AIS报文异常动态信息甄别方法
Discriminating method of abnormal dynamic information in AIS messages

针对船舶自动识别系统(AIS)报文中的异常动态信息,提出一种基于概率推理的包括先验知识提取、证据建模、证据合成与权重系数优化4个步骤的识别方法,运用似然度建模方法将经过人工识别的AIS数据中的速度、航向角和轨迹位置信息转化为0～1之间的证据信度,并用证据推理(ER)规则合成,以验证过的AIS数据作为输入,采用非线性优化方法修正证据权重系数,利用武汉天兴洲大桥水域轮渡与武桥水域货船的AIS数据进行验证试验。试验结果表明:在优化权重系数下武汉天兴洲大桥水域轮渡的正确数据、错误数据、总体数据识别准确率分别为91.67%、97.62%、92.63%; 以总体偏差最小为目标时,武桥水域货船的正确数据、错误数据、总体数据识别准确率分别为91.79%、89.87%、91.65%; 以正确数据偏差最小为目标时,武桥水域货船的正确数据、错误数据、总体数据识别准确率分别为93.18%、49.95%、90.03%。可见,基于ER规则的AIS动态信息甄别方法能针对不同的优化目标灵活调整证据权重系数,具有接近人工水平的识别准确率。
Aiming at the abnormal dynamic information in ship automatic identification system(AIS)messages, a recognition approach based on probabilistic inference with four steps including prior knowledge extraction, evidence modeling, evidence combination and weight coefficient optimization was proposed. Likelihood modeling approach was used to transform artificially identified velocity, course angle and track position information included in AIS data to evidence reliability between 0 and 1 that was composed by evidential reasoning(ER)rule. The verified AIS data was regarded as the input, and nonlinear optimization approach was used to modify the weight coefficient of evidence. The AIS data of ferry in Wuhan Tianxingzhou Bridge waters and the cargo ships in Wuqiao waters were used to carry out verification test. Test result shows that the recognition accuracies of correct data, incorrect data and total data for ferry in Wuhan Tianxingzhou Bridge waters under optimized weight coefficients are 91.67%, 97.62% and 92.63% respectively. When the minimum total deviation is goal, the recognition accuracies of correct data, incorrect data and total data for cargo ships in Wuqiao waters are 91.79%, 89.87% and 91.65% respectively. When the minimum deviation of correct data is goal, the recognition accuracies of correct data, incorrect data and total data for cargo ships in Wuqiao waters are 93.18%, 49.95% and 90.03% respectively. Obviously, the discriminating method of AIS dynamic information based on ER rule can flexibly adjust the weight coefficient of evidence with different optimized goals, and has the accuracy close to the artificial level. 10 tabs, 6 figs, 25 refs