Research And Realization Of Fault Diagnosis Methods In Process Industry Based On Improved SVM Algorithm

The fault monitoring and diagnosis system in industrial process is a significant part of process industry CIPS. The cost would be fatal both in manpower and property once the accident happened. Therefore, the fault monitoring and diagnosis of process has become an important research aspect of the controlling field. Monitoring and diagnosing the fault timely and precisely in process can not only reduce the accidents probability and increase the safety of process operation, but also reduce the production management cost and improve the product quality. The fault diagnosis method based on improved SVM in this paper is applying support vector machine—the youngest theory in artificial intelligence and machine learning—to fault diagnosis technology, which can fully make use of measurable process data and do not rely on the accurate mathematic model.The theory and methods of SVM were studied, and the problems that SVM dealing with the large-scale samples in fault diagnosis were discussed in this paper. Combining the large-scale samples reduction strategy with SVM, an improved SVM and new strategy of large-scale sample set reduction has been proposed in this paper. This new strategy can eliminate the samples called non-support vectors which are useless or wasted to the classifier. With this effort, the problems like slow learning, large storage demand, poor generation ability which caused by large-scale training sample set have been solved and the advantage of the modified SVM method has been proved through experiment. The constraint of reduction threshold introduced by the paper can deeply dip the information that contained in wrong classified samples when the classification accuracy of original classifier is low, and increase the classification accuracy of the final classifier.In order to compare the different sample set reduction methods, a new reduction algorithm evaluation has been proposed in this paper. The quality of reduction algorithm depends on whether it can ensure the final classification accuracy while shortening the time. For the problem that time and accuracy are not comparable in different order of magnitude, formulas of time reduction rate and accuracy decreasing rate have been proposed. The results of experiments prove that a reduction algorithm can be good evaluated by comparing the two parameters.To solve the multi-classification problem the fault diagnosis faced in actual process industry, the paper applies the improved SVM to multi-classification problem and two typical process industries. A large number of fault diagnosis experiments have been taken to TE process and the diagnosis method has been verified. Meanwhile, diagnosis experiments to the abnormal blast furnace conditions have been done and the results are excellent too, demonstrating the effectiveness and feasibility of the improved SVM based on process industry fault diagnosis method in practice.Integrating VC++6.0and SQL Server2000, a blast furnace fault diagnosis system based on improved SVM has been developed. The results of experiment prove that the system is basic satisfactory in all aspects.