基于深度学习的飞行器智能故障诊断方法

Deep Learning Theory with Application in Intelligent Fault Diagnosis of Aircraft

高温度、快转速、重载荷、大扰动和强冲击的复杂运行工况使得飞行器关键机械部件不可避免地发生故障。飞行器的关键机械部件故障特征往往具有微弱性、非线性、耦合性、不确定性以及因果关系复杂等特点。以“先进信号处理技术+特征提取及选择”为框架的传统智能方法难以有效承担飞行器故障检测任务,深度学习作为智能故障诊断领域中的新起之秀,能自主挖掘隐藏于原始数据中的代表性诊断信息,直接建立原始数据与运行状态间的精确映射联系,在很大程度上摆脱了对人工特征设计与工程诊断经验的依赖。介绍了深度置信网络、卷积神经网络、深度自动编码机和循环神经网这四种主流深度学习模型的基本原理,总结了深度学习在故障诊断领域中最新研究现状,描述了基于四种深度学习模型的故障诊断思路,并依次实现了其在机械部件智能诊断和预测中的应用。试验结果表明深度学习方法能有效建立监测数据与关键机械部件健康状态间的精确映射联系,实现准确的故障诊断和预测。

The key mechanical parts of aircraft will inevitably generate multifarious faults due to the severe working conditions with high temperature, fast speed, heavy load, large disturbance and strong impact. The faults of aircraft key parts often show some characteristics such as weakness, randomness, coupling, diversity, uncertainty and so on. Therefore, using the traditional methods based on advanced signal processing techniques, feature extraction and feature selection, it is a great challenge to diagnose the various faults of aircraft key parts. As a very promising tool in the field of intelligent fault diagnosis, deep learning can largely get rid of the dependence on manual feature design and engineering diagnosis experience, which can directly establish accurate mapping relationships between the raw data and various operation conditions. The basic theory of four kinds of popular deep learning models are briefly introduced, including deep belief network, convolutional neural network, deep auto-encoder and recurrent neural network. The recent research work of deep learning on fault diagnosis is summarized. These four deep models are respectively used for intelligent fault diagnosis and prognosis of mechanical parts. The results confirm that deep learning models are able to automatically capture the representative information from the massive measured data through multiple feature transformations, and directly establish the accurate mapping relationships between the raw data and various operation conditions.