Fault Detection Research For Multi-modal Process Based On Dynamic Local Data Integration

In recent years,with the rapid development of technology,the industrial processes have more and more complex production processes,so a particularly important question is how to use fault diagnosis technology to improve the safety and reliability of industrial process.Fault diagnosis method based on multivariate statistical analysis is an important branch of fault diagnosis areas.Because the method does not rely on mathematical models and is easy to obtain the required mass production process in the industrial field,it has important theoretical value and application value.In this thesis,the research status of fault diagnosis in multi-modal process is stated.Multimodal data sets have two main characteristics:one is that the data centers of each mode don't coincide;the other is that the data discrete degree is different,that is,the variance of each mode data is different.For the multi-modal problem in batch process of chemical industry,a new dynamic multiway local outlier factor method?DMLOF?is proposed to solve the multi-center problem in multi-modal process.Pointing at multi-modal data have the problem of multi center and the distinctly different modal variance,a method based on the local neighborhood standardization partial least squares?LNS-PLS?is proposed in this paper.Aiming at the quality-related problem of multimodal data in batch processes,a principal component near-neighbor standardized partial least squares?TNSPLS?method is proposed.A new improved partial least squares?IPLS?fault detection method for multimodal process is proposed for output-related and input-related faults in multimodal process in this paper.The four methods are applied to the penicillin fermentation process and have good results.Specifically,the main work and contributions are as follows:1.For the multi-modal problem in batch process of chemical industry,using moving window with local outlier factor?LOF?propose a new dynamic multiway local outlier factor method?DMLOF?for on-line fault detection of industry process,and the method can improve the performance of fault detection.Firstly,the approach unfolded the batch dataset into a two dimensional dataset,then in the time slice using local outlier factor algorithm with moving window technology calculated the local outlier factor statistics,and using the kernel density estimation?KDF?determined the control limits.Secondly,the new data was projected in the corresponding window after standardized,and the local outlier factor statistics of new data was determined and compared with control limits for fault detection.Finally,the results of simulation experiment of penicillin fermentation process showed the validity of the algorithm.2.Pointing at multi-modal batch data has the problem of multi center and the distinctly different modal variance,a method based on the local neighborhood standardization partial least squares?LNS-PLS?is proposed in this paper.Firstly,the local nearest neighbor standardization?LNS?method was used to transform the original data into Gaussian distribution in this paper.The PLS model was established and the control limit of T²and SPE was determined.Next,after the LNS standardized of the test data,the new Gaussian PLS monitoring indexes were calculated for process monitoring and fault detection.Finally,the effectiveness of the algorithm was verified by the simulation experiment of numerical example and penicillin fermentation process.The results show that the proposed method solve the problem of the neighbor set of fault samples spanning two modes and has better fault detection ability for multi-modal data.3.Aiming at the quality-related problem of multimodal data in batch processes,a principal component near-neighbor standardized partial least squares?TNSPLS?method is proposed.Firstly,the method established the partial least squares model.The principal component matrix and projection matrix was obtained by the model.The principal component matrix was standardized to get new principal component,and theT²_ns statistics,Q statistics and control limits were calculated.Then the principal components of test data were calculated by projection matrix,and standardized by their nearest neighbors.The test dataT²_nsand Q index were calculated for process monitoring and fault detection.Finally,the effectiveness of the algorithm was verified by the simulation experiment of numerical example and penicillin fermentation process.The results show that the proposed method has better fault detection ability for quality related multi-modal data.4.The projection of multimodal data into the principal component space of PLS still has a multimodal structure.In addition,the residual subspace of PLS can contain large variations,which is not appropriate to be monitored using Q-statistic.In order to solve these two problems,an improved PLS method is proposed.The new method deals with the principal component space of PLS to make it approximately obeyed the multivariate Gauss distribution.At the same time,the residual space was further decomposed into IPS and IRS.The validity of IPLS method was verified by numerical examples and penicillin fermentation process.The experimental results show that the proposed method has good fault detection ability for quality-related and quality-independent faults in multi-modal process.