首页 >

孪生网络在航天产品电性能测试方面的应用-信息融合大会增刊

王宇斐,刘骁佳,刘欢,曹立俊,罗志强   

  1. 上海航天精密机械研究所
  • 收稿日期:2022-02-17 修回日期:2022-03-25 出版日期:2022-03-30 发布日期:2022-03-30
  • 通讯作者: 王宇斐
  • 基金资助:
    基于xx装配车间的大数据决策优化技术

Application of Siamese neural network in electrical performance test of aerospace products

  • Received:2022-02-17 Revised:2022-03-25 Online:2022-03-30 Published:2022-03-30
  • Contact: Yu-Fei WANG

摘要: 航天产品电性能测试是在外加激励的情况下,测试航天产品不同部位的信号反馈情况,从而确定各部件的质量情况,是检验航天产品质量的重要手段之一。但是,测试得到的信号通道多、时序长、十分复杂,依赖传统信号处理方法难度较高,目前处理方式采用人工统计,效率低,易出错,亟需借助深度神经网络的方法来提高效率和准确率。本文将孪生神经网络引入航天产品电性能测试领域,将数据信息解析为图像信息,转化为图像识别问题;对几种主干网络模型进行对比分析,最终选择ResNet50作为主干网络。实验结果表明,在相同实验条件下,研究的算法能够有效地识别78个通道信号的波形信息,识别准确率达到90.385%,提高航天产品复杂测试信号的识别能力以及智能化水平。

关键词: 电性能测试, 孪生神经网络, 图像识别, 复杂信号, 主干网络

Abstract: Vibration synthesis data is one of the important means to inspect the quality of aerospace products by collecting data status of each functional module of aerospace assembly products to describe the quality of different parts of the products. In view of the low efficiency and insufficient accuracy of traditional manual statistical methods for complex vibration sig-nal recognition, an improved siamese neural network vibration signal recognition algorithm - VggVSNet is proposed. Due to the characteristics of multiple channels, long this paper parses the data information into image information and trans-forms it into an image recognition problem. In order to overcome the slow learning speed or even difficult learning of VGG16, the backbone network of twin neural networks, the generalization ability is improved by reducing the number of convolutional layers and adding Batch Normalization layer to avoid gradient disappearance. The experimental results show that under the same experimental conditions, the studied algorithm can effectively recognize the waveform infor-mation of 78 channels of vibration signals, and the recognition accuracy reaches 96.154%, which has obvious ad-vantages over the classical VGG16 network and improves the recognition ability of complex vibration signals as well as the intelligence level.

Key words: vibration signal, siamese neural network, image recognition, Complex signal, backbone

中图分类号: