中文摘要：

针对模拟电路性能的在线评价问题，既要考虑运算速度又要考虑评价的可靠性。而现场数据的采集常包含有错值，对于错值正确的处理直接关乎到评价结果的可靠性。运用标准支持向量回归机（LSSVR），结合鲁棒学习的优越性，设计修正双核径向基核函数（MDRBF）在线调节核宽度保证支持向量数目确定的精确性，利用改进的鲁棒学习算法处理包含错值的数据集，在线完成模拟电路输出预测与实际输出对比，获取预测误差。该方法利用鲁棒学习算法更新LSSVR权值来处理错值，同时应用增量、减量交互的学习方法兼顾历史数据，控制存储数据总量，完成鲁棒KSSVR（RLSSVR）模型的在线更新。实验以高校模拟电路实验为依托，采用近两年内由精密仪器设备测评所得的小功率放大器的8项技术指标构建训练集，进行RLSSVR在线评价。实验表明，所提出的方法能有效处理错值所带来的回归偏差，性能优于传统LSSVR法、8-SVR法及WLSSVR法，与精密仪器性能评价结果较为接近，且有较优的运算速度，适于在线推广。

英文摘要：

Focusing on the issue of analog circuit performance online evaluation, the calculation speed and evaluation relia- bility should be considered. The data acquired from industrial field usually possess nonlinear feature, time varying feature and often contain fault values;correctly processing the fault values directly relates to the reliability of the evaluation re- sults. A novel online evaluation strategy based on modified robust least square support vector regression （RLSSVR） is pro- posed for the analog circuit performance evaluation. More specially, the modified double-kernel radial basis function （MDRBF） is first employed to interfuse more flexibility to the kernel,such as on line adjusting the kernel bandwidth, which can guarantee the accuracy of the support vector number of the LSSVR. The modified robust learning algorithm is applied to process the data set that includes fault values. The real outputs and the predicted outputs of the analog circuit arc compared online, and the predicted error is obtained. The proposed scheme makes use of the robust learning algorithm to train the weights of the LSSVR model iteratively to process the fault values, and then the trained RLSSVR model is up- dated online with the incremental and decremental interaction learning method,which takes both the history data and the amount of the control storage data into consideration. In the experiment ,on the basis of the college analog electronic exper- iments, the eight technical indexes of the low power amplifiers obtained with the precision instrument evaluation in recent two years were adopted to construct the training set,and the RLSSVR online evaluation was carried out. Simulation experi- ment reveals that the proposed method can effectively proces the regressive deviation caused by the faults values. The evaluation performance and the testing speed of the proposed method are superior to those of the traditional methods, such as LSSVR, e-SVR and weighted LSSVR （WLSSVR） ;the evaluation results are close to the prec