基于IKPCA-GA-BP算法的心电情绪识别研究

Study on ECG emotion recognition based on IKPCA-GA-BP algorithm

目的 为了提高使用心电信号特征参数进行情绪识别的准确率与效率,提出一种基于改进核主成分分析的遗传算法:优化遗传算法反向传播(improved kernel principal component analysis-genetic algorithm-back propagation, IKPCA-GA-BP)神经网络。方法 首先,以心电传感器记录的数据为试验样本,利用自适应选取γ值的改进核主成分分析算法,对二进样条小波变换提取的多组特征进行数据降维和数据重构,得到综合变量;其次,建立反向传播神经网络模型,并利用遗传算法优化网络的初始权值和偏置值;最后,通过改变模型训练样本与测试样本比例,对比分析IKPCA-GA-BP算法与传统识别算法的情绪分类效果。结果 该算法在保证准确率达到96%的前提下,可在1 s左右识别相关情绪。另外,对于悲伤情绪的识别,大多数模型表现得并不理想,而IKPCA-GA-BP算法可实现接近100%的准确率。结论 心电信号中,P波、QRS波群和T波包含许多有助于情绪识别的信息(例如R-R间期、P波振幅等),但这些信息并不能直接用于试验分析,需要通过有效的组合与处理才能发挥最大作用。此外,在高兴、轻松、悲伤和恐惧这四种情绪中,多数识别算法通常较难准确辨别悲伤情绪。

Objective To improve the accuracy and efficiency of emotion recognition by using characteristic parameters of ECG signals,a genetic algorithm called improved kernel principal component analysis-genetic algorithm-back propagation(IKPCA-GA-BP)neural network based on improved kernel principal component analysis is proposed.Methods First,taking the data recorded by ECG sensors as test samples,the improved kernel principal component analysis algorithm with adaptive selection ofγvalues was used to perform data dimension reduction and data reconstruction on multiple groups of features,which had been extracted by the binary spline wavelet transform,to obtain comprehensive variables.Secondly,the back propagation neural network model was established,and the genetic algorithm was used to optimize the initial weights and bias values of the network.Finally,by changing the proportion of model training samples to test samples,the effects of emotion classification separately based on IKPCA-GA-BP algorithm and traditional recognition algorithm were comparatively analyzed. Results By this algorithm, related emotions could be recognized in about 1 s on the premise of an ensured accuracy rate of 96%. In addition, most models did not perform well in identifying sadness emotion, however, IKPCA-GA-BP algorithm obtained an accuracy rate of nearly 100%. Conclusion In ECG signals, P-wave, QRS complex and T-wave contain a lot of information that is helpful for emotion recognition (for example, R-R interval, P-wave amplitude, etc.). However, this information can not be directly used for experimental analysis, and requires effective combination and processing to maximize its efficacy. Furthermore, among four emotions: happiness, relaxation, sadness and fear, accurate recognition of sadness often challenges most recognition algorithms.