基于上下文自编码的船舶行为语义表征

考虑船舶行为的时序相关性，提出了一种基于上下文自编码的船舶行为语义表征(SRCAE)模型；提取船舶经度、纬度、航速、航向等行为特征参量，建立了行为特征序列；借助连续词袋模型将行为特征序列划分为中心船舶行为和上下文船舶行为，利用深度自编码网络构建了船舶上下文行为的语义表征模型，将得到的中心船舶行为编码作为表征向量输出，通过聚类算法构建船舶行为词典；选取长江口南槽交汇水域作为研究对象，利用船舶自动识别系统产生的数据对提出的模型和方法进行了验证。分析结果表明：所提出的SRCAE模型能有效表征船舶行为之间的上下文联系，与传统自编码器和长短期记忆网络自编码器等模型相比SRCAE模型具有更低的表征误差；分别采用k均值(k-Means)、高斯混合模型(GMM)与核k均值(Kernel k-Means)3种聚类算法提取船舶行为词典，与原始数据相比SRCAE模型产生的表征向量更易于区分不同船舶行为模式，其中k-Means效果最优，轮廓系数、卡林斯基-哈拉巴斯指数和戴维森堡丁指数指标分别达到了0.384、18.308、0.531，共产生转向加速、转向减速、直行加速、直行减速等30种复合行为，有效提取了不同行为模式下船舶行为词组合关系。

Semantic representation of ship behavior based on context autoencoder

Considering the temporal correlation of ship behavior, a semantic representation model based on the context autoencoder (SRCAE) was proposed for ship behavior. The behavioral feature parameters, such as the longitude, latitude, speed, as well as the course, were extracted to establish the behavioral feature sequence. The behavioral feature sequence was divided into the center ship behavior and context ship behavior via the continuous bag-of-words (CBOW) model. The deep autoencoder (AE) networks were utilized to construct the semantic representation model of context ship behavior, and the encoded center ship behavior obtained from the model was output as the representation vector. The clustering algorithm was employed to establish the ship behavior dictionary. The South Passage Intersection Water of the Yangtze Estuary was selected as the research object, and the data from the automatic identification system (AIS) for ships were employed for verification of the proposed model and method. Analysis results show that the context relationships between ship behaviors can be effectively represented by the proposed SRCAE model, and the representation error of the SRCAE model is lower than that of the traditional AE model and long short-term memory autoencoder (LSTMAE) model. Three clustering algorithms, namely, k-means, Gaussian mixture model (GMM), and kernel k-means, were used to extract the ship behavior dictionary. Compared with the original data, the representation vectors generated by the SRCAE model are easier to distinguish different ship behavior patterns, among which the effect of k-means is the best, and the Silhouette coefficient (SC), Calinski-Harabasz index (CHI), and Davies-Bouldin index (DBI) of k-means reach 0.384, 18.308, and 0.531, respectively. A total of 30 types of composite behaviors are generated, such as steering acceleration, steering deceleration, straight-ahead acceleration, straight-ahead deceleration, and so on and the combination relationships of ship behavior words under different behavior patterns are effectively extracted.