Study On Soft Sensor Method For Coal Mill Air Flow Based On Support Vector Machine

Somebody once said:"who conquered the sensor, he conquered the world." Visibly, how important it is to measure in the Industrial Age. However, along with the increase of the measuredquantity number, so that some of the state can’t be measured by the existing measuring technology and hardware instrument accurately, and we must find another way to solve this problem. As soon as the soft sensor technology is put forward,it gets a great attention, and quickly start to develop and popularize in the two-way of theory and application.The accurate and reliable measurement of the coal mill primary air flow is a key important factor that for the stable operation of the thermal power unit control system and improving the boiler combustion efficiency. However, by the restriction of the site condition where the apparatus installed, so that the hardware meters measuring value is far between the actual values. Sometimes the deviation is so large that the operators have to control the air flow manually. Combining with the practical use, this paper adopts the soft sensor method to solve the problem, and also takes a research on secondary variables selection, data pretreatment, mathematical model establishing and training as well as testing etc. Inder to determine the secondary variables, we used the method through the mechanism analysis, equipments connection relationship analysis, and combined with the experience of equipment operation and related coefficient analysis etc.And the secondary variables sample values were get data processing through the moving average filtering and data normalization. Finally, the after training soft sensor model was tested by the way of off-line and on-line. Based on support vector machine (SVM) regression, we established the air flow soft sensor model, while analyzed and studied the kernel function and punish factor choices. Through the simulation experiment, we selected the radial basis function (RBF) as the kernel function, and by the cross validation method we determined the best punishment factor for the ε-SVR algorithm. Finally, use the C language and MFC we designed the soft sensor application package with the man-machine interface.The SVR model is tested by the DCS real-time measured value, and the result shows that use soft sensor method can predict accurate result than using any kind of existing hardware flowmeters, and can adapt to the changes in the unit conditions.