基于ReliefF算法与偏最小二乘方法的鸽子转向运动解码

Pigeon Turning Motion Decoding Based on Relief F Algorithm with Partial Least Squares Algorithm

针对微电极阵列记录的局部场电位(LFP)信号包含大量的噪声和冗余信息,而且信号特征维数高,从而影响解码正确率的问题,结合Relief F算法与偏最小二乘(PLS)方法,解码了动物的转向运动行为。设计了鸽子的十字迷宫目标导向实验,采集鸽子弓状皮质尾外侧(NCL)LFP神经信号,提取信号的特征,利用Relief F算法对各个特征赋予相应的权重值,根据权重阈值选取合适的特征构成特征子集;并用PLS对特征子集提取主成分,最后用支持向量机(SVM)进行解码,并将解码结果与单独使用Relief F算法和PLS算法比较。结果:LFP信号经Relief F-PLS特征提取后,五组鸽子实测数据的解码正确率分别达到95.00%、80.00%、95.00%、92.50%、85.71%,高于单独用Relief F或PLS算法的解码正确率,而且所提取的特征数更少。说明Relief F-PLS方法结合了Relief F和PLS的优点,提高了解码正确率;而且提取的特征数更少,有效地去除原始特征中的干扰特征和冗余特征,验证了该算法的有效性,为相关研究探索了一条新路径。

The local field potential(LFP)signals recorded by the microelectrode array contained a lot of noise and redundant information,and the signal feature dimension was high,which affected the accuracy of decoding.The method that combining ReliefF algorithm and Partial Least Squares(PLS)to decode the animal s turning behavior was applied.The plus maze goal-directed experiment was designed,to collect LFP neural signals from NCL of pigeons.The characteristics of the LFP signal was extracted,and then used the ReliefF to assign the corresponding weight value to each feature,chosen the corresponding feature according to the weight of the feature.Next,the principal component was extracted from the feature subset by PLS,used the support vector(SVM)to decode the pigeon s turning motion and compared it with the ReliefF and the PLS.After the LFP signals was extracted by ReliefF-PLS,the decoding accuracy of the five pigeons real data sets were 95.00%,80.00%,95.00%,92.50%and 85.71%respectively.Compared with ReliefF and PLS,ReliefF-PLS not only improved the decoding accuracy,but also extracted fewer features.The ReliefF-PLS method that combined the advantages of ReliefF and PLS,improved the decoding accuracy greatly,and extracted fewer features.It removed the interference characteristics and redundant features effectively,verified the validity of the algorithm,and explored a new path is for the related research.