针对BCI研究中样本采集代价较大,如何充分利用未标记的样本来提升识别性能的问题,本研究将仿生模式识别算法(BPR)与增量半监督学习算法结合,以Bagging算法框架为基础,提出了一种新的增量半监督的仿生模式识别算法(BPR-ISSL).以脑机接口国际大赛公布的数据集对该算法进行了离线分析以及模拟在线的实验,并且使用作者在线采集的实际脑电数据进行了实际在线实验,比较分析了BPR-ISSL与作者之前提出的改进仿生模式识别算法,以及BPR-ISSL与增量半监督SVM、增量半监督BP的识别性能.实验结果表明:在训练样本较充足的情况下,BPR-ISSL识别准确率优于其它3种算法,平均准确率要高出3％左右；并且通过对标准差的计算,其在抗过学习和稳定性上也体现出了明显的优势.%In order to solve the problem that how to make full use of the unlabeled samples to promote recognition accuracy due to high cost acquisition of data in BCI research, biomimetic pattern recognition (BPR) was combined with semi-supervised learning algorithm on the basis of the Bagging algorithm framework, and a novel Incremental Semi-Supervised Biomimetic Pattern Recognition (BPR-ISSL) was proposed in this work. We conducted off-line analysis and simulative on-line experiments on the datasets from previous brain-computer interface competitions and real on-line experiments on actual on-line EEG data to test the accuracy and efficiency of the proposed algorithm. Comparing with those of improved BPR, SVM-ISSL and BP-ISSL, the BPR-ISSL showed that higher accuracy (3% higher in average) in reference to those of the other three algorithms, especially with the sufficient samples. Furthermore, it has a better anti-over-learning ability and stability according to standard deviation.
展开▼
