CN-122020273-A - Ship track prediction method, system, equipment, medium and terminal

CN122020273ACN 122020273 ACN122020273 ACN 122020273ACN-122020273-A

Abstract

The invention belongs to the technical field of propagation track prediction, and discloses a ship track prediction method, a system, equipment, a medium and a terminal, aiming at the problems of heterogeneity of marine traffic data and long-term prediction error accumulation, firstly, the capturing capacity of a model on track space-time continuity is improved through mutual fusion of multi-mode embedding and position coding; secondly, a causal self-attention mechanism and a partition modeling strategy are used, and accurate capture of space-time dependency relationship is realized through dynamic feature division and mask constraint; and finally, guiding the current iteration of the model by using a historical clustering result to adjust the weight, thereby enabling the model to bias a characteristic region with higher data quality. In addition, the invention introduces a fuzzy loss function, so that the influence of noise on the model is reduced, and the aim of improving the overall prediction performance is fulfilled.

Inventors

SHEN XIANHAO
ZHUANG YI
ZHOU YUYANG
Yang Diejun

Assignees

桂林理工大学

Dates

Publication Date: 20260512
Application Date: 20251031

Claims (10)

1. The ship track prediction method is characterized by comprising the following steps of: Step 1, constructing a ship track prediction model based on dynamic clustering and a space-time perception transducer architecture; step 2, cleaning and converting AIS data of the automatic identification system, and extracting multi-mode characteristics of longitude, latitude, speed and course; step3, mapping each feature to a high-dimensional semantic space through an independent embedding function to form a fused track representation; step 4, normalizing the time sequence in the dynamic clustering module, and selecting high-density points as an initial clustering center by adopting kernel density estimation; step 5, updating a clustering center according to the shape similarity measurement combining the dynamic time warping distance and the information entropy, and storing the clustering center in a history clustering memory module; Step 6, judging whether the cluster is stable or not, if not, continuing iteration until convergence; Step 7, inputting the clustering label data into a space-time perception transducer architecture, and introducing a position code capable of being learned; Step 8, realizing space-time feature dynamic division by using a causal self-attention mechanism and a partition modeling strategy; Step 9, introducing a fuzzy loss function and training a model by adopting a AdamW optimizer; and 10, outputting a ship track prediction result.
2. The method of claim 1, wherein the multi-modal embedding layer designs independent embedding functions for longitude, latitude, speed and heading, maps them to a unified semantic space, and generates final feature representations by weighted fusion, each feature weight being adaptively adjustable by training to enhance the model's ability to distinguish between multi-dimensional dynamic features.
3. The method of claim 1, wherein the kernel density estimation is used to determine an initial cluster center by calculating a spatial density value of each AIS data point and selecting a number of data points with highest density as the initial cluster center, thereby improving stability and convergence rate of the clusters.
4. The method of claim 1, wherein the shape similarity measure suppresses noise effects while achieving elastic matching between time sequences by weighted fusion of dynamic time warping distance and sequence joint entropy to enhance robustness of clusters under abnormal trajectories.
5. The method of claim 1, wherein the learnable position codes map time steps into code vectors that are consistent with embedded vector dimensions by linear transformation and input a transform encoder after adding to the embedded vectors, enabling the model to explicitly model the temporal continuity of the trajectory.
6. The method of claim 1, wherein the fuzzy loss function forms a smooth probability distribution by gaussian kernel density estimation of a true trajectory distribution, and utilizes Kullback-Leibler divergence metric to predict a difference of the distribution from the true distribution to achieve continuity and robustness of the predicted output.
7. A ship track prediction system that implements the ship track prediction method according to claim 1, comprising: the model building module is used for building a prediction model based on dynamic clustering and a space-time perception transducer architecture; the preprocessing module is used for cleaning and converting AIS data and extracting multi-mode characteristics; The feature mapping module is used for embedding each feature into a high-dimensional semantic space; The dynamic clustering module is used for carrying out time sequence standardization, cluster center initialization and updating; The cluster stability judging module is used for detecting the cluster convergence state and outputting a cluster label; The transducer analysis module is used for executing causal self-attention calculation on the labeled space-time characteristics; the parameter optimization module is used for introducing a fuzzy loss function and optimizing model parameters based on AdamW algorithm; and the prediction output module is used for outputting a ship track prediction result.
8. The system of claim 7, wherein the dynamic clustering module comprises a kernel density estimation unit, a similarity calculation unit, and a history cluster memory unit, wherein the similarity calculation unit performs similarity calculation based on a weighted combination of dynamic time warping distance and information entropy.
9. A spatio-temporal perceptual transducer device for modeling ship trajectory data, comprising: the multi-mode embedding unit is used for independently embedding and weighting and fusing longitude, latitude, navigational speed and course characteristics; a leachable position encoding unit for mapping the time step to a high-dimensional space; the causal self-attention unit is used for restraining attention weight through a lower triangular mask matrix and carrying out track modeling only by relying on historical information; and the partition modeling unit is used for dynamically dividing the characteristic space to adapt to the space-time distribution non-uniformity.
10. The ship intelligent navigation prediction device based on the dynamic clustering and Transformer combined architecture is characterized by comprising a data acquisition module, an embedded processing module, a cluster analysis module, a Transformer reasoning module and a prediction display module, wherein the device realizes intelligent prediction and track visual output of future positions of a navigation ship by fusing cluster labels and learnable time sequence features.

Description

Ship track prediction method, system, equipment, medium and terminal Technical Field The invention belongs to the technical field of propagation track prediction, and particularly relates to a ship track prediction method, a system, equipment, a medium and a terminal based on space-time perception transformation and dynamic clustering enhancement. Background Waterway transportation is taken as a core component of a modern traffic network, and the intelligent development of the waterway transportation is important to the improvement of global shipping efficiency and safety. With the popularization of automatic ship identification systems (AIS), real-time acquisition of a large amount of space-time data provides a new opportunity for ship track prediction. The data can be used for real-time traffic monitoring and route congestion relief, collision risks can be avoided in advance by predicting ship behaviors, and the data becomes a key technical support of an intelligent shipping system. However, the dynamic complexity of offshore environments (e.g., line diversity, weather changes) and the heterogeneity of AIS data (e.g., noise interference, sampling non-uniformity) make high-precision, long-thread trajectory prediction still pose serious challenges. In recent years, deep learning techniques have been used to inject new vigor into ship trajectories. Early studies were based on a hybrid architecture of traditional timing methods and recurrent neural networks. For example, ARIMA-LSTM method, which decomposes linear and nonlinear characteristics by moving average filtering to model and predict respectively, when it depends on a fixed time window, it IS difficult to capture long-term dependence, IS-STGCNN method, which extracts interactive characteristics of ship by space-time diagram convolution, but lacks the structure of preset static diagram to adapt to dynamic space relation, and bilinear self-encoder method, although it can generate future track, has insufficient robustness to noise data and does not consider diversity of navigation modes. The comprehensive analysis of the existing method has the urgent need of breaking through two bottlenecks, namely that the first bottleneck is insufficient in dynamic space-time dependence modeling, and the second bottleneck is long-term prediction performance decay. In view of the above analysis, the technical problem that needs to be solved in the prior art is that the dynamic space-time dependency modeling in the prior art is insufficient and the performance decay is predicted for a long time. Disclosure of Invention Aiming at the problems existing in the prior art, the invention provides a ship track prediction method, a ship track prediction system, a ship track prediction device, a ship track prediction medium and a ship track prediction terminal, which aim to improve the accuracy of ship track prediction and increase the anti-interference capability on noise by integrating a space-time perception transducer architecture and a dynamic cluster enhancement mechanism. The invention is realized in such a way that a ship track prediction method comprises the following steps: Step 1, constructing a ship track prediction model based on dynamic clustering and a space-time perception transducer architecture, and initializing a global model; Step 2, preprocessing AIS data, including data cleaning and data conversion, and extracting multi-mode characteristics of ship tracks, including longitude, latitude, speed and course; step 3, mapping each feature to a semantic space through a multi-mode embedding layer, and then inputting the features into a dynamic clustering module; step 4, in the dynamic clustering module, carrying out standardization processing on all time sequence data, eliminating scale differences among different time sequences, and then initializing a clustering center; Step 5, evaluating the similarity between each time sequence and the clustering center through shape similarity measurement, distributing the time sequences to the clustering center most similar to the time sequences according to the similarity, then updating the clustering center, and inputting the updated time sequences to a history clustering memory module; Step 6, judging whether the clusters are stable, if so, outputting each cluster label, and if not, repeating the step 5 until the clusters are stable; Step 7, inputting the data with the clustering labels into a space perception transducer architecture, and introducing a position code capable of being learned; Step 8, dynamically dividing a feature space by using a causal self-attention mechanism and a partition modeling strategy to adapt to the space-time distribution non-uniformity of AIS data; Step 9, carrying out a series of cyclic training such as layer normalization, back propagation, residual error connection and the like on the data after the data is subjected to a causal autonomous attention mechanism, introducing a fuzzy loss functi