A Novel Study Of Analog Circuit Online Performance Evaluation Based On Support Vector Regression

Against the fault values defects of traditional analog circuit performance evaluation, and poor real-time performance, difficult to apply to online evaluation, this paper studies the analog circuit performance online evaluation strategy based on support vector regression machine(LSSVR). The main contents include:online clustering algorithm and sub-model connection method research, the application of robust LSSVR^ FLSSVR and incremental or decremental interactive updating model in performance evaluation, PSO optimize algorithm parameters. Details are as follows:(1) This paper puts forward analog circuit online performance evaluation strategy based on clustering algorithm and FLSSVR. Considering the small sample limitations of SVM, this strategy employed for executing the task of circuit model partition relying on data features by fuzzy clustering algorithm, thus training analysis, effectively reducing the dependence on small samples of SVM.(2) This paper puts forward a novel strategy for evaluation of analog circuit online performance based modified robust LSSVR. More specially, the modified multi-kernel function is first employed to interfuse more flexibility to the kernel in line such as the bandwidths which can guarantee accuracy of the support vector numbers based on the LSSVR. In addition, the modified robust learning algorithm is applied to deal with the data set includes fault values. The residuals are formed from real outputs and predicted outputs of the analog circuit online. The scheme trains the weights of LSSVR model iteratively to deal with fault value by robust learning algorithm and then the trained RLSSVR model is updated by means of incremental learning or decremental learning interaction which take the interests of both history data and the control storage data into consideration, besides verified the effectiveness of the proposed method by simulation experiments.(3) This paper puts forward analog circuit online performance evaluation strategy based on improved membership function of FLSSVR. This strategy assigned each sample a value of membership according to the important degree by FLSSVR, to realize active suppression to the fault values and disturbance. Traditional membership function based on distance does not accurately reflect the relationship between the sample data, missing the abnormal sample easily, resulting in abnormal samples and normal samples have the same membership values. In this paper, we modified membership function with k neighbors thought.(4) This paper studies the parameter optimization method of FLSSVR and multi-model connection method applied in fuzzy clustering algorithm. Although online clustering algorithm can effectively solve the small sample problem of SVM, lack of sub-model connection method, this paper puts forward switch and weighted combination sub-model connection method applied to analog circuit performance evaluation strategy, besides verified the effectiveness of the proposed method by simulation experiments.