The Application Of Support Vector Machine In Industrial Inferential Measurements

Because neural network is based upon empirical risk minimization and asymptotic theories, it is suitable to deal with situations where the amount of samples is tremendous and even infinite. However, in the industrial practice, finite and little amount of samples is usually present. So a novel promising machine learning technique specifically developed for analyzing little amount of samples, SVM (Support Vector Machine), will be more suitable in practical industrial application.SVM has gained increasing attentions recently for its excellent properties and performance. Tremendous theoretical advancements have been achieved in comparison to few application cases so far which are limited to speech recognition, image processing and text classification etc. Especially there is a lack of industrial application. So in this thesis, we attempt to apply SVM in industrial practice and we focus our attention on industrial inferential measurement.While coping with specific engineering problems, several varied algorithms have been proposed and applied in this thesis. The main contributions of this thesis are as follows:1. Review the principles, development and state-of-art techniques of statistical learning theory and support vector machine. In the mean time, the concepts of industrial inferential measurement are also briefly introduced.2. After reviewing the popular techniques utilized in water quality monitoring and analyzing their disadvantages, a novel method to rapidly detect COD (Chemical Oxygen Demand) in polluted water with UV (ultra-violet) spectroanalysis is proposed. This method utilizes an on-line algorithm based upon LSSVM (least square support vector machine), which can build adaptive models to predict the COD values of unknown water samples quickly and accurately. In the modeling process, every training sample is also assigned a prior weight to take their significance to the final predictive model into account. Practical application results show that this adaptive algorithm can build better estimation models than the traditional modeling techniques, such as PLS, MLR and BP neural network etc. Meanwhile, the new COD model shows a good correlation between COD estimated values and COD analysis values.3. The methodology based upon LSSVM for on-line inferential measurement has been proposed and applied in parameter estimation problem encountered in a continuous catalytic reforming unit. To deal with model deviation problem, moving windows framework is incorporated into inferential measurement methodthat is based upon LSSVM. To reduce computational load further, a recursive LSSVM algorithm is utilized. On the other hand, every training sample is also assigned a prior weight by taking their respective contribution to the final predictive model into account. Theoretical analysis and application results show that this algorithm can construct and modify on-line soft sensor model rapidly and circumvent model deviation, a common problem present in inferential measurement, efficiently. Simulation results also show that the computational toad of this proposed algorithm is lower than standard SVM and LSSVM algorithms. Finally, simulation results show that soft sensors based on this algorithm have much better predictive ability than those constructed by using traditional soft sensor modeling techniques, such as RBF neural network and BP neural network etc.4. To overcome the common problems, difficulty of determining the optimal structure and slow training process, present in BP neural network, a novel non-iterative training algorithm for multilayer feedforward neural network has been proposed. Because this training algorithm is only based on linear least squares, and no iterative learning process is needed, its computational load is relatively small. On the other hand, this algorithm determines the best weights of network and the minimal number of hidden nodes automatically according to the demand of model precision and the performance of tentative model. Compared with the well-known BP (back-propa...