CAJ | 학술논문

为了进一步提高欠定盲源分离算法中混合矩阵估计方法的性能，提出了一种基于加权最小二乘支持向量机( SVM)的欠定盲源分离混合矩阵估计方法。该方法利用信号的方向角度特征估计出有效信源信号个数，然后采用加权最小二乘支持向量机方法获得初始权值，每次将其中一个权值对应的样本点作为测试样本，其余点作为训练样本，依次对样本的误差变量进行更新，再根据权值计算公式实现所有权值的更新，进而确定最优分类平面，实现对观测信号的最优分类，最终估计出混合矩阵。实验结果表明，新算法是有效的，其平均误差是基于K-均值方法误差的0.2倍左右，是基于SVM算法平均误差的0.5倍左右。
위료진일보제고흠정맹원분리산법중혼합구진고계방법적성능，제출료일충기우가권최소이승지지향량궤( SVM)적흠정맹원분리혼합구진고계방법。해방법이용신호적방향각도특정고계출유효신원신호개수，연후채용가권최소이승지지향량궤방법획득초시권치，매차장기중일개권치대응적양본점작위측시양본，기여점작위훈련양본，의차대양본적오차변량진행경신，재근거권치계산공식실현소유권치적경신，진이학정최우분류평면，실현대관측신호적최우분류，최종고계출혼합구진。실험결과표명，신산법시유효적，기평균오차시기우K-균치방법오차적0.2배좌우，시기우SVM산법평균오차적0.5배좌우。
To further improve the performance of the underdetermined blind source separation algorithm,an algorithm based on weighted least square support vector machine( WLS-SVM) is proposed. Firstly,it esti-mates the number of source signal using the characteristic of frequency domain signal. Secondly, it uses WLS-SVM to obtain the initial weight values. The sample point corresponding to one of the weight values is used as the test sample every time,and the other is used as training sample. The error variable is upda-ted sequentially,and then all weight values are updated according to the weight calculation formula to de-termine the optimal classification plane and realize optimal classification of observed signals to estimate the mixed matrix. Simulation results prove that the proposed algorithm has smaller error compared with tradi-tional algorithm. The error of proposed algorithm is twenty percent of that of K-means based method,and fifty percent of that of SVM-based method.