ADS-B Air Traffic for Anomalous Trajectory Detection

The "ADS-B Air Traffic for Anomalous Trajectory Detection" is a dataset of flight recordings from the OpenSky network, used to evaluate anomaly detection methods in the domain of aviation data. The recordings in the dataset were all collected from flights cruising over the "LPPC" and "EGGX" airspace. The "LPPC" airspace is a flight information region (FIR) on the west coast of Portugal, which encompasses more than 600,000 square kilometers. The recordings for the "LPPC" FIR were limited to a 4-hour time frame on January 1, 2020, between 10:00 AM and 2:00 PM. The "EGGX" airspace is an FIR on the west coast of Ireland, which encompasses more than 2,300,000 square kilometers. The recordings for the "EGGX" FIR were limited to a 4-hour time frame on January 2, 2020, between 10:00 AM and 2:00 PM. In total, 42 unique flights were recorded in the selected time frames, each one containing at-least 1500 ADS-B broadcast messages, sampled each second (i.e., at-least 25-minute recordings). • Train: 25 flights from the "LPPC" FIR, with a total of 83620 broadcast ADS-B messages (23 hours, 13 minutes, and 40 seconds) • Validation: 10 flights from the "LPPC" FIR, with a total of 40613 broadcast ADS-B messages (11 hours, 16 minutes, and 53 seconds) • Test: 7 flights from the "EGGX" FIR, with a total of 15392 broadcast ADS-B messages (4 hours, 16 minutes, and 32 seconds) To simulate the anomalous flight behavior, we inject a single window of data during each flight in the testing set, after 10 minutes of flight recording, for ten consecutive minutes (window of 600 ADS-B messages). In total, four different types of injections are performed, as follows: Noise - The anomalies are synthetically generated by adding random samples, drawn from a normal (Gaussian) distribution to the original values. The rest of the injections are spoofed trajectories, created by data extracted from different real flight trajectories. To create a continuous spoofed trajectory, we calculate the cumulative sum of message differences from the source flight and add them to the target flight messages. The injections are as follows: Manoeuver (Flight AIB232E) - The flight recording of an A380 Airbus pilot in Germany that traced the outline of an enormous Christmas tree during a test flight, making circular maneuvers at more than 40,000ft. Landing (Flight TAP070) - The recording of a commercial flight landing in LPPT (Lisbon) Airport, on January 1, 2020. Departing (Flight UAL65) - The recording of a commercial flight departing from LPPT (Lisbon) Airport, on January 1, 2020.

"用于异常轨迹检测的ADS-B航空交通数据集"是一套源自OpenSky网络的飞行记录数据集，用于航空数据领域的异常检测方法评估。数据集中的所有记录均采集自在"LPPC"与"EGGX"空域巡航的航班。其中"LPPC"空域为葡萄牙西海岸的飞行情报区（Flight Information Region，简称FIR），覆盖面积超过60万平方千米。LPPC飞行情报区的记录时段限定为2020年1月1日的4小时窗口，即当日10:00至14:00。"EGGX"空域为爱尔兰西海岸的飞行情报区，覆盖面积超过230万平方千米。EGGX飞行情报区的记录时段限定为2020年1月2日的4小时窗口，即当日10:00至14:00。本次选取的时段内共记录到42架次独立航班，每架次至少包含1500条自动相关监视广播（Automatic Dependent Surveillance-Broadcast，简称ADS-B）广播报文，采样间隔为1秒，对应单架次至少25分钟的飞行记录时长。
- 训练集：取自LPPC飞行情报区的25架次航班，总计包含83620条ADS-B广播报文，总时长为23小时13分40秒
- 验证集：取自LPPC飞行情报区的10架次航班，总计包含40613条ADS-B广播报文，总时长为11小时16分53秒
- 测试集：取自EGGX飞行情报区的7架次航班，总计包含15392条ADS-B广播报文，总时长为4小时16分32秒
为模拟异常飞行行为，我们在测试集的每架次航班飞行记录时长达到10分钟后，插入一段连续10分钟的异常数据窗口（对应600条ADS-B报文）。本次实验共设置4种不同的异常注入类型，具体如下：
噪声型异常：通过向原始报文数据添加服从正态（高斯）分布的随机采样值，合成生成异常数据。其余三种异常均为伪造飞行轨迹，基于不同真实飞行轨迹提取的数据构建。具体构建方式为：计算源航班报文的差值累积和，并将其叠加至目标航班的报文中，以生成连续的伪造飞行轨迹。四种注入类型分别为：
1. 机动动作型（航班号AIB232E）：取自德国一架空客A380的试飞记录，该航班在4万英尺以上高空完成了环绕巨型圣诞树轮廓的环形机动飞行动作
2. 着陆型（航班号TAP070）：取自2020年1月1日一架商业航班在里斯本LPPT机场的着陆过程记录
3. 起飞型（航班号UAL65）：取自2020年1月1日一架商业航班从里斯本LPPT机场的起飞过程记录

搜集汇总

背景与挑战

背景概述

该数据集包含葡萄牙和爱尔兰特定空域的飞行记录，用于异常轨迹检测研究。数据包括42个独特航班，分为训练、验证和测试集，并模拟了4种不同类型的异常飞行行为。