迁移学习在变工况方向舵故障诊断中的应用

Application of Transfer Learning in Rudder Fault Diagnosis Under Variable Operating Conditions

为解决飞行器方向舵在复杂多变的工况条件下的故障诊断准确性问题,提出了一种结合卷积神经网络(CNN)和基于网络的深度迁移学习(NDTL)的NDTL-CNN故障诊断方法。首先,搭建了飞行器方向舵的故障仿真模型,采集不同工况条件、健康状态下的多维传感器数据;然后,设计了CNN,其自适应地从定工况数据中深度提取特征,能够有效捕获方向舵的故障特征信号;最后,对定工况下的预训练CNN进行模型微调,将其迁移到变工况数据中进行故障诊断。实验结果表明:所提方法在短时间内将变工况下CNN的诊断精度提高了15%,最终NDTL-CNN的诊断精度为97.7%,达到了在复杂多变的工况条件下精确辨识方向舵的健康状态。

In order to solve the low fault diagnosis accuracy of aircraft rudders under complex and dynamically changing operating conditions,an NDTL-CNN fault diagnosis method that combined convolutional neural network(CNN)with network-based deep transfer learning(NDTL)is proposed.Firstly,a simulation model of aircraft rudder fault is established to collect multi-dimensional sensor data under different operating conditions and health states;Then,a CNN is designed to adaptively extract deep features from the fixed operating conditions data,which can effectively capture the fault feature signals of the rudder;Finally,the pre-trained CNN under the fixed operating conditions is fine-tuned and transferred to variable operating conditions for fault diagnosis.The experimental results show that the proposed method improves the diagnostic accuracy of the CNN under variable operating conditions by 15%in a short time and the final diagnostic accuracy of the NDTL-CNN reaches 97.7%,which is capable of accurately recognizing the rudder′s health state under complex and dynamically changing operating conditions.