现代电子技术

传统的脑电时频信号分类方法仅关注脑电信号的局部特性，没有充分探究不同脑区之间的关联特征，无法全面捕捉大脑活动的前后时间关联，存在对大脑理解不够全面、分类准确率不高等问题。文中基于加州大学尔湾分校提供的酗酒脑电数据集，运用相位锁定值（PLV）构建功能脑网络，研究了α、β、γ、θ四个子频段和全频段的EEG脑网络拓扑特征，探究了酗酒者与健康对照者EEG脑网络的拓扑属性差异。同时，提出一种基于Fisher特征筛选和改良麻雀算法优化的BiLSTM分类算法（Fisher⁃ISSA⁃BiLSTM），运用Fisher准则进一步筛选了脑网络特征，通过双向长短时记忆网络充分分析了脑电信号前后时序的关联性，运用改良的麻雀搜索算法（ISSA）优化BiLSTM的超参数。在原麻雀算法的基础上，引入Sobol映射的初始化方式，提升了麻雀种群的分布质量；加入搜索因子，避免算法过早地陷入局部最优；引入自适应方向因子dti, j，优化了麻雀跟随者位置的更新方向。相比其他超参数优化算法，文中算法分类所需时间减少了约4%~5%，分类准确率达92.6%，相比传统的LSTM分类算法提升了约20%，对于运用脑电信号识别酗酒患者具有一定的实际意义。

The traditional time⁃frequency EEG (electroencephalogram) classification methods predominantly focus on local characteristics of EEG signals. They neglect inter⁃regional brain connectivity, and fail to fully capture the temporal correlation of brain activities, which result in incomplete understanding of the brain and low classification accuracy rate. On the basis of the UCI alcoholism EEG dataset, this study constructed functional brain networks by phase⁃locking value (PLV), investigated the EEG brain network topological characteristics across sub⁃bands (α, β, γ, and θ) and full⁃band, and explored the topological attribute differences of EEG brain networks between alcoholics and healthy controls. A classification algorithm, which integrates Fisher feature selection with bidirectional long short⁃term memory (BiLSTM) network which is optimized by improved sparrow search algorithm (ISSA), is proposed, and it is named Fisher⁃ISSA⁃BiLSTM. The Fisher criterion is used to further screen the characteristics of brain networks, the BiLSTM is used to fully analyze the correlation of the timing sequence of the EEG signals, and the ISSA is used to optimize the superparameter of BiLSTM. On the basis of the SSA, the initialization method of Sobol mapping is introduced to enhance the distribution quality of the sparrow population. Additionally, a search factor is incorporated to prevent the algorithm from falling into local optima prematurely. An adaptive direction factor dti, j is also introduced to optimize the update direction of the location of the sparrow followers. In comparison with the other hyperparameter optimization algorithms, the classification duration of the proposed algorithm is reduced by approximately 4% to 5%, and its classification accuracy rate reaches 92.6%, which is about 20% higher than that of the traditional LSTM classification algorithm. To sum up, the proposed algorithm has a certain practical significance for identifying alcoholics by using EEG signals.