Track prediction method for ship in distress
By adopting the model structure of the sequencing curl neural network in the trajectory prediction of the ship's trajectory and the passage of the passage, the learning difficulty of multi -channel characteristics input and long sequence input is solved. Sensitivity of features.
Patent Information
- Application Number
- CN202411313097.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-20
- Publication Date
- 2025-01-28
AI Technical Summary
The prior art has difficulties in learning neural networks under multi-channel feature input in ship trajectory prediction, and it is difficult to effectively deal with long sequence input and variable meteorological conditions.
Adopt a model structure of the self -attention layer and channel attention layer combining the timing convolutional neural network, strengthen the effects of meteorological characteristics through the self -attention layer, and enhance the cross -channel feature learning ability through the channel attention layer, and process the long sequence input.
It improves the accuracy of the ship's trajectory prediction, reduces the comparative form error compared with real data, and enhances the learning ability of multi -characteristic input.
Smart Images
Figure CN119377664A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of ship trajectory prediction, and in particular to a method for predicting the trajectory of a ship in distress. Background Art
[0002] With the development of economy and trade and the implementation of marine development strategy, ocean shipping, fishing, offshore oil exploration and other activities are becoming more frequent, and the frequency of marine hazards and accidents is also increasing. When a maritime disaster occurs, in order to minimize the loss of life and property, it is necessary to accurately and quickly estimate the search area, thereby improving the success rate of maritime search and rescue.
[0003] There are many problems in the process of predicting the trajectory of a distressed ship. The increasing amount of ship data has brought new challenges to data management and application. More accurate analysis of data with both time and space has increased the difficulty of data query and processing. The trajectory data has a certain scale, and data sparsity is inevitable. Some ships have less data in other sea areas, but more in a certain sea area. For the ship trajectory, if the sampling rate is low, two trajectory points far apart will also increase the uncertainty of the trajectory data. The navigation of the distressed ship is mainly determined by the current and tide. In addition, the changeable weather conditions will also have a significant impact on the trajectory of the ship. In addition to the ship itself, it is difficult for a third party to know the size of the external force on the ship and how it acts. In short, trajectory prediction is both difficult and hot.
[0004] At present, most of the ship trajectory methods based on artificial intelligence use recurrent neural network structure and convolutional neural network as the basic structure of the model. However, recurrent neural network is prone to the problem of disappearance of long-distance dependency when facing long sequence prediction tasks; convolutional neural network needs to increase the convolution kernel size and the number of network layers when processing long sequence input, which puts higher requirements on the computing power of the model. In addition, both lack focus on multi-channel features and it is difficult to distinguish the impact of each channel feature on the prediction results. Therefore, in order to address the problems existing in the current mainstream methods and solve the problem of difficulty in cross-channel feature learning after multi-feature input, a new ship trajectory prediction method needs to be designed. Summary of the invention
[0005] In order to solve the problems existing in the prior art, the present invention provides a trajectory prediction method for a ship in distress, which is used to solve the problem of neural network learning difficulties under the condition of multi-channel feature input.
[0006] To solve the above problems, the technical solution of the present invention is as follows: A method for predicting the trajectory of a ship in distress comprises the following steps:
[0007] S1 constructs a data set; mainly based on historical ship position data, combined with relevant meteorological conditions data as additional features, and coded as input data;
[0008] The S2 dataset normalizes different features to accelerate the convergence of the model;
[0009] S3 divides the data set into training set, test set, and validation set;
[0010] S4 builds the model structure;
[0011] The output result of the S5 model structure is the ship position information at the prediction time.
[0012] Furthermore, in S1, the meteorological condition data includes wind direction, wind speed, waves, ocean currents, air temperature, sea temperature, precipitation, storm surge, etc.
[0013] Furthermore, in S1, a fixed time interval is adopted and a sliding window algorithm is used to divide the input data into equal-length intervals. For missing data, linear interpolation is used to complete them.
[0014] Furthermore, in S3, the ratio of training set, test set, and validation set is 8:1:1.
[0015] Furthermore, in S4, the model structure is constructed by first constructing the self-attention layer. The self-attention calculation method in Transformer is used to map the data vector into feature matrices Q, K, and V respectively through three attention calculation matrices, and the output results are calculated according to the feature matrices. The calculation is shown in the public notice:
[0016]
[0017] Further, in S4, the model structure is constructed and then the channel attention layer is constructed. The original input size of the channel attention layer is C×H×W. The original input is reshaped into a first feature matrix. The size of the first feature matrix is C×(H×W). The original input is reshaped and transposed into a second feature matrix. The size of the second feature matrix is (H×W)×C. The first feature matrix and the second feature matrix will form a third feature matrix with a size of C×C through matrix multiplication. The third feature matrix is restored to its size by matrix multiplication with the original input, and a residual connection is used to retain the original input information.
[0018] Furthermore, in S4, a model structure is constructed. After two layers of attention operations, the features will be input into a temporal convolutional neural network. The temporal convolutional neural network is a multi-layer structure, and each layer consists of three layers: a temporal convolution layer, a batch normalization layer, and a ReLu activation layer.
[0019] Furthermore, the temporal convolution layer includes causal convolution, dilated convolution and residual structure.
[0020] Furthermore, the causal convolution only depends on the value of the next layer at time t and before for the value of the previous layer at time t.
[0021] Furthermore, the dilated convolution skips part of the input and increases the receptive field of the convolution.
[0022] Compared with the prior art, the present invention has the following beneficial effects:
[0023] The method proposed in this application strengthens the impact of various meteorological characteristics on ship behavior, while solving the problem of neural network learning difficulties under multi-channel feature input, and enhancing the learning ability for long sequence input, thereby improving the accuracy of the prediction and reducing the mean square error compared with the real data. BRIEF DESCRIPTION OF THE DRAWINGS
[0024] Figure 1 This is a structural diagram of the model of the present invention;
[0025] Figure 2 This is a schematic diagram of the self-attention calculation of the present invention;
[0026] Figure 3 This is the channel attention structure diagram of the present invention. DETAILED DESCRIPTION
[0027] A method for predicting the trajectory of a ship in distress comprises the following steps:
[0028] S1. Building a dataset
[0029] The historical position data of the ship and the relevant meteorological conditions data are taken as additional feature inputs. In addition to the historical position data of the ship, meteorological conditions data such as wind direction, wind speed, waves, ocean currents, air temperature, sea temperature, precipitation, storm surges, etc. are taken as additional feature inputs. A fixed time interval is adopted and the sliding window algorithm is used to divide the data into equal length intervals. For missing data, linear interpolation method is used to fill it.
[0030] S2, the data set normalizes different features to accelerate the convergence of the model.
[0031] S3. Divide the dataset into training set, test set, and validation set in a ratio of 8:1:1.
[0032] S4. Construct the model structure. The model structure is as follows Figure 1 shown.
[0033] The first is a self-attention layer, which uses the self-attention calculation method in Transformer, as shown in the self-attention calculation diagram. The data vector is mapped to feature matrices Q, K, and V through three attention calculation matrices, and the output result is calculated based on the feature matrix. The calculation is shown in the public diagram:
[0034]
[0035] After strengthening the meteorological features as input, the number of input data channels is too large, and simple self-attention calculation cannot fully learn cross-channel features. In order to strengthen the differentiated learning of different channel features, this patent adds a channel attention layer afterwards. As shown in the channel attention structure diagram, the original input size is C×H×W. The original input is reshaped and transposed to form two feature matrices with sizes of C×(H×W) and (H×W)×C. The two feature matrices will form a feature matrix of size C×C after matrix multiplication. The feature matrix is matrix multiplied with the original input to restore the size. Finally, a residual connection is used to retain the original input information.
[0036] After two layers of attention operations, the features will be input into the temporal convolutional neural network. The temporal convolutional neural network is a multi-layer structure, and each layer consists of three layers: a temporal convolution layer, a batch normalization layer, and a ReLu activation layer.
[0037] The temporal convolution layer is the core structure of the temporal convolutional neural network, which is mainly composed of causal convolution, dilated convolution and residual structure. The causal convolution only depends on the value of the next layer at time t and before for the value of the previous layer at time t. The difference from the traditional convolutional neural network is that the causal convolution cannot see future data. It is a unidirectional structure, not bidirectional, and is a strict time constraint model.
[0038] The dilated convolution skips part of the input and increases the receptive field of the convolution. The modeling length of time in traditional convolutional neural networks is limited by the size of the convolution kernel. If you want to capture longer dependencies, you need to stack many layers linearly. The dilated convolution selectively calculates part of the input, which reduces the number of network layers and the possibility of gradient disappearance.
[0039] The residual structure connection is also used to avoid information loss that may be caused by deepening the network. The feature extraction capability enhanced by the temporal convolutional neural network can enhance the feature capture capability for long sequence inputs with the same amount of computation, allowing the model to learn the impact of longer sequences on current information.
[0040] S5. The output result of the model is the ship position information at the prediction time, and the loss function is the mean square error loss function.
[0041] The above specific implementation methods are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to examples, a person skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the scope of the technical solution of the present invention, which should be included in the scope of the claims of the present invention.
Claims
1. A method for predicting the trajectory of a ship in distress, characterized in that: The following steps are involved: S1 constructs a data set; mainly based on historical ship position data, combined with relevant meteorological conditions data as additional features, and coded as input data; The S2 dataset normalizes different features to accelerate the convergence of the model; S3 divides the data set into training set, test set, and validation set; S4 builds the model structure; The output result of the S5 model structure is the ship position information at the prediction time.
2. The prediction method according to claim 1, characterized in that: In S1, the meteorological condition data includes wind direction, wind speed, waves, ocean currents, air temperature, sea temperature, precipitation, storm surge, etc.
3. The prediction method according to claim 2, characterized in that: In S1, a fixed time interval is used to divide the input data into equal-length intervals using a sliding window algorithm. For missing data, linear interpolation is used to complete it.
4. The prediction method according to claim 3, characterized in that: In S3, the ratio of training set, test set, and validation set is 8:1:
1.
5. The prediction method according to claim 4, characterized in that: In S4, the model structure is constructed by first constructing the self-attention layer. The self-attention calculation method in Transformer is used to map the data vector into feature matrices Q, K, and V respectively through three attention calculation matrices, and the output results are calculated according to the feature matrices. The calculation is shown in the public notice:
6. The prediction method according to claim 5, characterized in that: In S4, the model structure is constructed and then the channel attention layer is constructed. The original input size of the channel attention layer is C×H×W. The original input is reshaped into a first feature matrix with a size of C×(H×W). The original input is reshaped and transposed into a second feature matrix with a size of (H×W)×C. The first feature matrix and the second feature matrix will form a third feature matrix with a size of C×C through matrix multiplication. The third feature matrix will be restored to its size by matrix multiplication with the original input, and a residual connection will be used to retain the original input information.
7. The prediction method according to claim 6, characterized in that: In S4, the model structure is constructed. After two layers of attention operation, the features will be input into the temporal convolutional neural network. The temporal convolutional neural network is a multi-layer structure, and each layer consists of three layers: a temporal convolution layer, a batch normalization layer, and a ReLu activation layer.
8. The prediction method according to claim 7, characterized in that: The temporal convolution layer includes causal convolution, dilated convolution and residual structure.
9. The prediction method according to claim 8, characterized in that: The causal convolution depends only on the value of the previous layer at time t and before time t in the next layer.
10. The prediction method according to claim 9, characterized in that: The dilated convolution skips part of the input and increases the receptive field of the convolution.