基于深度学习的航空器异常飞行状态识别

Abnormal Flight States of Aircraft Identification Based on Deep Learning Method

摘要:

飞行设备快速存取记录仪（Quick Access Recorder, 以下简称QAR）保留了原始航班各类重要飞行参数在内的航行信息，使研究分析航空器实时状况和保障飞行质量成为可能。针对QAR数据高维大样本的特点，在如今大数据背景下，除了传统机理建模分析航空器飞行状态外，采用深度学习的方式建立基于数据驱动的航空器飞行状态识别模型，理论与实用意义兼具。通过对真实QAR飞行数据的研究，开发了基于深度稀疏受限玻尔兹曼机的异常飞行状态识别程序。首先利用小波降噪技术对原始飞行数据进行预处理清洗，在一系列典型飞行参数上提取经典时域特征以及小波奇异熵等信息熵特征构成特征集。在此基础上，分别利用经典的线性主元分析技术和深度稀疏玻尔兹曼机对特征集进行有效降维，最后采用四折交叉验证方式，通过高斯过程分类器实现对飞行状态的辨识。实验结果显示，基于深度受限玻尔兹曼机〖CD*2〗高斯过程分类的飞行状态识别具有较高分类准确性。

Abstract:

The quick access recorder(QAR) retains the navigational information of all important flight parameters of the original flight, making it possible to analyze aircraft realtime conditions and ensure flight quality. According to the characteristics of highdimensional largescale QAR data, under the background of Big Data, different from the traditional mechanism modeling and analysis of aircraft flight state, the paper uses deep learning to establish a datadriven aircraft flight state recognition model. Based on the study of real QAR flight data, an abnormal flight state recognition program based on the Sparse Restricted Boltzmann Machine is developed. First of all, we use the wavelet denoising translation method to preprocess the original flight data. And then, we select a series of typical flight parameters, extract the classical timedomain features of these parameters and the mixed entropy feature like Wavelet Singular Entropy to form the feature set. Then we use the Principal Component Analysis technique and the Sparse Restricted Boltzmann Machine to effectively reduce the feature set. Finally, we use fourfold cross validation method. We put the training set into the Gaussian process classifier as a last step. The experimental results show that the flight state recognition based on the Sparse Restricted Boltzmann MachineGaussian process classification has high classification accuracy.