Research On Intelligent Detection Method For Abnormal Measured Data In Batch Processes

The batch processes are important manufacturing approaches in modern industry.Batch manufacture is changed from simple to complex,which is greatly promoted because of the increasing demand of high-quality product in the modern society.Field instrumentations or transmitters provide rich measured data of process in batch manufacture,which provides a foundation for online monitoring and optimization control based on data-driven.However,due to the impact of field instrumentation or transmitter performance degradation,electromagnetic interference and so on,abnormal measured data of batch process will occur,which reducing the quality of measured data and restricting the accuracy improvement of data-driven process modeling.Thus,it has important theoretical significance and application value to study intelligent detection methods for abnormal measured data in batch processes,and to provide reliable process data for data-driven process modeling,on-line monitoring and optimization control.The dissertation has studied intelligent detection methods for abnormal measured data in batch processes by fully analyzing the characteristics of batch processes and process measured data.The researches are mainly completed as follows:1.An unsupervised multi-scale sequential partition b.ased phase partition method for batch processes was presented.First,an unsupervised multi-scale fuzzy clustering analysis method was proposed to obtain the membership by solving the clustering objective functions constrained by distances between the fuzzy clustering and the sequence.Then,a sequential phase partition.method was proposed to capture the partition positions between main and the transition phase according to the membership change of the time-varying dataset.After that,the cooperative global and local phase partition rules were set,and the optimal phase number and the partition results can be adjusted adaptively according to the sum of quadratic error index and the partition performance combination index.In addition,the online phase partition results of measured data for batch processes can be obtained by the presented support vector data description based phase partition models.Finally,method effectiveness was illustrated by a sequential electronic hand-written image and the penicillin fermentation process.The experiment results showed that the proposed phase partition method,which can solve the problem of over-dependence on the prior knowledge,capture the time-varying characteristic accurately and realize the reliable and accurate online phase partition;2.A trajectory synchronization method of lifting wavelet package transform(LWPT)and dynamic time warping(DTW)was proposed.First,based on LWPT,a method to decompose the trajectory of data at multiple levels of high and low frequency,was presented to extract the trajectory information in different frequency bands.Then,DTW was introduced to synchronize the trajectory of different frequency bands along the direction of similar feature.After that,LWPT based trajectory synthesized method was presented to synthesize the information of the batch process measured data at multiple levels of high and low frequency.Method effectiveness was illustrated by a penicillin fermentation process.The experiment results show that the proposed method can reduce the impact of the Gibbs phenomenon on data trajectory synthesis and avoid the complicated Fourier transform convolution operations,which can increase synchronization speed and improve synchronization accuracy;3.A dynamic hypersphere structure change support vector data description based detection method of abnormal measured data was proposed.First,the dynamic hypersphere structure change support vector data description,a method to build the measured data based hypersphere and set the quantification rules of the structure changes,was presented to obtain the important structure of the hypersphere.Then,the training data based static hypersphere and the testing data based dynamic hypersphere were built,and the structure changes of hyperspheres were captured.Finally,the mapping rules between dynamic hypersphere structure change and abnormal data detection were established,and the abnormal threshold was constructed according to the relationship between online and historical measured data of process,which can achieve online detection of abnormal data in batch processes.4.Effectiveness of proposed methods were demonstrated by a semiconductor etch process and a penicillin fermentation process.The results show that the dynamic detection criteria of proposed methods had high fault detection rates and decreased the false alarm ratesThe research on the intelligent detection for the abnormal measured data in batch processes has obtained valuable achievements in the aspects of both methods and algorithms.The proposed method is correct and effective,and have a prosperous future in biopharmaceuticals,fine chemicals,light industry,food,and agriculture,which can effectively and stably improve product quality and promote production saf-ety.Through further improvement and promotion of the research results,it is possible to promote the reliability of on-line monitoring,optimization control and other related methods and technologies in practical industrial applications,and to efficiently and stably improve product quality and production safety.