Industrial Process Monitoring:An Approach Based On Wavelets And Statistics

The safety of production procedure and consistency of product quality are always two themes of the process industry. To avoid disastrous accidents and decrease fluctuations of product quality so that products are competitive, the process conditions must be under closely monitoring and faults should be timely detected. However, the large scale and complex structure of industrial process, as well as the uncertainty of real environment, makes the theoretic research and implementation of process monitoring system as one of challenges in the field of process control. The model-based monitoring methods are much maturer in theory in comparison with the model-free one, but the former has great difficulties when applied in the real plants due to the previously mentioned characteristics of industrial process. Therefore, the topic of this thesis focuses on the model-free monitoring method, which is more practicable for its data-driven or knowledge-based characteristics. Two primary mathematical tools used in our proposed monitoring strategy are wavelets and statistics. Both of them are model-free methods when applied to the process monitoring. Specifically, the wavelet theory deals with process data as a signal processing or function approximation tool, and statistics (here we actually refer to the multivariate statistical analysis) is used to build the process monitoring model which is realized by statistically analyze process data. The wavelets and statistics play different but complementary roles in the whole process monitoring strategy. The main contributions of this thesis are as follows: ?The elementary concepts and scope of process monitoring are systematically introduced. The model-based and model-free monitoring methods are compared, and the latter is shown more suitable and practicable for industrial applications. ?A wavelet-based rnultiscale hierarchical error-compensation algorithm xIv is presented to treat the multi-rate issues of process variables. The proposed method is used to reconstruct a low sampling rate signal to a higher one. The reconstruction errors are given and derived results are justified. This algorithm has the advantages of noise-free, high reconstruction accuracy and explicit physical background. ?The character and requirement of the filtering of process data are illustrated, and the ideas of wavelet thresholding filtering and robust wavelet decomposition are introduced. Further, key techniques of multiscale on-line process data decomposition are studied and the implementation of shift-invariant decomposition and interval wavelet decomposition algorithms are presented. (~) The probability density function of process variable is estimated by using wavelet thresholding density estimator (WTDE). The parameters of WTDE as the grid structure, smoothing parameter, and thresholds are selected by the proposed criteria. The iterative Q-Q plot is used to eliminate the influence of gross errors to the density estimation procedure. ?The characteristics and fault detection behavior of PCA are explored, and the relations of expectations of ~ and SPE statistics to the statistical parameters of process data are presented. These relationships reveal the influence factors to the detection behavior of PCA and can be used to differentiate the process change from fault. Some default conclusions conflicting with the real root cause and thus l...