遗传算法与修正的自适应矩估计优化循环神经网络的心音分类方法

Heart sound classification method combining GA with rectified Adam for RNN optimization

针对传统的循环神经网络(RNN)在识别分类心音信号方面具有梯度爆炸、梯度消失和短期记忆的问题,该文提出了无需心音分段的结合遗传算法(GA)与修正的自适应矩估计(RAdam)优化RNN的心音分类模型。该模型的优势是将GA和RAdam优化器以串联的方式融合到RNN中,以达到改进RNN的作用。首先,利用GA的选择、变异和遗传操作,优化RNN的输入层节点数,获取心音特征向量的最优个体的初始解。其次,根据最优个体中的权重、偏置矩阵,赋予模型初始权值和阈值,获得初始权重最优解,整个模型共享参数。最后,联合改进的学习率自适应优化算法,优化RNN模型。结果表明,结合经典的梅尔(Mel)倒频谱系数方法提取心音信号的特征向量,心音信号分类准确率达到90.29%,相比于未优化的RNN模型,准确率提高了17.79%。

Aiming at the problems of gradient explosion,gradient disappearance,and short-term memory in traditional recurrent neural networks(RNN)for identifying and classifying heart sound signals,a heart sound classification model is proposed combining genetic algorithm(GA)and rectified Adam(RAdam)optimized RNN without heart sound segmentation.The advantage of this model is that it integrates GA and RAdam optimizer in series into a RNN to improve its performance.Firstly,the selection,mutation and genetic operation of the GA are used to optimize the number of nodes in the input layer of the RNN,and the initial solution of the optimal individual of the heart sound feature vector is obtained.Secondly,according to the weight and bias matrix in the optimal individual,the initial weight and threshold are assigned to the model,and the optimal solution of the initial weight is obtained,and the entire model shares parameters.Finally,combined with the improved learning rate adaptive optimization algorithm,the RNN model is optimized.The results show that combining the classical Mel-frequency cepstral coefficient method to extract the eigenvectors of the heart sound signal,the classification accuracy of the heart sound signals reaches 90.29%,which is 17.79 percentage points higher than that of the unoptimized RNN model.