A method for predicting offshore ship situation distribution

Abstract

The present invention discloses a method for predicting the distribution of offshore ship situation, which comprises the following steps: obtaining the longitude and latitude of the corresponding ship in the offshore range according to the ship automatic identification system; generating a ship single-channel scatter plot based on the longitude and latitude in a single time period; obtaining N ship single-channel scatter plots and arranging them in time sequence to obtain a single-channel scatter plot sequence; using the single-channel scatter plot sequence as the input of the trained distribution prediction model to obtain the prediction scatter plots after different times; decoding the prediction scatter plot to obtain the offshore ship situation distribution prediction result. The present invention uses a deep learning model to perform high-precision and high-efficiency ship situation prediction, which provides strong support for navigation safety and shipping management.

Description

Technical Field

[0001] The present invention relates to the field of ship traffic management, and in particular to a method for predicting offshore ship situation distribution. Background Art

[0002] Vessel traffic management in offshore and port areas is of great significance for ensuring shipping safety and improving navigation efficiency. With the rapid development of the shipping industry, the number of ships continues to increase, making the navigation environment in offshore and port areas increasingly complex. Therefore, accurately predicting the distribution density and situation of ships in the future is crucial for optimizing navigation routes, avoiding collision risks, and improving the overall efficiency of shipping. However, existing prediction methods are often limited by data processing capabilities and model complexity, and it is difficult to meet the needs of high-precision prediction of ship situation. Summary of the invention

[0003] In view of the above-mentioned deficiencies in the prior art, the present invention provides a method for predicting offshore ship situation distribution, which solves the problem that the prior art is difficult to meet the demand for high-precision prediction of ship situation.

[0004] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is:

[0005] A method for predicting offshore ship situation distribution is provided, which comprises the following steps:

[0006] S1. Obtain the latitude and longitude of the corresponding ship in the offshore range according to the ship automatic identification system;

[0007] S2, generating a single channel scatter plot of ships based on the longitude and latitude in a single time period;

[0008] S3, obtaining N single-channel scatter plots of ships and arranging them in time sequence to obtain a single-channel scatter plot sequence;

[0009] S4, using the single-channel scatter plot sequence as the input of the trained distribution prediction model to obtain prediction scatter plots after different time periods;

[0010] S5. Decode the prediction scatter plot to obtain the prediction result of offshore ship situation distribution;

[0011] The scatter plots include single-channel scatter plots of ships and predicted scatter plots. The scatter plots are black and white, and white represents ships.

[0012] Furthermore, the specific method for generating a ship single channel scatter plot includes the following sub-steps:

[0013] S2-1. Without considering the arc of the earth, a rectangle is set in the offshore area and used as the prediction area; the longitude and latitude coordinates of the upper left corner of the rectangle are P1 (lon1, lat1), and the longitude and latitude coordinates of the lower right corner are P2 (lon2, lat2);

[0014] S2-2. A pure black image with a channel number of 1 and a size of h×w is synthesized, and the ships in the rectangle are mapped to the pure black image with white dots according to the height-to-width ratio to obtain a single-channel scatter plot of ships with a size of h×w.

[0015] Furthermore, the calculation expressions for the height H and width W of the prediction area are:

[0016] H=110×|lat1―lat2|

[0017]

[0018] Where cos(.) is the cosine function.

[0019] Furthermore, the duration of a single time period is the same as the time interval for uploading the longitude and latitude of the ship by the ship automatic identification system.

[0020] Furthermore, the distribution prediction model is a U-net neural network model, including six sequentially connected downsampling modules, six upsampling modules, and M detection heads;

[0021] Each downsampling module halves the input image size and doubles the number of channels; the input of the first downsampling module is the input of the U-net neural network model; the output of the sixth downsampling module is the input of the first upsampling module;

[0022] The first five upsampling modules double the input image size and halve the number of channels; the sixth upsampling module reduces the number of channels to 1 / k, where M = N / k;

[0023] The output of the nth downsampling module is connected with the output of the 7th-nth upsampling module in the channel dimension and serves as the input of the 7th-nth upsampling module; the output of the sixth upsampling module is the input of M detection heads; n = 1, 2, 3, 4, 5;

[0024] M detection heads correspond to the prediction scatter plots at M different prediction time points.

[0025] Furthermore, the value of N is 192, and the data format of the single-channel scatter plot sequence is [P 192 ,P 191 ,…,P i ,…,P1],P i is the single channel scatter plot of the i-th ship; the value of k is 6; the value of M is 64.

[0026] Furthermore, each downsampling module includes a first 1×1 convolutional layer, a first BN layer, a first ReLU activation layer, a first 3×3 convolutional layer, a second BN layer, a second ReLU activation layer, a second 1×1 convolutional layer, a third BN layer, a concatenation layer, a third ReLU activation layer and a 3×3 maximum pooling layer connected in sequence; another input of the concatenation layer is the output of the fourth BN layer; the input of the fourth BN layer is the output of the third 1×1 convolutional layer; the input of the first 1×1 convolutional layer and the input of the third 1×1 convolutional layer are jointly the input of the downsampling module to which they belong; the output of the 3×3 maximum pooling layer is the output of the downsampling module to which they belong.

[0027] Furthermore, each upsampling module includes a second 3×3 convolutional layer, a fifth BN layer, a third 3×3 convolutional layer, a sixth BN layer and a 2×2 convolutional layer connected in sequence; the input of the second 3×3 convolutional layer is the input of the upsampling module to which it belongs; and the output of the 2×2 convolutional layer is the output of the upsampling module to which it belongs.

[0028] Furthermore, during the training process of the distribution prediction model, the initial learning rate is 0.01, and the subsequent learning rate is adjusted by the time decay method; the batch size is 4; the number of training iterations is 150; the loss function is:

[0029]

[0030]

[0031] Among them, loss all is the total loss value of the distribution prediction model; m is the mth detection head; loss m is the loss value of the mth detection head; head pq Represents the pixel value at the position (p,q) in the output feature map of the detection head; P pq is the pixel value at the position (p,q) in the label scatter plot; h and w are the height and width of the ship single channel scatter plot, respectively.

[0032] Furthermore, the specific method of decoding the prediction scatter plot and obtaining the prediction result of offshore ship situation distribution is as follows:

[0033] Calculate the height and width of the physical area corresponding to the predicted scatter graph according to the coordinates of the white points in the predicted scatter graph and the size of the predicted scatter graph;

[0034] According to the height and width of the physical area corresponding to the predicted scatter plot, calculate the offset distance of the white point relative to the upper left corner of the ship's single channel scatter plot;

[0035] The offset distance is converted into offset longitude and latitude to obtain the predicted position of offshore ships, and then the offshore ship situation distribution prediction results are obtained.

[0036] The beneficial effects of the present invention are:

[0037] 1. This method uses deep learning models to predict ship situations with high accuracy and efficiency, providing strong support for navigation safety and shipping management.

[0038] 2. This method achieves a concise and intuitive representation of complex navigation data by abstracting the ship position snapshot into a two-dimensional scatter plot, which not only simplifies the data processing process but also improves the efficiency of data processing.

[0039] 3. This method accurately reflects key features (such as the number of ships, distribution density, and heading) in a two-dimensional scatter plot. After being carefully encoded, these features provide rich information for the training and prediction of deep learning models, thereby improving the accuracy of predictions. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 is a schematic diagram of the process of this method;

[0041] Figure 2 This is a schematic diagram of the structure of the distribution prediction model;

[0042] Figure 3 It is a structural diagram of the downsampling module;

[0043] Figure 4 It is a structural diagram of the upsampling module;

[0044] Figure 5 It is a schematic diagram of the connection in the channel dimension;

[0045] Figure 6 A schematic diagram of the ship position obtained by the ship automatic identification system in the embodiment;

[0046] Figure 7 Schematic diagram for generating a single channel scatter plot of a ship;

[0047] Figure 8 Schematic diagram for decoding the prediction scatter plot. DETAILED DESCRIPTION

[0048] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0049] like Figure 1As shown, the offshore ship situation distribution prediction method includes the following steps:

[0050] S1. Obtain the latitude and longitude of the corresponding ship in the offshore range according to the ship automatic identification system;

[0051] S2, generating a single channel scatter plot of ships based on the longitude and latitude in a single time period;

[0052] S3, obtaining N single-channel scatter plots of ships and arranging them in time sequence to obtain a single-channel scatter plot sequence;

[0053] S4, using the single-channel scatter plot sequence as the input of the trained distribution prediction model to obtain prediction scatter plots after different time periods;

[0054] S5. Decode the prediction scatter plot to obtain the prediction result of offshore ship situation distribution;

[0055] The scatter plot includes a single-channel scatter plot of ships and a predicted scatter plot. The scatter plot is a black and white plot, and the white dots represent ships.

[0056] The specific method of generating a ship single channel scatter plot includes the following sub-steps:

[0057] S2-1. Without considering the arc of the earth, a rectangle is set in the offshore area and used as the prediction area; the longitude and latitude coordinates of the upper left corner of the rectangle are P1 (lon1, lat1), and the longitude and latitude coordinates of the lower right corner are P2 (lon2, lat2);

[0058] S2-2. A pure black image with a channel number of 1 and a size of h×w is synthesized, and the ships in the rectangle are mapped to the pure black image with white dots according to the height-to-width ratio to obtain a single-channel scatter plot of ships with a size of h×w.

[0059] The calculation expressions for the height H and width W of the prediction area are:

[0060] H=110×|lat1―lat2|

[0061]

[0062] Where cos(.) is the cosine function. The units of H and W are both kilometers, 40000 is the circumference of the earth's equator (kilometers), and 110 represents the kilometers per 1 dimension.

[0063] In this embodiment, the length of a single time period is the same as the time interval for the ship automatic identification system to upload the ship's longitude and latitude. That is, assuming that the time interval for the ship automatic identification system to upload the ship's longitude and latitude is 1 minute, the length of a single time period is also 1 minute.

[0064] like Figure 2As shown, the distribution prediction model is a U-net neural network model, including six sequentially connected downsampling modules, six upsampling modules, and M detection heads;

[0065] Each downsampling module halves the input image size and doubles the number of channels; the input of the first downsampling module is the input of the U-net neural network model; the output of the sixth downsampling module is the input of the first upsampling module;

[0066] The first five upsampling modules double the input image size and halve the number of channels; the sixth upsampling module reduces the number of channels to 1 / k, where M = N / k;

[0067] The output of the nth downsampling module is connected with the output of the 7th to nth upsampling modules in the channel dimension and serves as the input of the 7th to nth upsampling modules; the output of the sixth upsampling module is the input of M detection heads; n = 1, 2, 3, 4, 5; the connection in the channel dimension is as follows Figure 5 As shown;

[0068] M detection heads correspond to the prediction scatter plots at M different prediction time points.

[0069] like Figure 3 As shown, each downsampling module includes a first 1×1 convolutional layer, a first BN layer, a first ReLU activation layer, a first 3×3 convolutional layer, a second BN layer, a second ReLU activation layer, a second 1×1 convolutional layer, a third BN layer, a concatenation layer, a third ReLU activation layer and a 3×3 maximum pooling layer connected in sequence; another input of the concatenation layer is the output of the fourth BN layer; the input of the fourth BN layer is the output of the third 1×1 convolutional layer; the input of the first 1×1 convolutional layer and the input of the third 1×1 convolutional layer are both the inputs of the downsampling module to which they belong; the output of the 3×3 maximum pooling layer is the output of the downsampling module to which they belong.

[0070] like Figure 4 As shown, each upsampling module includes a second 3×3 convolutional layer, a fifth BN layer, a third 3×3 convolutional layer, a sixth BN layer and a 2×2 convolutional layer connected in sequence; the input of the second 3×3 convolutional layer is the input of the upsampling module to which it belongs; and the output of the 2×2 convolutional layer is the output of the upsampling module to which it belongs.

[0071] During the training process of the distribution prediction model, the initial learning rate is 0.01, and the subsequent learning rate is adjusted by the time decay method; the batch size (batch_size) is 4; the number of training iterations is 150; the loss function is:

[0072]

[0073]

[0074] Among them, loss all is the total loss value of the distribution prediction model; m is the mth detection head; loss m is the loss value of the mth detection head; head pq Represents the pixel value at the position (p,q) in the output feature map of the detection head; P pq is the pixel value at the position (p,q) in the label scatter plot; [(100-m)÷100] is the coefficient of the mth detection head, which indicates that the longer the prediction time is, the smaller the contribution to the loss is. It also means that the longer the time is, the less reliable the prediction is, which is in line with common sense.

[0075] In this embodiment, the value of N is 192, and the data format of the single-channel scatter plot sequence is [P 192 ,P 191 ,…,P i ,…,P1],P i is the scatter plot of the i-th ship single channel; the value of k is 6; the value of M is 64. h and w are the height and width of the scatter plot of the ship single channel respectively. The values ​​of h and w are both 512.

[0076] In this embodiment, the output of the distribution prediction model is a 64-channel prediction scatter plot of 64×512×512 (that is, 64 single-channel black-background white-dot scatter plots), where the first channel represents the ship situation distribution after 1 minute, the second channel represents the ship situation distribution after 2 minutes, and the 64th channel represents the ship situation distribution after 64 minutes.

[0077] In the specific implementation process, Figure 2 and Figure 8 As shown, the predicted scatter plot output by each detection head of the distribution prediction model is a heat map. During the decoding process, the coordinates of each peak in the predicted scatter plot are taken as the white point coordinates of the predicted scatter plot.

[0078] Through the above steps, 64 single-channel black-background white-dot scatter plots are obtained, and each white dot is processed as follows:

[0079] 1. Assume that the coordinates of the white point in the predicted scatter plot are (x, y), and the height and width of the scatter plot are known to be (512, 512). Calculate the height and width of the physical area corresponding to the scatter plot as (H, W) (in kilometers);

[0080] 2. Calculate the offset kilometers of the white point relative to the upper left corner of the ship single channel scatter plot. Assume that the lateral offset is dx (unit: kilometer) and the longitudinal offset is dy (unit: kilometer). The calculation formulas are: dx = (x ÷ 512) × W, dy = (y ÷ 512) × H;

[0081] 3. Convert the offset kilometers into offset longitude and latitude. Assume that the horizontal offset longitude is dlon and the vertical offset latitude is dlat. The calculation formulas are: dlat=dx÷110;

[0082] 4. The latitude and longitude coordinates of the final white point are (lon, lat), and the calculation formula is: lon = lon1 + dlon, lat = lat1 + dlat, where (lon1, lat1) is the latitude and longitude coordinates of the upper left corner of the physical area corresponding to the predicted scatter plot.

[0083] In one embodiment, Figure 6 As shown in the figure, ships are randomly distributed in the sea area. Each ship is equipped with AIS (Automatic Identification System), which can accurately locate the longitude and latitude of each ship. Each triangle corresponds to a ship. Every minute, each ship uploads its longitude and latitude and type information to the central server for subsequent operations. For the convenience of intuitive display, as shown in the figure, Figure 7 As shown, each triangle representing a ship becomes a white dot in the scatter plot.

[0084] In summary, the present invention uses deep learning models to perform high-precision and high-efficiency ship situation prediction, providing strong support for navigation safety and shipping management.

Claims

1. A method for predicting offshore ship situation distribution, characterized in that: The following steps are involved: S1. Obtain the latitude and longitude of the corresponding ship in the offshore range according to the ship automatic identification system; S2, generating a single channel scatter plot of ships based on the longitude and latitude in a single time period; S3, obtaining N single-channel scatter plots of ships and arranging them in time sequence to obtain a single-channel scatter plot sequence; S4, using the single-channel scatter plot sequence as the input of the trained distribution prediction model to obtain prediction scatter plots after different time periods; S5. Decode the prediction scatter plot to obtain the prediction result of offshore ship situation distribution; The scatter plot includes a single-channel ship scatter plot and a predicted scatter plot. The scatter plot is black and white, with white representing ships. The specific method of generating a ship single channel scatter plot includes the following sub-steps: S2-1. Without considering the arc of the earth, a rectangle is set in the offshore area and used as the prediction area; the longitude and latitude coordinates of the upper left corner of the rectangle are P1 (lon1, lat1), and the longitude and latitude coordinates of the lower right corner are P2 (lon2, lat2); S2-2, synthesize a pure black image with a channel number of 1 and a size of h×w, map the ships in the rectangle to the pure black image with white dots according to the height-to-width ratio, and obtain a single-channel scatter plot of ships with a size of h×w; The specific method of decoding the prediction scatter plot and obtaining the prediction results of offshore ship situation distribution is as follows: Calculate the height and width of the physical area corresponding to the predicted scatter graph according to the coordinates of the white points in the predicted scatter graph and the size of the predicted scatter graph; According to the height and width of the physical area corresponding to the predicted scatter plot, calculate the offset distance of the white point relative to the upper left corner of the ship's single channel scatter plot; The offset distance is converted into offset longitude and latitude to obtain the predicted position of offshore ships, and then the offshore ship situation distribution prediction results are obtained.

2. A method for predicting offshore ship situation distribution according to claim 1, characterized in that: The calculation expressions for the height H and width W of the prediction area are: H=110×|lat1-lat2| Where cos(.) is the cosine function.

3. A method for predicting offshore ship situation distribution according to claim 1, characterized in that: The duration of a single time period is the same as the time interval for the ship's automatic identification system to upload the ship's longitude and latitude.

4. A method for predicting offshore ship situation distribution according to claim 1, characterized in that: The distribution prediction model is a U-net neural network model, which includes six sequentially connected downsampling modules, six upsampling modules, and M detection heads; Each downsampling module halves the input image size and doubles the number of channels; the input of the first downsampling module is the input of the U-net neural network model; the output of the sixth downsampling module is the input of the first upsampling module; The first five upsampling modules double the input image size and halve the number of channels; the sixth upsampling module reduces the number of channels to 1 / k; The output of the nth downsampling module and the output of the 6th-nth upsampling modules are connected in the channel dimension and used as the input of the 7th-nth upsampling modules; the output of the 6th upsampling module is the input of M detection heads; n = 1, 2, 3, 4, 5; M detection heads correspond to the prediction scatter plots of M different prediction time points; The value of N is 192, and the data format of the single-channel scatter plot sequence is [P 192 ,P 191 ,…,P i ,…,P1],P i is the single channel scatter plot of the i-th ship; the value of k is 6; the value of M is 64.

5. A method for predicting offshore ship situation distribution according to claim 4, characterized in that: Each downsampling module includes a first 1×1 convolutional layer, a first BN layer, a first ReLU activation layer, a first 3×3 convolutional layer, a second BN layer, a second ReLU activation layer, a second 1×1 convolutional layer, a third BN layer, a concatenation layer, a third ReLU activation layer and a 3×3 maximum pooling layer connected in sequence; the other input of the concatenation layer is the output of the fourth BN layer; the input of the fourth BN layer is the output of the third 1×1 convolutional layer; the input of the first 1×1 convolutional layer and the input of the third 1×1 convolutional layer are both the input of the downsampling module to which they belong; the output of the 3×3 maximum pooling layer is the output of the downsampling module to which they belong.

6. A method for predicting offshore ship situation distribution according to claim 4, characterized in that: Each upsampling module includes a second 3×3 convolutional layer, a fifth BN layer, a third 3×3 convolutional layer, a sixth BN layer and a 2×2 convolutional layer connected in sequence; the input of the second 3×3 convolutional layer is the input of the upsampling module to which it belongs; and the output of the 2×2 convolutional layer is the output of the upsampling module to which it belongs.

7. A method for predicting offshore ship situation distribution according to claim 4, characterized in that: During the training process of the distribution prediction model, the initial learning rate is 0.01, and the subsequent learning rate is adjusted by the time decay method; the batch size is 4; the number of training iterations is 150; the loss function is: Among them, loss all is the total loss value of the distribution prediction model; m is the mth detection head; loss m is the loss value of the mth detection head; head pq Represents the pixel value at the position (p,q) in the output feature map of the detection head; P pq is the pixel value at the position (p,q) in the label scatter plot; h and w are the height and width of the ship single channel scatter plot, respectively.

Images