Data-Based Fault Diagnosis And Prediction For Batch Processes

Incidents in complex industrial processes can not only seriously affect the operations of the processes, but also result in tremendous loss of personnel and wealth. With the increasing requirements on safety and reliability of manufacturing processes, fault diagnosis and prediction attracts more and more attention and has become a hot research topic in the process control. Recently, with the urging requirement in the market of multi-type and high-quality products, the manufacturing of high-value-added products produced mostly through batch processes has become increasingly important in many industries. Consequently, the safety in batch processes has also become a critical issue. In relevant approaches, data-based multivariate statistical analysis techniques show particular advantages in dealing with the high-dimensional and coupling data, which makes them specially and increasingly attractive. Multivariate statistical modeling, online monitoring, fault diagnosis and fault prediction have been under wide investigation for batch processes.Based on Principal Component Analysis (PCA) and Fisher Discriminant Analysis (FDA), this dissertation develops a series of fault diagnosis and fault prediction methods for solving problems in batch processes:1. According to the multi-phase characteristics of batch processes, a phase identification and fault diagnosis method is proposed in this dissertation. The proposed method makes use of the changes of principal component number and loading matrixes to identify stable phases and transition phases automatically. Consequently, focusing on their different data nature, different statistical models are respectively developed as well as the corresponding online monitoring strategy.2. To overcome the weakness of MPCA fault diagnosis, FDA method is studied which takes into account both normal and fault data for modeling. For the cases with insufficient fault data, a diagnosis strategy is developed based on Bootstrap and phase-based Recursive Multi-way Fisher Discriminant Analysis (RMFDA) to improve the diagnosis precision. For nonlinear processes, a Recursive Kernel Fisher Discriminant Analysis (RKFDA) strategy is proposed for the nonlinear fault diagnosis.3. According to the gradually-changing property of faults over batches, a fault prediction method is developed based on Fisher Eigenvector Difference. In the proposed method, the statistical values of the next batch are estimated by the autoregressive models built using Fisher Eigenvector Difference. By comparing the estimated values with the control limits, the gradually-changing faults can be predicted.4. By combining the proposed fault diagnosis and prediction methods, a fault diagnosis framework is set up with its primary functions specified. A fault diagnosis platform is also implemented for injection modeling based on PCA and FDA to verify and illustrate the proposed statistical methods.The successful simulations, experiments and applications of the proposed approaches to batch process systems demonstrate the effectiveness of the present methods, which, thus, enrich the achievement of statistical modeling, online monitoring, fault diagnosis and prediction for batch processes.The successful development of the platform also has practical significance, since it can be used not only for injection modeling process but also for other complex industrial processes.