Research On Fault Detection Method Based On Kernel Principal Component Analysis Under Multi-block Modeling Framework

In modern industry,how to ensure stable process operation and efficient productivity are two key problems that need to be solved urgently,and the hot spot in industrial control fault detection technology is an effective way to achieve this goal.With the rapid development of sensor technology and distributed control systems,data-driven fault detection methods are widely used in plant-wide control systems.The multi-block integrated model proposed in recent years can effectively detect industrial processes under distributed and complex conditions.In this thesis,the fault detection of industrial processes based on kernel principal component analysis in the framework of multi-block modeling is presented in full as follows:(1)In order to solve the problem that the traditional fault detection method based on kernel principal component analysis needs to determine the chunking threshold according to the system mechanism model when establishing the chunking model of mutual information variables and the complexity of modeling due to the large number of chunks,a kernel principal component fault detection method based on the clustering chunking of correlation matrix is proposed.The algorithm uses the mutual information between variables to construct the correlation matrix,and then uses the spectral clustering algorithm to cluster the variables based on their correlation vector approximation,and finally uses the Calinski-Harbasz index as the evaluation value of the clustered chunks,which not only avoids the uncertainty of the chunks based on the empirical judgment threshold,but also makes the variables in the same sub-block have strong correlation.Further,a fault detection model based on kernel principal component analysis is established for all sub-blocks,and the detection results of each subblock are fused using Bayesian inference,and the proposed algorithm is analyzed by numerical cases and Tennessee-Eastman process simulation experiments and compared with similar multi-block model detection algorithms to verify the proposed algorithm.(2)To address the problem that the large amount of normal information in the correlation matrix clustering chunking kernel principal component fault detection method tends to overwhelm the abnormal information that triggers faults,resulting in poor fault detection,a kernel principal component fault detection method based on the autoencoder error model and correlation matrix clustering chunking is proposed.The algorithm uses the process normal operating condition data to train the autoencoder model,then extracts the reconstruction error based on the trained model for the test data,amplifies the abnormal information that triggers faults,adopts the modeling strategy based on correlation matrix clustering chunking,and finally fuses the detection results of each sub-chunk using Bayesian inference.The proposed algorithm is analyzed by numerical cases and Tennessee-Eastman process simulation experiments and compared with other multi-block model detection methods to validate the proposed algorithm.(3)In response to the traditional fault detection algorithm based on kernel principal analysis that only considers the sample observation information and does not fully exploit the implicit features of the observation,a multi-block kernel principal fault detection method based on observation feature extraction is proposed based on the temporal dimension of the data sample.The method extracts the accumulated error features and the rate of change features of the observation information from the observation information perspective,and then divides the data into three sub-blocks based on the extracted features combined with the original observation information,and then builds a kernel principal fault detection model for each sub-block and fuses the detection results of each sub-block by Bayesian inference.The results are further proposed as a weighting strategy based on the unified contribution graph framework for the multi-block kernel principal element algorithm to identify fault sources.The proposed algorithm is validated by Tennessee-Eastman simulation experiments and a practical application.