基于GA-BP神经网络模型的石墨烯发声器研究

Study of Graphene Loudspeaker Based on GA-BP Neural Network Model

为了解决传统石墨烯发声器物理模型计算量大、精度不高的问题,提出了一种基于GA遗传算法的BP神经网络模型,该算法模型具有更高的精度和适应性。首先,介绍了石墨烯热声发声原理以及实验设计,建立了GA-BP神经网络模型。其次,对模型的参数进行了调节,并对比了基于dropout、基于正则化、基于正则化和GA遗传算法的三种神经网络模型。随后在模型中输入石墨烯发声器的正弦激励幅值、频率以及测量距离,使用GA遗传算法对隐藏层的权值和偏置进行全局寻优,将寻优结果代入BP神经网络,最终预测出给定条件下的声压级。结果表明,在均使用正则化的条件下,BP神经网络预测准确度为98.05%,均方差为0.23;GA-BP神经网络预测准确度达到98.62%,均方差仅为0.14。优化后精准度提高了0.57%,均方差降低了41.36%,展现出更加优异的准确性和适应性。该研究为预测多类特征传感器的非线性输出结果提供了一种高精准度、高适应性的方案。

In order to solve the problem that the traditional physical model of graphene sound generators requires a large amount of calcu lation and has low accuracy,a BP neural network model based on GA genetic algorithm is proposed,which has higher accuracy and adapta bility.Firstly,the principle and the experimental design of graphene thermoacoustic are introduced and GA-BP neural network model is es tablished.Secondly,the parameters of the model are adjusted,and three different neural network models based on dropout,regularization,both regularization and GA genetic algorithms are compared.Thirdly,sinusoidal excitation amplitude,frequency and measurement distance of the graphene loudspeaker are inputted and GA genetic algorithm is used to globally optimize the weight and bias of the hidden layer.Fi nally the optimization results are brought into the BP neural network to predict the sound pressure level.Under the condition that regulari zation is used,the prediction accuracy of BP neural network is 98.05%and the mean square error is 0.23.At the same time,the prediction accuracy of GA-BP neural network reaches 98.62%,and the mean square error is only 0.14.After optimization,the accuracy is increased by 0.57%and the mean square error is reduced by 41.36%,showing better accuracy and adaptability.This work provides a high precision and high adaptability scheme for predicting the nonlinear output of multi-class feature sensors.