汽车工程 ›› 2025, Vol. 47 ›› Issue (8): 1596-1606.doi: 10.19562/j.chinasae.qcgc.2025.08.015

• • 上一篇    

基于GAN-FCNN数据增强的特种车辆声品质预测研究

钱堃(),杜习康,王言夫,段继英,刘珂,谭璟,沈政华,赵剑   

  1. 大连理工大学机械工程学院,大连 116081
  • 收稿日期:2025-01-07 修回日期:2025-02-11 出版日期:2025-08-25 发布日期:2025-08-18
  • 通讯作者: 钱堃 E-mail:qiankun_nvh@163.com
  • 基金资助:
    中央高校基本科研业务费专项资金项目(DUT22RC(3)002)和中国博士后科学基金(2019M650657)

Research on Sound Quality Prediction of Special Vehicles Enhanced with GAN-FCNN Data

Kun Qian(),Xikang Du,Yanfu Wang,Jiying Duan,Ke Liu,Jing Tan,Zhenghua Shen,Jian Zhao   

  1. School of Mechanical Engineering,Dalian University of Technology,Dalian 116081
  • Received:2025-01-07 Revised:2025-02-11 Online:2025-08-25 Published:2025-08-18
  • Contact: Kun Qian E-mail:qiankun_nvh@163.com

摘要:

针对特种车辆声品质预测时,面临采集噪声样本成本较高、样本处理后只能获得少量样本集、缺乏充足的噪声样本,从而影响各类预测模型训练时的模型精度问题,本文建立GAN-FCNN网络,使用4层全连接层构建生成器与判别器进行对抗训练,生成伪样本集。将增强样本集分别带入到LASSO线性回归模型和RF、BP及其PSO优化模型进行回归预测,通过验证,模型预测精度与性能均得到提升。相比于传统的过采样算法,GAN-FCNN网络有更高的准确率,更适用于特种车辆声品质预测模型建立时进行样本扩充。

关键词: 数据增强, 特种车辆, 声品质, 对抗神经网络, 过采样

Abstract:

For the special vehicle sound quality prediction, the cost of collecting noise samples is high, and only a small number of sample sets can be obtained after sample processing, lacking sufficient noise samples, which affects the model accuracy during the training of various prediction models. In this paper, a GAN-FCNN network is established, and a four-layer fully connected layer is used to construct a generator and discriminator for adversarial training, and a pseudo-sample set is generated. The enhanced sample set is introduced into the LASSO linear regression model and the RF, BP and PSO optimization models respectively for regression prediction. Through verification, the prediction accuracy and performance of the models are improved. Compared with the traditional oversampling algorithm, the GAN-FCNN network has higher accuracy, which is more suitable for sample expansion in the establishment of special vehicle sound quality prediction model.

Key words: data enhancement, special vehicle, sound quality, GAN, oversampling