运动想象脑信号的深度置信网络分类优化
Deep Belief Networks Classification Optimization of Motor Imagination Brain Signals
-
摘要: 关于脑机接口（BCI）系统中的运动想象（MI）脑信号的特征提取一直是一个难题。相较于SSVEP、AEP和P300等其他BCI模式，MI的分类准确率相对较低，缺乏有效的识别方案。本文提出了一种结合深度置信网络（Deep Belief Network，DBN）和麻雀搜索算法（Sparrow search algorithm，SSA）的特征提取和分类识别算法SSA-DBN。融合SSA与DBN的优势，可以在保持较低计算复杂度的同时，提高特征提取和分类识别的准确率。本研究首先利用完全自适应噪声集合经验模态分解（CEEMDAN）方法提取信号的固有模态函数（IMF）特征。然后，将筛选出的适合分类识别的IMF分量与希尔伯特黄变换（HHT）方法相结合，提取出不同导联时频信号的特征空间向量，并进行叠加平均。最后，将特征向量输入到SSA-DBN算法进行分类处理。为确保公平性，在验证算法性能时，选取了具有代表性的BCI Competition IV Dataset 2a数据集，同时对比了其他算法的表现，并详细说明了调参方法。为了避免过拟合问题，可以考虑使用更大规模的数据集进行测试，如PhysioNet或BCI2000等公开数据集库。在实验结果展示时，除了展示SSA-DBN算法的性能，还展示了其他算法在相同数据集和评价指标下的性能，以便进行公平的对比。结果显示，SSA-DBN算法在四分类任务中的平均准确率最高达到了86.35%，较其他方法有显著提升，为提高MI信号脑机接口的系统性能提供了一种新的有效途径。Abstract: The extraction of motor imagery （MI） features in brain computer interface （BCI） systems has always been a challenge. Compared to other BCI modes such as SSVEP， AEP， and P300， the classification accuracy of MI is relatively low and there is a lack of effective recognition schemes. This article proposes a feature extraction and classification recognition algorithm SSA-DBN that combines Deep Belief Network （DBN） and Sparrow search algorithm （SSA）. Integrating the advantages of SSA and DBN can improve the accuracy of feature extraction and classification recognition while maintaining low computational complexity. This study first utilizes the fully adaptive noise set empirical mode decomposition （CEEMDAN） method to extract the intrinsic mode function （IMF） features of signals. Then， the selected IMF components suitable for classification and recognition are combined with the Hilbert Huang Transform （HHT） method to extract the feature space vectors of different lead time-frequency signals and perform superposition averaging. Finally， input the feature vectors into the SSA-DBN algorithm for classification processing. To ensure fairness， a representative BCI Competition IV dataset 2a was selected during the validation of algorithm performance. The performance of other algorithms was compared， and the parameter tuning method was explained in detail. In order to avoid the overfitting problem， you can consider using a larger data set for testing， such as PhysioNet or BCI2000 and other public data set libraries. When presenting the experimental results， in addition to demonstrating the performance of the SSA-DBN algorithm， the performance of other algorithms under the same dataset and evaluation indicators was also demonstrated for fair comparison. The results show that the SSA-DBN algorithm has the highest average accuracy of 86.35% in four classification tasks， which is significantly improved compared to other methods. It provides a new and effective way to improve the system performance of MI signal brain computer interface.
下载: