A method and system for predicting ship stern flow

Abstract

The application discloses a ship stern flow prediction method and system, and relates to the technical field of aerospace. The method comprises the following steps: constructing a neural network with a frequency domain attention mechanism; training the neural network by using ship stern flow data at a historical moment, so as to obtain a trained neural network; the trained neural network is used for predicting ship stern flow data at a future moment; obtaining ship stern flow data before a current moment of a final stage of carrier aircraft landing; and predicting by using the trained neural network. The application can realize accurate prediction of ship stern flow data at the final stage of carrier aircraft landing.

Description

Technical Field

[0001] This invention relates to the field of aerospace technology, and in particular to a method and system for predicting stern currents. Background Technology

[0002] Stern currents are formed by airflow passing over the hull of a moving ship. The strongest stern current interference occurs within a few hundred meters of the stern of an aircraft carrier. When carrier-based aircraft land, flight quality in this area is significantly affected, resulting in large lateral and vertical trajectory deviations, and in severe cases, potentially leading to a collision. During landing, stern currents are a crucial factor interfering with landing control and affecting landing accuracy. Therefore, forecasting stern currents at the end of landing is an essential skill.

[0003] With the development of deep learning, using neural networks to predict sequential data can effectively extract features from the data, allowing the data of each component of a ship's stern flow to be treated as time-series data. Currently, commonly used neural network structures for sequential data prediction generally include recurrent neural networks (RNNs) and attention-based neural networks. RNNs iteratively process the sequential data through recurrent neural units, calculating the prediction results. However, RNNs are prone to problems such as vanishing and exploding gradients during training, making it difficult for their prediction accuracy to meet the requirements of practical applications. Attention mechanisms are a widely used network structure in deep learning. By introducing attention mechanisms, neural networks can better capture the correlations between data, improving the prediction accuracy of sequential data. However, the high time complexity of attention-based neural networks makes it difficult for the real-time performance of prediction methods to meet the requirements of practical applications. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for predicting stern flow, which can accurately predict stern flow data during the terminal phase of carrier-based aircraft landing.

[0005] To achieve the above objectives, the present invention provides the following solution:

[0006] A method for predicting stern currents, comprising:

[0007] A neural network incorporating a frequency domain attention mechanism is constructed. The neural network comprises: a fully connected layer, a normalization layer, an encoding tree module, and a decoder module. The encoding tree module includes three encoder modules, which serve as the root node, left leaf node, and right leaf node of the encoding tree module, respectively. Each encoder module includes a frequency domain self-attention module, a residual connection, a moving average module, and a one-dimensional convolutional module. The decoder module includes a frequency domain self-attention module, a frequency domain cross-attention module, a one-dimensional convolutional layer, and a moving average module. The frequency domain self-attention module and the frequency domain cross-attention module have the same structure, both including: three fully connected layers, inverse Fourier transform, frequency domain sampling, and Fourier transform.

[0008] The neural network is trained using historical stern flow data to obtain a trained neural network; the trained neural network is used to predict stern flow data at future times.

[0009] The system acquires stern flow data up to the current moment during the final stage of carrier-based aircraft landing and uses the resulting neural network for prediction.

[0010] Optionally, the stern flow data includes sequence data of the stern flow in three dimensions: longitudinal, transverse, and vertical.

[0011] Optionally, training the neural network using historical stern flow data to obtain the trained neural network specifically includes:

[0012] The stern flow data at historical moments are divided into training, validation, and test sets;

[0013] The stern flow data for each historical moment is split to obtain sample data. The sample data includes: stern flow data for 48 time steps as input and stern flow data for the next 8 time steps as output to be predicted.

[0014] The neural network is trained using the split training set.

[0015] Optionally, the step of acquiring the stern flow data up to the current moment of the terminal phase of the carrier-based aircraft's landing and using the resulting neural network for prediction specifically includes:

[0016] The stern flow data prior to the current moment first passes through a fully connected layer and a normalization layer to obtain the feature representation of the sequence data;

[0017] The features represent the periodic and trend data output by the encoder at the root node of the coding tree module.

[0018] The periodic and trend data are processed by encoders represented by the left and right leaf nodes, respectively, and the periodic and trend data at the left and right leaf nodes are output respectively.

[0019] The periodic and trend terms at the left and right leaf nodes are added together to obtain the feature representation of the sequence data after the coding tree;

[0020] The encoded sequence data feature representation is processed by the decoder module to output the predicted periodic and trend terms;

[0021] The final prediction result for the stern flow sequence data is obtained by adding the predicted periodic term and trend term.

[0022] A ship stern flow prediction system, comprising:

[0023] A neural network construction module is used to construct a neural network incorporating a frequency domain attention mechanism. The neural network includes: a fully connected layer, a normalization layer, an encoding tree module, and a decoder module. The encoding tree module includes: three encoder modules, which serve as the root node, left leaf node, and right leaf node of the encoding tree module, respectively. Each encoder module includes: a frequency domain self-attention module, a residual connection, a moving average module, and a one-dimensional convolutional module. The decoder module includes: a frequency domain self-attention module, a frequency domain cross-attention module, a one-dimensional convolutional layer, and a moving average module. The frequency domain self-attention module and the frequency domain cross-attention module have the same structure, both including: three fully connected layers, inverse Fourier transform, frequency domain sampling, and Fourier transform.

[0024] A neural network training module is used to train the neural network using historical stern flow data to obtain a trained neural network; the trained neural network is used to predict stern flow data at future times.

[0025] The stern flow prediction module is used to acquire stern flow data up to the current moment during the final stage of carrier-based aircraft landing; and then uses the resulting neural network for prediction.

[0026] A ship stern flow prediction system includes: at least one processor, at least one memory, and computer program instructions stored in the memory, which implement the method when the computer program instructions are executed by the processor.

[0027] A storage medium having computer program instructions stored thereon, which, when executed by a processor, implement the method.

[0028] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0029] This invention provides a method and system for predicting ship stern flow, which constructs a neural network incorporating a frequency domain attention mechanism. This method combines Fourier analysis with a neural network to predict ship stern flow, enabling the neural network to extract features of sequence data in the frequency domain. This results in more accurate predictions of the sequence data and meets the real-time requirements of practical prediction problems. Attached Figure Description

[0030] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0031] Figure 1 A schematic diagram of the stern current prediction method provided by the present invention;

[0032] Figure 2 This is a schematic diagram of the neural network structure that incorporates the frequency domain attention mechanism provided by the present invention. Detailed Implementation

[0033] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0034] The purpose of this invention is to provide a method and system for predicting stern flow, which can accurately predict stern flow data of carrier-based aircraft at the terminal phase of landing by introducing a neural network with a frequency domain attention mechanism.

[0035] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0036] like Figure 1 As shown, the stern current prediction method provided by the present invention includes:

[0037] S101, construct a neural network incorporating a frequency domain attention mechanism; the neural network includes: a fully connected layer, a normalization layer, an encoding tree module, and a decoder module; the encoding tree module includes: three encoder modules; the three encoder modules serve as the root node, left leaf node, and right leaf node of the encoding tree module, respectively; each encoder module includes: a frequency domain self-attention module, a residual connection, a moving average module, and a one-dimensional convolution module; the decoder module includes: a frequency domain self-attention module, a frequency domain cross-attention module, a one-dimensional convolution layer, and a moving average module; the frequency domain self-attention module and the frequency domain cross-attention module have the same structure, both including: three fully connected layers, inverse Fourier transform, frequency domain sampling, and Fourier transform;

[0038] The calculation formula for the fully connected layer network is y = Wx + b; in the normalization layer, the mean vector μ and variance vector σ of the input data are calculated first, and the formula for calculating the i-th dimension of the mean vector is... The formula for calculating the i-th dimension of the variance vector is: Where n is the dimension of the output data of the fully connected layer. Where ⊙ represents the Hadamard product operation, γ represents the scaling parameter vector, and β represents the translation parameter vector.

[0039] After passing through each layer of the network, the data is activated by the tanh function before being input into the next layer. The formula for the tanh activation function is:

[0040] The frequency domain attention module takes vectors x and z as input. First, these vectors are input into the fully connected layer, where v = W. v x, k = W k x, q = W q z; then, the vectors q, k, and v are subjected to a fast Fourier transform to obtain the frequency domain vector f. q =FFT(q), f k =FFT(k), f v =FFT(v); This function obtains the frequency domain vector by random sampling. Then, attention calculation is performed in the frequency domain. Using the calculated attention matrix A Perform weighted calculations. right Frequency domain component completion is performed, and the final output of the frequency domain attention module is obtained using inverse fast Fourier transform. The calculation formula is as follows:

[0041] S102, the neural network is trained using historical stern flow data to obtain a trained neural network; the trained neural network is used to predict stern flow data at future times; the stern flow data includes sequence data of the longitudinal, lateral and vertical dimensions of the stern flow.

[0042] S102 specifically includes:

[0043] The stern flow data at historical moments are divided into training, validation, and test sets; these sets are then divided in a 7:1:2 ratio to obtain the training, validation, and test sets.

[0044] The stern flow data for each historical moment is split to obtain sample data. The sample data includes: stern flow data for 48 time steps as input and stern flow data for the next 8 time steps as output to be predicted.

[0045] The neural network is trained using the split training set. The training process is as follows:

[0046] The split samples in the training set are randomly shuffled, and 16 samples are selected to form a batch and input into the neural network. The neural network outputs the prediction result of the future stern flow data of the batch of samples. The prediction result and the actual result are input into the loss function, and gradient backpropagation is performed based on the calculation result of the loss function to update the parameters of each module of the neural network. The loss function is selected as the mean squared error function, and the gradient backpropagation uses the Adam optimizer with an initial learning rate of 0.001.

[0047] A training cycle is defined as feeding all samples from the training set into the neural network once, and the training process lasts for 30 cycles. In each cycle, an inference is performed on all samples in the validation set. The neural network's predictions for all samples in the validation set and the actual data in the validation set are input into the absolute error function. The average of the absolute errors in the three dimensions is used as the evaluation metric for the current neural network.

[0048] Record the evaluation metrics on the validation set during training and save the neural network parameter file after each training cycle.

[0049] Finally, the neural network parameter file saved at the moment when the evaluation index is optimal is selected as the weight of the neural network model used to predict future stern flow data.

[0050] The stern flow data from the test set is input into a trained neural network with a frequency domain attention mechanism to predict the stern flow data at future times.

[0051] The test set is split into 16 samples each time, which are then fed into the neural network. The neural network outputs the prediction results of the future stern flow data of the batch of samples.

[0052] S103: Obtain the stern flow data of the carrier-based aircraft up to the current moment during the final stage of landing; and use the resulting neural network for prediction.

[0053] S103 specifically includes:

[0054] The stern flow data prior to the current moment first passes through a fully connected layer and a normalization layer to obtain the feature representation of the sequence data;

[0055] The features represent the periodic and trend data output by the encoder at the root node of the coding tree module.

[0056] The periodic and trend data are processed by the encoders represented by the left and right leaf nodes, respectively, and the periodic and trend data at the left and right leaf nodes are output respectively. The calculation formulas for the encoders represented by the left and right leaf nodes are as follows:

[0057] The periodic and trend terms at the left and right leaf nodes are added together to obtain the feature representation of the sequence data after the coding tree;

[0058] The encoded sequence data feature representation is processed by the decoder module, which outputs the predicted periodic and trend terms; the calculation formula of the decoder module is as follows:

[0059] The final prediction result for the stern flow sequence data is obtained by adding the predicted periodic term and trend term. The calculation formula is y = y t +y s .

[0060] As a specific embodiment, the present invention also provides a ship stern flow prediction system, comprising:

[0061] A neural network construction module is used to construct a neural network incorporating a frequency domain attention mechanism. The neural network includes: a fully connected layer, a normalization layer, an encoding tree module, and a decoder module. The encoding tree module includes: three encoder modules, which serve as the root node, left leaf node, and right leaf node of the encoding tree module, respectively. Each encoder module includes: a frequency domain self-attention module, a residual connection, a moving average module, and a one-dimensional convolutional module. The decoder module includes: a frequency domain self-attention module, a frequency domain cross-attention module, a one-dimensional convolutional layer, and a moving average module. The frequency domain self-attention module and the frequency domain cross-attention module have the same structure, both including: three fully connected layers, inverse Fourier transform, frequency domain sampling, and Fourier transform.

[0062] A neural network training module is used to train the neural network using historical stern flow data to obtain a trained neural network; the trained neural network is used to predict stern flow data at future times.

[0063] The stern flow prediction module is used to acquire stern flow data up to the current moment during the final stage of carrier-based aircraft landing; and then uses the resulting neural network for prediction.

[0064] In order to execute the method corresponding to Embodiment 1 above and achieve the corresponding functions and technical effects, the present invention also provides a ship stern flow prediction system, including: at least one processor, at least one memory, and computer program instructions stored in the memory, wherein the method is implemented when the computer program instructions are executed by the processor.

[0065] Based on the above description, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention. The aforementioned computer storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory, random access memory, magnetic disks, or optical disks.

[0066] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0067] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for predicting stern currents, characterized in that, include: Construct a neural network that incorporates a frequency domain attention mechanism; The neural network comprises: a fully connected layer, a normalization layer, an encoding tree module, and a decoder module; the encoding tree module comprises: three encoder modules; the three encoder modules serve as the root node, left leaf node, and right leaf node of the encoding tree module, respectively; each encoder module comprises: a frequency domain self-attention module, a residual connection, a moving average module, and a one-dimensional convolution module; the decoder module comprises: a frequency domain self-attention module, a frequency domain cross-attention module, a one-dimensional convolution layer, and a moving average module; the frequency domain self-attention module and the frequency domain cross-attention module have the same structure, both comprising: three fully connected layers, inverse Fourier transform, frequency domain sampling, and Fourier transform; The neural network is trained using historical stern flow data to obtain a trained neural network; the trained neural network is used to predict stern flow data at future times. Acquire the stern flow data of the carrier-based aircraft up to the current moment during the final stage of landing; and use the resulting neural network for prediction. The process of acquiring stern flow data up to the current moment before the terminal phase of carrier-based aircraft landing, and then using the resulting neural network for prediction, specifically includes: The stern flow data prior to the current moment first passes through a fully connected layer and a normalization layer to obtain the feature representation of the sequence data; The features represent the periodic and trend data output by the encoder at the root node of the coding tree module. The periodic and trend data are processed by encoders represented by the left and right leaf nodes, respectively, and the periodic and trend data at the left and right leaf nodes are output respectively. The periodic and trend terms at the left and right leaf nodes are added together to obtain the feature representation of the sequence data after the coding tree; The encoded sequence data feature representation is processed by the decoder module to output the predicted periodic and trend terms; The final prediction result for the stern flow sequence data is obtained by adding the predicted periodic term and trend term.

2. The method for predicting stern currents according to claim 1, characterized in that, The stern flow data includes sequence data of the stern flow in three dimensions: longitudinal, lateral, and vertical.

3. The method for predicting stern currents according to claim 1, characterized in that, The process of training the neural network using historical stern flow data to obtain the trained neural network specifically includes: The stern flow data at historical moments are divided into training, validation, and test sets; The stern flow data for each historical moment is split to obtain sample data. The sample data includes: stern flow data for 48 time steps as input and stern flow data for the next 8 time steps as output to be predicted. The neural network is trained using the split training set.

4. A ship stern current prediction system, used to implement the ship stern current prediction method as described in any one of claims 1-3, characterized in that, include: The neural network building module is used to construct neural networks that incorporate frequency domain attention mechanisms. The neural network comprises: a fully connected layer, a normalization layer, an encoding tree module, and a decoder module; the encoding tree module comprises: three encoder modules; the three encoder modules serve as the root node, left leaf node, and right leaf node of the encoding tree module, respectively; each encoder module comprises: a frequency domain self-attention module, a residual connection, a moving average module, and a one-dimensional convolution module; the decoder module comprises: a frequency domain self-attention module, a frequency domain cross-attention module, a one-dimensional convolution layer, and a moving average module; the frequency domain self-attention module and the frequency domain cross-attention module have the same structure, both comprising: three fully connected layers, inverse Fourier transform, frequency domain sampling, and Fourier transform; A neural network training module is used to train the neural network using historical stern flow data to obtain a trained neural network; the trained neural network is used to predict stern flow data at future times. The stern flow prediction module is used to acquire stern flow data up to the current moment during the final stage of carrier-based aircraft landing; and then uses the resulting neural network for prediction.

5. A ship stern current prediction system, characterized in that, include: The method comprises at least one processor, at least one memory, and computer program instructions stored in the memory, which, when executed by the processor, implement the stern flow prediction method as described in any one of claims 1-3.

6. A storage medium storing computer program instructions thereon, characterized in that, When the computer program instructions are executed by the processor, the stern flow prediction method as described in any one of claims 1-3 is implemented.

Images