Ship Trajectory Prediction Method Based on Spatiotemporal Feature Fusion

Abstract

This application relates to the technical field of trajectory prediction. Specifically, it discloses a ship trajectory prediction method based on spatio-temporal feature fusion. It uses a neural network model based on deep learning to process the AIS information at each time in the known ship trajectory to extract the temporal dimension features and spatial dimension features of the ship trajectory. Then, through the explicit modeling analysis of the implicit association between the temporal dimension features and spatial dimension features of the ship trajectory, the implicit key clues between the two are mined, and based on this, the alignment and interaction fusion of the temporal dimension features and spatial dimension features of the ship trajectory are realized. Furthermore, the future trajectory of the ship is predicted based on the spatio-temporal fusion features of the ship trajectory. In this way, the correlation between the ship's sailing speed, course change and geographical location can be deeply mined, so as to achieve more accurate feature alignment and interaction fusion, and then improve the accuracy of ship trajectory prediction.

Description

Technical Field

[0001] This application relates to the technical field of trajectory prediction, and more specifically, to a ship trajectory prediction method based on spatio-temporal feature fusion. Background Art

[0002] In the context of the booming development of the current maritime transportation industry, ship trajectory prediction is crucial for ensuring maritime traffic safety, optimizing the allocation of shipping resources, and improving port operation efficiency. Accurate trajectory prediction can avoid potential collision risks in advance and reasonably plan ship routes, thus effectively reducing congestion and improving the overall shipping efficiency. In recent years, with the wide application of the Automatic Identification System (AIS) for ships, a large amount of ship navigation data has been accumulated, providing rich data support for the development of ship trajectory prediction technology, and the trajectory prediction method based on spatio-temporal feature fusion has become a research hotspot.

[0003] For example, the invention patent with the publication number CN117933492A proposes a long-term ship trajectory prediction method based on spatio-temporal feature fusion. It converts the observed values of AIS information at each time in the known ship trajectory into corresponding high-dimensional embedding vectors, uses the ship trajectory prediction model TCNformer to extract temporal dimension features and spatial dimension features from the high-dimensional feature vector sequence, and further fuses the temporal dimension features and spatial dimension features, and predicts the trajectory information at the next time point according to the spatio-temporal fusion features until the prediction task is completed. In this way, the problems of the complexity of the ship motion mode and the heterogeneity of AIS data are effectively solved, and the long-term prediction of ship trajectories is realized.

[0004] However, in the prior art, it directly adopts the splicing fusion method to fuse the temporal dimension features and spatial dimension features of ship trajectories, and fails to fully explore the complex internal relationship between time and space features. Specifically, during the ship's navigation, the changes in speed and heading in the temporal dimension are closely related to the geographical location in the spatial dimension. The direct splicing fusion method cannot deeply capture this internal relationship, which may lead to inaccurate prediction results. In addition, due to the influence of factors such as ship navigation speed and route planning, the change frequencies and scales of time and space features are not consistent. Direct splicing fusion will cause problems such as information redundancy or loss of association in the fused features, making the fused features unable to comprehensively and accurately reflect the actual situation of ship navigation, thereby affecting the accuracy of trajectory prediction.

[0005] Therefore, an optimized ship trajectory prediction method based on spatio-temporal feature fusion is expected. Summary of the Invention

[0006] To solve the above technical problems, this application is proposed. An embodiment of this application provides a ship trajectory prediction method based on spatio-temporal feature fusion, which uses a neural network model based on deep learning to process the AIS information at each time in the known ship trajectory to extract the temporal dimension features and spatial dimension features of the ship trajectory. Then, through the implicit correlation explicit modeling analysis of the temporal dimension features and spatial dimension features of the ship trajectory, the implicit key clues between the two are mined, and based on this, the alignment and interactive fusion of the temporal dimension features and spatial dimension features of the ship trajectory are realized. Furthermore, the future trajectory of the ship is predicted based on the spatio-temporal fusion features of the ship trajectory. In this way, the correlation between the ship's sailing speed, heading change and geographical location can be deeply mined, so as to achieve more accurate feature alignment and interactive fusion, and then improve the accuracy of ship trajectory prediction.

[0007] According to one aspect of this application, a ship trajectory prediction method based on spatio-temporal feature fusion is provided, which includes:

[0008] Obtain the known ship trajectory;

[0009] Based on the AIS information at each time in the known ship trajectory, extract the temporal dimension features and spatial dimension features of the known ship trajectory;

[0010] Perform spatio-temporal feature interaction fusion based on implicit correlation explicit modeling on the temporal dimension features and the spatial dimension features to obtain spatio-temporal fusion features;

[0011] Based on the spatio-temporal fusion features, predict the ship trajectory information.

[0012] Compared with the prior art, the ship trajectory prediction method based on spatio-temporal feature fusion provided by this application uses a neural network model based on deep learning to process the AIS information at each time in the known ship trajectory to extract the temporal dimension features and spatial dimension features of the ship trajectory. Then, through the implicit correlation explicit modeling analysis of the temporal dimension features and spatial dimension features of the ship trajectory, the implicit key clues between the two are mined, and based on this, the alignment and interactive fusion of the temporal dimension features and spatial dimension features of the ship trajectory are realized. Furthermore, the future trajectory of the ship is predicted based on the spatio-temporal fusion features of the ship trajectory. In this way, the correlation between the ship's sailing speed, heading change and geographical location can be deeply mined, so as to achieve more accurate feature alignment and interactive fusion, and then improve the accuracy of ship trajectory prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] The above and other objects, features, and advantages of the present application will become more apparent by describing the embodiments of the present application in more detail with reference to the accompanying drawings. The accompanying drawings are used to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation to the present application. In the accompanying drawings, the same reference numerals generally represent the same components or steps.

[0014] Figure 1 It is a flowchart of a ship trajectory prediction method based on spatio-temporal feature fusion according to an embodiment of the present application.

[0015] Figure 2 It is a schematic diagram of data flow of a ship trajectory prediction method based on spatio-temporal feature fusion according to an embodiment of the present application.

[0016] Figure 3 It is a flowchart of sub-step S3 of a ship trajectory prediction method based on spatio-temporal feature fusion according to an embodiment of the present application.

[0017] Figure 4 It is a flowchart of sub-step S32 of a ship trajectory prediction method based on spatio-temporal feature fusion according to an embodiment of the present application. Detailed implementation manners

[0018] As shown in the present application and the claims, unless the context clearly indicates otherwise, words such as "a", "an", "one", and / or "the" are not specifically singular and may also include plural. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of the clearly identified steps and elements, and these steps and elements do not constitute an exclusive list. The method or device may also include other steps or elements.

[0019] Although the present application makes various references to certain modules in the system according to the embodiments of the present application, however, any number of different modules can be used and run on the user terminal and / or server. The modules are only illustrative, and different aspects of the system and method can use different modules.

[0020] Flowcharts are used in the present application to illustrate the operations performed by the system according to the embodiments of the present application. It should be understood that the operations before or below are not necessarily executed precisely in order. On the contrary, as needed, various steps can be executed in reverse order or simultaneously. At the same time, other operations can also be added to these processes, or one or several steps can be removed from these processes.

[0021] Next, exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. It should be understood that the present application is not limited by the exemplary embodiments described here.

[0022] It should be noted that in this application, all actions of obtaining data are carried out on the premise of complying with the corresponding data protection regulations and policies of the country where it is located and obtaining authorization from the owner of the corresponding device.

[0023] As mentioned in the above background art, Patent CN117933492A proposes a long-term ship trajectory prediction method based on spatio-temporal feature fusion. It converts the AIS information observation values at each time in the known ship trajectory into corresponding high-dimensional embedding vectors, uses the ship trajectory prediction model TCNformer to extract the time-dimensional features and space-dimensional features from the high-dimensional feature vector sequence, and further fuses the time-dimensional features and space-dimensional features to predict the trajectory information at the next time point until the prediction task is completed. In this way, the problems of the complexity of the ship motion mode and the heterogeneity of AIS data are effectively solved, and the long-term prediction of the ship trajectory is realized.

[0024] However, in the prior art, the above method simply relies on splicing and fusion to integrate the time features and space features of the ship trajectory, and it is difficult to deeply reveal the complex internal relationship between the time and space features. During ship navigation, the speed and course changes in the time dimension are closely related to the geographical location in the space dimension, and the direct splicing method cannot effectively capture this internal connection, which may greatly reduce the accuracy of the prediction result. In addition, due to the constraints of factors such as ship navigation speed and route planning, there are differences in the change frequency and scale of time and space features. Direct splicing and fusion are prone to information redundancy or loss of association in the fused features, and cannot comprehensively and accurately present the real situation of ship navigation, ultimately having an adverse impact on the accuracy of trajectory prediction. To address the above technical problems, this application proposes an optimized ship trajectory prediction method based on spatio-temporal feature fusion. It uses a neural network model based on deep learning to process the AIS information at each time in the known ship trajectory to extract the time-dimensional features and space-dimensional features of the ship trajectory. Then, through the implicit association explicit modeling analysis of the time-dimensional features and space-dimensional features of the ship trajectory, the implicit key clues between the two are mined, and based on this, the alignment and interaction fusion of the time-dimensional features and space-dimensional features of the ship trajectory are realized. Furthermore, based on the spatio-temporal fusion features of the ship trajectory, the future trajectory of the ship is predicted. In this way, the correlation between the ship navigation speed, course change and geographical location can be deeply mined, so as to achieve more accurate feature alignment and interaction fusion, and then improve the accuracy of ship trajectory prediction.

[0025] Figure 1 It is a flowchart of the ship trajectory prediction method based on spatio-temporal feature fusion according to an embodiment of this application. Figure 2Schematic diagram of data flow for the ship trajectory prediction method based on spatio-temporal feature fusion according to an embodiment of the present application. As Figure 1 and Figure 2 shown, the ship trajectory prediction method based on spatio-temporal feature fusion includes the steps of: S1, obtaining known ship trajectories; S2, extracting the time dimension features and space dimension features of the known ship trajectories based on the AIS information at each time in the known ship trajectories; S3, performing spatio-temporal feature interaction fusion based on implicit association explicit modeling on the time dimension features and the space dimension features to obtain spatio-temporal fusion features; S4, predicting ship trajectory information based on the spatio-temporal fusion features.

[0026] In the above ship trajectory prediction method based on spatio-temporal feature fusion, in step S1, known ship trajectories are obtained. It should be understood that in practical applications, the Automatic Identification System (AIS) of ships is an important source for obtaining ship trajectory data. The AIS system can collect various types of information of ships in real time, including the position information of ships (represented by latitude and longitude), the operating speed (SOG), the course over ground (COG), etc., which are continuously updated over time to form known navigation trajectory records of ships. In addition, since the quality of the original AIS data is uneven and contains various problems. For example, the data may contain information with speed values that are significantly inconsistent with the actual situation (such as a ship's speed exceeding the normal navigation speed range), data generated when the ship is in a moored or anchored state, and AIS observation information within 1 nautical mile of the coastline. These data not only increase the burden of data processing but may also interfere with subsequent trajectory analysis and prediction. Therefore, in order to improve the accuracy and efficiency of ship trajectory prediction, it is necessary to clean and preprocess the collected AIS data. The specific operations include deleting AIS information with unrealistic speed values, removing information of moored or anchored ships, and deleting AIS observation information within 1 nautical mile of the coastline. Furthermore, the known AIS data of the ship to be predicted is extracted from the AIS data after data cleaning and preprocessing to form the known ship trajectories.

[0027] Specifically, on the one hand, it is necessary to select appropriate AIS data sources. This not only includes eliminating obviously incorrect or abnormal data points, but also involves checking data integrity. For example, if the AIS signal of a certain ship is lost during a certain period, other means (such as satellite communication, radar, etc.) need to be used to supplement the missing information to ensure the continuity of the trajectory. On the other hand, during the data acquisition process, it is also necessary to conduct consistency tests on data from different sources. Due to the differences in the working principles and technical parameters of various sensors, directly merging and using them may lead to error accumulation. Therefore, before integrating multi-source data, necessary calibration work should be carried out to ensure that all data can be compared and analyzed in the same coordinate system. For example, using high-precision Geographic Information System (GIS) tools to convert the position information from different sensors into a unified map projection system can effectively reduce the deviation caused by inconsistent coordinate systems.

[0028] In addition to technical considerations, the issue of data security cannot be ignored. With the increase in the amount of AIS data and the expansion of its application scope, protecting privacy has become an important issue. Especially when it comes to sensitive areas or specific types of ships, in order to meet research needs while complying with relevant laws and regulations and avoiding infringement of personal privacy, anonymization measures can be taken at the beginning of data collection, only retaining the necessary navigation information for academic research or commercial use, and removing any identifiers that may identify specific individuals.

[0029] Furthermore, considering the complexity and dynamic characteristics of the marine environment, it is particularly important to establish a real-time updated database. Such a database can not only reflect the latest changes in ship positions and states in a timely manner, but also provide support for rapid response in case of emergencies. For example, in the face of bad weather conditions or emergency collision avoidance requirements, the latest and most accurate data can be quickly called to assist in decision-making. To this end, it is necessary to develop an efficient data synchronization mechanism to ensure seamless information exchange between land-based receiving stations and mobile terminals, and the entire process should minimize delays as much as possible.

[0030] It should be noted that obtaining the trajectories of known ships is a continuous process, and it is necessary to update data regularly and adjust the analysis strategy. As time goes by, new shipping lanes are constantly opened up, and the navigation patterns on existing shipping lanes may also change. Therefore, establishing a dynamic data update mechanism to keep up with the latest AIS data in a timely manner and adjusting the analysis methods and technical routes according to the actual situation is particularly important. This not only helps to maintain the timeliness of research results, but also ensures that all decisions made based on this are always close to the actual situation, maximizing the practical value of ship trajectory prediction.

[0031] In the above-mentioned ship track prediction method based on the fusion of time and space features, the step S2 extracts the time dimension features and space dimension features of the known ship track based on the AIS information at each time in the known ship track. It should be understood that during the navigation process, the position, speed, heading and other states of the ship change continuously over time. By analyzing the speed and heading information at different times, the time nodes and frequencies of the ship's acceleration, deceleration, turning and other behaviors can be understood, and the subsequent movement trend of the ship can be predicted. At the same time, by analyzing the position information, navigation direction, etc. of different points in the ship's navigation track, the navigation characteristics and preferences of the ship can be understood, thereby providing a spatial dimension information reference for track prediction.

[0032] Specifically, for the AIS information at each time in the known ship trajectory, firstly, various attributes (such as position, speed, heading, etc.) in the AIS information are converted into high-dimensional vector representations through embedded coding technology to capture the potential associations between these attributes and form a sequence of high-dimensional feature vectors of AIS information. Then, since the time dimension characteristics of the ship trajectory mainly focus on the dynamic changes of the ship at different time points. In this regard, the TCN (Temporal Convolutional Networks) network can be used to process the sequence of high-dimensional feature vectors of the AIS information. Through its causal convolution structure, the convolution operation is performed on the time dimension of the sequence of high-dimensional feature vectors of the AIS information, which can effectively extract the local dependencies in the sequence data while maintaining the time order, so that the time dimension characteristics reflecting the ship motion mode can be extracted from the sequence of high-dimensional feature vectors of the AIS information. At the same time, the spatial dimension characteristics of the ship trajectory focus on the distribution and movement trend of the ship in the geographic space. To this end, the powerful parallel computing capability of the Transformer network can be used to process the sequence of high-dimensional feature vectors of the AIS information using a self-attention mechanism to capture the long-term correlations in the known AIS observation data, and to mine the correlations and movement trends of ships in different geographical locations, thereby extracting spatial dimension features that reflect the spatial distribution characteristics of ships. It is worth mentioning that the method for extracting the time dimension features and spatial dimension features of the known ship trajectory involved in this application can adopt the method disclosed in Chinese patent CN117933492A, and of course other equivalent feature extraction methods can also be used, which is not limited to this application.

[0033] In the above-mentioned ship trajectory prediction method based on spatio-temporal feature fusion, in step S3, spatio-temporal feature interaction fusion based on implicit association explicit modeling is performed on the time dimension feature and the space dimension feature to obtain spatio-temporal fusion features. It should be understood that in this application, it is considered that the time dimension feature reflects the dynamic change of the ship's navigation state over time, with obvious time series and periodicity; while the space dimension feature focuses on the positional relationship and movement trajectory of the ship in the geographical space. Due to the influence of factors such as ship navigation speed and route planning, the change frequencies and scales of its time dimension feature and space dimension feature are not consistent. For example, in different sea areas, the navigation speed of the ship may be different, which leads to differences in the distance traveled by the ship in the space dimension within the same time interval in the time dimension. Traditional simple splicing fusion methods often ignore this heterogeneity between features, resulting in unsatisfactory feature fusion effects, and thus reducing the model's prediction ability for ship trajectories. In response to this, this application proposes a spatio-temporal feature interaction fusion method based on implicit association explicit modeling, which establishes a deep connection between the two by mining the implicit key clues between the time dimension feature and the space dimension feature, and then performs alignment interaction between the time dimension feature and the space dimension feature on this basis to obtain spatio-temporal fusion features. Among them, Figure 3 is a flowchart of sub-step S3 of the ship trajectory prediction method based on spatio-temporal feature fusion according to an embodiment of the present application. As Figure 3 shown, step S3 includes the steps: S31, performing feature principal component analysis on the time dimension feature to obtain a set of time dimension feature principal component components; S32, respectively performing implicit association analysis on each time dimension feature principal component component in the set of the space dimension feature and the time dimension feature principal component components to construct a {space dimension deep implicit feature, anchored time dimension principal component deep implicit feature} feature pair; S33, performing spatio-temporal feature interaction fusion on the {space dimension deep implicit feature, anchored time dimension principal component deep implicit feature} feature pair to obtain the spatio-temporal fusion features.

[0034] Specifically, step S31 is expressed by the formula:

[0035]

[0036] Among them, is the covariance matrix of the time dimension feature, represents the matrix composed of the set of time dimension feature principal component components obtained by performing eigenvalue decomposition on represents transpose, represents The diagonal matrix composed of the set of time - dimension principal - component eigenvalues obtained by performing eigenvalue decomposition, and respectively represent the first and the th eigenvalues in the set of time - dimension principal - component eigenvalues, 、 and respectively represent the first, the second and the th time - dimension feature principal - component components in is the number of the time - dimension feature principal - component components, represents the principal - component analysis network, represents the time - dimension feature.

[0037] That is, in the normal case, the present application considers that the spatial - dimension feature mainly reflects the position relationship of the ship, and the correlation within its features is relatively clear, that is, the features are relatively stable and dependent on the geographical - space structure. While the time - dimension feature mainly reflects the dynamic changes of the ship's navigation state, and the correlation within its features is relatively complex, and may contain more redundant information. Therefore, the present application first performs principal - component analysis on the time - dimension feature, and transforms the time - dimension feature into a more representative set of time - dimension feature principal - component components, so as to extract the core trend and periodic changes in the time - dimension feature, and separate the redundancy and noise, thereby reducing the complexity and computational amount of subsequent interaction analysis.

[0038] Figure 4 is a flowchart of sub - step S32 of the ship track prediction method based on spatio - temporal feature fusion according to an embodiment of the present application. As Figure 4 shown, the step S32 includes steps: S321, respectively extract the deep implicit features of each time - dimension feature principal - component component in the set of the spatial - dimension feature and the time - dimension feature principal - component components to obtain a set of spatial - dimension deep implicit features and time - dimension principal - component deep implicit features; S322, perform implicit key - clue anchoring encoding on the set of the spatial - dimension deep implicit features and the time - dimension principal - component deep implicit features to obtain the {spatial - dimension deep implicit features, anchored time - dimension principal - component deep implicit features} feature pair.

[0039] More specifically, in a specific example of the present application, the step S321 includes: respectively perform implicit - feature extraction based on point - convolution encoding on each time - dimension feature principal - component component in the set of the spatial - dimension feature and the time - dimension feature principal - component components to obtain the set of the spatial - dimension deep implicit features and the time - dimension principal - component deep implicit features, which is expressed by the formula:

[0040]

[0041]

[0042] Among them, represents the spatial dimension feature, represents the point convolution operation, represents the activation function, represents the deep implicit feature of the spatial dimension, represents the th time dimension feature principal component component in the set of the time dimension feature principal components, , , and respectively represent the 1st, 2nd, th th

[0043] That is, in the present application, through point convolution encoding processing, the deep implicit features of the spatial dimension feature and each time dimension feature principal component component are further extracted. In particular, point convolution does not focus on the locality of the original feature space of the feature, but directly performs information fusion within the global range of the feature through cross-dimensional weight mapping, and further enhances the feature expression ability through a non-linear activation function, so as to be able to efficiently extract deep-level and strongly representative implicit features, and obtain a set of deep implicit features of the spatial dimension and deep implicit features of the time dimension principal components.

[0044] More specifically, in a specific example of the present application, the step S322 includes: calculating the feature correlation factor between the deep implicit feature of the spatial dimension and each deep implicit feature of the time dimension principal components in the set of the deep implicit features of the time dimension principal components, and selecting the deep implicit feature of the time dimension principal components corresponding to the largest feature correlation factor as the deep implicit feature of the time dimension principal components anchored by the deep implicit feature of the spatial dimension, so as to construct the {deep implicit feature of the spatial dimension, anchored deep implicit feature of the time dimension principal components} feature pair, which is expressed by the formula:

[0045]

[0046]

[0047] Among them, represents calculating the inner product, represents the smoothing coefficient, Denotes the calculation of the norm, Denotes the index corresponding to the maximum value, Denotes the index of the time - dimension principal - component deep implicit feature corresponding to the maximum eigen - correlation factor, Denotes the anchored time - dimension principal - component deep implicit feature, Denotes the feature pair {spatial - dimension deep implicit feature, anchored time - dimension principal - component deep implicit feature}.

[0048] That is, in order to filter out the noise and redundant components in the set of the time - dimension principal - component deep implicit features, the present application further performs implicit correlation analysis on the set of the spatial - dimension deep implicit features and the time - dimension principal - component deep implicit features to capture the implicit key clues between the spatial - dimension features and the time - dimension features, so as to achieve semantic alignment between features. In a specific implementation, by calculating the correlation between the spatial - dimension deep implicit feature and each time - dimension principal - component deep implicit feature, the time - dimension principal - component deep implicit feature most relevant to the spatial - dimension deep implicit feature is extracted as the key clue, and it is used as the anchor point to construct the feature pair {spatial - dimension deep implicit feature, anchored time - dimension principal - component deep implicit feature}, thereby endowing the feature interaction with a clear structural prior through the alignment operation and eliminating the ambiguity between features.

[0049] Specifically, in a specific example of the present application, the step S33 includes: inputting the feature pair {spatial - dimension deep implicit feature, anchored time - dimension principal - component deep implicit feature} into the feature fine - grained interaction network to obtain the spatio - temporal fusion feature. More specifically, the step S33 includes: using the spatial - dimension deep implicit feature as the query vector, using the spatial - dimension feature as the value vector, using the anchored time - dimension principal - component deep implicit feature as the key vector, and inputting the query vector, the key vector, and the value vector into the feature attention interaction module based on the Transformer architecture to obtain the spatial - dimension - time - dimension attention interaction feature; performing linear weighted fusion on the spatial - dimension - time - dimension attention interaction feature and the time - dimension feature principal - component component corresponding to the anchored time - dimension principal - component deep implicit feature to obtain the spatio - temporal fusion feature, which is represented by the formula:

[0050]

[0051] Where, Denotes the normalized exponential function, And Denote different weight parameters respectively, Denotes the feature scale scaling parameter, Denotes the vector multiplication operation, Denotes the said The corresponding principal component component of the time dimension feature represents the spatio-temporal fusion feature

[0052] Here, the present application uses a feature fine-grained interaction network to process the anchored feature pairs, so as to mine more complex non-linear feature associations between the spatial dimension feature and the time dimension feature on the basis of the existing semantic alignment. In a specific implementation, the feature fine-grained interaction network is based on the self-attention mechanism. By calculating the attention weights between the anchored feature pairs, the association strength between different features is dynamically adjusted to capture a more refined feature interaction pattern between the spatial dimension feature and the time dimension principal component feature, improving the accuracy and efficiency of feature interaction analysis. In addition, the feature fine-grained interaction network further enhances the integrity of feature representation by introducing the original spatial dimension feature and the anchored principal component component of the time dimension feature, obtaining the spatio-temporal fusion feature, which helps to more comprehensively understand and predict the navigation behavior of the ship.

[0053] Specifically, in a preferred example of the present application, the step S33 includes: First, perform feature power-law distribution alignment constraint on the deep implicit feature of the spatial dimension and the deep implicit feature of the anchored principal component of the time dimension to obtain an {optimized deep implicit feature of the spatial dimension, optimized deep implicit feature of the anchored principal component of the time dimension} feature pair, which is expressed by the formula as:

[0054]

[0055]

[0056] where represents the exponential function with the natural constant as the base and are respectively the feature values at the and th positions in the vectors and represents the optimized deep implicit feature of the spatial dimension represents the optimized deep implicit feature of the anchored principal component of the time dimension

[0057] Then, input the {optimized deep implicit feature of the spatial dimension, optimized deep implicit feature of the anchored principal component of the time dimension} feature pair into the feature fine-grained interaction network to obtain the spatio-temporal fusion feature

[0058] Here, in view of the alignment ambiguity caused by source uncertainty between the spatial dimension and time dimension features, the present application introduces a weakening and blurring mechanism for weak blurring power-law expansion. That is, the 3 / 8 exponent is used as the precursor prior expansion to perform power-law prior responsive blurring convergence on the feature parameter boundary alignment condition, and the 1 / 4 exponent is used as the main body alignment expansion to strictly constrain the relaxation of the feature power-law distribution alignment decay. Thus, in the case where the alignment boundary condition constraints within the correlation effective range are ill-defined, the value correlation mechanism is used to avoid the prior ambiguity of the system behavior under a single mechanism, so as to correct the semantic distribution consistency blurring within the alignment interval, and further improve the mining intuitiveness of the implicit fine-grained interaction correlation features between the deep implicit features of the spatial dimension and the deep implicit features of the anchored time dimension principal components.

[0059] In the above ship trajectory prediction method based on spatio-temporal feature fusion, in step S4, based on the spatio-temporal fusion features, ship trajectory information is predicted. It should be understood that predicting ship trajectory information based on the spatio-temporal fusion features is a prior art. For example, it can use the principle disclosed in Chinese Patent CN117933492A to predict ship trajectory information. Of course, it can also use other principles to predict ship trajectory information, and this is not limited by the present application. For example, an LSTM (Long Short-Term Memory Network) can be used to construct a prediction model to utilize the excellent long-term memory ability of the LSTM model to capture the long-term dependence relationship of the ship's navigation behavior in the spatio-temporal fusion features, so as to generate a prediction result for the next time point based on the known spatio-temporal state of the ship trajectory. By repeating this process until the predetermined prediction length is reached. It should be noted that after each prediction is completed, the newly generated trajectory information can be fed back into the model as the input for the next stage of prediction, so as to achieve rolling prediction.

[0060] Specifically, when processing ship trajectory data, the fusion of time dimension features and spatial dimension features makes it possible to understand complex navigation patterns. The LSTM model, through its unique gating mechanism, can effectively learn and remember important information in a long time series while ignoring irrelevant details, which makes it an ideal choice for dealing with such problems.

[0061] The core of the LSTM model lies in its cell state, which is an information flow running through the entire chain structure, allowing information to flow in a controlled manner. At each time step, the LSTM decides which new information will be added to this state and which old information should be forgotten. This mechanism makes the LSTM particularly good at dealing with data with long-time span correlations, such as information about the position, speed, and direction of a ship changing over time at sea.

[0062] When designing the architecture of the LSTM model, different numbers of layers and units of LSTM layers can be selected according to specific task requirements and available computing resources. Generally speaking, increasing the number of LSTM layers can improve the model's ability to capture complex patterns, but it will also increase the training time and the risk of overfitting. Therefore, it is necessary to weigh the pros and cons according to the actual situation. In addition, a fully connected layer (Dense Layer) can be considered to be added after the LSTM layer to further extract high-level features and output the final prediction results. In this process, the selection of activation functions is also very important. For example, using ReLU (Rectified Linear Unit) as the activation function of the hidden layer can speed up the training, while Sigmoid or Softmax is often used in the output layer.

[0063] It is worth noting that although the LSTM model has powerful long-term memory capabilities, in practical applications, attention still needs to be paid to preventing the problems of gradient vanishing or explosion. These problems may cause the model to be difficult to converge or fall into a local optimal solution. For this reason, some technical means can be taken, such as setting reasonable initial weights, using methods like Gradient Clipping to alleviate these problems.

[0064] Finally, when generating the prediction result for the next time point based on the known spatio-temporal state of the ship's track, the LSTM model not only depends on the input data at the current moment but also refers to the memory information of multiple previous time steps. This means that for each ship, the LSTM model can infer the possible future sailing path based on the ship's sailing history over a period of time in the past, combined with the real-time obtained spatial position and motion parameters. This characteristic makes the LSTM particularly powerful in dealing with prediction tasks in a dynamically changing environment, which helps to discover potential risks in advance, optimize the route planning, and thus improve the overall shipping efficiency and safety.

[0065] In summary, the ship track prediction method based on spatio-temporal feature fusion according to the embodiments of the present application is elucidated. It uses a neural network model based on deep learning to process the AIS information at each time in the known ship track to extract the time dimension features and space dimension features of the ship track. Then, through the implicit association and explicit modeling analysis of the time dimension features and space dimension features of the ship track, the implicit key clues between the two are mined, and based on this, the alignment and interaction fusion of the time dimension features and space dimension features of the ship track are realized. Furthermore, the future track of the ship is predicted based on the spatio-temporal fusion features of the ship track. In this way, the correlation between the ship's sailing speed, course change, and geographical location can be deeply mined, so as to achieve more accurate feature alignment and interaction fusion, and then improve the accuracy of ship track prediction.

[0066] The basic principles of the present invention have been described above in conjunction with specific embodiments. However, it should be noted that the advantages, benefits, effects, etc. mentioned in the present invention are only examples and not limitations. It cannot be considered that these advantages, benefits, effects, etc. are essential for each embodiment of the present invention. Additionally, the specific details of the above embodiments are only for the purposes of illustration and facilitating understanding, rather than limitations. The above details do not limit the present invention to necessarily adopting the above specific details for implementation.

[0067] In the above embodiments, the descriptions of each embodiment have their own focuses. For parts not detailed or recorded in a certain embodiment, reference may be made to the relevant descriptions of other embodiments. In the several embodiments provided by the present invention, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative. For example, the unit division is only a logical function division, and there may be other division methods in actual implementation. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0068] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and without departing from the spirit or basic characteristics of the present invention, the present invention can be implemented in other specific forms. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting. The scope of the present invention is defined by the appended claims rather than the above description. Therefore, all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be encompassed within the present invention. Any associated drawing marks in the claims should not be regarded as limiting the claimed rights.

[0069] Finally, it should be noted that the above description has been given for the purposes of illustration and description. In addition, the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A ship trajectory prediction method based on spatio-temporal feature fusion, characterized in that, Including: Obtaining known ship trajectories; Based on the AIS information at each time in the known ship trajectories, extracting the time-dimensional features and space-dimensional features of the known ship trajectories. Among them, various attributes in the AIS information are converted into high-dimensional vector representations to form a sequence of AIS information high-dimensional feature vectors. The various attributes include position, speed, and heading. A TCN network is used to process the sequence of AIS information high-dimensional feature vectors to extract time-dimensional features reflecting the ship's motion pattern, and a Transformer network is used to perform self-attention mechanism processing on the sequence of AIS information high-dimensional feature vectors to extract space-dimensional features reflecting the ship's spatial distribution characteristics; Performing spatio-temporal feature interaction fusion based on implicit association explicit modeling on the time-dimensional features and the space-dimensional features to obtain spatio-temporal fusion features; Predicting ship trajectory information based on the spatio-temporal fusion features; Performing spatio-temporal feature interaction fusion based on implicit association explicit modeling on the time-dimensional features and the space-dimensional features to obtain spatio-temporal fusion features, including: Performing principal component analysis on the time-dimensional features to obtain a set of principal component components of the time-dimensional features; Performing implicit association analysis on each principal component component of the space-dimensional features and the set of principal component components of the time-dimensional features to construct a feature pair of {deep implicit features of the space dimension, anchored deep implicit features of the principal components of the time dimension}; Performing spatio-temporal feature interaction fusion on the feature pair of {deep implicit features of the space dimension, anchored deep implicit features of the principal components of the time dimension} to obtain the spatio-temporal fusion features; Performing implicit association analysis on each principal component component of the space-dimensional features and the set of principal component components of the time-dimensional features to construct a feature pair of {deep implicit features of the space dimension, anchored deep implicit features of the principal components of the time dimension}, including: Respectively extracting the deep implicit features of each principal component component of the space-dimensional features and the set of principal component components of the time-dimensional features to obtain a set of deep implicit features of the space dimension and deep implicit features of the principal components of the time dimension; Calculating the feature correlation factors between the deep implicit features of the space dimension and each deep implicit feature of the principal components of the time dimension in the set of deep implicit features of the principal components of the time dimension, and selecting the deep implicit feature of the principal components of the time dimension corresponding to the largest feature correlation factor as the anchored deep implicit feature of the principal components of the time dimension of the deep implicit features of the space dimension to construct the feature pair of {deep implicit features of the space dimension, anchored deep implicit features of the principal components of the time dimension}.

2. The ship trajectory prediction method based on spatio-temporal feature fusion according to claim 1, characterized in that Respectively extracting the deep implicit features of each principal component component of the space-dimensional features and the set of principal component components of the time-dimensional features to obtain a set of deep implicit features of the space dimension and deep implicit features of the principal components of the time dimension, including: Perform implicit feature extraction based on point convolution encoding on each of the temporal dimension feature principal component components in the set of the spatial dimension feature and the temporal dimension feature principal component components to obtain the set of the spatial dimension deep implicit features and the temporal dimension principal component deep implicit features.

3. The method for predicting ship trajectories based on spatio-temporal feature fusion according to claim 2, wherein, Perform spatio-temporal feature interaction and fusion on the {spatial dimension deep implicit feature, anchored temporal dimension principal component deep implicit feature} feature pair to obtain the spatio-temporal fusion feature, including: Input the {spatial dimension deep implicit feature, anchored temporal dimension principal component deep implicit feature} feature pair into a feature fine-grained interaction network to obtain the spatio-temporal fusion feature.

4. The method for predicting ship trajectories based on spatio-temporal feature fusion according to claim 3, characterized in that Input the {spatial dimension deep implicit feature, anchored temporal dimension principal component deep implicit feature} feature pair into a feature fine-grained interaction network to obtain the spatio-temporal fusion feature, including: Use the spatial dimension deep implicit feature as the query vector, the spatial dimension feature as the value vector, and the anchored temporal dimension principal component deep implicit feature as the key vector, and input the query vector, the key vector, and the value vector into a feature attention interaction module based on the Transformer architecture to obtain the spatial dimension - temporal dimension attention interaction feature; Perform linear weighted fusion on the spatial dimension - temporal dimension attention interaction feature and the temporal dimension feature principal component component corresponding to the anchored temporal dimension principal component deep implicit feature to obtain the spatio-temporal fusion feature.

5. The method for predicting ship trajectories based on spatio-temporal feature fusion according to claim 4, characterized in that, Perform spatio-temporal feature interaction and fusion on the {spatial dimension deep implicit feature, anchored temporal dimension principal component deep implicit feature} feature pair to obtain the spatio-temporal fusion feature, including: Perform feature power-law distribution alignment constraint on the spatial dimension deep implicit feature and the anchored temporal dimension principal component deep implicit feature to obtain the {optimized spatial dimension deep implicit feature, optimized anchored temporal dimension principal component deep implicit feature} feature pair; Input the {optimized spatial dimension deep implicit feature, optimized anchored temporal dimension principal component deep implicit feature} feature pair into a feature fine-grained interaction network to obtain the spatio-temporal fusion feature.

Images