AUV (Autonomous Underwater Vehicle) drift trajectory prediction method for optimizing bidirectional long-short-term memory-time variable attention model based on nonlinear Kepler optimization algorithm

Abstract

The invention provides a multivariable AUV (Autonomous Underwater Vehicle) drift trajectory prediction method based on an NKOA-BiLSTM-TVA model. According to the method, drift trajectory prediction of an AUV is realized by utilizing trajectory data of the AUV and marine environment factors. According to the method, a time variable attention layer (TVA) which crosses a time dimension and a variable dimension is designed and is fused with the time variable attention layer (TVA). According to the method, the KOA algorithm is initialized through reverse learning, a nonlinear convergence factor changing along with time is fused into the original KOA algorithm to improve the KOA algorithm, hyper-parameters of the BiLSTM-TVA model are optimized through the NKOA algorithm, and the search breadth and depth of the algorithm are improved through the NKOA algorithm on the basis of the original algorithm. The AUV drift trajectory prediction method researched by the invention is of great significance in predicting the water outlet point of the wrecked AUV, salvaging and recycling the wrecked AUV, researching fault causes in the later period and the like, and is of great significance in protecting data on the AUV.

Description

Technical Field

[0001] The present invention relates to the fields of deep learning and AUV trajectory prediction, and in particular to an AUV drift trajectory prediction method based on an NKOA-BiLSTM-TVA model. Background Art

[0002] Autonomous underwater vehicle (AUV) is an unmanned underwater autonomous vehicle that can perform a variety of underwater tasks and is widely used in marine engineering, military and civilian fields. AUV can work for a long time without supervision and conduct large-scale ocean exploration and data collection. This is of great significance for studying the marine environment, marine life, marine chemistry, etc. AUV can reach deep sea areas that humans cannot easily reach and provide valuable first-hand data. Applications in the military field, such as performing reconnaissance missions, conducting submarine patrols, and detecting enemy submarines. They can complete tasks in complex and dangerous environments and reduce personnel risks. Due to its importance, there are research topics such as AUV trajectory prediction, path planning, trajectory control, and trajectory tracking based on AUV. The safety issue of AUV has always been a hot topic of research. When an AUV fails during the execution of a mission, it is of great significance to predict the last possible water outlet or sinking point of the AUV, salvage and recovery, and later study the cause of the failure and solution measures. Researchers make safety warnings and rescue preparations in advance to ensure that the AUV can be discovered and recovered in time, which is also of great significance for protecting the data on the AUV. The principle of AUV drift trajectory prediction is to predict its future trajectory through its historical trajectory data and marine environmental factors. The research on these issues involves trajectory prediction and time series prediction. Therefore, it is very important to study an effective AUV drift trajectory prediction method.

[0003] In the early research of trajectory prediction, traditional non-neural network prediction methods include prediction based on dynamic modeling, prediction based on physical model, prediction based on Kalman filter and fitting polynomial. However, due to the complex ocean environment of AUV movement, traditional prediction methods perform poorly when considering factors such as complex ocean environment. With the development of deep learning, trajectory prediction methods based on neural networks have achieved better prediction results. Recurrent neural networks (RNNs) show good performance in processing time series data. They are often used in fields such as machine translation and speech recognition, and are also applied to trajectory prediction tasks. However, with the increase in the complexity of prediction tasks, RNNs face problems such as long-term dependence, gradient disappearance and gradient explosion. LSTM is an improved recurrent neural network. New modules are added to address the problems of RNNs, effectively solving the limitations of RNNs in time series prediction tasks. The internal structure of LSTM contains gating mechanisms such as forget gate, input gate and output gate. The forget gate processes the current input and the output of the hidden state at the previous moment. It controls which information is forgotten and which information is retained; the input gate controls the input of new information; the output gate controls the output of information in the cell state. Recurrent neural networks based on RNN improvements and models that integrate attention mechanisms have emerged and achieved better prediction results than RNN in trajectory prediction tasks.

[0004] In the existing AUV trajectory prediction methods, the model usually only uses the historical trajectory of the AUV to predict the trajectory at future moments, while ignoring the impact of complex marine environmental factors on the AUV trajectory. At the same time, the existing methods only use a simple attention mechanism to extract the dependent features of the trajectory in the time dimension, and do not utilize the feature dependencies between the trajectory variable dimensions. Therefore, it is necessary to design an effective attention mechanism to solve these problems. In the AUV trajectory prediction task, finding the optimal hyperparameters of the model is the key to improving the prediction accuracy. Traditional methods use a large number of experiments to find the optimal hyperparameters, but this is not only inefficient, but also easy to cause the model to fall into local optimality. The Kepler Optimization Algorithm (KOA) is a newly proposed algorithm for solving optimization problems in 2023. The idea of ​​the algorithm comes from Kepler's laws of planetary motion. The hyperparameters of the AUV trajectory prediction model are optimized by improving the KOA algorithm, and the performance of the model in predicting the drift trajectory of the AUV is improved by finding the optimal hyperparameters of the model. Summary of the invention

[0005] The present invention aims to study a prediction method for predicting the drift trajectory of AUV, which is applied in the field of AUV and ship trajectory prediction. This method uses a multivariate AUV drift trajectory prediction model NKOA-BiLSTM-TVA to predict the drift trajectory of AUV and compares the prediction results with those of other AUV trajectory prediction models.

[0006] The specific implementation of the invention is as follows:

[0007] (1) Obtain the trajectory data of the AUV and the ocean current environment data of the AUV activity area.

[0008] (2) The data preprocessing process includes missing value filling and data normalization.

[0009] (3) The data preprocessing process includes missing value filling and data normalization.

[0010] The missing value filling methods include: using the moving average method to fill the missing values ​​by taking the average value of the data points within a certain time window.

[0011] The data normalization method includes: in order to accelerate the convergence speed of the model and improve the accuracy of model prediction, the maximum and minimum normalization method is used to normalize the data and scale the data to a range between [0, 1].

[0012] (4) The normalized data set is processed using the sliding window method.

[0013] (5) CNN is introduced to process spatial features, and the temporal information between trajectories is further learned through the BiLSTM network. At the same time, a cross-time dimension and variable dimension fusion attention mechanism layer, the time dimension attention layer, is used to extract the dependencies between the time dimension and variable dimension features in the historical trajectory. The KOA algorithm is initialized based on reverse learning, and a nonlinear convergence factor is added to the algorithm to improve the Kepler algorithm. The model hyperparameters are optimized through the NKOA algorithm.

[0014] (6) Combining the time variable attention layer, BiLSTM network and NKOA algorithm, a NKOA-BiLSTM-TVA model is proposed for AUV drift trajectory prediction, and the prediction results of AUV drift trajectory are compared with those of BiLSTM network, DCNN-BiLSTM network, CNN-BiLSTM-Attention network and TPA-BiLSTM network. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the examples of the present invention and do not constitute a limitation of the present invention.

[0016] Figure 1 . LSTM model structure diagram of the present invention

[0017] Figure 2 .BiLSTM model structure diagram of the present invention

[0018] Figure 3.Structure diagram of the multi-head attention mechanism of the present invention

[0019] Figure 4 .TVA layer structure diagram of the present invention

[0020] Figure 5 .BiLSTM-TVA model structure diagram of the present invention

[0021] Figure 6 .Structure diagram of the NKOA-BiLSTM-TVA model of the present invention

[0022] Figure 7 .Results of BiLSTM-TVA of the NKOA optimization model of the present invention

[0023] Figure 8 Comparison of the predicted trajectory and the actual trajectory of the NKOA-BiLSTM-TVA model in longitude

[0024] Fig. 9 Comparison of the predicted trajectory and the actual trajectory of the NKOA-BiLSTM-TVA model in latitude

[0025] Fig.10 Comparison of the predicted trajectory and the actual trajectory of the NKOA-BiLSTM-TVA model in terms of height

[0026] Fig.11 Comparison of the predicted trajectory and the actual trajectory of the NKOA-BiLSTM-TVA model in three-dimensional space DETAILED DESCRIPTION

[0027] The following is combined with Figure 1-11 , the present invention is further described in detail. The specific implementation steps are as follows:

[0028] Step 01. The drift trajectory data of the AUV is measured by a variety of navigation sensors and positioning systems, and its trajectory information includes longitude, latitude and depth, which are used to describe the exact position of the AUV in the underwater three-dimensional space. In addition, the data also includes the pitch angle (pitch), roll angle (roll) that control the heading and attitude of the AUV, and the marine environmental parameters measured by the scientific sensors carried by the AUV. The factors affecting the trajectory of the AUV in the marine environment include the salinity, temperature, water flow and ocean currents of seawater. The values ​​of environmental factors such as ocean currents can be calculated based on the historical information of marine hydrological data to calculate the situation of ocean current environmental factors when the AUV has an accident. In the experiment of the present invention, the data selected for predicting the drift trajectory of the AUV include longitude, latitude, depth, pitch angle, and roll angle. The selected marine environmental factors include the salinity of seawater and the speed of ocean currents. The longitude, latitude, depth, pitch angle and roll angle in the trajectory are predicted by these data.

[0029] Step 02. Perform data preprocessing on the collected data. The data preprocessing methods include: missing value filling processing and data normalization operations.

[0030] Step 0201. Use the moving average method to fill in missing values ​​by taking the average value of data points within a certain time window.

[0031] Step 0202. Use the maximum and minimum normalization method to normalize the data and scale the data to the range between [0, 1].

[0032]

[0033] Among them, X′ is the normalized data, X is the original trajectory data, and X max is the maximum value of the data, X min is the minimum value of the data.

[0034] Step 03. The Kepler Optimization Algorithm is a recently proposed algorithm for solving optimization problems based on Kepler's laws of planetary motion. The present invention first uses reverse learning to initialize the KOA algorithm and integrates a time-varying nonlinear convergence factor to optimize the hyperparameters of the model in the present invention. The process of NKOA is described as follows:

[0035] (1) Initialization:

[0036] Initialization is required at the beginning of the algorithm. The KOA algorithm randomly initializes N planets in the search space as candidate solutions. Each planet consists of d dimensions and uses them to represent the dimension of the problem to be optimized. The formula for random initialization is as follows:

[0037]

[0038] Among them, i∈[1,N], j∈[1,d], X i, represents the i-th candidate solution, and They respectively represent the upper and lower bounds of the parameter range of the jth problem to be optimized in the ith candidate solution, and r is a random number between 0 and 1.

[0039] (2) Definition of Gravity

[0040] The change of the position and speed of the planet depends on the gravitational force between the sun and the planet. Sun With Planet X i The gravitational force between them can be expressed as:

[0041]

[0042] Among them, e i is the eccentricity, M Sun, m i are the masses of the Sun and the planets, are their normalized values, λ is a small number to prevent the denominator from being zero, μ is the gravitational constant, and r1 is a random number between 0 and 1. is the Euclidean distance R between the sun and the planet at time t i (t) is the normalized value, μ(t) is an exponential function of time t, which is used to limit the search accuracy and is defined as follows:

[0043]

[0044] Among them, γ is a constant, μ0 is the initial value, t and T max They are the current iteration number and the maximum iteration number of the algorithm respectively.

[0045] (3) Planetary velocity update

[0046] The magnitude of a planet's velocity is closely related to its distance from the sun. When a planet is farther from the sun, the velocity between them is relatively small; when the distance between them becomes smaller, the velocity between them becomes larger. Based on this principle, the formula for updating the planet's velocity can be defined as:

[0047]

[0048] in, is the speed of planet i at time t, r3 and r4 are random numbers in [0,1], and is a randomly generated vector between 0 and 1, and F is a parameter factor that can change the direction of planetary motion. F can prevent the algorithm from falling into a local optimal solution. and represents the initial solution randomly selected from the population, a i It is the semi-major axis of the elliptical orbit of object i at time t, as defined by Kepler's third law.

[0049] (4) Get rid of local optimal solution

[0050] In the solar system, most planets rotate counterclockwise around the sun, but a few rotate clockwise. The algorithm uses this phenomenon to set a parameter factor F that changes the search direction, which can prevent the algorithm from falling into a local optimal solution. This also enables the algorithm to better find the optimal solution to the problem in the entire search space, improving the effectiveness and robustness of the algorithm.

[0051] (5) Update the position of the planet

[0052] With the planet's velocity update formula, the planet's position update formula can be expressed as:

[0053]

[0054] (6) Update the distance to the sun

[0055] The algorithm includes two stages: exploration and development. These two stages simulate the process of how the distance between the planet and the sun changes over time as the planet orbits the sun. When the planet is close to the sun, the algorithm is in the development stage, which mainly optimizes the development operator; when the planet is far away from the sun, the algorithm is in the exploration stage, which mainly optimizes the exploration operator. This process depends on a tuning parameter h. If its value is relatively large, the search space is expanded through the exploration operator to find a better solution; if its value is relatively small, the optimal solution is further sought in the area near the optimal solution through the development operator. Its principle can be expressed by the formula:

[0056]

[0057] Among them, h is the adjustment parameter used to control the distance between the planet and the sun, γ is a random number that obeys the normal distribution, and α2 is a cyclic control parameter.

[0058] (7) Location Update Strategy

[0059] In order to ensure that the sun and planets are in the optimal position, the algorithm adopts a position update strategy, which can be defined by the formula:

[0060]

[0061] (8) Improve KOA

[0062] Opposition-Based Learning (OBL) is a population initialization method. For an individual solution, it can generate a reverse solution in the solution space, so that the algorithm considers both the current solution and the reverse solution during the optimization process. If the reverse solution is better than the current solution, then the reverse solution can be selected as a new candidate solution. By considering the initial solution and its reverse solution at the same time, the search process can cover a wider area of ​​the solution space. This diversity helps reduce the risk of the algorithm falling into a local optimal solution. The reverse solution introduced by reverse learning may be located in an unexplored favorable area in the search space, which enhances the global search capability of the algorithm and accelerates the convergence depth of the algorithm.

[0063] For the jth parameter of the ith planet, its reverse solution X can be generated according to the following formula i,j ':

[0064]

[0065] If the reverse solution is better than the current solution, then the reverse solution can be selected as a new candidate solution. In this way, the algorithm can quickly move to more promising search areas.

[0066] The factor F in the KOA algorithm can change the search direction, thereby preventing the algorithm from falling into the local optimal solution. However, its principle is only to change the direction of planetary motion to prevent the algorithm from falling into the local optimal solution. In order to further improve the performance of the algorithm, this paper combines the KOA algorithm with a nonlinear convergence factor that changes with time. Such a fusion enables the algorithm to effectively and more accurately find the global optimal solution in both the exploration and development stages. The formula of the nonlinear convergence factor can be expressed as:

[0067]

[0068] Where e is the Euler constant, ν, λ, is the constant term coefficient, θ is a constant term random number between 0 and 1, t and T max They are the current iteration number and the maximum iteration number of the algorithm respectively.

[0069] From the above formula, we can see that the nonlinear convergence factor a decreases nonlinearly with the increase in the number of iterations. At the beginning of the algorithm, the planet can move at a longer distance, so that the algorithm can better find the global optimal solution in the exploration phase; as the number of iterations increases, the distance of the planet's movement slows down, and the algorithm can more accurately find the optimal solution in the local range in the development phase. The NKOA algorithm helps the algorithm balance the breadth and depth of the search in the exploration and development phases by integrating a nonlinear convergence factor, thereby improving the performance and accuracy of the algorithm. After adding the nonlinear convergence factor to the planet position update formula, the planet position update formula is:

[0070]

[0071] Step 04. Recurrent Neural Networks (RNN) is a neural network structure that can better process sequence data. LSTM is an improved recurrent neural network. The structure of LSTM is as follows: Figure 1 .

[0072] Among them, I t is the data input at time t, f gt It is the forget gate, gt is the input gate, o gt is the output gate, h t is the hidden layer state at the current moment, C t is the output at the current moment, is the temporary state of the cell. Their calculation formula can be expressed as:

[0073] f gt =σ(W fh ·h t-1 +W fI I t +b f )

[0074] i gt =σ(W ih ·h t-1 +W iI I t +b i )

[0075] o gt =σ(W oh ·h t-1 +W oI I t +b o )

[0076] h t =o gt *tanh(C t )

[0077]

[0078] Among them, W is the weight coefficient matrix to be learned by the network, b is the bias parameter; σ(z) is the sigmoid activation function. The BiLSTM model can use the sequence data of historical moments as well as the sequence data information of future moments to more effectively mine the feature relationship between historical and future moment data, which is conducive to improving the prediction accuracy of the model. The structure of BiLSTM is as follows: Figure 2 .

[0079] Step 05. The attention mechanism can give higher weights to important features in the trajectory, thereby improving the prediction effect of the model. The self-attention mechanism has been widely used in computer vision, natural language processing, time series prediction and other fields. The main formula of scaled dot product self-attention is as follows:

[0080]

[0081] Among them, Q, K, and V represent Query, Key, and Value, respectively. They are calculated by linear transformation of the same input matrix X. k is the dimension of Key, and the structure of the multi-head attention mechanism is as follows Figure 3 .

[0082] The multi-head attention mechanism generates different attention matrices by stacking the self-attention mechanism multiple times, and then learns different weight parameters. The multi-head attention mechanism can be described by the following formula:

[0083] MHA(Q,K,V)=Concat(Head1,…,Head n )W o

[0084] Among them, each Head i is a single attention mechanism, which can be expressed as:

[0085] Head i =Attention(XW i Q ,XW i K ,XW i V )

[0086] The TVA designed in the present invention includes two parts: time dimension attention and variable dimension attention. The time dimension attention mainly extracts the dependencies between trajectories at different time steps, while the variable dimension attention is mainly used to extract the dependencies between different trajectory dimensions in the trajectory. The input data is first processed by the multi-head attention mechanism, and then the result is passed to the Normalization layer. The attention layer and the Normalization layer are connected by a residual connection, and the processed result is input into the MLP layer, and the residual connection is used again. The final result is output through the Normalization layer. The structure of the TVA layer is as follows: Figure 4 .

[0087] Time dimension attention: In the time dimension attention stage, the input of the TVA layer is a given two-dimensional tensor O C ∈R S×D , S is the input time step, D is the number of input trajectory variable dimensions, O C i,: represents all variable dimensions at the i-th time step, O C :,d represents all time steps in the dth variable dimension, O C is the output of the BiLSTM hidden layer. The process of extracting trajectory feature dependencies in the time dimension through time dimension attention can be defined as:

[0088] O T =LayerNorm(MHA T (O C :,d ,O C :,d ,O C :,d )+O C :,d )

[0089] O T =LayerNorm(MLP(O T )+O T )

[0090] Where d∈[1,D], MHA T () is to extract feature dependencies in the time dimension through a multi-head attention mechanism. T is the output of the attention in the time dimension. Here, all variable dimensions share the MHA parameters, and the Normalization layer is used to normalize each feature. MLP is a forward fully connected layer.

[0091] Variable dimension attention: Variable dimension attention takes the output of time dimension attention and the original features as input. Variable dimension attention extracts feature dependencies for all time dimensions in a sliding window in each variable dimension; through variable dimension attention, the dependencies between different variable dimensions in the trajectory can be fully extracted. Through variable dimension attention, more weight is given to trajectory variables that have a more important impact on the future AUV trajectory. At the same time, the ocean environment can be fully utilized to predict the future trajectory of the AUV, thereby improving the accuracy of trajectory prediction.

[0092]

[0093] O V =LayerNorm(MLP(O V )+O V )

[0094] Among them, MHA V () is to extract feature dependencies on the variable dimension through the multi-head attention mechanism, and all time dimensions share MHA. V It is the output after time dimension and variable dimension dependency extraction.

[0095] The process of extracting TVA layer feature dependency can be expressed as:

[0096] Y=O V =TSA(O C )

[0097] Among them, O C ,Y represent the input and output of the TVA layer respectively.

[0098] Step 06. The BiLSTM-TVA model designed by the present invention includes an input layer, a CNN layer, a BiLSTM layer, a TVA layer and an output layer. First, the AUV trajectory data is preprocessed in the input layer, and then a data set is constructed according to the sliding window principle. It is then input into the CNN layer, and the CNN layer extracts the spatial features in the AUV trajectory. After the CNN network extracts the spatial features between the trajectories, the BiLSTM network is used to extract the sequence features between the trajectories. The output of the BiLSTM hidden layer is used as the input of the TVA layer, and the TVA layer considers both the time dimension and the variable dimension. Then, the hyperparameters of the BiLSTM-TVA model are optimized by the NKOA algorithm, and the optimized hyperparameters are used for model training and learning. The mean square error MSE between the predicted trajectory and the actual trajectory is calculated, and the MSE value is fed back to the NKOA algorithm as the fitness value for algorithm iteration. Finally, a set of optimal hyperparameters for the model is obtained. The structure of the BiLSTM-TVA model is as follows: Figure 5 , the process of NKOA algorithm optimizing the hyperparameters of BiLSTM-TVA model is as follows Figure 6 .

[0099] Step 07. The data set used in the present invention has 10,000 records, and the time interval for collecting each data record is 1 minute. The first 80% of the data set is used as the training set, and the last 20% is used as the test set. The normalized training set and test set are trained and predicted by the window sliding method based on timestamps in the model training. The model first determines the optimal hyperparameters through the NKOA algorithm, and then trains the model on the training set using the determined hyperparameters, and then directly tests the model effect on the test set using the trained model. In the experiment, the sliding window size for model training and test data selection is 8, and the model only predicts five AUV trajectory features. In order to numerically describe the accuracy of AUV trajectory prediction and the prediction effect of the model, two error evaluation indicators are set in the experiment, namely the root mean square error (RMSE) and the mean absolute error (MAE). Different error evaluation indicators use different calculation methods to reflect the differences, and the calculation methods are as follows:

[0100]

[0101] Where N is the sample size of trajectory data, y n is the nth true trajectory value, is the trajectory result predicted by the model. For these two evaluation indicators, when their values ​​are close to 0, the smaller the prediction error is, the more accurate the model is, that is, the higher the model prediction structure accuracy is, the smaller its value is.

[0102] Step 08. The experiment uses the NKOA algorithm to optimize the model's hyperparameters, including learning rate, batch size, number of epochs, number of neurons in the first and second layers of the BiLSTM hidden layer, and number of fully connected layers. Set the number of iterations of the NKOA algorithm to 10. Figure 7 The optimization process of the learning rate, batch size, number of epochs, and number of neurons in the first and second layers of BiLSTM hidden layers and fully connected layers is given. After 6 iterations, the fitness value of the NKOA algorithm finally stabilized at 0.00088. Table 1 gives the specific configuration and optimization results of the network model hyperparameters. Each hyperparameter eventually converges to the optimal solution. The learning rate parameter range is [0.001, 0.01], and finally converges to 0.0023; the batch size range is [8, 128], and finally converges to 16; the epoch parameter range is [10, 200], and finally converges to 82; the number of hidden layers in the first layer of the BiLSTM network ranges from [1, 30], and finally converges to 8; the number of hidden layers in the second hidden layer ranges from [1, 30], and finally converges to 7; the number of neurons in the fully connected layer ranges from [1, 100], and finally converges to 65.

[0103] Table 1 Results of optimizing model hyperparameters using the NKOA algorithm

[0104]

[0105] After obtaining the hyperparameter results of the NKOA algorithm optimization model, the model is first trained on the training set using the optimal hyperparameters, and the trained model is saved. Then, the trained model is used to perform a model trajectory prediction test on the test set. The present invention compares the experimental results of the NKOA-BiLSTM-TVA model with those of the BiLSTM, DCNN-LSTM, CNN-BiLSTM-Attention and TPA-BiLSTM models on the test set. The comparison results verify the prediction performance of the model in AUV drift trajectory prediction. The results of the model NKOA-BiLSTM-TVA in the present invention predicting the longitude of the AUV trajectory on the test set are shown in the figure. Figure 8 ; The results of predicting the latitude of the AUV trajectory on the test set are as follows Fig. 9 ; The results of predicting the AUV trajectory depth on the test set are as follows Fig.10 ;

[0106] Table 2 shows the MAE and RMSE results of the trajectory prediction of the five models. It can be seen from Table 2 that the method NKOA-BiLSTM-TVA in the present invention has the best prediction effect on the AUV drift trajectory. Due to the small change trend in the longitude and latitude of the AUV drift trajectory, the prediction results of the TPA-BiLSTM model and the NKOA-BiLSTM-TVA model in longitude and latitude are similar, but the prediction effect of NKOA-BiLSTM-TVA on the other three variables is better.

[0107] Table 2 Comparison of RMSE and MAE values ​​of each model on the AUV trajectory dataset.

[0108]

[0109] The experimental results prove that the NKOA-BiLSTM-TVA model in the present invention has good network parameter optimization ability and high prediction accuracy in the AUV drift trajectory prediction task.

Claims

1. A method for predicting AUV drift trajectories based on a nonlinear Kepler optimization algorithm to optimize a bidirectional long short-term memory-time variable attention model, characterized in that: Obtain the AUV's trajectory data and the ocean current environment data of the AUV's activity area; Preprocess the acquired AUV trajectory data and construct a dataset for AUV trajectory prediction in combination with ocean current environmental factors; Based on the constructed data set, the NKOA-BiLSTM-TVA model is first constructed, and then the drift trajectory of the AUV is predicted by the NKOA-BiLSTM-TVA model; Based on the predicted AUV trajectory result, the error calculation is performed with the actual AUV trajectory to verify the prediction effect of the present invention on the AUV trajectory.

2. The AUV drift trajectory prediction method based on the NKOA-BiLSTM-TVA model as claimed in claim 1, characterized in that: The data preprocessing process includes missing value filling processing and data normalization.

3. The AUV drift trajectory prediction method based on the NKOA-BiLSTM-TVA model as claimed in claim 2, characterized in that: The missing value filling method includes: using a moving average method to fill the missing values ​​by taking the average value of data points within a certain time window. The data normalization method includes: in order to accelerate the convergence speed of the model and improve the accuracy of model prediction, a maximum and minimum value normalization method is used to normalize the data, and the data is scaled to a range between [0, 1].

4. The AUV drift trajectory prediction method based on the NKOA-BiLSTM-TVA model as claimed in claim 1, characterized in that: The normalized data is processed through the sliding window method, and the AUV trajectory at future moments is predicted by utilizing the AUV historical trajectory information.

5. The AUV drift trajectory prediction method based on the NKOA-BiLSTM-TVA model as claimed in claim 1, characterized in that: The spatial features of the trajectory will be processed by introducing CNN, and the temporal information between trajectories will be further learned through the BiLSTM network. This setting comprehensively considers the impact of marine environmental factors on AUV trajectory prediction.

6. The AUV drift trajectory prediction method based on the NKOA-BiLSTM-TVA model as claimed in claim 1, characterized in that: Through a fusion attention mechanism layer across time dimension and variable dimension - the time dimension attention layer, the dependencies between the time dimension and variable dimension features in the historical trajectory are extracted respectively. Combined with the BiLSTM network to predict the trajectory of the AUV, the model can focus on the time points and spatial positions that have an important impact on trajectory prediction, thereby improving the prediction precision and accuracy.

7. The AUV drift trajectory prediction method based on the NKOA-BiLSTM-TVA model as claimed in claim 1, characterized in that: The KOA algorithm is initialized based on reverse learning, and a nonlinear convergence factor is added to the algorithm to improve the Kepler algorithm. The model hyperparameters are optimized by the NKOA algorithm, which improves the generalization ability and prediction accuracy of the model.

8. The AUV drift trajectory prediction method based on the NKOA-BiLSTM-TVA model as claimed in claim 1, characterized in that: The NKOA-BiLSTM-TVA model for AUV drift trajectory prediction is proposed by combining the time variable attention layer, BiLSTM network and NKOA algorithm, which can effectively predict the trajectory of AUV crashed in the ocean.

9. The AUV drift trajectory prediction method based on the NKOA-BiLSTM-TVA model as claimed in claim 1, characterized in that: The AUV drift trajectory is predicted using the BiLSTM network, DCNN-BiLSTM network, CNN-BiLSTM-Attention network, TPA-BiLSTM network, and NKOA-BiLSTM-TVA network. Based on the prediction result data, the error between the predicted AUV trajectory and the actual trajectory is calculated. The beneficial effects of the present invention are: (1) When predicting the AUV trajectory, the present invention fully combines the AUV trajectory information and ocean factors and takes the ocean environment as the input of the model. Before applying the BiLSTM network, the CNN network is used to extract the spatial features between the trajectories. (2) The present invention proposes a fusion attention layer across time dimension and variable dimension, namely, a time-variable attention layer. This attention layer can be used to extract the dependency between time dimension features and variable dimension features in historical trajectories. In addition, it can also be applied to other model structures. (3) The present invention first initializes the KOA algorithm based on reverse learning and adds a nonlinear convergence factor to the algorithm to improve the Kepler algorithm. The model hyperparameters are optimized by the NKOA algorithm, thereby improving the generalization ability and prediction accuracy of the model. (4) The present invention proposes an AUV trajectory prediction NKOA-BiLSTM-TVA model that combines a time-variable attention layer, a BiLSTM network, and an NKOA algorithm. This model provides a new idea and method for the research and application of underwater vehicle trajectory prediction.

Images