Civilian ship trajectory behavior analysis and optimization method based on deep learning

Abstract

The present invention relates to a civilian ship trajectory behavior analysis optimization method based on deep learning, which belongs to the field of ship trajectory analysis and solves the problem that only AIS data can not be used to analyze the civilian ship trajectory behavior in the prior art. The method includes: obtaining the AIS data of the civilian ship to be analyzed, the weather information of the location of the civilian ship to be analyzed, and the image information of the sea area where the civilian ship to be analyzed is located; pre-processing the AIS data, weather information and image information of the sea area to obtain four groups of vectors, namely: AIS information vector, weather information vector, maritime traffic congestion vector and civilian ship draft vector, and inputting the four groups of vectors into the trained civilian ship trajectory analysis model to obtain the civilian ship trajectory analysis result. Accurate analysis of the trajectory of civilian ships is achieved to avoid collisions with other ships or obstruction of navigation.

Description

Technical Field

[0001] The present invention relates to the field of trajectory analysis, and in particular to a civilian ship trajectory behavior analysis optimization method based on deep learning. Background Art

[0002] Traditional civil ship trajectory analysis mainly relies on ship observation data and manual records. The currently more advanced ship automatic identification system (AIS) data collection can only collect real-time data on the ship's position, speed, and heading. Its specific display form is text, with a relatively single dimension, and this single data information is not enough for subsequent enhanced analysis of civil ship trajectory behavior. Therefore, there is an urgent need for a systematic method with more diverse dimensions and more comprehensive trajectory analysis capabilities. Summary of the invention

[0003] In view of the above analysis, an embodiment of the present invention aims to provide a civilian ship trajectory behavior analysis optimization method based on deep learning, so as to solve the problem that the existing data dimension is single and the trajectory of civilian ships can only be analyzed manually.

[0004] The purpose of the present invention is mainly achieved through the following technical solutions:

[0005] The present invention provides a civilian ship trajectory behavior analysis and optimization method based on deep learning, comprising the following steps:

[0006] Obtaining AIS data of the civilian ship to be analyzed, weather information of the location of the civilian ship to be analyzed, and image information of the sea area where the civilian ship to be analyzed is located;

[0007] Preprocessing the AIS data, the weather information and the image information of the sea area to obtain four groups of vectors, wherein the four groups of vectors are: an AIS information vector, a weather information vector, a maritime traffic congestion vector and a civilian ship draft vector;

[0008] The four groups of vectors with the same length after preprocessing are input into the trained civilian ship trajectory analysis model to obtain the civilian ship trajectory analysis results; the civilian ship trajectory analysis model is a deep fusion network model including a double-layer bidirectional long short-term memory neural network and a convolutional neural network.

[0009] Furthermore, the marine traffic congestion vector is obtained based on the image information preprocessing of the sea area, including:

[0010] According to the image information of the sea area where the civilian ship to be analyzed is located, the YOLOv5 algorithm is used to obtain the traffic congestion at sea;

[0011] The maritime traffic congestion is characterized and constructed into a maritime traffic congestion vector.

[0012] Furthermore, the formula for obtaining the maritime traffic congestion is:

[0013]

[0014] Where ρ is the busyness performance value; n is the number of civilian ships in the target detection frame detected by the YOLOv5 algorithm; (D i_x -D j_x ) represents the x-coordinate difference between the i-th and j-th civilian ship target detection frames in target detection; (D i_y -D j_y ) represents the y-coordinate difference between the i-th and j-th civilian ship target detection frames in target detection.

[0015] Furthermore, the civilian ship draft vector is obtained based on the AIS data and the image information of the sea area through preprocessing, including:

[0016] Calculate the draft of the civilian ship to be analyzed based on the AIS data of the civilian ship to be analyzed and the image information of the sea area where the civilian ship to be analyzed is located;

[0017] The characteristics of the draft of the civilian ship to be analyzed are enhanced to construct a civilian ship draft vector.

[0018] Furthermore, the draft of the civilian ship to be analyzed is calculated, including:

[0019] S21, according to the image information of the sea area where the civilian ship to be analyzed is located, determine the serial number K of the longest civilian ship in the image, and the method is:

[0020]

[0021] in, is the x-coordinate value of the upper left corner of the target detection frame where the k-th ship is located; is the x-coordinate value of the upper right corner of the target detection frame where the k-th ship is located; k is the serial number of all ships in the image; Ais length is the length of the longest civilian ship in the image in the AIS data;

[0022] S22. Calculate the length of the civilian ship to be analyzed in the x direction in the target detection frame, and the formula is:

[0023]

[0024] Among them, Ais length_ID is the length of the civilian ship to be analyzed in the AIS data; Ais length_K Len is the length of the longest civilian ship in the image in the AIS data; K_picThe length of the longest civilian ship in the image in the x direction in the target detection frame;

[0025] S23, determining the serial number I of the civilian ship to be analyzed in the image, the method is as follows:

[0026] I=Min(Len ID_pic -Len k_pic )

[0027] Where k is the serial number of all ships in the image;

[0028] S24. Calculate the draft W of the civilian ship to be analyzed, and the formula is:

[0029]

[0030] W=Ais hight_I -T

[0031] Where T represents the distance between the top of the civilian ship to be analyzed and the sea surface; (y I_pic_1 ,y I_pic_3 ) represent the y coordinate values ​​of the upper left corner and the lower left corner of the detection frame where the civilian ship to be analyzed is located; Ais hight_I Indicates the height of the civilian ship to be analyzed in the AIS data.

[0032] Furthermore, the four groups of vectors are all 1*7 vectors; among them,

[0033] The speed, longitude coordinates, latitude coordinates, overall heading, length and two supplementary word vectors in the AIS data <pad>Construct a 1*7 AIS information vector;

[0034] Construct a 1*7 weather information vector using the wave height, wave size, wind speed, wind direction, tidal direction, rainfall and tidal speed in the weather information;

[0035] After enhancing the features of maritime traffic congestion and the draft of the civilian ship to be analyzed, 1*7 vectors are constructed respectively.

[0036] Furthermore, the maritime traffic congestion and the real-time draft of the civilian ship to be analyzed are enhanced according to the following enhanced formula:

[0037] N p =α,α=1,2,3,4,5,6 or 7

[0038] N T =β,β=1,2,3,4,5,6 or 7

[0039] Among them, N p N is the number after the busyness is expanded; T is the number after the draft is expanded; α and β are the hyper parameters of busyness and draft respectively;

[0040] When α,β is less than 7, add word vector <pad>To construct a 1*7 vector.

[0041] Furthermore, the civilian ship trajectory analysis results obtained include:

[0042] The four groups of 1*7 vectors after preprocessing are passed through a double-layer bidirectional long short-term memory neural network to obtain a double-layer bidirectional long short-term memory neural network prediction probability value;

[0043] The four groups of 1*7 vectors after preprocessing are passed through a convolutional neural network to obtain a convolutional network output prediction probability value;

[0044] Use the following formula to perform integrated analysis to obtain the final probability value, and take the highest probability value as the civilian ship trajectory analysis result;

[0045]

[0046] Among them, Last Output is the final probability value, The predicted probability value is a two-layer bidirectional long short-term memory neural network, Conv Output Output predicted probability values ​​for the convolutional network.

[0047] Furthermore, the double-layer bidirectional long short-term memory neural network prediction probability value obtained by the double-layer bidirectional long short-term memory neural network includes:

[0048] The four groups of 1*7 vectors after preprocessing are input into a double-layer bidirectional long short-term memory neural network as input sequences;

[0049] The output layer vector is calculated by the bidirectional LSTM layer, where the LSTM layer draft output vector is adjusted using the weight balance formula. The adjustment formula is:

[0050]

[0051] Among them, C t_4 is the hidden layer input when the other vectors input the draft output vector; f t ,i t is the weight calculation formula; C t-1 It is the input of the previous hidden layer; belongs to the current input; α, β are hyperparameters;

[0052] Pass the output layer vector through a fully connected layer, and then apply the softmax function to output the labeled prediction probability value

[0053] Furthermore, obtaining the convolutional network output prediction probability value through the convolutional neural network includes:

[0054] Splitting the preprocessed four groups of 1*7 vectors into 28 eigenvalues ​​to obtain a first group of input vectors;

[0055] Randomly arrange the 28 eigenvalues ​​to obtain a second set of input vectors;

[0056] Using a residual network to perform convolution analysis on the first and second groups of input vectors respectively;

[0057] Apply a fully connected layer to the convolution output vector, and then apply the softmax function to output the convolution network labeled output probability value Conv Output .

[0058] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects:

[0059] 1. The present invention uses AIS data to compare and match the image information of nearby positioning points collected by monitoring satellites and coastal information collection equipment, so as to more accurately realize the positioning of civilian ships in specific application scenarios, the busyness of the sea surface where they are located, and the characterization of the draft of the ships, so as to accurately determine the trajectory type and avoid collisions with other ships or obstructions to navigation.

[0060] 2. The navigation trajectory of a civil ship is a very complex behavior, involving multiple factors, and there are many couplings and correlations between them. The present invention adopts a deep fusion network architecture to complete the mapping relationship between features and target values. By performing trajectory prediction through a deep fusion network, it can better learn the movement laws and data characteristics of the ship, and can determine the trajectory of the ship more quickly and accurately.

[0061] 3. In a deep fusion network architecture of the present invention, the draft output vector of the double-layer bidirectional long short-term memory neural network is adjusted to reduce the influence weight of other vectors on the draft vector, which can speed up the network training speed and fitting performance.

[0062] In the present invention, the above-mentioned technical solutions can also be combined with each other to achieve more preferred combination solutions. Other features and advantages of the present invention will be described in the subsequent description, and some advantages can become obvious from the description, or can be understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained through the contents particularly pointed out in the description and the drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] The drawings are only for the purpose of illustrating particular embodiments and are not to be considered limiting of the present invention. Like reference symbols denote like components throughout the drawings.

[0064] Figure 1 Flowchart of the optimization method for trajectory analysis of civil ships based on deep learning

[0065] Figure 2 The text vector conversion output method used in the present invention is shown in FIG.

[0066] Figure 3 Schematic diagram of four sets of vectors constructed for the present invention

[0067] Figure 4 Image information of the vicinity of the location of the civilian ship to be analyzed collected by coastal information collection equipment or high-resolution monitoring satellites

[0068] Figure 5 Civilian ship trajectory analysis model used in the present invention DETAILED DESCRIPTION

[0069] The preferred embodiments of the present invention are described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not used to limit the scope of the present invention.

[0070] A specific embodiment of the present invention, as Figure 1 As shown, a civilian ship trajectory behavior analysis and optimization method based on deep learning is disclosed, comprising the following steps:

[0071] S11, obtaining AIS data of the civilian ship to be analyzed, weather information of the location of the civilian ship to be analyzed, and image information of the sea area where the civilian ship to be analyzed is located;

[0072] Specifically, the AIS data of civilian ships is obtained using an AIS receiving device or from an AIS data provider. The AIS data is processed by text parsing to extract the required key information, such as the position coordinates, time axis, speed and heading of the civilian ship.

[0073] Furthermore, weather information about the location of the civilian ship to be analyzed is obtained based on the monitoring satellite data, such as wave height, wave size, wind speed, tidal speed, rainfall, wind direction and tidal direction.

[0074] Furthermore, coastal information collection equipment or high-resolution monitoring satellites are used to photograph and sample scenes near the location of the civilian ship to be analyzed, thereby obtaining image information of the location of the civilian ship to be analyzed.

[0075] S12, preprocessing the AIS data, the weather information and the image information to obtain four groups of vectors, wherein the four groups of vectors are: an AIS information vector, a weather information vector, a maritime traffic congestion vector and a civilian ship draft vector;

[0076] Specifically, the AIS data and weather information obtained in step S11 are stored in text form.

[0077] It should be noted that the amount of text information in this embodiment is relatively large, and since there is no objective function, network model, etc. in the previous text information processing, it is impossible to perform a Word2Vec type training on it. Here, a semantic information conversion method is proposed based on the actual application scenario of civil ships, such as Figure 2 As shown, semantic text information can be quickly converted into a one-dimensional vector, reducing the difficulty of training and improving the error tolerance rate.

[0078] Furthermore, the semantic information conversion method is to convert the speed, longitude coordinates, latitude coordinates, overall heading, length and two supplementary word vectors in the AIS data <pad>Build as Figure 3 The 1*7 AIS information vector is shown.

[0079] It should be noted that the general direction of navigation is set as an arc offset from the latitude.

[0080] Furthermore, the semantic information conversion method is to construct the following semantic information: wave height, wave size, wind speed, tidal speed, rainfall, wind direction and tidal direction in the obtained weather information Figure 3 The 1*7 weather information vector shown.

[0081] It should be noted that the values ​​of wind direction and tidal direction are expressed in radians.

[0082] Furthermore, the maritime traffic congestion and the draft of civilian ships are enhanced and then 1*7 vectors are constructed respectively.

[0083] For example, according to Figure 4 The image information near the location of the civilian ship to be analyzed is collected by coastal information collection equipment or high-resolution monitoring satellites, and YOLOv5 is used to analyze the image information to check the number of civilian ships in this sea area and judge the busyness of maritime traffic. The formula is:

[0084]

[0085] Where ρ is the busyness performance value; n is the number of civilian ships in the target detection frame detected by the YOLOv5 algorithm; (D i_x -D j_x ) represents the x-coordinates of the i-th and j-th civilian ship target detection frames in target detection; (D i_y -D j_y ) represents the y coordinates of the i-th and j-th civilian ship target detection frames in target detection.

[0086] In formula (1), the distances between the center points of all ships are summed up, and the inverse is multiplied by an incremental value related to the number of ships. Through this formula, it can be achieved that when the number of ships in the target frame is large and the distance is small, the busyness performance value will be larger.

[0087] In addition, in this embodiment, image analysis information is combined with AIS data to perform morphological analysis on a specific ship to obtain the specific draft of the civilian ship to be analyzed. The AIS data contains the longitude coordinates, latitude coordinates, and specific length of the ship. It is difficult to accurately match the ship information in the AIS data with the specific ship in the image by traditional methods. This embodiment proposes an exemplary maximum matching principle. The calculation obtains the draft of the civilian ship to be analyzed, and the specific steps include:

[0088] S21, according to the image information of the sea area where the civilian ship to be analyzed is located, determine the serial number K of the longest civilian ship in the image, and the method is:

[0089]

[0090] in, is the x value of the upper left corner coordinate of the kth ship in the target detection frame; is the x value of the upper right corner coordinate of the kth ship in the target detection frame; k is the serial number of all ships in the image; Ais length is the length of the longest civilian ship in the image in the AIS data;

[0091] S22. Calculate the length of the civilian ship to be analyzed in the x direction in the target detection frame, and the formula is:

[0092]

[0093] Among them, Ais length_ID is the length of the civilian ship to be analyzed in the AIS data; Ais length_K Len is the length of the longest civilian ship in the image in the AIS data; K_pic is the length of the longest civilian ship in the image in the x direction in the target detection frame;

[0094] S23, determining the serial number I of the civilian ship to be analyzed in the image, the method is as follows:

[0095] I=Min(Len ID_pic -Len k_pic )

[0096] Where k is the serial number of all ships in the image;

[0097] S24. Calculate the draft W of the civilian ship to be analyzed, and the formula is:

[0098]

[0099] W=Ais hight_I -T

[0100] Where, T represents the distance between the top of the ship and the sea surface; (y I_pic_1 ,y I _ pic_3 ) represent the y value of the upper left corner and the y value of the lower left corner of the civilian ship to be analyzed in the detection frame; Ais hight_I Indicates the height of the civilian ship to be analyzed in the AIS data.

[0101] Specifically, in the above steps, the longest civilian ship in the picture is first found, and by finding the maximum value of the entity length in the AIS data, the largest ship in the actual scene of the picture is matched with the AIS data, and its length pixel value in the picture is obtained. In this way, through step S33, the length pixel value of the ship to be matched in the picture can be obtained, and then the verification box query can be performed in YOLOv5 to obtain the corresponding position of the civilian ship to be analyzed in the actual application scene, and then the height of the top of the ship above the water surface can be obtained, and finally the actual draft of the civilian ship to be analyzed can be obtained.

[0102] It should be noted that in this embodiment, the busyness of the sea surface and the draft of the civilian ship to be analyzed are extremely critical factors. Among them, areas with busy maritime traffic may require civilian ships to directly follow specific navigation tracks, abide by navigation rules, and avoid collisions with other ships or obstructions to navigation. In addition, the real-time draft of the ship is a key factor in the selection of navigation waters. The same ship with cargo and without cargo chooses to sail in different waters.

[0103] Therefore, in this embodiment, the busyness and the ship draft are subjected to feature enhancement operations, which are characterized by using separate vectors, and their quantities are enhanced.

[0104] Specifically, the maritime traffic congestion and the real-time draft of the civilian ship to be analyzed are enhanced according to the following enhancement formula:

[0105] N p =α,α=1,2,3,4,5,6,7

[0106] N T =β,β=1,2,3,4,5,6,7

[0107] Among them, N p N is the number after the busyness is expanded; T is the number after the draft is expanded; α and β are the hyperparameters of busyness and draft respectively.

[0108] When α, β is less than 7, add word vector <pad>To construct a 1*7 vector.

[0109] For example, Figure 3 As shown in the figure, when α is 3, the constructed maritime traffic busyness vector is busyness, busyness, busyness, <pad> 、 <pad> 、 <pad> 、 <pad>.

[0110] For example, Figure 3 As shown, when β is 2, the draft vector constructed is draft, draft, <pad> 、 <pad> 、 <pad> 、 <pad> 、 <pad>.

[0111] It should be noted that in the above steps, the definition of each data that needs to be collected and its certain characterization are completed. By collecting a large amount of the above data, the trajectories of civilian ships whose paths can be manually judged as risky are marked. The marked options are divided into three categories: continue to maintain the current trajectory, need to make appropriate adjustments, and please adjust as soon as possible.

[0112] S13. Input the four groups of vectors with the same length after preprocessing into the trained civilian ship trajectory analysis model to obtain the civilian ship trajectory analysis results; the civilian ship trajectory analysis model is a deep fusion network model of a double-layer bidirectional long short-term memory neural network and a convolutional neural network.

[0113] Specifically, the civilian ship trajectory analysis model is as follows: Figure 5 As shown, the initial input feature values ​​are four groups of vectors with the same length after preprocessing.

[0114] Furthermore, a two-layer bidirectional long short-term memory neural network (Bi_LSTM) is used to combine and analyze the four groups of vectors to obtain the predicted probability value

[0115] Specifically, the preprocessed four groups of 1*7 vectors are input into a double-layer bidirectional long short-term memory neural network as input sequences.

[0116] The output layer is calculated through the bidirectional LSTM layer.

[0117] It should be noted that, since the influence between the draft output vector and the other three vectors is small, in this embodiment, it is necessary to use the weight balance formula to adjust the LSTM layer draft output vector to reduce its influence on the other vectors in the Bi_LSTM network. The specific adjustment formula is:

[0118]

[0119] Among them, C t_4 is the hidden layer input when the other vectors input the draft output vector; f t ,i t is the weight calculation formula; C t-1 It is the input of the previous hidden layer; Belongs to the current input; α, β are hyperparameters and need to be adjusted manually.

[0120] According to the above formula, the influence weights of other vectors on the draft vector can be reduced by directly reducing the values ​​of α and β. Since the influence relationship of the draft vector is difficult to obtain through training, the influence transmission of the draft to other vectors can be directly set to a very low value. This is used to speed up the network training speed and fitting performance.

[0121] Furthermore, the output layer vector is passed through a fully connected layer and then the softmax function is applied to output the labeled prediction probability value That is, the classification probability of the three situations: continue to maintain the current trajectory, need to make moderate adjustments, and make adjustments as soon as possible.

[0122] Furthermore, since the data set vectors collected in this embodiment may also have certain influences, such as the weather vector x 1 =[wave height, sea waves, wind speed, wind direction, tidal direction, rainfall, tidal speed], there may be mutual influence between the wave height and the sea waves, and the above-mentioned Bi_LSTM network alone cannot satisfy this characteristic relationship. Therefore, in this embodiment, vector splitting, shuffling and reorganization are adopted, and then convolution operation is performed to enhance the mutual influence relationship.

[0123] Specifically, the four groups of 1*7 vectors after preprocessing are used to obtain the predicted probability value Conv through a convolutional neural network. Output ,include:

[0124] Splitting the four groups of vectors into 28 eigenvalues ​​to obtain a group of input vectors;

[0125] Randomly arrange the 28 eigenvalues ​​to obtain another set of input vectors;

[0126] The two sets of input vectors are convolutionally analyzed using a residual network;

[0127] Apply a fully connected layer to the convolution output vector, and then apply the softmax function to output the convolution network labeled output probability value Conv Output , that is, the classification probability of continuing to maintain the current trajectory, needing to make moderate adjustments, and requiring adjustments as soon as possible.

[0128] Furthermore, the following formula is used for integrated analysis to obtain the final predicted probability value:

[0129]

[0130] Among them, Last Output is the final predicted probability value, The predicted probability value is a two-layer bidirectional long short-term memory neural network, Conv Output Output probability values ​​for the convolutional network.

[0131] The highest probability of classification among the three situations obtained from the integrated analysis, namely, continuing to maintain the current trajectory, needing to make appropriate adjustments, and requiring adjustments as soon as possible, is taken as the result of the civilian ship trajectory analysis.

[0132] It should be noted that both the two-layer bidirectional long short-term memory neural network and the convolutional neural network require pre-training.

[0133] The training samples of the two-layer bidirectional long short-term memory neural network include four sets of input vectors and labels. The labels are: continue to maintain the current trajectory, need to make appropriate adjustments, please adjust as soon as possible, and their respective probability values. The cross entropy loss function is used to train the two-layer bidirectional long short-term memory neural network. Each output probability value is compared with the probability value in the training sample, and the prediction error value is calculated. When the accuracy reaches the set value or the number of training times exceeds the maximum number, the training ends.

[0134] The convolutional neural network training sample includes a set of input vectors obtained by splitting four sets of input vectors into 28 eigenvalues, and labels. The labels are "continue to maintain the current trajectory", "need to make appropriate adjustments", "please make adjustments as soon as possible" and their respective probability values. The convolutional neural network is trained using the cross entropy loss function. The probability values ​​of the three situations output each time are compared with the probability values ​​in the training samples, and the prediction error value is calculated. When the accuracy reaches the set value or the number of training times exceeds the maximum number, the training ends.

[0135] The four groups of input vectors include AIS information vector, weather information vector, maritime traffic congestion vector and civilian ship draft vector, which are obtained by the same preprocessing steps as in the analysis process.

[0136] The probability values ​​of the three situations in the label can be obtained based on expert experience analysis.

[0137] Through the above formula, the output characteristics and function mapping relationship of the double-layer bidirectional long short-term memory neural network and the convolutional neural network can be integrated and analyzed. Through the deep fusion network for trajectory prediction, the movement law and data characteristics of the ship can be better learned, and the trajectory of the ship can be determined more quickly and accurately, thereby increasing the accuracy and fault tolerance of the entire network.

[0138] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by any technician familiar with the technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.< / pad> < / pad> < / pad> < / pad> < / pad> < / pad> < / pad> < / pad> < / pad> < / pad> < / pad> < / pad> < / pad>

Claims

1. A civilian ship trajectory behavior analysis and optimization method based on deep learning, characterized in that: The steps include: Obtaining AIS data of the civilian ship to be analyzed, weather information of the location of the civilian ship to be analyzed, and image information of the sea area where the civilian ship to be analyzed is located; The AIS data, the weather information and the image information of the sea area are preprocessed to obtain four groups of vectors, wherein the four groups of vectors are: an AIS information vector, a weather information vector, a maritime traffic congestion vector and a civilian ship draft vector; obtaining the civilian ship draft vector includes: calculating the draft of the civilian ship to be analyzed according to the AIS data of the civilian ship to be analyzed and the image information of the sea area where the civilian ship to be analyzed is located; and enhancing the features of the civilian ship draft to be analyzed to construct a civilian ship draft vector; wherein the calculation of the civilian ship draft to be analyzed includes: S21, according to the image information of the sea area where the civilian ship to be analyzed is located, determine the serial number K of the longest civilian ship in the image, and the method is: ； in, is the x-coordinate value of the upper left corner of the target detection frame where the k-th ship is located; is the x-coordinate value of the upper right corner of the target detection frame where the k-th ship is located; k is the serial number of all ships in the image; is the length of the longest civilian ship in the image in the AIS data; S22. Calculate the length of the civilian ship to be analyzed in the x direction in the target detection frame, and the formula is: ； in, is the length of the civilian ship to be analyzed in the AIS data; The length of the longest civilian ship in the image as measured by AIS data; The length of the longest civilian ship in the image in the x direction in the target detection frame; S23, determining the serial number I of the civilian ship to be analyzed in the image, the method is as follows: ； Where k is the serial number of all ships in the image; S24. Calculate the draft W of the civilian ship to be analyzed, and the formula is: ； ； in, Indicates the distance between the top of the civilian ship to be analyzed and the sea surface; Respectively represent the upper left corner and lower left corner of the detection frame where the civilian ship to be analyzed is located Coordinate value; Indicates the height of the civilian ship to be analyzed in the AIS data; The four groups of vectors with the same length after preprocessing are input into the trained civilian ship trajectory analysis model to obtain the civilian ship trajectory analysis results; the civilian ship trajectory analysis model is a deep fusion network model including a double-layer bidirectional long short-term memory neural network and a convolutional neural network.

2. The method according to claim 1, characterized in that: Obtaining a maritime traffic congestion vector based on the image information preprocessing of the sea area includes: obtaining the maritime traffic congestion vector using a YOLOv5 algorithm based on the image information of the sea area where the civilian ship to be analyzed is located; The maritime traffic congestion is characterized and constructed into a maritime traffic congestion vector.

3. The method according to claim 2, characterized in that: The formula for obtaining the maritime traffic congestion degree is: ； in, is the busyness performance value; n is the number of civilian ships in the target detection frame detected by the YOLOv5 algorithm; Indicates the x-coordinate difference between the i-th and j-th civilian ship target detection frames in target detection; Represents the y-coordinate difference between the i-th and j-th civilian ship target detection frames in target detection.

4. The method according to claim 1, characterized in that: The four sets of vectors are vector; where The speed, longitude coordinates, latitude coordinates, overall heading, length and two supplementary word vectors in the AIS data <pad>Build AIS information vector;< / pad> The wave height, wave size, wind speed, wind direction, tidal direction, rainfall and tidal speed in the weather information are constructed Weather information vector; The maritime traffic congestion and the draft of the civilian ships to be analyzed are enhanced and constructed vector.

5. The method according to claim 4, characterized in that: The maritime traffic congestion and the real-time draft of the civilian ship to be analyzed are enhanced according to the following enhanced formula: ； ； in, The number after the busyness is expanded; The amount after the draft is expanded; are the hyperparameters of busyness and draft, respectively; when When it is less than 7, add word vector <pad>To build vector.< / pad> 6. The method according to claim 1, characterized in that: The civilian ship trajectory analysis results obtained include: The four groups after pretreatment The vector is passed through a double-layer bidirectional long short-term memory neural network to obtain a double-layer bidirectional long short-term memory neural network prediction probability value; The four groups after pretreatment The vector is passed through the convolutional neural network to obtain the predicted probability value of the convolutional network output; Use the following formula to perform integrated analysis to obtain the final probability value, and take the highest probability value as the civilian ship trajectory analysis result; ； in, is the final probability value, is the predicted probability value of the two-layer bidirectional long short-term memory neural network, Output predicted probability values ​​for the convolutional network.

7. The method according to claim 6, characterized in that: The method of obtaining the predicted probability value of the double-layer bidirectional long short-term memory neural network by using the double-layer bidirectional long short-term memory neural network includes: The four groups after pretreatment Vector, as input sequence, is input into a two-layer bidirectional long short-term memory neural network; The output layer vector is calculated by the bidirectional LSTM layer, where the LSTM layer draft output vector is adjusted using the weight balance formula. The adjustment formula is: ； in, It is the hidden layer input when the other vectors input the draft output vector; is the weight calculation formula; It is the input of the previous hidden layer; belongs to the current input; is a hyperparameter; Pass the output layer vector through a fully connected layer, and then apply the softmax function to output the labeled prediction probability value .

8. The method according to any one of claims 6 or 7, characterized in that: The convolutional neural network output prediction probability value obtained by the convolutional neural network includes: The four groups after pretreatment The vector is split into 28 eigenvalues ​​to obtain the first set of input vectors; Randomly arrange the 28 eigenvalues ​​to obtain a second set of input vectors; Using a residual network to perform convolution analysis on the first and second groups of input vectors respectively; Apply a fully connected layer to the output vector after convolution, and then apply the softmax function to output the convolution network labeled output probability value .

Images