基于改进生成对抗网络的飞参数据异常检测方法

Flight parameter data anomaly detection method based on improved generative adversarial network

针对民用飞行器安全性、可靠性要求严苛,实际民航运营中飞行参数的异常样本稀少,整体样本不平衡且缺少标注的问题,研究深度学习与生成对抗网络技术,提出基于改进生成对抗网络的飞参数据飞行级异常检测方法.该方法不依赖样本数量与标签,实现无监督学习的检测方法.针对飞参数据,输入正常数据样本,应用易收敛的WGAN-GP改进型生成对抗网络模型,模拟生成正常数据样本,计算输入数据与模拟正常数据的巴氏距离,实现对异常数据的检测.通过美国国家航空航天局模拟飞参数据的人工合成数据集以及真实运营环境下采集的快速存取记录器数据构建的飞参数据集,开展试验验证.结果表明,与常用无监督模型相比,提出方法在部分异常检测性能指标上有显著提升.

Deep learning and generative adversarial network technology were analyzed aiming at the stringent safety and reliability requirements of civil aircrafts, the scarcity of abnormal samples of flight parameter data in actual civil aviation operations, the imbalance of the overall sample and the lack of labeling. A flight-level anomaly detection method for flight parameter data was proposed based on improved generative adversarial network. The method does not rely on the number of samples and labels, and realizes the detection method of unsupervised learning. Normal data samples were input for flight parameter data. The easy-to-converge WGAN-GP improved generative adversarial network model was applied to simulate normal data samples. The Bhattacharyya distance between the input data and the simulated normal data was calculated, and the detection of abnormal data was realized. The test verification was conducted through the artificial synthesis data set of NASA simulated flight parameter data and the flight parameter data set constructed by the quick access recorder data collected in the real operating environment. Results show that the proposed method has a significant improvement in some anomaly detection performance indicators compared with the commonly used unsupervised model.