Study On Support Vector Machine And Its Applications To Soft Measurement For Several Important Parameters In Pulping Process

In order to save energy, reduce emission, protect environment and improvethe quality of paper, it is necessary to further improve the level of automationand informationization in pulp and paper industry. Measurement of someparameters such as Kappa number during pulping, Black liquor consistency inwashing process and alkali recovery has always been a hotspot in pulp and paperindustry, and it has been one of typical difficulties in enterprise automation andinformationization development. Support Vector Machine is a kernel functionlearning machine following the principle of the structural risk minimization,which possesses the advantages of complete theory, global optimization, goodadaptability and better generalization. It is now a studying hotspot ofinternational industrial automation. Compared with the traditionalmachine-learning method, Support Vector Machine has great potential indevelopment and application. Surrounding with several application problems ofonline soft-measurement based on Support Vector Machine for some importantprocess parameters in pulp, the paper combines the process techniques andknowledge, and improves some algorithms to solve problems of slow speed, lowprecision and lack of online revise in application of the current soft-measurementmodel.Simulation and application results show the obvious effect of theimproved algorithms in the paper. The main contributions of this paper aresummed up as follows:(1)The research of soft measurement model for Kappa number based onthe on-line adaptive SVM. Pointed to the disadvantages such as low precision, poor online adaptive capacity of the former soft measurement model for Kappanumber in the pulp cooking process, a SVM-based online adaptive model ispresented. The cooking is conducted in several finite localities, so FuzzyC-means algorithm is adopted to divide sample points into several sectionsfollowing up the partition principals that make the training sample of each classmost similar and the training sample of different class most dissimilar. Then thesimilar soft measurement model is established. Through subdividing a maincategory of soft measurement model into some small categories, the staticprecision of the soft measurement model enhanced correspondingly. The meritsand demerits of common soft measurement methods for Kappa number at homeand abroad and the domestic applications are discussed, which shows that moreresearches are focused on improving the modeling static precision instead of thedynamic precision. Studies on the process analysis and the current cookingsituation show that there have been great changes taken place in domesticcooking operating conditions. Because of the variety of the domestic cookingoperating conditions, there is no doubt that the precision reduces during theprocess of modeling if the model has no ability of online self-adaptation. Ingeneral, Support Vector Regression is offline and batching, which isinappropriate for Kappa number measurement in cooking with constantoperating regime variety, so the algorithm should be improved to set up modelbased on delta self-adaptation Support Vector Machine adequate for operatingregime variety. That is to say, during the process of using the model, newsamples are collected constantly and the model is relearned repeatedly based onthe prime one. The algorithm injects the new operating regime factors into theoriginal model appropriately. In order to reduce the share of computer memoryand improve the model speed, it is necessary to select the appropriate samplereplacement strategy. Considering that cooking is a typical slowly time-variedprocess, the paper adopts the sample replacing strategy which combines themethod based on Support Vector data description and the method based onmoving time window to satisfy the algorithm and performance of cookingprocess. Compared with general methods, on one hand, the model makes full useof the historical training results and decreases follow-up time significantly; onthe other hand, the history data is no longer to be saved to reduce the storage space required by the algorithm. Simulation and application results show that thedelta on-line self-adaptation Support Vector Machine model based on datamining is suitable for the measurement of Kappa number in cooking process.(2)The research of soft measurement model for Black liquor consistencyin the evaporation process of alkali recovery based on the on-line adaptive errorcompensation LS-SVM.According to the conditions that the export Black liquorconsistency is difficult to measure in the evaporation process of alkali recovery,this paper presents the Online adaptive error compensation LS-SVM model.Through studying and analyzing the alkali recovery process and the principle ofmultiple-effect evaporator, it is known that Black liquor consistency in theevaporation process of alkali recovery is influenced by many factors such asinitial diluted concentration and flow of black liquor, total effective difference intemperature of multiple-effect evaporator, the major components, thetemperature of cooling water, ambient temperature, the heat transfer coefficientand so on. Black liquid evaporation usually consists of multiple-effectevaporators, and each evaporator consists of vapors, black liquor andcondensation water. The evaporation process has a serious non-linearperformance due to liquid column pulsation in evaporator's internal pipelinesand oscillation phenomenon caused by temperature, concentration and flow ofblack liquor. Owing to long size and high capacity of evaporator's internalpipeline, one end of system may take a long time to respond to physic quantumchanges of the other end, so the actual influencing factors of Black liquorconsistency is more complex. Present models of the soft measurement for Blackliquor consistency takes pressure and temperature as accessorial variables andexport Black liquor consistency as main variable while ignoring other factors. Asa result the accuracy of the model is not high. Residual adaptive compensationsystem established in this paper considers other factors. The paper improves theprecision of the model in two ways: firstly, the algorithm of on-line Least SquareSupport Vector Machine is adopted in model training and on-line correction tomake the model adapt the current work situation; secondly, the tache of residualadaptive compensation is added to make the result of soft measurement moreaccurate by on-line modifying method. The common Least Square SupportVector Machine algorithm is simple and has more calculation and high velocity, but it loses the inner sparsity and robustness of Support Vector Machinemodeling. If the algorithm is not been improved, it will be unusable in actual dueto increasing training examples, long running time and large memory occupancy.Online adaptive Least Square Support Vector Machine removes the mostlongstanding sample while adding a new sample, and corrects the model onlinein delta mode to reduce the arithmetic complexity and save computer memory.The residual compensation adaptive model is established with multiple linearregression method, and its adaptive efficiency method is built using the currentdata and discarding the earliest error. Simulation and application results showthat the soft-measurement model for export Black liquor consistency offers highaccuracy and it has been able to adapt to changes in working conditions.(3)The research of soft measurement model for Black liquor consistencyin washing process based on the fuzzy on-line adaptive LS-SVM.The research ofits soft measurement model is relatively less. The former models have poorgeneralization capability and low accuracy. In view of the shortcomings of theformer models, a new model is proposed based on the fuzzy on-line adaptiveLS-SVM. Online measurement for Black liquor consistency in washing processis difficult. The analysis results of the technology of washing process show thatBlack liquor consistency is influenced by pulp flow, pulp thickness, wash slurrywater, vacuum degree, drum speed and so on. Therefore, it is difficult to executethe real-time measurement by traditional methods. Considering the actualsituation of the factory, the modeling parameters are determined in the conditionwithout increasing the instrument or affecting the normal production. Throughthe on-site analysis and discussions with senior engineers, it is determined to useinlet Black liquor consistency, inlet Black liquor flow and water flow in washingprocess as the accessorial variables, and the first period Black liquor consistencyas the leading variable to set up the soft measurement model for the first periodBlack liquor consistency. The soft measurement model based on the on-site datacollection can only toughly reflect the substantial change trend of the industryobject, so there are inescapable errors existing in this model. When the model isput into use, the object characteristics and working point may change with timeprogressing because of its nonlinear, time-varying and uncertain properties, sothat the errors of parameter measurement may increase with the original data sampling of models. Online adaptive Least Square Support Vector Machine canadapt to the changes of the working conditions and correct the models online.Samples from washing process contain different levels of noise. In ordering toimprove the accuracy of the model, one sample point is signified by one degreeof Support Vector defined in order to show the weight of the sample points in themodel training process. The sample replacement strategy of the model is toreplace the sample point which is the minimum sample point of the Lagrangemultiplier absolute value with a new sample point. Simulation and applicationresults show that the soft-measurement model for Black liquor consistency issuitable for washing process.(4)In this thesis, checking the abnormal value based on SVM is discussedand studied. The support vector machine method can discover and solve theproblems in washing process such as the measuring instrument's calibrationinaccuracies, no-timely, instrument malfunction, larger process interference,benchmark drift, operator's mistakes, a periodically pipeline cleaning, theinstrument's value volatility caused by the pipeline stopper, the data conflictcaused by data distortion and so on. The paper discusses the scheme which canfind the abnormal data in leading variables by using SVM regression algorithm,and it is verified by example, then it is used in select abnormal data of leadingvariables during washing process.The paper takes the soft measurement for Kappa number during pulping,Black liquor consistency in alkali recovery and washing process as researchsubjects, presents a series of algorithms based on Support Vector Machine toimprove the on-line measurement precision, and proves the validity of theproposed algorithm through simulation and application. The research results areuseful as reference for other industry soft-measurements.