Research On Analog Circuit Fault Diagnosis Based On Support Vector Machines

Analog circuits are important parts of electronic equipments. To guarantee the reliability and maintainability, research is focused on the fault diagnosis of analog circuits. With the increasing of scale and integration of circuits, the traditional methods show the defects in accuracy of diagnosis and real-time operation in projects. As the characteristics of analog circuit fault diagnosis: limited fault samples, imbalanced data etc, support vector machines (SVM) is introduced into fault diagnosis of analog circuit in this thesis. Fault samples are obtained by faults simulation and used to train the SVM multi-classification which is used for diagnosis in projects.The purpose of this thesis is to select the appropriate SVM classification model, the method of multi-classification construction, and to improve them according to the characteristics of analog circuit fault diagnosis to construct the fault diagnosis machine based on SVM which is applied in projects.The main content of this thesis is as follows:The defects of the traditional methods and early intelligent methods of fault diagnosis and the advantage of SVM on fault diagnosis field are analyzed. The background knowledge, principles and algorithms of SVM, the SVM classification model, methods for evaluating and methods for multi-classification construction are expounded. The influence of the model parameters and kernel parameters to the performance of the classification is evaluated by experiments.To solve the problems of basic SVM multi-classification caused by imbalanced data and unseparated data, a new method to construct multi-classification is proposed in this thesis, which includes the"weighted penalty factor C"strategy and"K Nearest Neighbor"method. This new method would improve the diagnosis accuracy of minority class samples and reduce the unseparated area which would improve the performance of the multi-classification totally. Experiments on the standard circuits show that the improved multi-classification is better than the basic one on the diagnosis accuracy of minority class samples.The work in this thesis holds certain theoretical and practical value on the research of analog circuit fault diagnosis.