Study On Nonlinear Feature Extraction For Process Monitoring

With the development of modern science and technology, the process industry is on a much larger scale and process technique is becoming more complex, also, the degree of automation becomes increasingly higher. Meanwhile, the problems of industrial safety, energy conservation, emission reduction, and industrial structure optimization are much concerned. Traditional process methods of such as monitoring, production optimization and online measurement are based on mechanism model, so they are highly accurate and interpretable. In practice, however, it is difficult to obtain rigorous mechanism models because of complexity in process system. And the system is subject to random disturbances, noise interference and measurement error. Such factors also affects the accuracy of mechanism models. With the development of computer science and technology, large numbers of process data can be easily obtained and stored, also more complex scientific computing and more rapid database retrieval can be done. These, with the addition of development of statistical science, provide basic foundations for data-driven modeling, simulation and optimization.This thesis is carried out under the backgrounds mentioned above. It studies nonlinear feature extraction for process monitoring. Feature extraction is one of basic tasks of data-driven methods. If industial data are concerned, it intends to obtain the valuable information which describes the process essentially in complex original data. In general, the number of dimension extracted is lower than the original one. Feature extracting can also provide models which represent the relationship between inputs and outputs. Process of obtaining essential feature of industry can be regarded as either modeling of system or data preprocessing step which provide simpler data set for subsequent modeling, reducing the complexity of the model, and obtaining more precise calculation results. Taking into account the characteristics of industrial process data, the following aspects are studied in this thesis. 1. Classical linear feature extraction algorithm, principal component analysis, is combined with extreme learning machine classifier to obtain model for fault identification in fault diagnosis.2. A novel soft-sensing method is proposed and applied to gas-phase ethylene polymerization process. Two important indexes representing the product characteristics, melt index and density, are calculated. The proposed method constructs a predictive model based on the kernel mapping support vector regression.3. A soft-sensing algorithm based on principal curves and polynomial least squares is proposed and applied to predict water content of distillation unit of EDC.4. Principal curve method is utilized for process monitoring, and it achieves desired results in CSTR and other simulated data.5. Besides, this thesis introduces a few typical applications adopting the artificial neural networks as nonlinear feature extraction tools in process monitoring. These are taken as contrast against other nonlinear algorithms.Some innovative achievements are obtained, such as:1. While utilizing extreme learning machine classifier to identify faults, a voting strategy is taken to enhance the reliability of results.2. Improved particle swarm optimization algorithm is utilized to estimate two parameters while using SVR to build soft-sensor and the PSO-SVR model is proposed.3. While building models based on principal curves together with nonlinear polynomial least squares and utilizing them to soft-sening, correlation increment factor is designed and the nonlinear feature extracting can give consideration to maximizing relationship between inputs and outputs as in traditional PLS.4. In the application of principal curves to process monitoring, multi-scale principal curve algorithm is proposed based on multi-resolution analysis. This algorithm uses wavelet coefficients to model principal curves just as in MSPCA. The experimental results are consistent with theoretical expectations and indicate that the algorithm is effective.5. In the application of principal curves to process monitoring, neural network which used to simulate the mapping relationship can be replaced by interpolation. By doing so, the complexity computing of the NN and instability results can be avoided. The computing load is allocated to the stages of modeling and calculation evenly.