基于支持向量機(jī)的模擬電路性能在線評(píng)價(jià)策略研究
發(fā)布時(shí)間：2019-07-05 09:42
【摘要】：針對(duì)傳統(tǒng)的模擬電路性能評(píng)價(jià)方法對(duì)于錯(cuò)值處理的缺陷,以及較差的實(shí)時(shí)性,難以應(yīng)用于在線評(píng)價(jià),本文以支持向量回歸機(jī)(LSSVR)為基礎(chǔ),研究了基于支持向量機(jī)的模擬電路性能在線評(píng)價(jià)方法。主要內(nèi)容包括：在線聚類算法及子模型連接方法研究,魯棒LSSVR、模糊LSSVR(FLSSVR)及增減交互模型在性能評(píng)價(jià)中的應(yīng)用,粒子群算法優(yōu)化核參數(shù)。具體工作如下：(1)提出基于聚類分析的FLSSVR在線模擬電路性能在線評(píng)價(jià)策略。該策略考慮支持向量機(jī)小樣本依賴的局限性,采用模糊聚類算法將整個(gè)數(shù)據(jù)樣本集根據(jù)數(shù)據(jù)特征劃分為具有多個(gè)具有同類特征的子樣本集,從而進(jìn)行訓(xùn)練分析,有效降低了支持向量機(jī)對(duì)于小樣本的依賴性。(2)提出基于魯棒LSSVR的模擬電路性能評(píng)價(jià)新策略。本文運(yùn)用標(biāo)準(zhǔn)LSSVR,結(jié)合魯棒學(xué)習(xí)的優(yōu)越性,設(shè)計(jì)修正多核徑向基核函數(shù)在線調(diào)節(jié)核寬度保證支持向量數(shù)目確定的精確性,利用改進(jìn)的魯棒學(xué)習(xí)算法處理包含錯(cuò)值的數(shù)據(jù)集,在線完成模擬電路輸出預(yù)測(cè)與實(shí)際輸出對(duì)比,獲取預(yù)測(cè)誤差。該方法利用魯棒學(xué)習(xí)算法更新LSSVR權(quán)值處理錯(cuò)值,同時(shí)應(yīng)用增量、減量交互的學(xué)習(xí)方法兼顧歷史數(shù)據(jù),控制存儲(chǔ)數(shù)據(jù)總量,完成RLSSVR模型的在線更新,并通過(guò)仿真實(shí)驗(yàn)驗(yàn)證了所提方法的可行性。(3)提出基于改進(jìn)隸屬度函數(shù)的FLSSVR模擬電路性能在線評(píng)價(jià)策略。該策略利用FLSSVR對(duì)每個(gè)樣本根據(jù)其重要程度分別賦一個(gè)隸屬度值,從而實(shí)現(xiàn)對(duì)錯(cuò)值和干擾的有效抑制。傳統(tǒng)的基于距離的隸屬度函數(shù)不能準(zhǔn)確反映樣本數(shù)據(jù)間的相互關(guān)系,容易遺漏異常樣本,從而導(dǎo)致異常樣本與正常樣本具有相同的隸屬度值。本文結(jié)合k近鄰思想,對(duì)隸屬度函數(shù)進(jìn)行了修正。(4)研究FLSSVR參數(shù)優(yōu)化及適用于模糊聚類算法的多模型連接方法。在線聚類算法雖能有效的解決支持向量機(jī)小樣本的困擾,但同時(shí)欠缺有效的子模型連接方法,本文給出了應(yīng)用于模擬電路性能評(píng)價(jià)策略的開(kāi)關(guān)切換及加權(quán)組合子模型連接方法,并通過(guò)仿真實(shí)驗(yàn)驗(yàn)證了所提方法的有效性。
文內(nèi)圖片：
圖片說(shuō)明：圖３－１兩個(gè)不同類中樣本么間緊密度的差別逡逑Ｆｉｇ．３－１邋Ｔｈｅ邋ｄｉｆｆｅｒｅｎｃｅ邋ｏｆ邋ｃｌｏｓｅ邋ｄｅｇｒｅｅ邋ｂｅｔｗｅｅｎ邋ｔｗｏ邋ｓａｍｐｌｅｓ邋ｉｎ邋ｄｉｆｆｅｒｅｎｔ邋ｃｌａｓｓ逡逑
[Abstract]:Based on the support vector regression machine (LSSVR), the on-line evaluation method of the performance of the analog circuit based on the support vector machine is studied. The main contents include: on-line clustering algorithm and sub-model connection method, robust LSSVR, fuzzy LSSVR (FLSSVR) and increase and decrease interaction model in performance evaluation, and particle swarm optimization. The specific work is as follows: (1) The online evaluation strategy of FLSSVR on-line simulation circuit based on cluster analysis is put forward. The strategy takes into account the limitation of support vector machine small sample dependency, and adopts the fuzzy clustering algorithm to divide the whole data sample set into a plurality of sub-sample sets with similar characteristics according to the data characteristics, so that the training analysis is carried out, and the dependence of the support vector machine on the small samples is effectively reduced. (2) A new strategy for evaluating the performance of an analog circuit based on robust LSSVR is presented. In this paper, we use the standard LSSVR to combine the advantages of robust learning, to design the modified multi-core radial basis function to adjust the kernel width on-line to guarantee the accuracy of the number of support vectors, and to use the improved robust learning algorithm to process the data set containing the error value. And the on-line completion of the analog circuit output prediction is compared with the actual output to obtain a prediction error. The method uses the robust learning algorithm to update the LSSVR weight value processing error value, and simultaneously applies the learning method of the increment and decrement interaction to balance the historical data, controls the total amount of the stored data, completes the on-line updating of the RLSSVR model, and verifies the feasibility of the proposed method through the simulation experiment. (3) An on-line evaluation strategy of FLSSVR analog circuit based on improved membership function is proposed. The strategy utilizes FLSSVR to assign a membership value to each sample according to their importance, so as to realize the effective suppression of the error value and the interference. The traditional distance-based membership function can not accurately reflect the relationship between the sample data and easily miss the abnormal sample, thus leading to the abnormal sample having the same membership value as the normal sample. In this paper, the membership function is modified with the idea of k-nearest neighbor. (4) The optimization of FLSSVR parameters and the multi-model connection method for fuzzy clustering algorithm are studied. Although the on-line clustering algorithm can effectively solve the problem of the small sample of the support vector machine, but at the same time the effective sub-model connection method is lacking, the invention provides a switch switching and weighting combination sub-model connection method applied to the performance evaluation strategy of the analog circuit, The validity of the proposed method is verified by the simulation experiment.
【學(xué)位授予單位】：渤海大學(xué)
【學(xué)位級(jí)別】：碩士
【學(xué)位授予年份】：2015
【分類號(hào)】：TN710;TP18
，
本文編號(hào)：2510437
文內(nèi)圖片：
圖片說(shuō)明：圖３－１兩個(gè)不同類中樣本么間緊密度的差別逡逑Ｆｉｇ．３－１邋Ｔｈｅ邋ｄｉｆｆｅｒｅｎｃｅ邋ｏｆ邋ｃｌｏｓｅ邋ｄｅｇｒｅｅ邋ｂｅｔｗｅｅｎ邋ｔｗｏ邋ｓａｍｐｌｅｓ邋ｉｎ邋ｄｉｆｆｅｒｅｎｔ邋ｃｌａｓｓ逡逑
[Abstract]:Based on the support vector regression machine (LSSVR), the on-line evaluation method of the performance of the analog circuit based on the support vector machine is studied. The main contents include: on-line clustering algorithm and sub-model connection method, robust LSSVR, fuzzy LSSVR (FLSSVR) and increase and decrease interaction model in performance evaluation, and particle swarm optimization. The specific work is as follows: (1) The online evaluation strategy of FLSSVR on-line simulation circuit based on cluster analysis is put forward. The strategy takes into account the limitation of support vector machine small sample dependency, and adopts the fuzzy clustering algorithm to divide the whole data sample set into a plurality of sub-sample sets with similar characteristics according to the data characteristics, so that the training analysis is carried out, and the dependence of the support vector machine on the small samples is effectively reduced. (2) A new strategy for evaluating the performance of an analog circuit based on robust LSSVR is presented. In this paper, we use the standard LSSVR to combine the advantages of robust learning, to design the modified multi-core radial basis function to adjust the kernel width on-line to guarantee the accuracy of the number of support vectors, and to use the improved robust learning algorithm to process the data set containing the error value. And the on-line completion of the analog circuit output prediction is compared with the actual output to obtain a prediction error. The method uses the robust learning algorithm to update the LSSVR weight value processing error value, and simultaneously applies the learning method of the increment and decrement interaction to balance the historical data, controls the total amount of the stored data, completes the on-line updating of the RLSSVR model, and verifies the feasibility of the proposed method through the simulation experiment. (3) An on-line evaluation strategy of FLSSVR analog circuit based on improved membership function is proposed. The strategy utilizes FLSSVR to assign a membership value to each sample according to their importance, so as to realize the effective suppression of the error value and the interference. The traditional distance-based membership function can not accurately reflect the relationship between the sample data and easily miss the abnormal sample, thus leading to the abnormal sample having the same membership value as the normal sample. In this paper, the membership function is modified with the idea of k-nearest neighbor. (4) The optimization of FLSSVR parameters and the multi-model connection method for fuzzy clustering algorithm are studied. Although the on-line clustering algorithm can effectively solve the problem of the small sample of the support vector machine, but at the same time the effective sub-model connection method is lacking, the invention provides a switch switching and weighting combination sub-model connection method applied to the performance evaluation strategy of the analog circuit, The validity of the proposed method is verified by the simulation experiment.
【學(xué)位授予單位】：渤海大學(xué)
【學(xué)位級(jí)別】：碩士
【學(xué)位授予年份】：2015
【分類號(hào)】：TN710;TP18
，
本文編號(hào)：2510437
本文鏈接：http://sikaile.net/kejilunwen/dianzigongchenglunwen/2510437.html
最近更新
教材專著
