一种基于降维密度聚类的船舶异常轨迹识别方法

目的 有效分析和探索海洋船舶时空轨迹行为模式，提高船舶轨迹聚类的效率与质量，更好地检测真实船舶的异常行为。方法 针对当前船舶轨迹数据研究中存在的对多维特征信息利用不足、检测效率不高、检测精度较差等问题，提出一种精确度高、能自主识别分析多维特征的船舶异常轨迹识别方法。首先利用随机森林分类器评估多维特征重要性，构建轨迹特征的最优组合;然后提出一种降维密度聚类方法，将T–分布随机邻域嵌入(T–SNE)和自适应密度聚类(DBSCAN)模型结合，通过构建特征选择层和无监督聚类层实现对数据元素非线性关系的高效提取以及对聚类参数的智能选择;最后根据聚类结果构建类簇特征向量，计算距离阈值判别轨迹相似度，实现轨迹异常检测模型的构建。结果 以UCI数据集为例，降维密度聚类方法对4、13、30、64维特征数据集的F1分数能达到0.9 048、0.9 534、0.8 218、0.6 627，多个聚类指标均优于DBSCAN、K–Means等常见聚类算法的。结论 研究结果表明，降维密度聚类方法能有效提取数据多维特征结构，实现聚类参数自适应，弥补密度聚类中参数难以确定的问题，有效实现对多种类型船舶轨迹异常的识别。

Trajectory Anomaly Identification Method of Vessels Based on Dimensional-density Reduction Clustering

The work aims to effectively analyze and explore the space-time trajectory behavior patterns of ocean vessels, improve the efficiency and quality of vessel trajectory clustering, and better detect abnormal behaviors of real vessels. In allusion to existing problems in current vessel trajectory data research, such as insufficient utilization of multidimensional feature information, low detection efficiency, poor detection accuracy, etc., a high accuracy and multi-dimensional feature identification method for vessel abnormal trajectory was proposed. Firstly, random forest classifier was used to evaluate the importance of multidimensional features and construct the optimal combination of trajectory features. Then, a dimensional-density reduction clustering method was proposed to combine T-SNE and DBSCAN models. By constructing feature selection layer and unsupervised clustering layer, the nonlinear relation of data elements could be extracted efficiently and the clustering parameters could be selected intelligently. Finally, the cluster feature vector was constructed according to the clustering results, and the distance threshold was calculated to distinguish the trajectory similarity, and the trajectory anomaly detection model was constructed. With UCI datasets as examples, the F1 score of this method could reach 0.904 8, 0.953 4, 0.821 8 and 0.662 7 for datasets with 4, 13, 30 and 64 dimensional features, and many clustering indexes were superior to DBSCAN, K-means and other common clustering algorithms. The results show that this method can effectively extract multi-dimensional feature structure of data, realize clustering parameter self-adaptation, make up for the problem that parameters are difficult to be determined in density clustering, and effectively realize the identification of multiple types of ship trajectory anomalies.