一种针对船舶自动避碰算法召回测试的事故场景预测方法

Accident Scenario Prediction Method for Ship Automatic Collision Avoidance Algorithm Recall Testing

在对船舶自动避碰算法进行仿真测试过程中发现缺陷并改进之后,需进行召回测试以验证改进的有效性,而随着被测算法的水平逐步提升,测试场景逐渐复杂,进行召回测试时难以设计更全面的测试场景。对此,提出一种能根据历史测试数据对新场景进行事故预测的卷积神经网络模型。通过该模型筛选出的场景与历史上发生过的事故场景属于同类,能支持更全面且有针对性的召回测试。在特征建模过程中,对测试场景进行9通道建模,以将场景信息表达为模型能识别的张量结构;在样本收集过程中,提出平行场景的方法,进行1次测试可收集到多组样本。在多样复杂场景中采用该预测方法对一种事故率为6.2%的被测避碰算法进行试验,结果发现该方法训练的模型从所有新的随机场景中筛选出24.9%的样本,这些筛出的场景以91.8%的概率覆盖所有事故场景。

Simulated testing of the ship's automatic collision avoidance algorithms is used to identify errors and make improvements.Subsequent recall testing is required to verify the effectiveness of these improvements.However,as the performance of the tested algorithms gradually increases,the testing scenarios become more complex,making it difficult to design comprehensive testing scenarios for recall testing.To address this,a convolutional neural network model is proposed that predicts accidents in new scenarios based on historical test data.The scenarios selected by this model are similar to those in which accidents have occurred in the past,allowing for more comprehensive and targeted recall testing.In the feature modeling process,a 9-channel modeling approach is used to represent scenario information as a tensor structure recognizable by the model.During the sampling process,the parallel scenario method is introduced,allowing multiple sample sets to be collected from a single test.Using this prediction method,we test a collision avoidance algorithm with an accident rate of 6.2%in a variety of complex scenarios.The results show that the model trained with this method selected 24.9%of the samples from all new random scenarios,with these selected scenarios covering all accident scenarios with a probability of 91.8%.