Local Learning Strategy Based Improved Independent Component Analysis And Its Application In Multimode Process Monitoring

With the development of science technology and improvement of social productivity,modern industry is becoming larger and more complicated.The demands for the safety and reliability of system are also increasing.Meanwhile,large amounts of industrial process data are recorded in the computer control system,which provides research base for data-driven based fault diagnosis methods.Aiming at the multimode characteristic involved in process industry,this paper studies multimode process monitoring method based on improved independent component analysis(ICA)in the framework of local learning strategy.Firstly,aiming at the fault diagnosis of multimode process,the paper proposes a 6)nearest neighbor independent component analysis(KNNICA)method.This method standardizes the multimode data in each sample's 6)nearest neighborhood to remove the multimode property.Based on the data following one single mode,KNNICA establishes ICA statistical model to detect the process faults.The simulation results on continuous stirred tank reactor(CSTR)system and Tennessee Eastman(TE)system show that KNNICA is capable to monitor the multimode process effectively.Then,in order to extract the statistical information hidden in the multimode process,the paper develops two improved independent component analysis methods including the local entropy independent component analysis(LEICA)method and the local statistics independent component analysis(LSICA)method.LEICA introduces local learning strategy into information entropy theory.In LEICA,the multimode characteristic of process data is removed by estimating the local probability density of samples,and then local entropy is calculated to mine the internal information of observed data.Statistics pattern analysis(SPA)is also an effective tool to extract statistical information.Therefore,LSICA combines local probability density estimation and SPA to deal with multimode process data and ICA model is built in local statistics space to monitor the state of industrial process.The monitoring results on TE and CSTR systems demonstrate that,compared with ICA and KNNICA methods,LEICA and LSICA have the better fault detection effects.Finally,to detect the faults in multimode process with transitions,the paper studies an independent component analysis method based on weighted distance space(WDSICA).This method splits the observed data into various snapshots using moving window technique,and then maps the original variables into weighted distance space.In this way,the multimode differences including transitions are eliminated.Furthermore,distance component statistical model is established using ICA algorithm and two monitoring statistics are constructed to monitor the process faults.A multimode CSTR system with transitions is applied to verify the effectiveness of the WDSICA method.