船舶异常行为的一致性检测算法

为了准确检测船舶的操纵异常行为和降低异常行为误报警率, 提出了船舶异常行为的一致性检测算法; 在船舶轨迹点中引入能够体现操纵模式的特征, 以转向行为与变速行为度量了操纵行为相似性; 将空间位置相似性与操纵行为相似性进行组合, 定义了船舶综合行为相似性, 计算了单个轨迹点与训练轨迹序列中的最近邻特征点, 构建了一致性检测的样本序列; 为克服样本重叠的类分布情形, 改进了一致性检测算法的奇异值度量, 并用综合行为相似性计算样本间的非一致性得分, 利用单个轨迹点的随机性检验值判断该轨迹点与样本序列的分布一致性; 以琼州海峡实测AIS数据作为正常数据, 以计算机模拟随机产生异常轨迹和人工自定义操纵异常行为作为异常数据, 分别进行异常检测试验。试验结果表明: 随机产生的异常轨迹检测正确率为100%, 但是轨迹评价集中有一部分正常轨迹被错误划分成异常轨迹, 在指定置信度水平分别为99.0%和99.7%的情形下, 误报警率分别为0.6%和0.2%, 分别低于显著性水平0.01和0.003, 因此, 利用一致性检测算法能有效检测计算机产生的随机异常轨迹, 并可通过指定显著性水平严格控制检测误报警率, 能有效检测人工自定义的船舶变速与转向异常行为, 而且检测结果能随船舶行为改变而变化。

Conformal detection algorithm of anomalous behaviors of vessel

In order to accurately detect the anomalous maneuvering behaviors of ship and reduce the false alarm rate of anomalous behaviors, a conformal detection algorithm of anomalous behaviors was proposed.Some characteristics was introduced into the trajectory points of ship to reflect the maneuvering modes, and the similarity of maneuvering behavior was measured through altering course behavior and speed-changing behavior.The integrated behavior similarity of ship was defined by combining the spatial location similarity and the maneuvering behavior similarity, the nearest neighbor feature points of single track points on training trajectory sequence were calculated, and the conformal detection sample sequence was constructed.In order to overcome the sample overlapping situation of class distribution, the singular value measure of conformal detection algorithm was improved, the nonconformance score between the samples was calculated by the integrated behavior similarity, and the randomness test value of single track point wasused to determine the distribution conformance of trajectory point and sample sequence.The real AIS data of Qiongzhou Strait were taken as the normal data, the random anomalous trajectories were simulated by the computer, the artificial abnormal maneuvering behaviors were defined, and the abnormal detection tests were carried out.Experimental result shows that the detecting accuracy rate of random anomalous behaviors is 100%, but a part of normal trajectories in the set of evaluation trajectories are divided into anomalous trajectories by mistake.When the confidence levels are 99.0% and 99.7%, respectively, the false alarm rates are 0.6% and 0.2%, respectively, and less than the significance levels of 0.01 and 0.003, respectively.Therefore, the algorithm can effectively detect the random abnormal trajectory generated by the computer, strictly control the false alarm rate by using designated significance level, and detect artificial changing speed and altering course abnormal behaviors, and the test result changes with the change of ship behavior.