Analog Circuit Fault Diagnosis Based On Support Vector Machines

According to the fuzzy evaluation and classification results from nonlinearity, fault diversity, computational complexity and parameter tolerance of traditional analog circuit, a novel evaluation and classification strategy based on support vector machine(SVM) which combined with three types of machine learning is proposed in this thesis. The main content includes evaluation of analog circuit performance online, supervised K neighbor approach, unsupervised clustering algorithm and the global and local semi-supervised method. Detail works are described as follows:(1) Evaluation of analog circuit performance online. This thesis puts forward analog circuit performance evaluation online based adaptive least squares support vector regression(LSSVR), meanwhile, this online evaluation utilizes the modified radial basis kernel function(MRBF) which has more flexibility to the kernel function online such as the bandwidths tuning, in order to better improve the LSSVR regression performance.(2) Under the condition of evaluation, three forms of machine learning are respectively described in analog circuit fault diagnosis. The method relies on supervised machine learning and proposes a strategy based supervised WKNN combined with WLSSVM in analog circuit fault diagnosis. WLSSVM learning algorithm is used to deal with support vector selection problem in updating KNN algorithm, the concrete implemented method is: the weighted vector of WLSSVM classifier is introduced into the KNN distance formula, the characteristics is treated as the weight vector of distance formula, and then KNN distance formula is changed.(3) Supervised learning in second part is the foundation of machine learning. Supervised learning method can meet the requirement, nevertheless,the defect of data selection cannot be ignored, which has strong dependence on data with unknown type. However, the actual system must have data with unknown type, and the results of the unknown data should lead to lag barrier as well as lower classification precision and so on under the supervised diagnosis method. In order to solve the data defect of part(2), part(3) chooses unsupervised machine learning method to improve. A new solution based on brief SVM which combined with UC is proposed for analog circuit fault classification method. This solution firstly reduce the range of training samples randomly by using the improved SVM(ISVM) in order to decrease the computing complexity and computing time; secondly shrink the sample data once again by virtue of unsupervised clustering(UC), so as to realize effective suppression of random data.(4) To synthesis the proposed supervision and unsupervised method in supervised WKNN-WLSSVM and unsupervised UC-ISVM, not only consider the existed properties of known samples but also pay attention to the potentiality of unknown samples, in this way, analog circuit fault diagnosis should have a big research space. This part put forward to a machine learning method combined with the two methods above, a novel semi-supervised learning method based on the global and local sustain support vector machine in analog circuit fault diagnosis.