Research Of Large-Scale Process Monitoring Methods Based On Multi-Block Strategy

It is of great significance to develop timely and reliable process monitoring methods for industrial process systems with increasingly large-scale and complex structures.As one of the main production modes of modern industry,large-scale industrial process often contains multiple sub-operating units,it is usually characterized by strong local characteristics and large dimensionality of process variables,and the traditional monitoring method of building a global centralized monitoring model is often difficult to accurately reflect the operation status of the process.As an integrated monitoring framework that has attracted much attention in recent years,the multi-block strategy is widely used in the field of large-scale process monitoring because of its ability to effectively extract process characteristics and reduce the complexity of monitoring models.This paper combines the multi-block strategy to develop a process monitoring method for the above process monitoring problems.The main research contents are as follows:(1)Aiming at the limitations of traditional clustering methods that excessively rely on prior knowledge when mining process local information,a large-scale process monitoring method based on modified K-nearest neighbors-density peak clustering algorithm is proposed.First,considering the complex data characteristics in the process,the information distance is introduced to optimize the distance function of the original algorithm.The information distance matrix describing the correlation between variables is obtained,and the number of blocks is automatically determined by the decline trend of the decision value of each variable.Second,the improved variable allocation strategy is used to complete the block division,the local monitoring model based on PCA is established within each block.Then,the monitoring statistics in the feature space and residual space are built separately,and the monitoring results are fused by Bayesian strategy to realize online process monitoring.(2)Aiming at the problem of traditional data-driven block division method that can easily cause pseudo correlation of variables within the same block and increase the difficulty of the identification of the root cause variable,a large-scale process monitoring method based on weighted digraph is proposed.First,comprehensively considering the causal relationship of variables in the process structure and the correlation of statistical information,process weighted digraph model is constructed based on process flow diagrams and normalized mutual information,the block division is completed through a community network partitioning algorithm.Second,a kernel principal component analysis method is introduced to establish a monitoring model within each block,so that the model had good monitoring performance for nonlinear process,and the local monitoring results are fused through Bayesian strategy to improve the accuracy of process monitoring.In order to achieve more accurate and efficient fault diagnosis,the set of variables related to the fault is obtained by weighted contribution rate,and then the local digraph of the fault is constructed to identify the source variables and analyze the propagation path.(3)Aiming at the problem that the traditional multi-block strategy ignores the complex correlation between the blocks in the actual process when establishing the local monitoring model,thus affecting the performance of the model,a large-scale process monitoring method based on the information fusion of neighborhood is proposed.Firstly,the block division is achieved based on the process knowledge,and the topological connection matrix used to describe the correlation between process blocks is obtained according to the process structure.Second,the attention of different block information is adjusted adaptively by the self-attention mechanism,and the block feature representation is constructed which integrates the neighborhood information.Then,the deep features of each block are obtained based on the nonlinear mapping of stacked autoencoder,and the monitoring statistics in the feature space and residual space are constructed separately.The analysis results show that the established monitoring model can achieve better monitoring performance after making full use of neighborhood process information to build fusion features.