面向飞行器电性能动态特性的信号特征识别与分类方法研究

Recently the analysis of aircraft surveillance signal data is a hot issue in the field of aircraft fault detection and health management.In fact it is a typical pattern recognition and classification problem. With the development of testing technology, it is necessary to solve such problems as large amount of test data, high feature dimension and low efficiency of traditional identification methods. In this project, the deep learning network model is adopted to identify and classify the electrical signals. Firstly the establishment of standard sample database using fuzzy clustering and expert knowledge will be proposed. Secondly, the RBM deep belief network construction using deep neuron model will put forward the classification method of migration method and fuzzy tuning optimization of online ensemble learning. With the process of supervised fine learning, the dynamic response of equipment electrical performance which is from shallow learning to deep learning has been completed. It also reveals the characteristics of electrical signal fault events, so that the system can identify the normal events and abnormal events accurately, and reduce the rate of manual intervention and false positives as much as possible. Finally, this methods of evaluating the similarity of electrical signals and the algorithms for verification and evaluation of learning algorithms are summarized. It can provide a reliable analysis method for aircraft testing, and ensure the development of aircraft. It can provide technical support for intelligent fault diagnosis and health management technology.

飞行器监测信号数据分析问题是近期飞行器故障检测与健康管理领域的研究热点问题，也是一个典型的模式识别与分类问题。随着测试技术的提高，需要解决测试数据量大，特征维数高，传统识别方法效率低的问题。本项目采用深度学习网络模型的方法对电信号进行特征识别与分类。首先提出模糊聚类与专家知识结合的标准样本库建立方法，采用受限玻尔兹曼机深度置信网络构建深层神经元模型，进一步提出迁移学习与模糊微调优化的方法和在线集成学习的分类方法。伴随着从无监督初步学习到有监督精细学习的过程，完成对设备电性能动态响应信号数据从浅层学习到深层学习的飞跃 。并揭示出电信号故障事件的特征，使系统对正常事件与异常事件进行精确的识别，尽可能的减少人工干预和误报率。最后，总结出评价电信号波形相似度和学习算法验证与评估的方法。可为飞行器测试提供可靠的分析方法，保证飞行器研制工作顺利进行，为智能化故障诊断与健康管理技术提供技术支撑。

本课题针对飞行器信号监测中的分类与识别问题展开研究，主要完成了以下工作：①基于飞行器供电系统负载电信号数据集，从异常信号识别与故障诊断方向展开研究。采用基于数据驱动的诊断方法，对随机森林算法进行研究并加以改进，仿真与实测实验验证了该方法在准确度上出极大的优越性；②针对飞行器供电系统故障诊断问题，本课题还展开了深度神经网络分类算法研究。通过采用多尺度残差卷积神经网络使分类器的分类性能得到了较大提高；此外，针对小样本信号数据集中，借助于迁移学习，有效解决了小样本问题对算法精度提出的巨大挑战；采用了深度聚类变分自编码器算法，有效提升了数据信号的聚类精度；最后、采用了基于深度循环神经网络的预测方法进行了故障预测实验；③本课题针对基于深度迁移学习的飞行器热控系统故障诊断与预测算法进行了研究。该算法分别包括基于联合多阶域分布差异的深度学习网络模型、基于卷积降噪自编码器关系网络的少样本学习算法以及基于WaveNet卷积神经网络的时间序列预测模型，有效解决了飞行器热控系统的故障诊断与预测问题，相较已有算法，本课题所采用方法大大提高了算法精度，具有重要的实际工程价值。综上所述，针对航空航天飞行器所搭载的复杂电子系统的故障诊断与健康管理技术，本课题分别从传统机器学习算法、深度学习算法以及针对小样本问题的迁移学习算法等多个方面深入研究，通过在多个数据集上进行验证，结果表明了所采用的方法在故障诊断及预测问题上的准确性和优越性，具有较强的工程应用价值。