A historical track network composition identification planning method and system

Abstract

The present invention provides a method and system for identifying and planning the network composition of historical tracks. The method includes a data acquisition and processing step, a clustering processing step, a cluster track segmentation and path search step, a track matching and secondary clustering step, and a track optimization processing step. By cleaning and integrating the historical ship AIS data using relevant algorithms, and then analyzing it using a clustering algorithm, the method generates tracks, splices tracks, and optimizes the overall track for the classified areas. This method can improve computing efficiency and avoid glitches.

Description

Technical Field

[0001] The present invention relates to the field of shipping information technology, and in particular to a historical track network composition identification and planning method and system. Background Art

[0002] As one of the most important functions of the shipping data platform, trajectory planning is the link that connects various elements of the shipping industry, such as cargo transportation and port operations, and plays a connecting role in the shipping space system. In the current mainstream shipping platform, trajectory planning mainly takes two forms: planning between ports and trajectory planning at arbitrary longitudes and latitudes. Compared with planning trajectory routes of arbitrary longitudes and latitudes in global waters, planning between ports is relatively simple: select high-quality routes between ports and take the median, and add certain business logic restrictions to select relatively reasonable port-to-port trajectories. However, trajectory planning between arbitrary longitudes and latitudes is also very necessary: ​​ships may encounter emergencies such as weather changes and ship anomalies during navigation. Therefore, in real-time navigation, ships need to change their original tracks during navigation; secondly, trajectory planning often involves related factors such as passing through special areas. When planning important tracks, it is often necessary to change the route within a pre-defined specific area. The existing planning between arbitrary coordinate points has the disadvantages of long calculation time, the need to search all tracks within a certain nautical mile of the selected longitude and latitude, low efficiency, the selected tracks are prone to "burrs" and low quality, and the selected tracks are the original tracks after screening, which cannot ensure the quality and representativeness of the selected tracks. Summary of the Invention

[0003] To address existing track planning issues such as long computation time, low efficiency, and poor quality of selected tracks, the present invention provides a historical track network composition identification and planning method. This method uses relevant algorithms to clean and integrate historical ship AIS data, then applies a clustering algorithm for analysis. Track generation, track splicing, and overall track optimization are then performed on the classified areas. This method improves computational efficiency and avoids glitches. The present invention also relates to a historical track network composition identification and planning system.

[0004] The technical solutions of the present invention are as follows:

[0005] A historical track network composition identification and planning method, characterized by comprising the following steps:

[0006] A data collection and processing step is to collect historical AIS data of ships sailing in a specific area and remove track data of adjacent ships whose heading is greater than a first threshold and whose AIS missing rate is greater than a second threshold;

[0007] In the clustering processing step, the Birch algorithm is used to cluster and label the processed track data, and the clusters passed by each track with cluster labels are detected and the network within the sea area is generated with the center point of each cluster as the network node;

[0008] In the intra-cluster track segmentation and path search step, the tracks are segmented according to the adjacent relationships of the clusters that each track with cluster labels passes through. Then, the distances between adjacent network nodes are calculated as the edge weights of the graph on the intra-sea network generated by the center point of each cluster. The Dijkstra algorithm is used to calculate the shortest path from the starting point to the end point of the network node to obtain the node sequence.

[0009] In the track matching and secondary clustering step, the tracks except the starting point and the end point are matched / joined according to the node sequence, and the matched / joined tracks are then subjected to secondary clustering using the QUICKBUNDLES algorithm;

[0010] In the track optimization processing step, the coordinates of the starting and ending points of the real positions are added to the beginning and end of each track after secondary clustering, and the long-short sequence neural network is used to optimize the track training to obtain the track planning.

[0011] Preferably, in the clustering processing step, the processed track data is clustered and labeled by directly generating labels using the field that identifies the track, or classifying according to the departure time and MMSI code of two adjacent records and labeling them as different categories when the MMSI codes are different; clustering is performed based on the two dimensions of longitude and latitude, so that the track data is divided according to the covered area, and the Birch algorithm is used to mark the center point of each cluster.

[0012] Preferably, in the clustering processing step, the trajectory data with cluster labels are traversed, the clusters passed by each track are detected, and the center point of each cluster is used as the network node. If the track passing between two adjacent clusters exceeds a third threshold, an edge is added between the corresponding network nodes to generate an intra-sea network, and the intra-sea network is a directed graph.

[0013] Preferably, in the intra-cluster track segmentation and path search steps, when the tracks are segmented according to the adjacent relationship of the clusters through which the tracks with cluster labels pass, the track segments within a cluster that come from the same neighboring cluster to another same neighboring cluster are divided into one category; when using the Dijkstra algorithm to calculate the shortest path from the starting point to the end point of the network node, if the starting point or the end point is not the cluster center point, the cluster where the starting point or the end point is located is first calculated, and the cluster to which the coordinates have the shortest Euclidean distance to the center point of the cluster is determined, and the center point of the cluster to which the coordinates belong is temporarily designated as the current starting point or end point.

[0014] Preferably, in the track matching and secondary clustering steps, the tracks except the starting point and the end point are matched / spliced ​​according to the node sequence, and for the clusters passing through the middle, one or a group of tracks are matched in the track fragment library, and an index from the track to the network nodes along the way is established; when the network node cannot match a complete track, the node sequence of length n is divided into two node sequences of length 1 and n-1, which are matched separately, and combined according to the principle of shortest distance at the splicing point. If the match is still not successful, the layer-by-layer order of 2 and n-2, 3 and n-3 is tried in sequence. If all fail, the track segments in each cluster are recursively combined according to the distance between the two points at the splicing point is less than the fourth threshold.

[0015] Preferably, in the track optimization processing step, a long-short sequence neural network is used, and the starting point of the real position, the end point of the real position and the track generated by secondary clustering are taken as input. The track is optimized and completed by training a number of times exceeding the fifth threshold, and a track plan including the starting point and end point of the real position is obtained on the basis of the track generated by secondary clustering.

[0016] A historical track network composition identification and planning system is characterized by comprising a data acquisition and processing module, a clustering processing module, an intra-cluster track segmentation and path search module, a track matching and secondary clustering module, and a track optimization processing module connected in sequence.

[0017] The data collection and processing module collects historical AIS data of ships sailing in a specific area and removes track data of adjacent ships whose headings are greater than a first threshold and whose AIS missing rate is greater than a second threshold;

[0018] The cluster processing module uses the Birch algorithm to cluster and label the processed track data, detects the clusters passed by each track with cluster labels, and generates a network within the sea area with the center point of each cluster as the network node;

[0019] The intra-cluster track segmentation and path search module segments the tracks according to the adjacent relationships of the clusters that each track with cluster labels passes through, and then calculates the distance between adjacent network nodes as the edge weight of the graph on the intra-sea network generated by the center point of each cluster. The Dijkstra algorithm is used to calculate the shortest path from the starting point to the end point of the network node to obtain a node sequence;

[0020] The track matching and secondary clustering module matches / joins the tracks except the starting point and the end point according to the node sequence, and then performs secondary clustering on the tracks generated by matching / joining using the QUICKBUNDLES algorithm;

[0021] The track optimization processing module adds the coordinates of the start point and the end point of the real position to the beginning and end of each track after secondary clustering, and uses a long-short sequence neural network to optimize the track training to obtain track planning.

[0022] Preferably, in the clustering processing module, the processed track data is clustered and labeled by directly generating labels using the field that identifies the track, or classifying them according to the departure time and MMSI code of two adjacent records and labeling them as different categories when the MMSI codes are different; clustering is performed from the two dimensions of longitude and latitude, so that the track data is divided according to the covered area, and the Birch algorithm marks the center point of each cluster; traversing the trajectory data with cluster labels, detecting the clusters passed by each track and using the center point of each cluster as a network node, if the track passing between two adjacent clusters exceeds a third threshold, adding edges between the corresponding network nodes to generate an intra-sea network, and the intra-sea network is a directed graph.

[0023] Preferably, the intra-cluster track segmentation and path search module segments the tracks according to the adjacent relationship of the clusters through which the tracks with cluster labels pass, and divides the track segments within a cluster that come from the same neighboring cluster to another same neighboring cluster into one category; when using the Dijkstra algorithm to calculate the shortest path from the starting point to the end point of the network node, if the starting point or the end point is not the cluster center point, the cluster where the starting point or the end point is located is first calculated, and the cluster to which the coordinates have the shortest Euclidean distance to the center point of the cluster is determined, and the center point of the cluster to which the coordinates belong is temporarily designated as the current starting point or end point.

[0024] Preferably, the track optimization processing module adopts a long-short sequence neural network, takes the starting point of the real position, the end point of the real position and the track generated by secondary clustering as input, optimizes and completes the track by training more than the fifth threshold number of times, and obtains a track plan including the starting point and end point of the real position based on the track generated by secondary clustering.

[0025] The beneficial effects of the present invention are:

[0026] The present invention provides a historical track network composition identification and planning method, which collects AIS historical data of ships sailing in a specific area and removes track data with adjacent ship headings greater than a first threshold and AIS loss rate greater than a second threshold to ensure that there are no round-trip or circling tracks. The processed track data are clustered and labeled using the Birch algorithm, and the clusters passed by each track with cluster labels are detected. The network within the sea area is generated with the center point of each cluster as the network node; the goal is to divide the track data according to the covered area. The Birch algorithm will mark a center point for each area (cluster) and use it as the imaginary coordinate of all track segments in the area. The total number of data after clustering remains unchanged (when the Birch parameter changes, some outliers may be removed), and each AIS historical data record With cluster labels, the tracks are divided and classified according to the cluster labels and saved. The Dijkstra algorithm is used to calculate the shortest path from the starting point to the end point to obtain a node sequence. Then, the tracks except the starting point and the end point are matched / joined according to the node sequence. The tracks of the starting point and the end point start and end from the inside of the cluster, so they need to be generated or directly simulated in the secondary level. For the clusters passing through the middle, one or a group of suitable tracks can be matched in the track fragment library, and the complete tracks are matched first. The QUICKBUNDLES algorithm is then used to perform secondary clustering on the tracks generated by matching / joining. For the large number of tracks between clusters, multiple voyages are analyzed and clustered into one through secondary clustering to facilitate subsequent track planning. The long-short sequence neural network (LSTM) is then combined to optimize the track. This invention primarily utilizes historical AIS data to improve the definition rules for important nodes within closed waters, redefines the waypoints within the waters, and uses the generated node and region division information to generate effective route planning routes. The Birch clustering algorithm is used to classify the longitude and latitude of each track, dividing the track data according to the coverage area. After the regional division, the network within the waters is regenerated based on the regional center points. The shortest path between nodes can then be quickly calculated using algorithms such as Dijkstra. Based on historical AIS data and ship dynamic data, the invention redefines the waypoints within the waters and constructs a route planning network based on the classified clusters. This network is restructured into a multi-layered network: within the current cluster, from the longitude and latitude within the cluster to the current cluster center point (this is the low-level network), and then tracks between clusters are generated (this is the high-level network), thereby improving computational efficiency. Furthermore, the entire completed track is optimized using an LSTM neural network to reduce glitches and other issues. This solution has been validated in areas such as the Bohai Bay in China.

[0027] The present invention also relates to a historical track network composition identification and planning system, which corresponds to the above-mentioned historical track network composition identification and planning method, and can be understood as a system for implementing the above-mentioned historical track network composition identification and planning method, including a data acquisition and processing module, a clustering processing module, an intra-cluster track segmentation and path search module, a track matching and secondary clustering module and a track optimization processing module connected in sequence. Each module works together to clean and integrate data through relevant algorithms, and then uses a clustering algorithm for analysis, and then generates tracks for the classified areas, splices tracks and finally optimizes the overall track. The system uses AIS historical track data to classify tracks and distinguish coverage areas within the sea area. Multiple tracks leading to other adjacent areas are stored in each area, and finally the tracks are spliced ​​and optimized using a deep learning neural network. The effect of track planning for any longitude and latitude in the sea area has been achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 It is a flow chart of the historical track network composition identification planning method of the present invention.

[0029] Figure 2 This is a schematic diagram of cluster labeling and network generation within the Bohai Bay of the present invention.

[0030] Figure 3 This is a schematic diagram of the route planning from Dalian to Tianjin according to the present invention. DETAILED DESCRIPTION

[0031] The present invention will be described below with reference to the accompanying drawings.

[0032] The present invention relates to a method for identifying and planning a historical track network. The flowchart of the method is as follows: Figure 1 As shown, the following steps are included:

[0033] 1. Data collection and processing steps: data collection and preprocessing, namely AIS data extraction / cleaning, collecting historical AIS data of ships sailing in a specific area and removing track data with adjacent ship headings greater than a first threshold and AIS missing rate greater than a second threshold.

[0034] Specifically, taking the Bohai Bay (Dalian to Yingkou) as an example, the trajectory data of ships sailing in the Bohai Bay area can be collected to obtain AIS historical data with latitude and longitude, mmsi, heading cog, issuance time postime, and pre_postime fields (the time span can be from one month to six months), and a single cargo category (otherwise the data volume will be too large). If there is a field that can identify the track, it is best to add it, such as leg_start_postime+leg_end_positme in the dynamics table. SQL results must be sorted in ascending order by the mmsi and postime fields. From the AIS historical data, tracks with an AIS loss rate greater than 0.1 (the second threshold) and a change in the adjacent heading cog greater than 45 (the first threshold) within the track are removed (to ensure that there are no round-trip or circling tracks). The above first and second thresholds are only preferred values, not exclusive limitations. Other values ​​can also be used, such as removing tracks with adjacent COG>60 and an AIS loss rate greater than 0.1.

[0035] 2. Clustering processing steps include using the Birch algorithm to cluster the track data and generate a network within the sea area. The Birch algorithm is used to cluster and label the processed track data, detect the clusters passed by each track with cluster labels, and generate a network within the sea area with the center point of each cluster as the network node.

[0036] 2.1. Clustering track data using the Birch algorithm

[0037] All pre-processed AIS historical data are labeled with traj according to the track they belong to. Two logics: directly generate using the field that can identify the track, or classify according to postime=pre_postime&&mmsi=mmsi of two adjacent records. Use the longitude and latitude dimensions of the cleaned data to make a cluster (the parameters can use the default values). The goal is to divide the track data according to the covered area. The Birch algorithm will mark a center point for each area (cluster). The center point is the arithmetic mean of all data points in the area, and it is used as the imaginary coordinate of all track segments in the area. The total number of data after clustering remains unchanged (when the birch parameters change, some outliers may be removed), and each AIS record is labeled with a clustering label. That is, label each track, cluster it with the longitude lon and latitude lat attributes, divide the area within the Bohai Bay and record the marked center point. Such as Figure 2 The diagram shows cluster labeling (cluster segmentation) and network generation within the Bohai Bay. The labels in the figure are cluster numbers.

[0038] 2.2. Generate a network within the sea area

[0039] Traverse the trajectory data with cluster labels and detect the clusters that each trajectory passes through. Take the center point of each cluster as the network node. If the number of tracks passing between two adjacent clusters exceeds the third threshold, for example, more than ten tracks pass through, then add edges between the corresponding network nodes to generate a network within the sea area, such as Figure 2 The network in the sea area is a directed graph. The generated network can then be used to quickly calculate the shortest path between network nodes using algorithms such as Dijkstra.

[0040] 3. Intra-cluster track segmentation and path search steps, including intra-cluster track segmentation and path search, are to segment the tracks according to the adjacent relationship of the clusters that each track with cluster labels passes through, and then calculate the distance between adjacent network nodes as the edge weight of the graph on the sea area network generated by the center point of each cluster. The Dijkstra algorithm is used to calculate the shortest path from the starting point to the end point of the network node to obtain the node sequence.

[0041] 3.1. Intra-cluster track segmentation

[0042] Tracks are segmented and stored according to cluster labels. The classification rule is to group track segments within a cluster that travel from one neighboring cluster to another neighboring cluster into one category, labeled "xx-xx-xx." For example, a track segment labeled "20-24-13" would include all track segments within cluster 24 that pass through clusters 20, 24, and 13, respectively.

[0043] 3.2 Path Search

[0044] On a network of cluster centers, the geographic distance (Haversine distance) between adjacent network nodes is calculated as the edge weight of the graph. Dijkstra's algorithm is then used to calculate the shortest path from the starting point to the end point, resulting in a node sequence. In most cases, the starting and end points are not exactly cluster centers. In this case, the clusters to which the starting and end points belong are first calculated (calculation method: the coordinates are assigned to the cluster with the closest Euclidean distance to the center point, a property guaranteed by the Birch algorithm). The center point of that cluster is temporarily designated as the current starting or end point. Dijkstra's algorithm is then used to calculate the shortest path from the starting point to the end point, resulting in a node sequence. Here, the sequence from Dalian to Tianjin is: 17-12-20-24-13.

[0045] 4. Track matching and secondary clustering steps, including track matching / splicing and track secondary clustering, are to match / splice the tracks except the starting point and the end point according to the node sequence, and then use the QUICKBUNDLES algorithm to perform secondary clustering on the tracks generated by matching / splicing.

[0046] 4.1 Track matching / stitching

[0047] According to the node sequence of the shortest path obtained, the clusters except the clusters where the starting point and the end point are located are matched and spliced. This embodiment matches the tracks except the clusters where the starting point and the end point are located according to the sequence (matching 12-20-24). The tracks of the starting point and the end point start or end from the inside of the cluster, so they need to be generated or directly simulated and completed at the secondary level. For the clusters passed through in the middle, one or a group of suitable tracks can be matched in the track fragment library, with priority given to matching complete tracks. For this purpose, an index from the track to the network nodes along the way is established to match the tracks that pass through these clusters in sequence according to any cluster sequence. When the network node sequence cannot match a complete track, multiple different track segments are matched according to the following logic: the sequence of length n is divided into two sequences of length 1 and n-1, and matched separately, and then combined according to the principle of shortest distance at the splicing point; if a match is still not successful, then try 2 and n-2, 3 and n-3, and so on. If all fail, the track segments within each cluster are recursively combined based on the fourth threshold of the distance between the two points at the splicing point being less than 30 kilometers.

[0048] 4.2 Secondary Clustering

[0049] The generated tracks (matched / concatenated) are clustered again using the QUICKBUNDLES algorithm in DIPY (with a threshold of 15). This step aims to analyze the large number of tracks between clusters and cluster them into a single track, facilitating subsequent track planning. For the sequence 12-20-24, for example, the tracks between 12-20 and 20-24 are extracted and clustered using the QUICKBUNDLES method. Based on the track distribution, one or more representative tracks are obtained.

[0050] 5. Track optimization processing step: add the start and end point coordinates of the real position to the beginning and end of each track after secondary clustering, and use the long-short sequence neural network (LSTM) to optimize the track training to obtain the track planning.

[0051] This step is to add the coordinates of the starting point and the ending point of the real position to the cluster center track obtained in (4.2) and redraw them using LSTM. Figure 2The obtained track sequence 12-20-24 uses the starting point coordinates to link the starting point of the real position of the track segment in cluster 17, and the end point coordinates to link the end point of the real position of the track in cluster 13. Using the LSTM network (hidden layer with 128 neurons), the starting point of the real position, the end point of the real position and the clustered track generated by the secondary clustering in (4.2) are input. The track is finally optimized and completed by training 3000 times (the fifth threshold). Based on the track generated by the secondary clustering, a track plan including the starting point and end point of the real position is obtained, as shown in the figure. Figure 3 The following diagram shows a route plan from Dalian to Tianjin. The start and end points are actual geographic locations, not cluster centers. This assumes that the LSTM-based trajectory from the start point to the first cluster (e.g., 17-12 in the example above) and the end point to the penultimate cluster (e.g., 24-13 in the example above) is reliable. This assumption is based on the assumption that there are sufficient Birch clusters to prevent single points from crossing land boundaries when connecting to other clusters.

[0052] The present invention also relates to a historical track network composition identification and planning system, which corresponds to the above-mentioned historical track network composition identification and planning method, and can be understood as a system for implementing the above-mentioned historical track network composition identification and planning method, including a data acquisition and processing module, a clustering processing module, an intra-cluster track segmentation and path search module, a track matching and secondary clustering module and a track optimization processing module connected in sequence. The modules work together to clean and integrate data through relevant algorithms, and then use clustering algorithms for analysis, and then generate tracks for the classified areas, track splicing and finally overall track optimization.

[0053] The data acquisition and processing module collects the AIS historical data of ships sailing in a specific area and removes the track data of adjacent ships with a heading greater than a first threshold and an AIS missing rate greater than a second threshold; the clustering processing module uses the Birch algorithm to cluster and label the processed track data, detects the clusters passed by each track with cluster labels, and generates an intra-sea network with the center point of each cluster as the network node; the intra-cluster track segmentation and path search module segments the track according to the adjacent relationship of the clusters passed by each track with cluster labels, and then generates an intra-sea network at the center point of each cluster. The distances between adjacent network nodes are calculated as the edge weights of the graph, and the Dijkstra algorithm is used to calculate the shortest path from the starting point to the end point of the network node to obtain a node sequence; the track matching and secondary clustering module matches / joins the tracks except the starting point and the end point according to the node sequence, and then performs secondary clustering on the tracks generated by matching / joining using the QUICKBUNDLES algorithm; the track optimization processing module adds the starting point and end point coordinates of the real position to the beginning and end of each track after the secondary clustering, and uses a long-short sequence neural network to optimize the track training to obtain track planning.

[0054] Furthermore, in the clustering processing module, the processed track data is clustered and labeled by directly generating labels using the field that identifies the track, or classifying them according to the departure time and MMSI code of two adjacent records and labeling them as different categories when the MMSI codes are different; clustering is performed from the two dimensions of longitude and latitude, so that the track data is divided according to the covered area, and the Birch algorithm marks the center point of each cluster; the trajectory data with cluster labels is traversed, the clusters passed by each track are detected, and the center point of each cluster is used as a network node. If the track passing between two adjacent clusters exceeds a third threshold, an edge is added between the corresponding network nodes to generate an intra-sea network, and the intra-sea network is a directed graph.

[0055] Furthermore, the intra-cluster track segmentation and path search module segments the tracks according to the adjacent relationship of the clusters through which the tracks with cluster labels pass, and classifies the track segments within a cluster that come from the same neighboring cluster to another neighboring cluster into one category; when using the Dijkstra algorithm to calculate the shortest path from the starting point to the end point of the network node, if the starting point or the end point is not the cluster center point, the cluster where the starting point or the end point is located is first calculated, and the cluster to which the coordinates have the shortest Euclidean distance to the center point of the cluster is determined, and the center point of the cluster to which the coordinates belong is temporarily designated as the current starting point or end point.

[0056] Furthermore, the track optimization processing module adopts a long-short sequence neural network, takes the starting point of the real position, the end point of the real position and the track generated by secondary clustering as input, optimizes and completes the track by training more than the fifth threshold number of times, and obtains a track plan including the starting point and end point of the real position based on the track generated by secondary clustering.

[0057] The historical track network composition identification planning method and system provided by the present invention mainly utilizes AIS historical data, improves the definition rules of important nodes in closed sea areas, redefines the road network nodes (WAY POINT) in the sea area, and uses the generated network nodes, regional division and other information to generate an effective track planning route. The Birch clustering algorithm is used to classify the longitude and latitude of each track, and the track data is divided according to the coverage area. After the area is divided, the network in the sea area is regenerated according to the regional center point. The shortest path between the nodes can then be quickly calculated according to algorithms such as Dijkstra. The present invention redefines the network nodes (way points) in the sea area based on AIS historical data and ship dynamic data and constructs a track planning road network based on the classified clusters. The road network is changed to a multi-level road network: from the longitude and latitude in the cluster to the center point of the current cluster (this is a low-level road network), the tracks between clusters are regenerated (this is a high-level road network). This is to achieve the purpose of improving computing efficiency. At the same time, the entire completed track is optimized twice through the LSTM neural network to make the track less prone to burrs and other problems, and the effect of track planning at any longitude and latitude in the sea area has been achieved.

[0058] It should be noted that the specific embodiments described above can enable those skilled in the art to more fully understand the present invention, but do not limit the present invention in any way. Therefore, although this specification has described the present invention in detail with reference to the drawings and embodiments, those skilled in the art should understand that the present invention can still be modified or replaced with equivalents. In short, all technical solutions and improvements that do not depart from the spirit and scope of the present invention should be included in the scope of protection of the patent for the present invention.

Claims

1. A historical track network composition identification and planning method, characterized in that: The following steps are involved: A data collection and processing step is to collect historical AIS data of ships sailing in a specific area and remove track data of adjacent ships whose heading is greater than a first threshold and whose AIS missing rate is greater than a second threshold; Clustering processing step: Birch algorithm is used to cluster and label the processed track data, detect the clusters passed by each track with cluster label, and use the center point of each cluster as the network node to generate the network within the sea area; In the intra-cluster track segmentation and path search step, the tracks are segmented according to the adjacent relationships of the clusters that each track with cluster labels passes through. Then, the distances between adjacent network nodes are calculated as the edge weights of the graph on the intra-sea network generated by the center point of each cluster. The Dijkstra algorithm is used to calculate the shortest path from the starting point to the end point of the network node to obtain the node sequence. The track matching and secondary clustering steps are as follows: the tracks except the starting point and the end point are matched / joined according to the node sequence, and the tracks generated by the matching / joining are then clustered again using the QUICKBUNDLES algorithm. Specifically, the tracks except the starting point and the end point are matched / joined according to the node sequence, and for the clusters passing through the middle, one or a group of tracks is matched in the track fragment library, and an index is established from the track to the network nodes along the way. When the network node cannot match a complete track, the node sequence of length n is divided into two node sequences of length 1 and n-1, and matched separately, and combined according to the principle of the shortest distance at the splicing point. If the match is still not successful, the layer-by-layer order of 2 and n-2, 3 and n-3 is tried in turn. If all fail, the track segments in each cluster are recursively combined according to the distance between the two points at the splicing point is less than the fourth threshold. In the track optimization processing step, the coordinates of the starting and ending points of the real positions are added to the beginning and end of each track after secondary clustering, and the long-short sequence neural network is used to optimize the track training to obtain the track planning.

2. The historical track network composition identification and planning method according to claim 1, characterized in that: In the clustering step, the processed track data is clustered and labeled. Labels are directly generated using the field that identifies the track, or two adjacent records are classified according to the departure time and MMSI code and labeled as different categories when the MMSI codes are different. Clustering is performed based on the longitude and latitude dimensions, thereby dividing the track data according to the covered area, and the Birch algorithm is used to mark the center point of each cluster.

3. The historical track network composition identification and planning method according to claim 2, characterized in that: In the clustering processing step, the trajectory data with cluster labels are traversed, the clusters passed by each track are detected, and the center point of each cluster is used as a network node. If the number of tracks passing between two adjacent clusters exceeds a third threshold, an edge is added between the corresponding network nodes to generate an intra-sea network, which is a directed graph.

4. The historical track network composition identification and planning method according to any one of claims 1 to 3, characterized in that: In the intra-cluster track segmentation and path search steps, when the tracks are segmented according to the neighboring relationship of the clusters through which the tracks with cluster labels pass, the track segments within a cluster that come from the same neighboring cluster to another neighboring cluster are classified into one category; when the Dijkstra algorithm is used to calculate the shortest path from the starting point to the end point of the network node, if the starting point or the end point is not the cluster center point, the cluster where the starting point or the end point is located is first calculated, and the cluster to which the coordinates have the shortest Euclidean distance to the center point of the cluster is determined, and the center point of the cluster to which the coordinates belong is temporarily designated as the current starting point or end point.

5. The historical track network composition identification and planning method according to any one of claims 1 to 3, characterized in that: In the track optimization processing step, a long-short sequence neural network is used, and the starting point of the real position, the end point of the real position and the track generated by secondary clustering are used as input. The track is optimized and completed by training a number of times exceeding the fifth threshold, and a track plan including the starting point and end point of the real position is obtained based on the track generated by secondary clustering.

6. A historical track network composition identification and planning system, characterized in that: It includes a data acquisition and processing module, a clustering processing module, an intra-cluster track segmentation and path search module, a track matching and secondary clustering module, and a track optimization processing module. The data collection and processing module collects historical AIS data of ships sailing in a specific area and removes track data of adjacent ships whose headings are greater than a first threshold and whose AIS missing rate is greater than a second threshold; The cluster processing module uses the Birch algorithm to cluster and label the processed track data, detects the clusters passed by each track with cluster labels, and generates a network within the sea area with the center point of each cluster as the network node; The intra-cluster track segmentation and path search module segments the tracks according to the adjacent relationships of the clusters that each track with cluster labels passes through, and then calculates the distance between adjacent network nodes as the edge weight of the graph on the intra-sea network generated by the center point of each cluster. The Dijkstra algorithm is used to calculate the shortest path from the starting point to the end point of the network node to obtain a node sequence; The track matching and secondary clustering module matches / joins the tracks except the starting point and the end point according to the node sequence, and then performs secondary clustering on the tracks generated by matching / joining using the QUICKBUNDLES algorithm. Specifically, the tracks except the starting point and the end point are matched / joined according to the node sequence, and for the clusters passing through the middle, one or a group of tracks is matched in the track fragment library, and an index from the track to the network nodes along the way is established. When the network node cannot match a complete track, the node sequence of length n is divided into two node sequences of length 1 and n-1, and matched respectively, and combined according to the principle of shortest distance at the splicing point. If the match is still not successful, the layer-by-layer order of 2 and n-2, 3 and n-3 is tried in sequence. If all fail, the track segments in each cluster are recursively combined according to the distance between the two points at the splicing point is less than the fourth threshold. The track optimization processing module adds the coordinates of the start point and the end point of the real position to the beginning and end of each track after secondary clustering, and uses a long-short sequence neural network to optimize the track training to obtain track planning.

7. The historical track network composition identification and planning system according to claim 6, characterized in that: In the clustering processing module, the processed track data is clustered and labeled. Labels are directly generated using the field that identifies the track, or two adjacent records are classified according to the departure time and MMSI code and labeled as different categories when the MMSI codes are different; clustering is performed based on the longitude and latitude dimensions, thereby dividing the track data according to the covered area, and the Birch algorithm marks the center point of each cluster; the trajectory data with cluster labels are traversed, the clusters passed by each track are detected, and the center point of each cluster is used as a network node. If the track passing between two adjacent clusters exceeds a third threshold, an edge is added between the corresponding network nodes to generate an intra-sea network, which is a directed graph.

8. The historical track network composition identification and planning system according to claim 6 or 7, characterized in that: The intra-cluster track segmentation and path search module segments tracks according to the adjacent relationships of the clusters through which each track with a cluster label passes, and classifies track segments within a cluster that come from the same neighboring cluster to another neighboring cluster into one category. When using the Dijkstra algorithm to calculate the shortest path from the starting point to the end point of a network node, if the starting point or the end point is not the cluster center, the cluster to which the starting point or the end point belongs is first calculated. The cluster to which the coordinate has the shortest Euclidean distance to the center point of the cluster is determined, and the center point of the cluster to which the coordinate belongs is temporarily designated as the current starting point or end point.

9. The historical track network composition identification and planning system according to claim 6 or 7, characterized in that: The track optimization processing module adopts a long-short sequence neural network, takes the starting point of the real position, the end point of the real position and the track generated by secondary clustering as input, optimizes and completes the track by training a number of times exceeding the fifth threshold, and obtains a track plan including the starting point and end point of the real position based on the track generated by secondary clustering.

Images