基于XGBoost模型的碳基扬声器的性能

Performance of Carbon-Based Speakers Based on XGBoost Model

为了解决研究碳基扬声器性能的物理模型复杂、计算量大、忽略一些影响因素导致理论结果不够准确的问题,研究了碳基扬声器的深度学习模型。介绍了碳基扬声器发声原理,并将碳基薄膜与其衬底的密度、热导率、比热容,以及热声效应过程中的输入交流电压有效值、输入交流电频率和测量距离作为输入特征,碳基扬声器的声压级作为输出特征,建立了XGBoost模型和基于dropout的反向传播(BP)神经网络模型。对这两个模型进行优化,发现当XGBoost模型的单个基学习器的深度为3、基学习器数量为523时,模型精度达到最高;当BP神经网络模型隐藏层数量为2、神经元数量为10时,模型精度达到最高。使用还原氧化石墨烯(rGO)、多壁碳纳米管(MWCNT)、激光划刻石墨烯(LSG)薄膜制备碳基扬声器,从表面微结构上观察到rGO薄膜的晶体质量最好;设计实验对三种碳基扬声器进行声压级测试,发现rGO薄膜制备的扬声器的声压级也最大。最后,将测试得到的数据划分出预测集,代入两种深度学习模型中分别进行求解,并将结果与解析热声模型结果进行对比,发现XGBoost模型在预测LSG薄膜扬声器输出声压级时,预测值和测量值的相对误差不超过2%,其预测精度和稳定性比BP神经网络模型和解析热声模型都要好,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度、高适应性的方案。

To solve the problems of physical model for studying the performances of carbon-based speakers,such as complicated model,large calculation,and inaccurate theoretical results due to some influencing factors being ignored,the deep learning models of carbon-based speakers were studied.The sound principle of carbon-based speakers was introduced.The density,thermal conductivity,and specific heat capacity of the carbon films and their substrates,and root mean square value of input AC voltage,input AC frequency and measured distance during thermoacoustic effects were taken as input characteristics.The sound pressure level of the carbon-based speaker was taken as the output characteristic.Then XGBoost model and back propagation(BP)neural network model based on dropout were established.The two models were optimized.It is found that the accuracy of XGBoost model reaches the highest when the depth of a single base learner of XGBoost model is 3 and the number of base learners is 523;and the accuracy of BP neural network model reaches the highest when the number of hidden layers in BP neural network model is 2 and the number of neurons is 10.Then,reduced graphene oxide(rGO),multiwalled carbon nanotube(MWCNT)and laser-scribed graphene(LSG)films were used to prepare carbon-based speakers.It is found that the crystal quality of rGO film is the best in the surface microstructure.And an experiment was designed to test the sound pressure level of three kinds of carbon-based speakers.It is found that the sound pressure level of the speaker prepared by rGO films is also the largest.Finally,the measured data were divided into prediction sets and substituted into two deep learning models for solving them respectively.The results were compared with the analytic thermoacoustic model.It is found that when predicting the output sound pressure level of LSG film speaker,the relative error between the predicted values of XGBoost model and measured values is less than 2%.The prediction accuracy and stability of XGBoost model are better than those of BP neural network model and analytic thermoacoustic model,showing more excellent accuracy and adaptability.The study provides a high-precision and high-adaptability scheme for predicting nonlinear output results of multi-class feature sensors.