Track-driven sea target identification method

Abstract

The invention relates to a track-driven sea target identification method, belongs to the field of track data mining, and solves the problems that the category of a sea target cannot be identified in real time and track data cannot be divided automatically in an existing method. Comprising the steps that multiple pieces of original AIS track data of a marine target are obtained and preprocessed, and multiple track point data sequences in each piece of AIS track data and a two-dimensional image corresponding to each piece of track data are obtained based on a preprocessing result; each piece of track data comprises a track number, a marine target category and a plurality of pieces of track point data; a sea target identification model is constructed, a training data set is constructed based on the track point data sequence and the corresponding two-dimensional image to train the sea target identification model, and a trained model is obtained; and obtaining to-be-identified maritime target track data in real time for preprocessing, obtaining a plurality of track point data sequences corresponding to the track data, and inputting the two-dimensional image into the trained model to obtain a maritime target category prediction result. And real-time accurate identification of marine target categories is realized.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of trajectory data mining, and particularly relates to a track-driven maritime target recognition method. BACKGROUND

[0002] Maritime target recognition is an important function in intelligent ship and shipping management systems. It can realize the recognition of maritime target type, navigation state, speed and other information through the analysis of Automatic Identification System (AIS) data. However, the traditional maritime target track recognition method mainly relies on manual analysis, which is inefficient and prone to errors.

[0003] With the development of artificial intelligence technology, especially the application of deep learning algorithm, the recognition method of maritime targets such as ships and warships has been significantly improved. However, the existing maritime target type recognition method uses complete long track for recognition, and it is difficult to recognize short track; the recognition of complete long track is a lagging recognition, and it cannot perform real-time type recognition. SUMMARY

[0004] In view of the above analysis, the embodiments of the present application aim to provide a track-driven maritime target recognition method to solve the technical problems that the existing maritime target type recognition method cannot recognize the maritime target type in real time, and cannot automatically segment the long track data for type recognition.

[0005] A track-driven maritime target recognition method, comprising the following steps:

[0006] Step S1, obtaining a plurality of original AIS track data of maritime targets and preprocessing, obtaining a data sequence of a plurality of track points in each AIS track data and a two-dimensional image corresponding to each AIS track data based on the preprocessing result; wherein each track data includes track number, maritime target type and a plurality of track point data;

[0007] Step S2, constructing a maritime target recognition model, and training the maritime target recognition model based on the track point data sequence and the corresponding two-dimensional image to obtain a trained maritime target recognition model;

[0008] Step S3, real-time preprocessing of AIS track data of the maritime target to be recognized, obtaining a plurality of track point data sequences and corresponding two-dimensional images corresponding to the AIS track data, inputting the trained maritime target recognition model, and obtaining a maritime target type prediction result.

[0009] Further, the data sequence q of the kth track point of the ith track is represented as follows:

[0010]

[0011] wherein, id is the track number of the AIS track data corresponding to the track point; t is the time stamp converted into hours; x, y are the plane rectangular coordinates of the track point; v is the sailing speed; φ is the angular acceleration; z v ,z α ,z ω are the Z-scores of the sailing speed, acceleration and angular acceleration, respectively;

[0012] obtaining a two-dimensional image corresponding to each AIS track data, comprising:

[0013] constructing a two-dimensional image in a two-dimensional space based on each AIS track data after preprocessing; wherein each track point corresponds to a pixel position in the image, and the pixel point information is the corresponding track point data after preprocessing;

[0014] adjusting all two-dimensional images to a uniform size.

[0015] Further, each track point data includes: latitude and longitude coordinate position, sailing speed, heading angle, time stamp; the original AIS track data set is preprocessed, that is, the track point data in the AIS track data is preprocessed, comprising:

[0016] data deduplication, outlier processing and missing value filling are performed;

[0017] convert the time stamp into hours t;

[0018] calculate the acceleration a based on the sailing speed v;

[0019] calculate the angular acceleration φ based on the heading angle;

[0020] convert the latitude and longitude coordinate position into plane rectangular coordinates x, y;

[0021] perform Z-score standardization on the sailing speed, acceleration and angular acceleration, and obtain the Z-scores of the sailing speed, acceleration and angular acceleration, respectively z v ,z α ,z ω .

[0022] Further, the maritime target recognition model comprises a track automatic segmentation module, a local recognition module, a global recognition module and a maritime target category evaluation module, and is obtained by the following method:

[0023] The training data set of the automatic track segmenting module, the local recognition module, the global recognition module, and the offshore target category evaluation module are respectively constructed; wherein the training data set of the automatic track segmenting module includes all track point data sequences, and the sample label is a state label; the training data set of the local recognition module is the track segment data obtained by the track segmenting module, and the label is an offshore target category label; the training data set of the global recognition module is a two-dimensional image corresponding to each AIS track data, and the label is an offshore target category label; the local recognition vector obtained by the local recognition module and the global recognition vector obtained by the global recognition module are combined to form a set recognition vector as the training data set of the offshore target category evaluation module, and the label is an offshore target category label;

[0024] The training data set of the track segmenting module is loaded, the automatic track segmenting module is trained by using a self-attention mechanism and KMeans clustering, a self-defined Loss loss function is used, an Adam optimizer is used for back propagation to update the weight of the track segmenting module, and a trained track segmenting module is obtained;

[0025] The training data set of the local recognition module is loaded, the local recognition module is trained by inputting samples into a 1DCNN and an LSTM at the same time, respectively, a cross-entropy and an MSE loss function are used for the 1DCNN and the LSTM, respectively, and an Adam optimizer is used for back propagation to update the weight of the track segmenting module, and a trained local recognition module is obtained;

[0026] The training data set of the global recognition module is loaded, the global recognition module is trained by using a cross-entropy loss function, an Adam is used for back propagation to update the parameter weight of the global recognition module, and a trained global recognition module is obtained;

[0027] The training data set of the offshore target category evaluation module is loaded, an ID3 algorithm is used to initialize a decision tree, until a preset maximum tree depth and a sample number in a tree node are less than a preset minimum sample number threshold, and a trained offshore target category evaluation module is obtained.

[0028] Based on the trained track segmenting module, the local recognition module, the global recognition module, and the offshore target category evaluation module, a trained offshore target recognition model is obtained.

[0029] The state label includes straight, turning, and anchoring.

[0030] Further, the automatic track segmenting module sequentially includes an input layer, an embedding layer, a first and a second multi-layer attention structure, and a KMeans clustering layer.

[0031] The input layer is used to receive each track point data sequence q corresponding to each track.

[0032] the embedding layer, configured to perform position encoding for each track point data sequence q;

[0033] the first and second multi-layer attention structure connected in sequence, configured to obtain a weighted sum vector O' of each track data res ;

[0034] the Kmeans clustering layer, configured to perform Kmeans clustering on the feature vector z res of each track point in O i,k to obtain an AIS segmentation result.

[0035] Further, the track automatic segmentation module is trained by loading a training data set of the track segmentation module, comprising:

[0036] Step 1, after each track point data sequence q in the training data set of the track segmentation module is input to the input layer and then to the embedding layer for position encoding, the first multi-layer attention structure is input.

[0037] Step 2, the first multi-layer attention structure calculates the attention score matrix by the self-attention and the learnable parameter matrix in the multi-head attention mechanism layer to obtain the Q, K, and V vectors of q, multiplies the attention score matrix by the scaling factor to obtain the scaled attention score matrix.

[0038] Step 3, the scaled attention score matrix is divided into 8 attention heads, and the softmax function is applied to obtain the attention weight matrix W1, W2, W3, …, W8 of each head.

[0039] Step 4, W1, W2, W3, …, W8 and V are multiplied to obtain the self-attention output vector O, including O1, O2, O3, …, O8, O1, O2, O3, …, O8 are spliced by the contract function and averaged along the splicing dimension to obtain the self-attention output vector O input to the feedforward network to obtain the feature vector O linear2 ;

[0040] Step 5, O linear2 is input to the LayerNorm layer to obtain the feature vector O ln , and the self-attention output vector O is connected to the output O ln through a residual connection to obtain the feature vector O res ;

[0041] Step 6, O res is input to the second multi-layer attention structure to obtain the weighted sum vector O' of each track datares ;

[0042] The seventh step is to weight and sum the value vectors in the V matrix using self-attention weights to obtain an AIS feature vector z of the kth time point in the ith track i,k ={z1, z2, …, z n} of each track point, and then perform PCA dimension reduction on the feature vector z i,k After the PCA dimension reduction, the Kmeans algorithm is used for clustering and segmentation to obtain the AIS segmentation result.

[0043] Until the loss function converges, the trained track automatic segmentation model is obtained.

[0044] Further, the feature vector z i,k The Kmeans algorithm is used for clustering and segmentation to obtain the AIS segmentation result, including:

[0045] The PCA principal component analysis is applied to the z i,k for dimension reduction;

[0046] The dimension-reduced z i,k is subjected to a Kmeans algorithm to perform clustering operation to obtain a corresponding classification vector c i,k for each z i,k {bit1, bit2, bit3}.

[0047] The global vector of each AIS track data is regarded as Z i {z i,k , c i,k}, and the z i,k is subjected to continuous time point detection, a preset continuous point threshold K, and a continuous value number N > K in a category, which is regarded as T i,l {z i,j , z i,j+1 , z i,j+2 , …, z i,j+N-1} is a local feature belonging to the l state in the ith AIS track data, and is divided into a segment.

[0048] Each AIS track data is divided into one or more AIS track segmentation data, and the part of the AIS track data that is not segmented is discarded, and each AIS track segmentation data corresponds to a state label.

[0049] Further, the local recognition module is constructed based on a 1DCNN network and an LSTM model, the training data set of the local recognition module is input into the local recognition module for training to obtain a local recognition vector, including:

[0050] respectively, and input into the 1DCNN network and the LSTM model at the same time; i,l {z i,j ,z i,j+1 ,z i,j+2 ,...,z i,j+N-1};

[0051] The 1DCNN network outputs a recognition vector a L ; and the LSTM model outputs a recognition vector b L ;

[0052] The weighted sum of a L and b L is obtained, and the local recognition vector R l is obtained, as shown below:

[0053] R L = a L ·W F +b L ·W S , L = 1, 2, 3

[0054] Wherein, W F is the weight of the 1DCNN network output recognition vector, and W S is the weight of the LSTM model output recognition vector.

[0055] Further, the global recognition module is constructed based on a CNN network, each AIS track data in the training data set of the global recognition module is input into the global recognition module for training, and a global recognition vector corresponding to each track data is obtained, including:

[0056] The uniform size of the two-dimensional image corresponding to each AIS track data is convolved and pooled, the output is applied to ReLU activation, output to the full connection layer, and then to the output layer, and the global recognition vector R4 is obtained.

[0057] Further, the offshore target category recognition module is constructed based on the decision tree initialized by the ID3 algorithm, the set recognition vector in the training data set of the offshore target category evaluation module is input into the offshore target category evaluation module for training, and the offshore target category prediction result is obtained, including:

[0058] First step, initialize the decision tree based on the ID3 algorithm; create a root node and set it as the entry point of the decision tree, and select information gain as the splitting criterion for feature selection;

[0059] Second step, input the local recognition vector R L corresponding to each AIS track data into the offshore target category evaluation module for training.and the set of global recognition vectors R4 is recognized as the recognition vector {R1, R2, R3, R4} as the input of the decision tree;

[0060] The third step is to calculate a gain value based on a splitting criterion for the current node, select a node with the maximum gain value for splitting, and select a node with the minimum impurity as the best split point for the selected split node, and create a child node at the best split point;

[0061] Until the preset maximum tree depth is reached and the number of samples in the tree node is less than the preset minimum sample number threshold, the leaf node is the marine target category prediction result.

[0062] Compared with the prior art, the present application can achieve at least one of the following beneficial effects:

[0063] 1. The method of the present application can process long tracks in real time and achieve automatic segmentation of tracks through the Kmeans clustering algorithm. Each segment of track data, whether long or short, can be identified in real time, thereby improving the processing capacity and recognition accuracy of long tracks and achieving real-time processing and segmentation identification of long tracks. The present application can process and identify short track and long track data in real time.

[0064] 2. The present application realizes automatic feature extraction of AIS track data and real-time identification of marine targets by combining the self-attention mechanism of the CCNet model and the deep learning algorithm, significantly improves the efficiency and accuracy of marine target category identification, and reduces the error rate of manual analysis. The efficiency and accuracy of marine target category identification are improved.

[0065] 3. The traditional method has difficulty in recognizing short tracks. The method of the present application effectively solves the problem of short track category identification by segmenting and extracting local features, so that short tracks can also obtain accurate category prediction results in real time. The short track recognition ability is optimized.

[0066] 4. By combining local recognition vectors and global feature vectors and using a decision tree for final marine target category prediction, the method of the present application not only improves the generalization ability for different types of marine targets, but also enhances the robustness to noise and outliers.

[0067] In the present application, the above technical solutions can be combined with each other to achieve more preferred combination solutions. Other features and advantages of the present application will be described in the subsequent specification, and some advantages will become apparent from the specification or by implementing the present application. The purpose and other advantages of the present application can be achieved and obtained from the contents specifically pointed out in the specification and the drawings. BRIEF DESCRIPTION OF DRAWINGS

[0068] The accompanying drawings are included to provide a further understanding of the embodiments and are incorporated in and constitute a part of this application. The drawings illustrate embodiments for purposes of exemplification. Those of ordinary skill in the art, and the practitioners of the art will appreciate that the aims of the present application are attained by less than all of the illustrations of the embodiments of the drawings.

[0069] Figure 1 A flow chart of a track-driven maritime target recognition method in an embodiment of the present application;

[0070] Figure 2 A schematic diagram of a maritime target recognition model structure in an embodiment of the present application;

[0071] Figure 3 A schematic diagram of a track segmentation model in an embodiment of the present application;

[0072] Figure 4 A schematic diagram of a feedforward network in an embodiment of the present application; DETAILED DESCRIPTION

[0073] The preferred embodiments of the present application will be described in detail below with reference to the drawings, which constitute a part of this application, and are used to explain the principles of the embodiments of the present application, but are not used to limit the scope of the present application.

[0074] It is found through research on maritime target track data that there are obvious differences in track characteristics of different maritime targets in three types of states, i.e., turning, straight running and anchoring. Based on this feature, the track is segmented based on an attention mechanism, maritime target category recognition is performed through local and global track features, the global and local recognition modules are integrated to improve the recognition accuracy of maritime target categories, and this is of great significance to improve maritime supervision in China.

[0075] One specific embodiment of the present application discloses a track-driven maritime target recognition method, as shown in Figure 1 The method comprises the following steps:

[0076] Step S1, a plurality of original AIS track data of maritime targets is acquired and preprocessed, a data sequence of a plurality of track points in each AIS track data and a two-dimensional image corresponding to each AIS track data are obtained based on the preprocessing result; wherein each track data comprises a track number, a maritime target category and a plurality of track point data;

[0077] Exemplarily, the maritime targets comprise fishing boats, cargo ships, cruise ships, cruise liners and the like; the original AIS track data refer to a maritime target track data set.

[0078] Step S2, a maritime target recognition model is constructed, a training data set is constructed based on the track point data sequence and the corresponding two-dimensional image, the maritime target recognition model is trained, and a trained maritime target recognition model is obtained;

[0079] Step S3, real-time acquisition of AIS track data of the to-be-identified marine target for preprocessing, obtaining a plurality of track point data sequences corresponding to the AIS track data and a corresponding two-dimensional image inputting the trained marine target identification model, and obtaining a marine target category prediction result.

[0080] The application realizes an efficient and accurate marine target category identification method by combining AIS track data preprocessing, feature extraction and deep learning model training, can identify and classify various types of marine targets in real time, thereby enhancing the ability of marine supervision and safety management.

[0081] Step S1, comprising steps S11-S12.

[0082] Step S11, acquiring a plurality of original AIS track data of the marine target and preprocessing.

[0083] Each track point data includes: longitude and latitude coordinate position, sailing speed, heading angle, time stamp; the original AIS track data set in each AIS track data is preprocessed, that is, the track point data in the AIS track data is preprocessed, including:

[0084] Data deduplication, outlier processing and missing value filling are performed;

[0085] Convert the time stamp to an hour unit t;

[0086] Calculate the acceleration alpha based on the sailing speed v;

[0087] Calculate the angular acceleration phi based on the heading angle;

[0088] Convert the longitude and latitude coordinate position to a plane rectangular coordinate x, y;

[0089] Z-score standardization is performed on the sailing speed, acceleration and angular acceleration, and Z-score z v ,z α ,z ω .

[0090] Longitude and latitude coordinate position: record the position of the marine target on the earth's surface, expressed in longitude and latitude values;

[0091] Heading angle: the direction of the current sailing of the marine target, expressed as the angle of rotation clockwise from the north direction;

[0092] Sailing speed: the distance moved by the marine target in unit time, expressed in knots (knots);

[0093] Marine target number: a unique number identifying the marine target;

[0094] Timestamp: the specific time point of the track point collection of the AIS track data, in seconds. In each AIS track data, the track data corresponding to one timestamp is one track point.

[0095] Exemplarily, the original AIS track data of the marine target in the present application comes from the AIS track data set provided by the "Golden Dolphin" Cup Algorithm Challenge.

[0096] Step S1 realizes the comprehensive preprocessing of multiple pieces of original AIS track data of the marine target, data preprocessing, specifically as follows:

[0097] (a) Data deduplication, for each AIS track data recorded continuously, compare the track information of each time point of the AIS track data, if the two timestamps, and the track data are completely the same, keep the AIS track data with the earliest timestamp, and delete the subsequent repeated data, in order to reduce data redundancy;

[0098] (b) Outlier processing, for the latitude and longitude coordinate position or sailing speed value exceeding (including lower than, higher than) the reasonable range threshold value, it is identified as an outlier;

[0099] Exemplarily, the reasonable range threshold value is defined as the range of the mean value plus or minus three times the standard deviation of the AIS track data.

[0100] For each outlier, according to its context data (i.e. the data points of the previous and subsequent one time stamp adjacent to the outlier), the value of the outlier is selected to be the average value of the data points of the previous and subsequent adjacent time stamp.

[0101] Exemplarily, considering the average value of the two adjacent points adjacent to the outlier, the average value of the adjacent points is calculated, as shown in formula (1):

[0102]

[0103] Where, V pre is the value of the previous time point of the outlier, and V next is the value of the subsequent time point of the outlier.

[0104] Exemplarily, according to actual needs, more adjacent points can also be considered, that is, all data points within a certain range before and after the outlier are included in the calculation, and the average value of the adjacent points is calculated, as shown in formula (2):

[0105]

[0106] Where, V i is the value of the i-th data point in the range near the outlier, and n is the total number of data points in the range near the outlier, each a data point.

[0107] The calculated average value is used to correct the value of the abnormal point, i.e., the original abnormal value is replaced with the average value.

[0108] The abnormal value correction operation is recorded in the metadata or log of the AIS track data, including the original abnormal value, the corrected value, and the adjacent points used.

[0109] (c) Missing value filling, identify missing values in the original AIS track data set, for each missing value, the original AIS track data is a time-continuous sequence, check whether the time stamp is continuous, non-continuous time stamp may indicate that there is a missing data.

[0110] If the time interval of the data points before and after it is small, linear interpolation is used to fill, as shown in formula (3):

[0111]

[0112] Where y is the missing value, y1 and y2 are the adjacent values before and after y, x is the missing value timestamp, x1 and x2 are the timestamps of the adjacent values y1 and y2;

[0113] t k -t k-1 Less than the preset threshold value, indicating that the time interval is small. Exemplarily, the preset threshold value is set to 15 minutes.

[0114] If the time interval is large, the mean value of the surrounding data (the mean value of the data points of the time stamps before and after) is used to fill.

[0115] Exemplarily, the time interval is 15-30 minutes, indicating that the time interval is large.

[0116] (d) Based on the sailing speed and heading angle of the track point, the acceleration and angular acceleration of each time point are calculated. The calculated speed data is applied to the difference operation to calculate the acceleration value of each time point, as shown in formula (4):

[0117]

[0118] Where t i is the timestamp converted into time in hours, v(t i ) is the corresponding speed at time t i ; v(t i+1 )-v(t i ) is the speed change of the two time points; t i+1 -t i is the time interval (hours) of the two time points.

[0119] Exemplarily, if the timestamp is in seconds, it is converted to hours, divided by 3600.

[0120] The angular acceleration ω(t i ) at each time point is calculated, as shown in equation (5):

[0121]

[0122] where φ(t i+1 ), φ(t i ) are the heading angles of the adjacent two time points, respectively.

[0123] The acceleration and angular acceleration are the rates of change of the sailing speed and heading angle of the maritime target, respectively.

[0124] (e) Coordinate conversion: the latitude and longitude position coordinates are converted into the plane coordinate system coordinates, i.e., the plane rectangular coordinates (x, y), by the map projection method, as shown in equations (6)-(7):

[0125] x = k * λ equation (6)

[0126]

[0127] where λ, are the longitude and latitude values of the latitude and longitude position coordinates, respectively, and k is the scaling factor.

[0128] When the longitude and latitude values are both 0, it is the O(0, 0) point of the plane coordinate system. The scaling factor k is used to adjust the proportion of the coordinates to ensure the accuracy of the converted coordinates.

[0129] (f) Z-score standardization of the sailing speed, acceleration and angular acceleration data, to calculate the Z-scores z v , z α and z ω of the sailing speed, acceleration and angular acceleration, respectively, as shown in equation (8):

[0130]

[0131] where μ v , μ α and μ ω are the mean values of the sailing speed, acceleration and angular acceleration of the original AIS track data set, respectively, and σ v , σ α and σ ω are the standard deviations of the sailing speed, acceleration and angular acceleration of the original AIS track data set, respectively.

[0132] The sailing speed, acceleration and angular acceleration are Z-score standardized to eliminate the influence of different dimensions, so that the mean of the data distribution is 0 and the standard deviation is 1, facilitating the calculation of subsequent steps.

[0133] In step S12, data sequences of a plurality of track points in each AIS track data and a two-dimensional image corresponding to each AIS track data are obtained based on the preprocessing result.

[0134] The data sequence of each track point is constructed, and the data of each track point includes: track number, time stamp t, plane rectangular position coordinate (x, y), heading angle φ, Z-score standardized sailing speed z v , Z-score standardized acceleration z α , and Z-score standardized angular acceleration z ω .

[0135] The data sequence q of the kth track point of the ith track is shown in formula (9):

[0136]

[0137] Where id is the track number of the AIS track data corresponding to the track point; t is the time stamp converted to hours; x and y are the plane rectangular coordinates of the track point; v is the sailing speed; φ is the angular acceleration; z v , z α , and z ω are Z-scores of the sailing speed, acceleration and angular acceleration, respectively.

[0138] The two-dimensional image corresponding to each AIS track data is obtained, including:

[0139] Based on each AIS track data after preprocessing, a two-dimensional image is constructed in a two-dimensional space; each track point corresponds to a pixel position in the image, and the pixel point information is the corresponding track point data after preprocessing;

[0140] All two-dimensional images are adjusted to a uniform size.

[0141] Each AIS track data includes a plurality of track points (one track point corresponds to one time stamp), and each track point corresponds to a data sequence q of 1 track point. Each AIS track data corresponds to a two-dimensional image.

[0142] Step S1 is the preprocessing of a plurality of original AIS track data of the marine target, which ensures the quality and consistency of the track data and the track point data, provides high-quality data for subsequent deep learning training, and lays a foundation for obtaining more accurate marine target category prediction results.

[0143] Step S2 includes S21-S24.

[0144] Construct and train a maritime target identification model. For example... Figure 2 As shown, the maritime target identification model includes an automatic track segmentation module, a local identification module, a global identification module, and a maritime target category evaluation module;

[0145] The automatic track segmentation module, local identification module, global identification module, and maritime target category assessment module are collaborative and sequentially executed components, collectively forming the complete workflow of the maritime target identification model. The automatic track segmentation module receives preprocessed track point data sequences and outputs track segmentation data as input to the local identification module. The local and global identification modules work in parallel, extracting local features from the local identification vector and global features from the global identification vector, respectively. These feature vectors are then used as input to the maritime target category assessment module. The maritime target category assessment module integrates the local and global identification vectors to perform the final maritime target category prediction.

[0146] It was obtained through training using the following method:

[0147] Training datasets are constructed for the automatic track segmentation module, local identification module, global identification module, and maritime target category evaluation module, respectively. The training dataset for the automatic track segmentation module includes all track point data sequences, with status labels as sample labels. The training dataset for the local identification module consists of track segmentation data obtained by the track segmentation module, labeled with maritime target category labels. The training dataset for the global identification module consists of two-dimensional images corresponding to each AIS track data, labeled with maritime target category labels. A set of identification vectors, composed of the local identification vectors obtained by the local identification module and the global identification vectors obtained by the global identification module, is used as the training dataset for the maritime target category evaluation module, labeled with maritime target category labels.

[0148] Load the training dataset of the track segmentation module, train the automatic track segmentation module using self-attention mechanism and KMeans clustering, use a custom Loss loss function, and use Adam optimizer to backpropagate and update the weights of the track segmentation module to obtain the trained track segmentation module.

[0149] Load the training dataset of the local identification module, and simultaneously input the samples into 1DCNN and LSTM to train the local identification module. 1DCNN and LSTM use cross-entropy and MSE loss functions respectively. Both use Adam optimizer to backpropagate and update the weights of the track segmentation module to obtain the trained local identification module.

[0150] Load the training dataset of the global identification module, train the global identification module using the cross-entropy loss function, and use Adam to backpropagate and update the parameter weights of the global identification module to obtain the trained global identification module.

[0151] Load the training dataset of the maritime target category evaluation module, initialize the decision tree using the ID3 algorithm, until the preset maximum tree depth is reached and the number of samples in the tree node is less than the preset minimum number of samples threshold, and obtain the trained maritime target category evaluation module.

[0152] Based on the trained track segmentation module, local identification module, global identification module, and maritime target category evaluation module, a trained maritime target recognition model is obtained.

[0153] The status labels include going straight, turning, and anchoring.

[0154] Step S21: Construction and training of the automatic track segmentation module.

[0155] like Figure 3 As shown, the automatic track segmentation module includes, in sequence, an input layer, an embedding layer, first and second multi-layer attention structures, and a KMeans clustering layer;

[0156] The input layer is used to receive the data sequence q of each track point corresponding to each track;

[0157] The embedding layer is used to perform position encoding for each waypoint data sequence q;

[0158] The first and second layers of attention structures, connected sequentially, are used to obtain the weighted sum vector O' of each track data. res ;

[0159] The K-means clustering layer is used to cluster O' res The feature vector z of each track point i,k Perform K-means clustering to obtain the AIS segmentation results.

[0160] The first and second multi-layer attention structures both include a self-attention and multi-head attention mechanism layer, a feedforward network layer, and a LayerNorm layer normalization.

[0161] The self-attention and multi-head attention mechanism layer is used to obtain the self-attention output vector corresponding to q;

[0162] The feedforward network layer is used to input the self-attention output vector into the feedforward network layer output feature vector O. linear2 ;

[0163] The LayerNorm layer is normalized, and Olinear2 Normalization using LayerNorm layers yields the output feature vector O. ln ;

[0164] The self-attention output is connected to output O via a residual. ln The above yields the eigenvector O. res ;

[0165] The output feature vector of the first multi-layer attention structure is input into the second multi-layer attention structure to obtain the weighted sum vector O' of each track data. res .

[0166] Loading the training dataset of the track segmentation module to train the automatic track segmentation module includes:

[0167] The first step is to process the data sequence q of each track point in the training dataset of the track segmentation module through the input layer and the embedding layer for position encoding, and then input it into the first multi-layer attention structure.

[0168] The second step involves the first multi-layer attention structure calculating the Q, K, and V vectors of q using the learnable parameter matrices in the self-attention and multi-head attention mechanism layers to obtain the attention score matrix. This attention score matrix is ​​then multiplied by a scaling factor. Obtain the scaled attention score matrix;

[0169] The third step is to divide the scaled attention score matrix into 8 attention heads and apply the softmax function to obtain the attention weight matrix W1, W2, W3, ..., W8 for each head;

[0170] Step 4: Multiply W1, W2, W3, ..., W8 with V to obtain the self-attention output vector O, including O1, O2, O3, ..., O8. Concatenate O1, O2, O3, ..., O8 using the contract function and perform average pooling along the concatenation dimension to obtain the self-attention output vector O. Input the input to the feedforward network to obtain the feature vector O. linear2 ;

[0171] Step 5: Place O linear2 Input LayerNorm normalization yields feature vector O ln The self-attention output vector O is then connected to the output O via a residual. ln The above yields the eigenvector O. res ;

[0172] Step 6: Place O res Input the second multi-layer attention structure to obtain the weighted sum vector O' of each track data. res ;

[0173] Step 7: Use self-attention weights to perform a weighted summation of the value vectors in the V matrix to obtain the AIS feature vector z at the k-th time point in the i-th track. i,k ={z1, z2, ..., z n}; The feature vector z of each track point i,k After performing PCA dimensionality reduction, K-means clustering was used to segment the data, and AIS segmentation results were obtained.

[0174] The trained automatic track segmentation model is obtained when the loss function converges.

[0175] (a) Input layer:

[0176] The input layer receives the data sequence q of each track point corresponding to each track and inputs it into the embedding layer.

[0177] (b) Embedding layer

[0178] The embedding layer is used to encode the position of each waypoint data sequence q. The position encoding vector has the same length as the waypoint data sequence vector q. Absolute position encoding is represented using sine and cosine functions, and the product of the two yields the relative position encoding.

[0179] For each track point data sequence q corresponding to a time t, the absolute position encoding for each dimension d is as shown in formula (10):

[0180]

[0181] Where pos is the position of the word in each dimension of the track point data sequence, i is the dimension index of the position encoding vector, and d is the dimension of the position encoding vector.

[0182] Multiply the absolute position codes PE(pos,2i) and PE(pos,2i+1) to obtain the relative position codes, which are used as the position codes for the data sequence q of the waypoints.

[0183] The location-encoded data sequence q is used as the input to the first multi-layer attention structure based on the CCNet model, enabling the multi-layer attention structure to utilize location information when processing the data sequence of waypoints.

[0184] (c) First and second multi-layer attention structure layers

[0185] The first multi-layer attention structure calculates the Query(Q), Key(K), and Value(V) vectors of q using the learnable parameter matrix in the self-attention and multi-head attention mechanism layers based on CCNet; and obtains the attention score matrix using the dot product of the Q vector and the K vector.

[0186] Multiply the attention score matrix by a scaling factor The scaled attention score matrix is ​​obtained; where d k The dimension is K;

[0187] The scaled attention score matrix is ​​divided into 8 attention heads, each of which learns a different representation subspace. For each attention head, the attention weight matrix W is obtained by applying the softmax function using its corresponding segmented attention score matrix. i (i = 1, 2, 3, ..., 8);

[0188] W i Multiplying (i = 1, 2, 3, ..., 8) with V yields the self-attention output vector O, including O1, O2, O3, ..., O8. O1, O2, O3, ..., O8 are concatenated using the `contract` function and average pooled along the concatenation dimension to obtain the final aggregated self-attention output vector O. This vector is then input into the feedforward network to obtain the feature vector O. linear2 ;

[0189] By learning multiple different representation subspaces of the input data in parallel and then aggregating these representations through average pooling, the model's ability to comprehensively understand and express information is enhanced, while improving generalization and reducing parameter redundancy.

[0190] O linear2 Input LayerNorm normalization yields feature vector O ln The self-attention output vector O is then connected to the output O via a residual. ln The above yields the eigenvector O. res ;

[0191] O res Input the second multi-layer attention structure to obtain the weighted sum vector O' of each track data. res .

[0192] (d) Feedforward network

[0193] The self-attention output vector O is input into the feedforward network, which contains two linear layers with the ReLU activation function in between.

[0194] like Figure 4 As shown, the first linear layer of the self-attention output O-input feedforward network;

[0195] The output vector O of the first linear layer linear1 ;O linear1 After passing through the ReLU activation function, the output is a nonlinear transformation; the output of the input to the second linear layer is O.linear2 , which is the final output of the feedforward network.

[0196] The output O of the feedforward network linear2 It serves as input for subsequent layer normalization and residual connection processing, and is used for training and feature extraction of the track segmentation model.

[0197] The feedforward network layer provides stability and expressive power for the automatic track segmentation module, while also promoting the effective flow of gradients in the deep network.

[0198] (e) LayerNorm normalization and residual connectivity

[0199] The output O of the feedforward network linear2 The output feature vector O is obtained by normalizing using LayerNorm. ln The self-attention output O is then added to the output feature vector O via a residual connection. ln The above yields the eigenvector O. res .

[0200] (f) The second-level attention structure and the first-level attention structure have the same specific structure.

[0201] The eigenvector O res As the input to the second multi-layer attention structure, the input is normalized through the self-attention and multi-head attention mechanism layers, the feedforward network, and the LayerNorm layer, finally yielding the output feature vector O'. res .

[0202] The role of the second-level attention structure is to deepen the network's extraction and understanding of data features while maintaining network stability, thereby enhancing the model's expressive power and generalization performance. The model can effectively learn and accumulate feature representations in a deeper network structure.

[0203] The value vectors in matrix V are weighted and summed using self-attention weights to obtain and output the feature vector z. i,k ={z1, z2, ..., z n Let} represent the feature vector of the i-th track at k time points.

[0204] O' res Let z be the weighted sum vector of a track, where each track point in the track data is represented by z. i,k ={z1, z2, ..., z n}

[0205] The weighted sum vector O' of a track data res In, it includes a weighted sum vector z of multiple waypoints. i,k .

[0206] (g) for zi,k Perform PCA principal component analysis for dimensionality reduction and K-means segment classification.

[0207] Candidate segments within a fixed time window N centered on the timestamps of high-attention data are identified, and their features are vectorized and dimensionality reduced by principal component analysis. Then, the K-means clustering algorithm is applied to cluster the track segments.

[0208] For each track point feature vector z of each AIS track data i,k Principal Component Analysis (PCA) was applied to reduce the dimensionality of the features, extract the principal components, and obtain the feature vector z for each track point after dimensionality reduction. i,k .

[0209] For example, the dimension reduction dimension β can be set to be less than z. i,k Dimensions.

[0210] For each dimensionality-reduced feature vector z i,k K-means clustering algorithm is used for segment classification. Since different categories of maritime targets exhibit different track state characteristics, the navigation state of the maritime target tracks is segmented to identify the target category. The dimensionality-reduced track features z... i,k Input the K-means algorithm and perform clustering operations.

[0211] Based on the motion state of the tracks, they are divided into three categories: straight, turning, and stationary. After K-means classification, since it is based on track points, the track features are arranged by time point k and track number i, using a category vector, with the corresponding category represented by a classification vector c. i,k {bit1, bit2, bit3} represents, where bit x The value is 0 or 1, and x is 1, 2, or 3.

[0212] c i,k {1,0,0} represents a straight line; c i,k {0,1,0} indicates a turn; c i,k 0,0,1} is the anchor.

[0213] The global vector of each track is considered as Z. i {z i,k ,c i,k}, will z i,k By using continuous point detection, a continuous point threshold K is set. If N consecutive values ​​of k > K in the same track within a category, then it is considered as T. i,l {z i,j ,z i,j+1 ,z i,j+2,...,z i,j+N-1} represents a representative local feature of track l belonging to this type in track i.

[0214] The automatic track segmentation module processes the feature vector z obtained based on the attention mechanism. i,k Automatic track segmentation is performed. Each continuous AIS track data is divided into different segments based on the movement status of the maritime target (straight ahead, turning, anchored). Each segment represents a continuous movement status.

[0215] The feature vector z of each track point i,k K-means clustering was used to segment the data, resulting in AIS segmentation results, including:

[0216] For z i,k Dimensionality reduction was performed using principal component analysis (PCA).

[0217] The dimensionality-reduced z i,k Perform the K-means algorithm to perform clustering operations and obtain each z i,k The corresponding classification vector c i,k {bit1, bit2, bit3};

[0218] The global vector for each AIS track data is considered as Z. i {z i,k ,c i,k}, for z i,k Continuous time point detection is performed, with a preset continuous point threshold K. If the number of consecutive values ​​of k in a category is N > K, then it is considered as T. i,l {z i,j ,z i,j+1 ,z i,j+2 ,...,z i,j+N-1} represents a local feature belonging to state l in the i-th AIS track data, which is divided into a segment;

[0219] Each AIS track data is divided into one or more AIS track segments, and the AIS data is divided into segments.

[0220] Unsegmented portions of the track data are discarded, and each AIS track segment corresponds to a status label.

[0221] The loss function of the automatic track segmentation module is defined as shown in formulas (11)-(13):

[0222] The three cluster centers μ of kmeans c Given that, in order to reduce the feature distance of similar track segments, the loss function loss1 is set as the feature vector z of each type. i,k Distance from cluster center μ cTo prevent poor track continuity, loss2 is the sum of continuous values, which is then normalized to obtain the loss. At the same time, backpropagation is used to update the model parameters.

[0223]

[0224] Where λ is a hyperparameter used to balance the weights of the two loss components, Loss1 and Loss2.

[0225] Where, N c For the eigenvector z i,k The vector length is given by N, where c is the class and C is the number of waypoints of the same class. For waypoints of the same class, N is the vector length. i NL is the sum of the continuous values ​​of the three types of waypoints for the i-th waypoint. i It is the sum of the track points of the i-th track.

[0226] The Adam optimizer is used to perform backpropagation to update the weights of the track segmentation module until the loss function converges, thus obtaining the trained track segmentation module.

[0227] Step S22: Construct a local recognition module and train the module.

[0228] The local identification module is built based on a 1DCNN network and an LSTM model. The training dataset for the local identification module is input into the module for training to obtain local identification vectors, including:

[0229] The local feature vector T of each sample in the training dataset of the local identification module is respectively... i,L {z i,j ,z i,j+1 ,z i,j+2 ,...,z i,j+N-1 The data is simultaneously fed into both the 1DCNN network and the LSTM model.

[0230] The 1DCNN network outputs a recognition vector a L The LSTM model outputs a recognition vector b. L ;

[0231] For a L and b L By performing a weighted summation, the local identification vector R is obtained. l As shown below:

[0232] R L =a L ·W F +b L ·W S L = 1, 2, 3

[0233] Among them, WF W is the weight of the output recognition vector of the 1DCNN network. S The weights are used to output the identification vectors for the LSTM model.

[0234] The AIS segmentation results (each segment) are used as samples, along with the corresponding maritime target category labels. The sample data is simultaneously input into a 1DCNN network and an LSTM model to identify maritime target categories and obtain local identification vectors.

[0235] Based on the feature vector T corresponding to different track states i,L With global eigenvector Z i ,

[0236] The feature vector T i,L The input is fed into a 1DCNN one-dimensional convolutional neural network, which consists of two convolutional layers and two pooling layers.

[0237] The layers are, in sequence, the first convolutional layer, the first pooling layer, the second convolutional layer, and the second pooling layer.

[0238] The feature map output by the 1DCNN network is reshaped into a three-dimensional tensor with shape (L, F, 1), where L is the length of the input sequence, F is the number of features, and 1 is the number of channels. Max pooling is applied to the output after ReLU activation, with a pooling window size of P = 2 and a stride of Q = 2.

[0239] The Adam optimization algorithm was selected, with a learning rate alpha = 0.001.

[0240] The input sample data is passed through the forward propagation path of a 1DCNN to calculate the predicted output.

[0241] The loss is calculated based on the predicted output and the true label, and the weights of the 1DCNN network are updated through backpropagation.

[0242] The loss function of the 1DCNN network is the cross-entropy loss function, which is suitable for multi-class classification tasks, as shown in Equation (14):

[0243]

[0244] Among them, y i,k Let p represent the true label of the i-th sample be k, and there are K label values ​​and N samples in total. i,k This represents the probability that the i-th sample is predicted to be the k-th label value.

[0245] By fitting the loss function, the inter-class distance was also increased to some extent.

[0246] Because three types of state feature vectors are input into a 1DCNN network, the training output is a three-state identification vector {a1, a2, a3} for maritime targets, where a L ={p i,1 ,p i,2 ,...,p i,k ,}

[0247] Input sample data into the LSTM model, which automatically performs forgetting and updating operations, and uses an internal gating mechanism to handle time dependencies. Each LSTM layer has 50 units, for a total of two layers. The reshaped feature tensor z... i The input is fed into a Long Short-Term Memory (LSTM) network model, which contains an LSTM layer with N neurons, followed by a Dropout layer. The Dropout rate p represents the probability that a neuron is lost.

[0248] p represents data between [0,1];

[0249] When p=0, it means there is no Dropout, and the output of all neurons will be preserved;

[0250] When p=1, the outputs of all neurons are discarded, and in this case, the network does not actually learn anything.

[0251] For example, the Dropout rate value in this invention is between [0.2, 0.5] and is set to 0.3. The specific value can be adjusted according to specific needs.

[0252] Connect a fully connected layer to output the obtained maritime target identification vector.

[0253] The LSTM model uses MSE as the loss function and the Adam optimization algorithm with a learning rate of 0.001. The final output of the three-state identification vectors for the maritime target is {b1, b2, b3}, where b... L ={p i,1 ,p i,2 ,...,p i,k ,}

[0254] Corresponding to different categories c i eigenvector z i Let W F W represents the weights corresponding to the output recognition vector of the 1DCNN network. S The weights corresponding to the identification vectors output by the LSTM model are both of dimension K.

[0255] Local identification vectors R are obtained under the three motion states of straight-ahead, turning, and anchoring. L a, output by the corresponding 1DCNN network L and the output b of the LSTM modelL We obtain the weighted sum.

[0256] The local recognition vector R is obtained through the trained local recognition module. L .

[0257] Step S23: The global identification module is constructed based on a CNN network. The two-dimensional image corresponding to each AIS track data in the training dataset of the global identification module is input into the global identification module for training, resulting in a global identification vector corresponding to each track data, including:

[0258] For each AIS track data corresponding to a uniform-sized two-dimensional image, convolution and pooling are performed, the output is activated by ReLU, and then output to a fully connected layer, and then to the output layer to obtain the global identification vector R4.

[0259] Based on the preprocessed AIS track data, each preprocessed track data corresponds to a two-dimensional image, and each track point corresponds to a pixel position in the two-dimensional image. Pixel information is regarded as track point information.

[0260] Adjust all images to a uniform size; that is, each track corresponds to the same number of track points.

[0261] Set the input layer to match the size of the standardized image;

[0262] Perform convolution and pooling operations on the input 2D image;

[0263] Apply the ReLU activation function to the output of the convolutional layer;

[0264] The output of the convolutional layer is fed to a fully connected layer to further extract feature vectors, using the softmax activation function.

[0265] The feature vectors are input into the output layer, and the probability of each class is output using the softmax activation function.

[0266] Define the loss function as the cross-entropy function. Choose an optimization algorithm, such as Adam, and set the learning rate for weight updates during model training. Repeat the forward propagation, loss calculation, backpropagation, and weight update steps.

[0267] Output the global identification vector R4, representing the category of maritime targets, obtained from the global two-dimensional image, denoted as: R4={p i,1 ,p i,2 ,...,p i,k ,}

[0268] Step S24: Construct a maritime target category identification module and train the module.

[0269] The maritime target category identification module is constructed based on a decision tree initialized using the ID3 algorithm. The set identification vectors from the training dataset of the maritime target category evaluation module are input into the maritime target category evaluation module for training, resulting in maritime target category prediction results, including:

[0270] Step 1: Initialize the decision tree based on the ID3 algorithm; create a root node, set it as the entry point of the decision tree, and select information gain as the splitting criterion for feature selection;

[0271] The second step is to process the local identification vector R corresponding to each AIS track data. L The set of identification vectors {R1, R2, R3, R4}, along with the global identification vector R4, is used as the input to the decision tree;

[0272] Step 3: For the current node, calculate the gain value based on the splitting criterion; select the node with the largest gain value for splitting; for the selected splitting node, select the node with the smallest impurity as the optimal splitting point; create child nodes at the optimal splitting point.

[0273] The prediction continues until the preset maximum tree depth is reached and the number of samples in the tree node is less than the preset minimum number of samples threshold. The leaf node then represents the prediction result of the marine target category.

[0274] For the feature vector set of the current node, calculate the gain value of each feature based on the splitting criterion;

[0275] Select the feature with the largest gain value as the splitting feature of the current node.

[0276] For a selected splitting feature, iterate through all its possible values, calculate the splitting point for each value, and select the point that minimizes impurity as the optimal splitting point.

[0277] Create child nodes at the optimal split point and divide the current node's data set into subsets based on the values ​​of the splitting features.

[0278] Repeat the splitting process for each child node until the termination condition is met.

[0279] Set the termination condition to reach the preset maximum tree depth d. max The number of samples in a node is less than the preset minimum sample number threshold. min Further splitting cannot reduce impurity beyond the preset minimum impurity reduction threshold. min .

[0280] The cost complexity pruning method is applied to select the optimal pruning point by adjusting the model complexity parameter λ.

[0281] The leaf nodes output the corresponding maritime target category, and the final maritime target category vector uses cn =c1,c2,…,c n Let n represent the number of categories of maritime targets.

[0282] For example, assuming the original AIS track data contains data on 10 categories of maritime targets, the final decision tree outputs a maritime target category vector C. 10 =C0,C1,......,C9, where each value is the predicted probability value of the maritime target category. The maritime target will be classified into the category that has the largest probability value.

[0283] Step S3: Real-time acquisition of AIS track data of the maritime target to be identified, preprocessing to obtain multiple track point data sequences and corresponding two-dimensional images corresponding to the AIS track data, inputting them into the trained maritime target identification model to obtain the maritime target category prediction probability value, and taking the one with the largest probability value as the maritime target prediction category result.

[0284] In summary, the trajectory-driven maritime target identification method of this invention has the following beneficial effects:

[0285] 1. The method of the present invention can process long tracks in real time and automatically segment tracks through the Kmeans clustering algorithm. Each track segment, whether long or short, can be identified in real time, thereby improving the processing capability and recognition accuracy of long tracks and realizing real-time processing and segmentation identification of long tracks. The present invention can process and identify both short and long track data in real time.

[0286] 2. This invention, by combining the self-attention mechanism of the CCNet model and deep learning algorithms, achieves automated feature extraction and real-time identification of maritime targets from AIS track data, significantly improving the efficiency and accuracy of maritime target category identification and reducing the error rate of manual analysis. This improves the efficiency and accuracy of maritime target category identification.

[0287] 3. Traditional methods face difficulties in identifying short tracks. The method of this invention effectively solves the problem of category identification for short tracks by segmenting the track and extracting local features, enabling accurate category prediction results for short tracks in real time. This optimizes the identification capability for shorter tracks.

[0288] 4. By combining local identification vectors and global feature vectors, and using decision trees for the final prediction of maritime target categories, the method of this invention not only improves the generalization ability for different types of maritime targets, but also enhances the robustness to noise and outliers.

[0289] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A track driven maritime target recognition method, characterized in that, The method comprises the following steps: Step S1, obtaining a plurality of original AIS track data of a marine target and preprocessing, obtaining a data sequence of a plurality of track points in each AIS track data and a two-dimensional image corresponding to each AIS track data based on the preprocessing result; wherein each track data comprises a track number, a marine target category and a plurality of track point data; Step S2, constructing a marine target identification model, and training the marine target identification model based on the track point data sequence and the corresponding two-dimensional image to obtain a trained marine target identification model; Step S3, real-time acquisition of AIS track data of a marine target to be identified for preprocessing, obtaining a plurality of track point data sequences corresponding to the AIS track data and corresponding two-dimensional images input into the trained marine target identification model, and obtaining a marine target category prediction result.

2. The method of claim 1, wherein, The data sequence q of the kth track point of the ith track is represented as follows: wherein id is the track number of the AIS track data corresponding to the track point; t is the time stamp in hours; x, y are the planar Cartesian coordinates of the track point; v is the sailing speed; φ is the angular acceleration; z v ,z α ,z ω are the Z-scores of the sailing speed, acceleration and angular acceleration, respectively; Obtaining the two-dimensional image corresponding to each AIS track data comprises: Based on each preprocessed AIS track data, a two-dimensional image is constructed in a two-dimensional space; wherein each track point corresponds to a pixel position in the image, and the pixel point information is the corresponding preprocessed track point data; Adjust all two-dimensional images to a uniform size.

3. The method of claim 1, wherein, Each track point data comprises a latitude and longitude coordinate position, a navigation speed, a heading angle, a timestamp; preprocessing each AIS track data in the original AIS track data set, i.e. preprocessing the track point data in the AIS track data, comprising: Data deduplication, outlier processing and missing value filling; Convert the timestamp to an hour unit t; Calculate the acceleration a based on the navigation speed v; Calculate the angular acceleration φ based on the heading angle; Convert the latitude and longitude coordinate position to a plane rectangular coordinate x, y; The sailing speed, acceleration and angular acceleration are Z-score standardized to obtain Z scores z of the sailing speed, acceleration and angular acceleration, respectively v α ω .​​ 4. The method of claim 1, wherein, The marine target identification model comprises a track automatic segmentation module, a local identification module, a global identification module and a marine target category evaluation module, which are trained by the following method: Respectively construct the training data set of the track automatic segmentation module, the local identification module, the global identification module and the marine target category evaluation module; wherein the training data set of the track automatic segmentation module comprises all track point data sequences, and the sample label is a state label; the training data set of the local identification module is the track segmentation data obtained by the track segmentation module, and the label is a marine target category label; the training data set of the global identification module is the two-dimensional image corresponding to each AIS track data, and the label is a marine target category label; the local identification vector obtained by the local identification module and the global identification vector obtained by the global identification module are combined to form a set identification vector as the training data set of the marine target category evaluation module, and the label is a marine target category label; Load the training data set of the track segmentation module, train the track automatic segmentation module using self-attention mechanism and KMeans clustering, use a custom Loss loss function, use the Adam optimizer to update the weights of the track segmentation module by back propagation, and obtain the trained track segmentation module; Load the training data set of the local recognition module, input the sample into 1DCNN and LSTM respectively, train the local recognition module, 1DCNN and LSTM use cross-entropy and MSE loss function respectively, both use Adam optimizer to update the weights of the track segmentation module by back propagation, and obtain the trained local recognition module; Load the training data set of the global recognition module, train the global recognition module using cross-entropy loss function, use Adam to update the parameter weights of the global recognition module by back propagation, and obtain the trained global recognition module; Load the training data set of the offshore target category evaluation module, use ID3 algorithm to initialize the decision tree, until the maximum tree depth and the number of samples in the tree node are less than the preset minimum sample number threshold, and obtain the trained offshore target category evaluation module. Based on the trained track segmentation module, local recognition module, global recognition module and offshore target category evaluation module, a trained offshore target recognition model is obtained. The state label includes straight, turn and anchor.

5. The method of claim 4, wherein, The track automatic segmentation module sequentially includes an input layer, an embedding layer, first and second multi-layer attention structures and a KMeans clustering layer. The input layer is used to receive each track point data sequence q corresponding to each track; The embedding layer is used for position encoding of the each track point data sequence q; The first and second multi-layer attention structures connected in sequence are used to obtain a weighted sum vector O' of each track data res ; The Kmeans clustering layer is configured to cluster O res The feature vector z of each track point in the AIS i,k The Kmeans clustering layer is configured to cluster O 6. The method of claim 5, wherein, Training the track automatic segmentation module using the training data set of the track segmentation module includes: Step 1, input each track point data sequence q in the training data set of the track segmentation module to the embedding layer for position encoding after the input layer, and then input the first multi-layer attention structure; In the second step, the first multi-layer attention structure calculates the attention score matrix by using the Q, K, and V vectors of q through the learnable parameter matrix in the self-attention and multi-head attention mechanism layer, and multiplies the attention score matrix by a scaling factor to obtain the scaled attention score matrix. Step 3, divide the scaled attention score matrix into 8 attention heads, and apply a softmax function to obtain an attention weight matrix W1, W2, W3, …, W8 for each head; Fourth step, multiply W1, W2, W3, …, W8 and V to obtain the self-attention output vector O, including O1, O2, O3, …, O8, splice O1, O2, O3, …, O8 using the contract function and average pool along the spliced dimension to obtain the self-attention output vector O input into the feedforward network to obtain the feature vector O linear2 ; Step 5, O linear2 input LayerNorm layer normalization to obtain the feature vector O ln , and the self-attention output vector O is connected to the output O through a residual connection ln , to obtain the feature vector O res ; Step 6, obtaining O res inputting the second multi-layer attention structure to obtain a weighted sum vector O res ; The seventh step is to weight and sum the value vectors in the V matrix using self-attention weights to obtain an AIS feature vector z of the kth time point in the ith track i,k = {z1, z2, …, z n}; The feature vector z i,k After PCA dimension reduction, the Kmeans algorithm is used for clustering to obtain the AIS segmentation result. Until the loss function converges, the trained track automatic segmentation model is obtained.

7. The method of claim 6, wherein, The feature vector z of each track point is calculated as follows: i,k Segmentation is performed using Kmeans algorithm clustering to obtain AIS segmentation results, including: For z i,k Apply PCA principal component analysis for dimensionality reduction; The reduced dimension z i,k The Kmeans algorithm is executed to perform clustering operation, obtaining each z i,k The corresponding classification vector c i,k {bit1, bit2, bit3}; The global vector for each AIS track data is considered as Z. i {z i,k ,c i,k }, for z i,k Continuous time point detection is performed, with a preset continuous point threshold K. If the number of consecutive values ​​of k in a category is N > K, then it is considered as T. i,l {z i,j ,z i,j+1 ,z i,j+2 ,...,z i,j+N-1 } represents a local feature belonging to state l in the i-th AIS track data, which is divided into a segment; Each AIS track data is divided into one or more AIS track segment data, and the part of the AIS track data that is not segmented is discarded. Each AIS track segment data corresponds to a state label.

8. The method of claim 4, wherein, The local recognition module is based on a 1DCNN network and an LSTM model, and the training data set of the local recognition module is input into the local recognition module for training to obtain a local recognition vector, including: respectively, and input into the 1DCNN network and the LSTM model simultaneously. i,l {z i,j ,z i,j+1 ,z i,j+2 ,...,z i,j+N-1} The 1D CNN network outputs a recognition vector a L ; the LSTM model outputs a recognition vector b L ; a L and b L are weighted and summed to obtain a local recognition vector R l as follows: R L = a L · W F + b L · W S , L = 1, 2, 3 wherein W F is the weight of the 1D CNN network output recognition vector, W S is the weight of the LSTM model output recognition vector.

9. The method of claim 4, wherein, The global recognition module is based on a CNN network, and each two-dimensional image corresponding to each AIS track data in the training data set of the global recognition module is input into the global recognition module for training to obtain a global recognition vector corresponding to each track data, including: For each AIS track data corresponding to the uniform size of two-dimensional image, convolution and pooling are carried out, the output is applied to ReLU activation, output to the full connection layer, and then to the output layer, and a global recognition vector R4 is obtained.

10. The method of claim 9, wherein, The offshore target category recognition module is based on the decision tree construction initialized by the ID3 algorithm, the set recognition vector in the training data set of the offshore target category evaluation module is input into the offshore target category evaluation module for training, and an offshore target category prediction result is obtained, including: The first step is to initialize the decision tree based on the ID3 algorithm; a root node is created and set as the entry point of the decision tree, and the information gain is selected as the splitting criterion for feature selection; Secondly, the local recognition vector R corresponding to each AIS track data L and the global recognition vector R4 are taken as the input of the decision tree. The third step is to calculate the gain value based on the splitting criterion for the current node; the node with the maximum gain value is selected for splitting; for the selected splitting node, the node with the minimum impurity is selected as the best splitting point; a child node is created at the best splitting point; Until the preset maximum tree depth is reached and the number of samples in the tree node is less than the preset minimum sample number threshold, the leaf node is the offshore target category prediction result.

Images