Key Technologies Research On Single Batch Modeling And Monitoring For Batch Industrial Processes

Modern industrial processes tend to be larger, complicated and highly automated features. If a fault occurs, the industrial processes may downtime or shutdown, and even result in a catastrophic event, in which the security of the person may be endangered. Therefore, in order to ensure safe and reliable operation of the production process and improve the quality of products, fault diagnose and monitoring technology has attracted more and more attentions of researchers in the academic and industrial communities. Among modern process industries, batch process industries, which mainly manufacture small-volume, multi-species and high value-added products, take an increasingly prominent place, due to their flexible to meet the changing demands of market. How to keep safe and reliable operation of the batch process as well as ensure continuous and stable product quality, which enables enterprises to improve their market competitiveness, is becoming a hot topic in process control filed.It is well known that batch industrial processes are with fairly complex characteristics and the data generated in the processes has also rich statistical properties. Since batch industrial processes are often characterized with multiple operations, uneven-length and non-Gaussian data, it makes that statistical analysis and online monitoring for these processes is a more challenging task. Taking the difference among batch data and processes characteristics into full account, this paper proposed a suite of comprehensive monitoring solutions for batch industrial processes based on single batch modeling method. The solutions aim to target these problems that include weak fault monitoring problem, uneven-length data problem, non-Gaussian distribution of the data produced under multi-operation modes, incomplete modeling data in the batch industrial processes.In order to effectively detect the weak fault caused by larger fluctuation (uncertainty) of initial conditions in the batch industrial processes, a novel multi-models weak fault monitoring method, called as single dynamic kernel principle component analysis (SDKPCA), is proposed. It firstly integrates kernel PCA and auto-regressive moving average exogenous (ARMAX) time series model by proposing single batch modeling thought to build SDKPCA for each batch of data at each stage. Then, hierarchical clusters are obtained through load matrix similarity among SDKPCA models. At different stages, multi-model structures are constructed along with the variation of the cluster number. This method takes the advantages of capturing dynamic and nonlinear properties in the processes as well as discriminating the difference among multi-batch data by using model refinement so that the weak fault that exists in the processes can be effectively detected.To target these problems that include severely uneven-length data in batch industrial processes and larger computational complexity on KPCA method in the case of larger amount samples set, a new batch processes modeling and monitoring method, called as feature-points-based single dynamic kernel principle component analysis (FP-based SDKPCA) is developed. At first, it extracts the same number feature points from each batch of data in processes by using three-step-feature-points extraction method. As a result, all of batch data are equalized. Then, these features points can be used to construct the monitoring model by adopting SDKPCA. Since the number of feature points is very fewer than that of total batch data, it substantially reduces the computation load of KPCA.To handle the correlation between processes changing over time caused by random operation or frequent operation in multi-operations and serious non-Gaussian distribution of data, an online monitoring method for multi-operation batch processes, called as local collection standardization and single dynamic kernel principal component analysis (LCS-SDKPCA), is presented. Since frequent operation and random operation in multi-operation processes make the correlation between processes over time, it is clearly that the modeling model should also be changed accordingly. Hence, the operation data with similar features in processes may be clustered to construct the model. However, the clustered data do not obey Gaussian distribution due to randomness of operation. To make the data follow Gaussian distribution, local collection standardization (LCS) is explored here so that multiple variable statistical model can be set up by using the data. Further, fault monitoring for multi-operations can be realized by using SDKPCA.It is prone to output infinite kinds of data due to randomness of operation in multi-operation. As a result, it is very difficult in collecting all kinds of data, which means incomplete modeling data. If data is unavailable under one certain condition, it is unable to establish a corresponding model since there is no this kind of data among historical modeling data. Therefore, when new operational data is coming, it is no doubt that the monitoring model is uncertainty without the corresponding model. To solve this problem, a fast monitoring method based on Up-Down model is proposed so that new kind of data can be effectively detected if it happened in process.The proposed monitoring methods are applied to the fault detection in a benchmark simulation of fed-batch penicillin production process and ladle furnace (LF) steelmaking production process. Experimental results demonstrated their effectiveness. This dissertation provides new methods and technologies for statistical modeling and monitoring the batch processes with frequent operation and serious uneven-length data, which is of great theoretical significance and distinctive values.