航空学报 > 2023, Vol. 44 Issue (S1): 727615-727615   doi: 10.7527/S1000-6893.2022.27615

拟VGG16网络的航空传感器故障检测分类

李忠智, 马金毅, 艾剑良, 董一群()   

  1. 复旦大学 航空航天系,上海 200433
  • 收稿日期:2022-06-01 修回日期:2022-06-15 接受日期:2022-06-22 出版日期:2023-06-25 发布日期:2022-06-24
  • 通讯作者: 董一群 E-mail:yiqundong@fudan.edu.cn
  • 基金资助:
    上海市青年科技英才扬帆计划(20YF1402500);上海市自然科学基金(22ZR1404500)

Fault detection and classification of aerospace sensors using deep neural networks finetuned from VGG16

Zhongzhi LI, Jinyi MA, Jianliang AI, Yiqun DONG()   

  1. Department of Aeronautics and Astronautics,Fudan University,Shanghai 200433,China
  • Received:2022-06-01 Revised:2022-06-15 Accepted:2022-06-22 Online:2023-06-25 Published:2022-06-24
  • Contact: Yiqun DONG E-mail:yiqundong@fudan.edu.cn
  • Supported by:
    Shanghai Sailing Program(20YF1402500);Natural Science Foundation of Shanghai(22ZR1404500)

摘要:

参考计算机视觉等领域的研究与应用进展,提出了拟图智能化故障诊断概念;拟照VGG16图像分类网络,提出了一种航空传感器故障检测与分类方法。首先,基于仿真、实飞等手段建立了航空传感器故障飞行数据库;该数据库包含4型大型客机、通航飞机在5种飞行状态的飞行数据,并可有效模拟气动数据、惯性测量单元等传感器的故障。其次,提出将航空器气动数据、惯性测量单元等传感器的测量数据堆叠成灰度图像数据格式;该图像保留了传感器测量数据的时间、空间耦合特征,将传感器故障检测与分类转换成为图像上的异常区域检测与分类问题。再次,提出了一种数据增强方法,将堆叠形成的传感器测量数据图像的维度增强为VGG16图像分类网络输入维度,并基于预训练的VGG16图像分类网络,采用微调优化网络模型,最终得到了拟图智能化航空传感器故障检测与分类深度神经网络。在多个航空器数据集上的实验结果表明,网络的平均测试准确度可以达到97.6%。最后,参考计算机视觉领域的深度神经网络可解释性分析方法,基于类激活映射图(CAM)对本文发展的传感器故障检测与分类网络进行了分析,初步阐明了网络内部各层卷积核节点特征提取运算的机理,提升了该网络故障检测与分类性能的可信度。

关键词: 航空传感器, 故障检测与分类, 拟图智能化, 深度神经网络, 可解释性分析

Abstract:

Following the research advances in machine vision, a concept defined as imagification-based intelligence has been proposed. A fault detection and classification method of fault detection and classification was developed using a deep neural network finetuned from VGG16, which originally was designed for image detection and classification problems. Firstly, a database corresponding to 4 aerospace vehicles and 5 flight conditions were constructed via both flight simulations and real-vehicle tests. Faults of aerospace sensors and inertial measurement unit are included in the database. Secondly, we proposed to stack the measurement data of aerospace sensors into a grayscale image, which was used to transfer the fault detection and classification problem into abnormal regions detection problem on the image. Thirdly, the size of the grayscale image was augmented to match the input size of VGG16, and the fault detection neural network was developed via finetuning the VGG16 directly. The experimental results on several aircraft show that the average test accuracy of the network reaches 97.6%. Following the explainability analysis methods in machine vision, Class Activation Mapping (CAM) results of different layers of the fault detection neural network were plotted, which solidifies the performance of the proposed neural network.

Key words: aerospace sensors, fault detection and classification, imagification-based intelligence, deep neural networks, explainability analysis

中图分类号: