global and single model,such as low prediction accuracy and inadequate local prediction ability.Meanwhile,the neural network is introduced in prediction model to further generalize the model's ability to handle nonlinear and highly correlated data.Furthermore,a state transition monitoring method based on FDA_kernel is proposed,which employs the Bayesian principle to integrate quality information and data distribution characteristics in the phases to identify online state switching.Thirdly,in terms of the time-vary and dynamic characters in industry process,the research about prediction model's update solution are conducted,which aims to strengthen the model's adaptive predictive capability.During the modeling process,the training data is easy to be saturated,and the credibility of new collected data and historical data have great influence on quality monitoring and prediction,thus,a novel adaptive quality prediction method based on Block-RPLS model is proposed.In this method,the local outlier factor is carried out to detect the outlier data in the updated data block,then,and the data block will be divided into stable data set and the deviated data set with the help of the outlier detection limit designed.Furthermore,the RPLS model's parameters will be updated by learning time-varying information iteratively and integrating the feature matrix which has modified by matrix similarity theory.Finally,the proposed methods are applied to the typical processes mentioned above,the fault diagnosis and quality prediction model are experimentally verified,and some relevant conclusions and future prospects are given.