A neural network-based three-dimensional active absorption method and device for wave generator

Abstract

The application provides a neural network-based three-dimensional active absorption method and device for a wave generator. By simplifying a three-dimensional active absorption transfer function, a three-dimensional active absorption problem is divided into a two-dimensional active absorption problem and a direction angle problem; two neural network models are further built to solve the two problems. Finally, the two models are integrated together, and the three-dimensional active absorption problem under the condition that the direction angle is unknown can be realized.

Description

Technical Field

[0001] This invention relates to a three-dimensional active absorption method and equipment for wave generators based on neural networks, belonging to the field of marine engineering. Background Technology

[0002] Wave generators are crucial equipment in marine physics simulation experiments for generating target waves. In recent years, to improve the utilization efficiency of wave pools, researchers have successively developed novel wave-generating systems with "L," "U," and "O" shaped layouts. However, during operation, these wave generators often exhibit significant secondary reflections from the wave-generating plates, severely impacting the stability of the wave field. Therefore, achieving integrated wave generation and absorption functions has become a core issue in the design of novel wave-generating devices. Currently, active absorption control technology is still under development and not yet fully mature. Especially in complex three-dimensional wavefield environments, the ability to identify the direction angle of the waves to be absorbed in active absorption systems is significantly insufficient, becoming a key technical bottleneck restricting the improvement of wave generator absorption performance.

[0003] Furthermore, with the rapid development of neural networks in recent years, their applications in various industries have become increasingly widespread. They possess the ability to automatically learn complex mapping relationships from high-dimensional data without requiring explicit physical models. This makes them an ideal tool for solving the active absorption problem in wave generators.

[0004] Therefore, this invention proposes a three-dimensional active absorption method and device for wave generators based on neural networks. By simplifying the three-dimensional active absorption transfer function, the three-dimensional active absorption problem is decomposed into solving two-dimensional active absorption and direction angle problems; furthermore, two neural network models are built to solve the above two problems respectively. Finally, the two models are integrated together to solve the three-dimensional active absorption problem when the direction angle is unknown. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a three-dimensional active absorption method and device for wave generators based on neural networks, used to achieve three-dimensional active absorption under directional angular position conditions in laboratory wave generation.

[0006] The technical solution of the present invention: A three-dimensional active absorption method for wave generators based on neural networks includes the following steps: Figure 1 As shown: S1. Derive the three-dimensional active absorption transfer function; S2. By simplifying the three-dimensional active absorption transfer function, the problem is decomposed into solving the two-dimensional active absorption problem and the orientation angle problem; S3. Analyze the errors caused by the simplification method; S4. Solve the two-dimensional active absorption problem by building a multilayer perceptron neural network model; S5. Predict wave direction angle by building a BO-CNN-BiLSTM model; S6. Integrate the multilayer perceptual neural network model trained in S4 and S5 with the BO-CNN-BiLSTM model to solve the three-dimensional active absorption problem when the orientation angle is unknown.

[0007] Furthermore, in S1, Figure 2 A three-dimensional active absorption schematic diagram of a wave generator is presented. First, a wave generator boundary is defined, consisting of multiple wave generators. The opposite side and both sides of the wave generator boundary are straight walls. The center of the wave generator boundary is the origin. O The direction perpendicular to the boundary of the wave generator and pointing to the opposite side is y In the positive direction of the axis, y Rotate 90 degrees clockwise in the positive direction of the axis. x Positive axis direction For time; the wave generator produces a traveling wave. Propagated into the pool, it generates a reflected wave upon encountering the boundary of the wave-generating plate. When a primary reflected wave encounters a wave-generating plate, it will undergo a secondary reflection, producing a secondary reflected wave. The secondary reflected wave will continue to propagate into the pool and affect the target wave field. To eliminate the secondary reflected wave, the wave generator needs to produce a compensating wave. The compensation wave and the secondary reflected wave must have the same amplitude but a 180° phase difference. Therefore, when the wave generator is in active absorption mode, the motion of the wave-generating plate consists of two parts: the motion that generates the traveling wave. and the motion that generates compensation waves At this time, the wave front surface obtained by the wave height sensor in front of the wavemaker is... The specific expression for the superposition of various wave components at the wave height sensor in front of the wave generator is as follows: (1) In the above formula, For traveling waves The corresponding non-propagation mode, To compensate for the wave The corresponding non-propagation mode.

[0008] Performing a Fourier transform on both sides of the above equation and taking its complex amplitude, we obtain: (2) in, (3) Where j is the imaginary unit; and These represent the motions that generate traveling waves. and the motion that generates compensation waves Complex amplitude value after Fourier transform This represents the number of non-propagational modes. and The hydrodynamic transfer function for the three-dimensional pusher is expressed as follows: (4) in, and Here is the two-dimensional pusher plate hydrodynamic transfer function; For real wavenumber, It is an imaginary wave number. Secondary reflected wave and y Axial direction angle.

[0009] Substituting formula (3) into formula (2), we get: (5) in, The three-dimensional active absorption transfer function is expressed as follows: (6) Two-dimensional active absorption transfer function The expression is: (7) Compared to , The expression includes wave angle information. However, wave angle information cannot be predicted in advance and needs to be detected in real-time within the wave field. Therefore, to handle the three-dimensional active absorption problem, it is necessary to... The angle term in the data is separated.

[0010] Furthermore, in S2, The expression is processed as follows: (8) Observing the above formula, we find that if we let .but The expression simplifies to: (9) further The expression simplifies to: (10) Let the above formula be The output is represented by x(t); For input, use u ( t () indicates input. Given a known quantity, and The wave height is obtained in real time by a wave meter in front of the wavemaker. Therefore, as long as the relationship between the input and output is established and the wave direction angle is detected in real time, the three-dimensional active absorption problem can be solved.

[0011] Furthermore, in S3, the error is due to the forced order in step S2. This is caused by the error introduced by the simplification method in the analysis, which is the error in the analysis of the assumption. The resulting impact. Figure 3 Given Q n Error analysis graph. The graph shows that when the frequency... f At frequencies <1Hz, the amplitude and phase errors caused by this assumption are less than 3.00%. In laboratory wave generation, the frequency... f A frequency of <1Hz can meet the requirements of most experimental conditions. This indicates that this simplified method is very suitable for laboratory three-dimensional active absorption.

[0012] Furthermore, in S4, regarding how to establish the relationship between the output and input: the traditional method is to construct a filter so that the filter's amplitude and phase response approximate the theoretical transfer function as closely as possible. This is then transformed into the time domain and implemented through the difference equation corresponding to the filter. However, to fit a filter such that its amplitude and phase response closely match the theoretical transfer function... Matching is a relatively complex problem, and traditional filter methods are not very efficient in terms of absorption.

[0013] Inspired by FIR filters, active absorption of wave generator displacement Input can be made from the current step. The historical step input is obtained through simple linear operations.

[0014] Furthermore, when there is only incident wave and no reflected wave in the water tank, the output... and input The sequences can all be calculated from theoretical values. Therefore, this study uses a multilayer perceptron neural network model, employing theoretical inputs and outputs as training data to train the weights and biases of each layer. Specifically, in a single training session, the correspondence between input and output is as follows: (11) In the above formula, This indicates the input for the current step; This indicates the input from the previous step; Indicated m Step input; Indicates the output of the current step; Indicates the time interval, i.e., the step size; The constructed multilayer perceptron neural network model consists of three parts: an input layer, a hidden layer, and an output layer, such as... Figure 4 As shown.

[0015] Furthermore, in S4, the output of the previous layer serves as the input of the current layer, and the process of obtaining the output from the input through the current layer's computation is represented as follows: (12) In the above formula, Indicates the input of the current layer; Indicates the output of the current layer; Indicates the weight of the current layer; Indicates the bias of the current layer; Furthermore, in S4, the multilayer perceptron neural network model selects the ReLU function as the nonlinear activation function; a regularization term is introduced as the loss function based on the mean squared error, and the Adam optimizer is used to update the parameters of the multilayer perceptron neural network model. During training, the training data is input into the multilayer perceptron neural network model for forward propagation calculation, and the corresponding loss function is calculated; then, the parameter gradient is calculated using the backpropagation algorithm, and the optimizer is used to update the parameters.

[0016] Furthermore, in S4, the parameters affecting the training effect of the multilayer perceptron model include the number of hidden layers, the number of neurons in each hidden layer, and the order. m And learning rate, order m That is, calculating the output of the current step. The number of historical step inputs required. Calculate the regression values ​​for the test set under different parameters. This is used to measure the predictive performance of a multilayer perceptron neural network model. Regression value This represents the correlation between the predicted output and the theoretical output. The closer the value is to 1, the smaller the error between the predicted output and the theoretical output. The closer the value is to 0, the greater the error between the predicted output and the theoretical output. The calculation formula is: (13) In the above formula, Indicates the first test set i One theoretical output; Indicates the first test set i One predicted output; This represents the mean of the theoretical output sequence in the test set.

[0017] Furthermore, in S4, the trained multilayer perceptron model is used to predict the data in the test set, and the corresponding two-dimensional active absorption wave generator displacement is predicted.

[0018] Furthermore, a deep learning wave direction angle prediction method using a Bayesian optimization-convolutional-bidirectional long short-term memory network, referred to as the BO-CNN-BiLSTM model, is employed to predict the wave direction angle. The CNN-BiLSTM model learns the relationship between wave height meter data and the corresponding angle. Simultaneously, Bayesian optimization (BO) is introduced to adaptively adjust the key hyperparameters of the CNN-BiLSTM model, achieving more efficient and accurate wave direction angle prediction with limited wave height meter data.

[0019] Furthermore, in S5, the wave height meter data is preprocessed in batches. Specifically, for each angle, the wave height meter has... w There are 10 sampling points. The sliding window size is defined as 10 ... T Step size is h ( h < T Then this w Each sampling point was divided into u The first sample consists of data from the first sampling point to the second sampling point. T The set consists of 1 sampling points; the second sample data is: the 1st h + 1 sampling point to the T + h+ A set consisting of 1 sampling point; the first i The sample data is: the ( ) i -1) h +1 sampling point to the T +( i -1) h The set consists of +1 sampling points. There are a total of 91 angle values ​​(starting from 0° and ending at 90°, with a step size of 1°), that is, a total of 91. u 91 samples. Each training iteration uses these 91 samples. u Randomly selected from the samples B Each sample is used as input data for a batch.

[0020] Furthermore, in S5, the input data of a batch is normalized to obtain pre-normalized input data. The specific expression is: (14) In the above formula, This represents a batch of input data; express The mean; express The variance.

[0021] Furthermore, in S5, Gaussian noise (mean 0, variance 0.01 times the maximum of the initially normalized input data) is added to the initially normalized input data to obtain the final normalized input data. The aim is to improve the stability of the model in complex environments.

[0022] Furthermore, in S5, the CNN model in the BO-CNN-BiLSTM model consists of convolutional layers, average pooling layers, and fully connected layers. The convolutional layers extract local features through convolution, specifically expressed as: (15) In the above formula, P This represents the input to the convolutional layer, i.e., the input data after final normalization. Matrix form; W This indicates the weights of the convolutional layer; b This indicates the bias of the convolutional layer; f Indicates the activation function; M Indicates input P The output after this convolutional layer. The activation function is the ReLU function, and its specific expression is: (16) After activation function processing, dimensionality reduction is performed using an average pooling layer to finally obtain the output of the fully connected layer. Z .

[0023] Furthermore, in S5, the BiLSTM model in the BO-CNN-BiLSTM model is composed of a forward LSTM model and a backward LSTM model.

[0024] For a forward LSTM model, it consists of a memory unit and three gating units, including a forget gate, an input gate, and an output gate. Each gating unit outputs a value between 0 and 1 through a sigmoid activation function, representing the degree to which information is allowed to pass through, with 0 indicating complete blocking and 1 indicating complete permission. Forget gate: Filters stored content and determines whether the stored data should be retained or discarded. The specific expression is as follows: (17) In the above formula, This represents the output of the fully connected layer of the CNN model at the current time step, and also serves as the input of the LSTM model at the current time step. This represents the output of the forget gate. It is the activation function, specifically the sigmoid function whose output range is 0-1, used to indicate the degree to which the door is open; This is the current time step input. The weight matrix between the forget gate and the forget gate It is a hidden state of historical time steps. The weight matrix between the forget gate and the forget gate It is the bias matrix of the forget gate; Input gate: Updates the cell state, determining whether the cell can remember new information. The specific expression is: (18) In the above formula, Indicates the output of the input gate. This is the current time step input. The weight matrix between the input gate and the input gate It is a hidden state of historical time steps. The weight matrix between the input gate and the input gate It is the bias matrix of the input gate; The expression for the candidate memory unit is: (19) In the above formula, It is a candidate memory unit. This is the current time step input. The weight matrix between candidate memory units, It is a hidden state of historical time steps. The weight matrix between candidate memory units, It is the bias matrix of candidate memory units; Current time step memory unit The update is performed using the output of the forget gate, the output of the input gate, and candidate memory units. The specific expression is as follows: (20) In the above formula, ⊙ represents the memory unit of the previous time step, and ⊙ represents the multiplication of corresponding elements; The output gate controls the output ratio of the memory unit, thereby obtaining the final hidden state output; the specific expression of the output gate is: (twenty one) In the above formula, This indicates the output of the output gate. This is the current time step input. The weight matrix between the output gate and the output gate It is a hidden state of historical time steps. The weight matrix between the output gate and the output gate It is the bias matrix of the output gate; Current time step hidden state : (twenty two) The hidden states at all time steps combined together constitute the output of the LSTM model after the input has passed through it. Equations (17)-(22) are the expressions for the forward LSTM model. Similarly, the expressions for the backward LSTM model are: (twenty three) (twenty four) (25) (26) (27) (28) In the above formula, , , , , , These represent the output of the forget gate, the output of the input gate, the candidate memory unit, the memory unit, the output of the output gate, and the hidden state of the backward LSTM model, respectively. , , , These represent the inputs in the backward LSTM model. The weights between the forget gate, input gate, candidate memory units, and output gate. , , , Indicates the corresponding bias; , , , These represent the hidden states at historical time steps in the feedforward LSTM model. The weights between the forget gate, input gate, candidate memory units, and output gate; Finally, the output vectors of the forward LSTM model and the backward LSTM model are concatenated to obtain the output after passing through the BiLSTM model: (29).

[0025] Furthermore, in S5, the Bayesian optimization algorithm is used to optimize the hyperparameters of the CNN-BiLSTM model. The computational flowchart of the BO-CNN-BiLSTM model is shown below. Figure 9 As shown. The hyperparameters to be optimized are: the number of convolutional kernels in the second convolutional layer of the CNN model, the kernel length in all convolutional layers of the CNN model, the number of neurons in a single LSTM model, the initial learning rate, and the dropout rate.

[0026] The specific tuning steps are as follows: S5.1, within the preset hyperparameter space, randomly select hyperparameter combinations to train the CNN-BiLSTM model; update and iterate the CNN-BiLSTM model parameters through the Adam optimizer and backpropagation algorithm to finally obtain the corresponding prediction target values, which constitute the observation dataset; S5.2, based on the observed dataset, a surrogate model is used to fit the objective function. A Gaussian model is chosen as the surrogate model to generate a probability distribution for predicting the objective function value. The prediction formula for the Gaussian model is: (30) (31) In the above formula, The combination of hyperparameters to be predicted; Hyperparameter combination The corresponding predicted value of the objective function; Hyperparameter combination The corresponding covariance; For the current hyperparameter combination The vector formed by the covariances of the various hyperparameter combinations in the observed dataset. The matrix formed by the covariances between the various hyperparameter combinations in the observation dataset; Hyperparameter combination Its own covariance; This is the variance term for noise. n The number of samples in the observation dataset, It is the identity matrix. p This is a vector composed of the target values ​​corresponding to each hyperparameter combination in the observation dataset. To represent the transpose of a matrix, Represents the inverse of a matrix; S5.3 utilizes the Gaussian regression process to evaluate the sampled hyperparameter points by improving the acquisition function through expectation, and then selects the optimal sample points as the next hyperparameter combination. The desired improvement to the acquisition function expression is: (32) In the above formula, To improve the acquisition function, It is the predicted value of the objective function, given by the Gaussian process; The current known optimal value, E For expectations; S5.4, Update the surrogate model, repeating S4.1-S4.3 within a preset number of iterations until the lowest test set loss value is found. RMSE The corresponding hyperparameter combination is used as the final hyperparameter combination and output.

[0027] Furthermore, in S5, the mean absolute mean square error is used. RMSE Mean absolute error MAE Correlation coefficient R 2 To evaluate the predictive performance of the model, the specific expression is: (33) (34) (35) In the above formula, Indicates sample u The corresponding theoretical value; This represents the average of the predicted values ​​for all samples in a batch. Indicates sample u The corresponding predicted value. MAE This reflects the model's average prediction error; RMSE It is more sensitive to some extreme errors, and the combination of the two types of errors better reflects the predictive performance of the model; R 2 It reflects the degree of model fit; the closer its value is to 1, the better the model's predictive performance.

[0028] Furthermore, in S6, the models trained in S4 and S5 are integrated together. The multilayer perceptron neural network model is responsible for solving the two-dimensional active absorption problem, while the BO-CNN-BiLSTM model provides it with orientation angle information, thereby realizing three-dimensional active absorption when the orientation angle is unknown.

[0029] A three-dimensional active absorption device for wave generators based on neural networks includes: at least one processor and a memory communicatively connected to the at least one processor; The memory stores instructions that can be executed by the processor, which are then executed by the at least one processor to cause the at least one processor to perform the aforementioned three-dimensional active absorption method for a wave generator based on a neural network.

[0030] Compared with the prior art, the beneficial effects of the present invention are: The three-dimensional active absorption method for wave generators based on neural networks provided by this invention has the following advantages compared with traditional methods: (1) A method to simplify the three-dimensional active absorption transfer function is proposed, which transforms the three-dimensional active absorption problem into a two-dimensional active absorption and orientation angle detection problem.

[0031] (2) For the two-dimensional active absorption problem, the traditional time-domain / frequency-domain solution is abandoned and a multilayer perceptron active absorption model is proposed. This model bypasses the complex filter design process and directly learns the relationship between the input sequence and the output sequence, which greatly simplifies the active absorption problem and improves the absorption efficiency.

[0032] (3) For the problem of direction angle detection, this invention proposes a BO-CNN-BiLSTM neural network model for predicting wave direction angle. This model adopts a deep learning architecture and directly learns the feature mapping relationship of complex wave direction from the time-domain sequence data collected by the wave height meter. Compared with traditional methods, it significantly improves the accuracy and computational efficiency of wave direction angle prediction, and provides richer means for hydrodynamic analysis.

[0033] (4) By combining the multilayer perceptron active absorption model with the BO-CNN-BiLSTM neural network model, a deep learning-based three-dimensional active absorption problem for wave generators is proposed. This model can be used to achieve three-dimensional active absorption under unknown orientation angles. Applying deep learning methods to the active absorption of wave generators opens up a new path for the active absorption problem. Attached Figure Description

[0034] Figure 1 This is a flowchart of the method.

[0035] Figure 2 This is a schematic diagram of the three-dimensional active absorption of the wave generator.

[0036] Figure 3 for Q n Error analysis diagram, where (a) is amplitude comparison, (b) is phase comparison, (c) is amplitude error, and (d) is phase error.

[0037] Figure 4 This is a schematic diagram of a multilayer perceptron neural network model.

[0038] Figure 5 These are the regression values ​​for the test set under different learning rates.

[0039] Figure 6 A comparison chart of theoretical and predicted displacements for the test set ( T p =1.0s).

[0040] Figure 7 A comparison chart of theoretical and predicted displacements for the test set ( T p =1.5s).

[0041] Figure 8 A comparison chart of theoretical and predicted displacements for the test set (T p =2.0s).

[0042] Figure 9 The flowchart shows the computation process of the BO-CNN-BiLSTM model.

[0043] Figure 10 The graph shows a comparison of the prediction performance of different models, where (a) is... θ =20°, (b) is θ =40°, (c) is θ =60°, (d) is θ =80°. Detailed Implementation

[0044] A three-dimensional active absorption method for wave generators based on neural networks includes the following steps: S1. Derive the three-dimensional active absorption transfer function; S2. By simplifying the three-dimensional active absorption transfer function, the problem is decomposed into solving the two-dimensional active absorption and orientation angle problem; S3. Analyze the errors caused by the simplification method; S4. Solve the two-dimensional active absorption problem by building a multilayer perceptron neural network model; S5. Predict wave direction angle by building a BO-CNN-BiLSTM model; S6. Integrate the models trained in S4 and S5 together to solve the three-dimensional active absorption problem when the orientation angle is unknown.

[0045] Furthermore, in S1, Figure 2 A three-dimensional active absorption diagram of a wave generator is presented. First, the bottom of the diagram represents the wave generator boundary, composed of multiple wave generator plates. The top and left / right sides are straight walls. The origin is located at the center of the wave generator plate boundary. O Vertically upwards is y Positive axis direction, right side is x Positive axis direction For time; the wave generator produces a traveling wave. The wave propagates into the pool and is reflected upon reaching the boundary. When a primary reflected wave encounters a wave-generating plate, it will undergo a secondary reflection, producing a secondary reflected wave. The secondary reflected wave will continue to propagate into the pool and affect the target wave field. To eliminate the secondary reflected wave, the wave generator needs to produce a compensating wave. The compensation wave and the secondary reflected wave must have the same amplitude but a 180° phase difference. Therefore, when the wave generator is in active absorption mode, the motion of the wave-generating plate consists of two parts: the motion that generates the traveling wave. and the motion that generates compensation waves At this time, the wave front surface obtained by the wave height sensor in front of the wavemaker is... The specific expression for the superposition of various wave components at the wave height sensor in front of the wave generator is as follows: (1) In the above formula, For traveling waves The corresponding non-propagation mode, To compensate for the wave The corresponding non-propagation mode.

[0046] Performing a Fourier transform on both sides of the above equation and taking its complex amplitude, we obtain: (2) in, (3) Where j is the imaginary unit; and These represent the motions that generate traveling waves. and the motion that generates compensation waves Complex amplitude value after Fourier transform This represents the number of non-propagational modes. and The hydrodynamic transfer function for the three-dimensional pusher is expressed as follows: (4) in, and Here is the two-dimensional pusher plate hydrodynamic transfer function; For real wavenumber, It is an imaginary wave number. Secondary reflected wave and y Axial direction angle.

[0047] Substituting formula (3) into formula (2), we get: (5) in, The three-dimensional active absorption transfer function is expressed as follows: (6) Two-dimensional active absorption transfer function The expression is: (7) Compared to , The expression includes wave angle information. However, wave angle information cannot be predicted in advance and needs to be detected in real-time within the wave field. Therefore, to handle the three-dimensional active absorption problem, it is necessary to... The angle term in the data is separated.

[0048] Furthermore, in S2, The expression is processed as follows: (8) Observing the above formula, we find that if we let .but The expression simplifies to: (9) further The expression simplifies to: (10) Let the above formula be The output is represented by x(t); For input, use u ( t () indicates input. Given a known quantity, and The wave height is obtained in real time by a wave meter in front of the wavemaker. Therefore, as long as the relationship between the input and output is established and the wave direction angle is detected in real time, the three-dimensional active absorption problem can be solved.

[0049] Furthermore, in S3, the error is due to the forced order in step S2. This is caused by the error introduced by the simplification method in the analysis, which is the error in the analysis of the assumption. The resulting impact.

[0050] Figure 3 The given water depth is 0.5m, and the direction angle is... Let the angles be 10°, 20°, and 30° respectively. and complete type (incorrect) When doing any processing The effect of amplitude and phase.

[0051] As can be seen from the figure, when the frequency f < 1Hz, and integrity The influence of amplitude and phase is minimal, with an error of less than 3.00%. However, the error gradually increases as the frequency exceeds 1 Hz. Changing the angle has no effect on the phase. Since the frequencies commonly used in laboratory wave generation are generally less than 1 Hz, therefore... The impact on the outcome is negligible.

[0052] Furthermore, in S4, regarding how to establish the relationship between the output and input: the traditional method is to construct a filter so that the filter's amplitude and phase response approximate the theoretical transfer function as closely as possible. This is then transformed into the time domain and implemented through the difference equation corresponding to the filter. However, to fit a filter such that its amplitude and phase response closely match the theoretical transfer function... Matching is a relatively complex problem, and traditional filter methods are not very efficient in terms of absorption.

[0053] Inspired by FIR filters, active absorption of wave generator displacement Input can be made from the current step. The historical step input is obtained through simple linear operations.

[0054] Furthermore, when there is only incident wave and no reflected wave in the water tank, the output... and input The sequences can all be calculated from theoretical values. Therefore, this study uses a multilayer perceptron neural network model, employing theoretical inputs and outputs as training data to train the weights and biases of each layer. Specifically, in a single training session, the correspondence between input and output is as follows: (11) In the above formula, This indicates the input for the current step; This indicates the input from the previous step; Indicated m Step input; Indicates the output of the current step; Indicates the time interval, i.e., the step size; The constructed multilayer perceptron neural network model consists of three parts: an input layer, a hidden layer, and an output layer, such as... Figure 4 As shown: Furthermore, in S4, the output of the previous layer serves as the input of the current layer, and the process of obtaining the output from the input through the current layer's computation is represented as follows: (12) In the above formula, Indicates the input of the current layer; Indicates the output of the current layer; Indicates the weight of the current layer; Indicates the bias of the current layer; Furthermore, in S4, the multilayer perceptron neural network model selects the ReLU function as the nonlinear activation function; a regularization term is introduced as the loss function based on the mean squared error, and the Adam optimizer is used to update the parameters of the multilayer perceptron neural network model. During training, the training data is input into the multilayer perceptron neural network model for forward propagation calculation, and the corresponding loss function is calculated; then, the parameter gradient is calculated using the backpropagation algorithm, and the optimizer is used to update the parameters.

[0055] Furthermore, in S4, the parameters affecting the training effect of the multilayer perceptron model include the number of hidden layers, the number of neurons in each hidden layer, and the order. m And learning rate, order m That is, calculating the output of the current step. The number of historical step inputs required. Calculate the regression values ​​for the test set under different parameters. This is used to measure the predictive performance of a multilayer perceptron neural network model. Regression value This represents the correlation between the predicted output and the theoretical output. The closer the value is to 1, the smaller the error between the predicted output and the theoretical output. The closer the value is to 0, the greater the error between the predicted output and the theoretical output. The calculation formula is: (13) In the above formula, Indicates the first test set i One theoretical output; Indicates the first test set i One predicted output; This represents the mean of the theoretical output sequence in the test set.

[0056] Furthermore, in S4, the trained multilayer perceptron model is used to predict the data in the test set, and the corresponding two-dimensional active absorption wave generator displacement is predicted.

[0057] First, calculate the period of the generated spectral peak. T p =1.0s, significant wave height H s =0.04m, water depth d The amplitude, angular frequency, and phase of each component wave of the JONSWAP irregular wave at a constant velocity of 0.5m are then calculated. The corresponding theoretical input and output sequences are then calculated separately. (Time interval) =0.004s, total time t =200s, the data from 0-140s is the test set, and the data from 140-200s is the test set.

[0058] Table 1 shows the results for different numbers and orders of hidden layers. m The following is obtained through calculation using a multilayer perceptron neural network. R 2 The value of . As can be seen from the table, for the number of hidden layers, as the number of hidden layers increases, R 2 Getting closer and closer to 1. R calculated with four and five hidden layers. 2 Basically the same. Regarding order... m , m Neither too small nor too large is effective. m R calculated when =500 2 Closest to 1.

[0059] Therefore, this invention selects 4 hidden layers in the neural network, with 256 neurons in each layer. 128 64 32; order m Take 500.

[0060] Table 1: Test set with different hidden layers and orders R 2 Distribution table

[0061] also, Figure 5 The regression values ​​for the test set under different learning rates are also provided. As can be seen from the figure, the regression value is closest to 1 when the learning rate is 0.001. Therefore, this model selects 0.001 as the final learning rate.

[0062] The trained multilayer perceptron active absorption model is used to predict the data in the test set, and the corresponding active absorption wave generator displacement is predicted.

[0063] To further illustrate the advantages of this method, Figure 6 , Figure 7 , Figure 8 The spectral peak periods are given respectively. T p The graph compares the theoretical and predicted displacements of the test wave-generating plate at 1.0s, 1.5s, and 2.0s. It shows that the predicted and theoretical displacements are essentially the same at different spectral peak periods, indicating that the proposed method is very effective.

[0064] Furthermore, a deep learning wave direction angle prediction method using a Bayesian optimization-convolutional-bidirectional long short-term memory network, referred to as the BO-CNN-BiLSTM model, is employed to predict the wave direction angle. The CNN-BiLSTM model learns the relationship between wave height meter data and the corresponding angle. Simultaneously, Bayesian optimization (BO) is introduced to adaptively adjust the key hyperparameters of the CNN-BiLSTM model, achieving more efficient and accurate wave direction angle prediction with limited wave height meter data.

[0065] Furthermore, in S5, the wave height meters are arranged parallel to the direction of the 0° regular wave crest line, and the spacing between two adjacent wave height meters is [missing information]. v =0.5m, total C =11 wave height meters, with a sampling interval of 11 for each wave height meter. .

[0066] Furthermore, in S5, the wave height meter data is preprocessed in batches. Specifically, for each angle, the wave height meter has... w =50,000 sampling points. Define the sliding window size as... T =200, step size is h =100. Then this w =50,000 sampling points were divided into u =499 samples. 91 angle values, that is, a total of 45409 (91) 499 samples. During each training session, samples are randomly selected from these 45409 samples. B =32 samples are used as input data in one batch.

[0067] Furthermore, in S5, the input data of a batch is normalized to obtain pre-normalized input data. The specific expression is: (14) In the above formula, This represents a batch of input data; express The mean; express The variance.

[0068] Furthermore, in S5, Gaussian noise (mean 0, variance 0.01 times the maximum of the initially normalized input data) is added to the initially normalized input data to obtain the final normalized input data. The aim is to improve the stability of the model in complex environments.

[0069] Furthermore, in S5, the CNN model in the BO-CNN-BiLSTM model consists of convolutional layers, average pooling layers, and fully connected layers. The convolutional layers extract local features through convolution, specifically expressed as: (15) In the above formula, P This represents the input to the convolutional layer, i.e., the input data after final normalization. Matrix form; W This indicates the weights of the convolutional layer; b This indicates the bias of the convolutional layer; f Indicates the activation function; M Indicates input P The output after this convolutional layer. The activation function is the ReLU function, and its specific expression is: (16) After activation function processing, dimensionality reduction is performed using an average pooling layer to finally obtain the output of the fully connected layer. Z .

[0070] Furthermore, in S5, the BiLSTM model in the BO-CNN-BiLSTM model is composed of a forward LSTM model and a backward LSTM model.

[0071] For a forward LSTM model, it consists of a memory unit and three gating units, including a forget gate, an input gate, and an output gate. Each gating unit outputs a value between 0 and 1 through a sigmoid activation function, representing the degree to which information is allowed to pass through, with 0 indicating complete blocking and 1 indicating complete permission. Forget gate: Filters stored content and determines whether the stored data should be retained or discarded. The specific expression is as follows: (17) In the above formula, This represents the output of the fully connected layer of the CNN model at the current time step, and also serves as the input of the LSTM model at the current time step. This represents the output of the forget gate. It is the activation function, specifically the sigmoid function whose output range is 0-1, used to indicate the degree to which the door is open; This is the current time step input. The weight matrix between the forget gate and the forget gate It is a hidden state of historical time steps. The weight matrix between the forget gate and the forget gate It is the bias matrix of the forget gate; Input gate: Updates the cell state, determining whether the cell can remember new information. The specific expression is: (18) In the above formula, Indicates the output of the input gate. This is the current time step input. The weight matrix between the input gate and the input gate It is a hidden state of historical time steps. The weight matrix between the input gate and the input gate It is the bias matrix of the input gate; The expression for the candidate memory unit is: (19) In the above formula, It is a candidate memory unit. This is the current time step input. The weight matrix between candidate memory units, It is a hidden state of historical time steps. The weight matrix between candidate memory units, It is the bias matrix of candidate memory units; Current time step memory unit The update is performed using the output of the forget gate, the output of the input gate, and candidate memory units. The specific expression is as follows: (20) In the above formula, ⊙ represents the memory unit of the previous time step, and ⊙ represents the multiplication of corresponding elements; The output gate controls the output ratio of the memory unit, thereby obtaining the final hidden state output; the specific expression of the output gate is: (twenty one) In the above formula, This indicates the output of the output gate. This is the current time step input. The weight matrix between the output gate and the output gate It is a hidden state of historical time steps. The weight matrix between the output gate and the output gate It is the bias matrix of the output gate; Current time step hidden state : (twenty two) The hidden states at all time steps combined together constitute the output of the LSTM model after the input has passed through it. Equations (17)-(22) are the expressions for the forward LSTM model. Similarly, the expressions for the backward LSTM model are: (twenty three) (twenty four) (25) (26) (27) (28) In the above formula, , , , , , These represent the output of the forget gate, the output of the input gate, the candidate memory unit, the memory unit, the output of the output gate, and the hidden state of the backward LSTM model, respectively. , , , These represent the inputs in the backward LSTM model. The weights between the forget gate, input gate, candidate memory units, and output gate. , , , Indicates the corresponding bias; , , , These represent the hidden states at historical time steps in the feedforward LSTM model. The weights between the forget gate, input gate, candidate memory units, and output gate; Finally, the output vectors of the forward LSTM model and the backward LSTM model are concatenated to obtain the output after passing through the BiLSTM model: (29).

[0072] Furthermore, in S5, the Bayesian optimization algorithm is used to optimize the hyperparameters of the CNN-BiLSTM model. The computational flowchart of the BO-CNN-BiLSTM model is shown below. Figure 9 As shown. The hyperparameters to be optimized are: the number of convolutional kernels in the second convolutional layer of the CNN model, the kernel length in all convolutional layers of the CNN model, the number of neurons in a single LSTM model, the initial learning rate, and the dropout rate.

[0073] The specific optimization steps are as follows: S5.1 Within the preset hyperparameter space, a combination of hyperparameters is randomly selected to train the CNN-BiLSTM model; the parameters of the CNN-BiLSTM model are updated and iterated through the Adam optimizer and backpropagation algorithm to finally obtain the corresponding prediction target value, which constitutes the observation dataset. S5.2, based on the observed dataset, a surrogate model is used to fit the objective function. A Gaussian model is chosen as the surrogate model to generate a probability distribution for predicting the objective function value. The prediction formula for the Gaussian model is: (30) (31) In the above formula, The combination of hyperparameters to be predicted; Hyperparameter combination The corresponding predicted value of the objective function; Hyperparameter combination The corresponding covariance; For the current hyperparameter combination The vector formed by the covariances of the various hyperparameter combinations in the observed dataset. The matrix formed by the covariances between the various hyperparameter combinations in the observation dataset; Hyperparameter combination Its own covariance; This is the variance term for noise. n The number of samples in the observation dataset, It is the identity matrix. p This is a vector composed of the target values ​​corresponding to each hyperparameter combination in the observation dataset. To represent the transpose of a matrix, Represents the inverse of a matrix; S5.3 utilizes the Gaussian regression process to evaluate the sampled hyperparameter points by improving the acquisition function through expectation, and then selects the optimal sample points as the next hyperparameter combination. The desired improvement to the acquisition function expression is: (32) In the above formula, To improve the acquisition function, It is the predicted value of the objective function, given by the Gaussian process; The current known optimal value, E For expectations; S5.4, Update the surrogate model, repeating S4.1-S4.3 within a preset number of iterations until the lowest test set loss value is found. RMSE The corresponding hyperparameter combination is used as the final hyperparameter combination and output.

[0074] Furthermore, in S5, the hyperparameter combination after Bayesian optimization is as follows: The first CNN convolutional layer has 0.5 kernels. 62; The second CNN convolutional kernel has 1 kernel. 62; The number of kernels in the third convolutional layer is 2. 62; (Mainly optimizes the number of kernels in the second CNN layer, the number of kernels in the first layer is 0.5 times that of the second layer, and the number of kernels in the third layer is 2 times that of the second layer); CNN convolution kernel length is 5; the number of neurons in a single LSTM is 256; the learning rate is 1e-3; the dropout rate is 0.2.

[0075] Furthermore, in S5, the mean absolute mean square error is used. RMSE Mean absolute error MAE Correlation coefficient R 2 To evaluate the predictive performance of the model, the specific expression is: (33) (34) (35) In the above formula, Indicates sample u The corresponding theoretical value; This represents the average of the predicted values ​​for all samples in a batch. Indicates sample u The corresponding predicted value. MAE This reflects the model's average prediction error; RMSE It is more sensitive to some extreme errors, and the combination of the two types of errors better reflects the predictive performance of the model; R 2 It reflects the degree of model fit; the closer its value is to 1, the better the model's predictive performance.

[0076] Furthermore, in S5, to better illustrate the advantages of the BO-CNN-BiLSTM model, the proposed BO-CNN-BiLSTM model, along with traditional CNN, CNN-LSTM, and CNN-BiLSTM models, were used to predict angle values. The parameters of all four trained models were stored. Then, irregular wave height data at four different angles (20°, 40°, 60°, and 80°) were generated, and the four trained models were used to predict the angle values. Figure 10A comparison of prediction performance is presented. The graph shows that when the angle is small, the prediction performance of the four models is not significantly different. As the angle increases, the differences in prediction performance between the different models begin to emerge. The CNN-BiLSTM model outperforms the CNN-LSTM model. This is because LSTM models can only process sequences sequentially (from front to back), while the BiLSTM model consists of two independent LSTM models, one for forward processing and one for backward processing, thus improving the overall performance of the model. Therefore, we chose to optimize the CNN-BiLSTM model using a Bayesian algorithm, resulting in our final model: BO-CNN-BiLSTM. The graph shows that this model performs well at smaller angles. θ When the angle is less than 60°, the predicted angle and the theoretical angle are basically the same. (Regarding the angle...) θ Even at an angle of 80°, the maximum prediction error does not exceed 3°. This further demonstrates the superior predictive performance of the model.

[0077] Furthermore, in S6, the models trained in S4 and S5 are integrated together. The multilayer perceptron neural network model is responsible for solving the two-dimensional active absorption problem, while the BO-CNN-BiLSTM model provides it with orientation angle information, thereby realizing the three-dimensional active absorption problem under the orientation angle position.

Claims

1. A three-dimensional active absorption method for wave generators based on neural networks, characterized in that, Includes the following steps: S1. Derive the three-dimensional active absorption transfer function; S2. By simplifying the three-dimensional active absorption transfer function, the problem is decomposed into solving the two-dimensional active absorption problem and the orientation angle problem; S3. Analyze the errors caused by the simplification method; S4. Solve the two-dimensional active absorption problem by building a multilayer perceptron neural network model; S5. Predict wave direction angle by building a BO-CNN-BiLSTM model; S6. Integrate the multilayer perceptual neural network model trained in S4 and S5 with the BO-CNN-BiLSTM model to solve the three-dimensional active absorption problem when the orientation angle is unknown.

2. The three-dimensional active absorption method for wave generators based on neural networks according to claim 1, characterized in that, In S1, firstly, a wave-generating plate boundary is defined, consisting of multiple wave-generating plates. The opposite side and both sides of the wave-generating plate boundary are straight walls; the center of the wave-generating plate boundary is the origin. O The direction perpendicular to the boundary of the wave generator and pointing to the opposite side is y In the positive direction of the axis, y Rotate 90 degrees clockwise in the positive direction of the axis. x Positive axis direction For time; the wave generator produces a traveling wave. Propagated into the pool, it generates a reflected wave upon encountering the boundary of the wave-generating plate. When a primary reflected wave encounters a wave-generating plate, it will undergo a secondary reflection, producing a secondary reflected wave. The secondary reflected wave will continue to propagate into the pool and affect the target wave field; in order to eliminate the secondary reflected wave, the wave generator needs to produce a compensating wave. The compensation wave and the secondary reflected wave must have the same amplitude but a 180° phase difference; therefore, when the wave generator is in active absorption mode, the motion of the wave-generating plate is divided into two parts: the motion that generates the traveling wave. and the motion that generates compensation waves At this time, the wave front surface obtained by the wave height sensor in front of the wavemaker is... The specific expression for the superposition of various wave components at the wave height sensor in front of the wave generator is as follows: (1) In the above formula, For traveling waves The corresponding non-propagation mode, To compensate for the wave The corresponding non-propagation mode; Performing a Fourier transform on both sides of the above equation and taking its complex amplitude, we obtain: (2) in, (3) Where j is the imaginary unit; and These represent the motions that generate traveling waves. and the motion that generates compensation waves Complex amplitude value after Fourier transform Indicates the number of non-propagating modes; and The hydrodynamic transfer function for the three-dimensional pusher is expressed as follows: (4) in, and Here is the two-dimensional pusher plate hydrodynamic transfer function; For real wavenumber, It is an imaginary wave number. Secondary reflected wave and y Axial direction angle; Substituting formula (3) into formula (2), we get: (5) in, The three-dimensional active absorption transfer function is expressed as follows: (6) Two-dimensional active absorption transfer function The expression is: (7) Compared to , The expression includes wave angle information; however, wave angle information cannot be predicted in advance and needs to be detected in real time within the wave field; therefore, to handle the three-dimensional active absorption problem, it is necessary to... The angle term in the data is separated.

3. The three-dimensional active absorption method for a wave generator based on a neural network according to claim 2, characterized in that, In S2, The expression is processed as follows: (8) Observing the above formula, we find that if we let ;but The expression simplifies to: (9) further The expression simplifies to: (10) Let the above formula be The output is represented by x(t); For input, use u ( t () indicates; input in Given a known quantity, and The wave height is obtained in real time by the wave meter in front of the wavemaker; therefore, as long as the relationship between the input and output is established and the wave direction angle is detected in real time, the problem of three-dimensional active absorption can be solved.

4. The three-dimensional active absorption method for a wave generator based on a neural network according to claim 3, characterized in that, In S3, the error is due to the forced command in step S2. The error caused by the simplification method in the analysis is the impact of the analysis hypothesis on the results. The resulting impact.

5. The three-dimensional active absorption method for a wave generator based on a neural network according to claim 4, characterized in that, In S4, a multilayer perceptron neural network model is used, employing theoretical inputs and outputs as training data to train the weights and biases of each layer. In a single training iteration, the correspondence between input and output is as follows: (11) In the above formula, This indicates the input for the current step; This indicates the input from the previous step; Indicated m Step input; Indicates the output of the current step; Indicates the time interval, i.e., the step size; The constructed multilayer perceptron neural network model consists of three parts: an input layer, a hidden layer, and an output layer. In S4, the output of the previous layer is used as the input of the current layer. The process of the input being processed by the current layer to obtain the output is represented as follows: (12) In the above formula, Indicates the input of the current layer; Indicates the output of the current layer; Indicates the weight of the current layer; Indicates the bias of the current layer; In S4, the nonlinear activation function chosen for the multilayer perceptron neural network model is the ReLU function; a regularization term is introduced as the loss function based on the mean squared error, and the Adam optimizer is used to update the parameters of the multilayer perceptron neural network model; during training, the training data is input into the multilayer perceptron neural network model for forward propagation calculation, and the corresponding loss function is calculated; then the backpropagation algorithm is used to calculate the parameter gradient, and the optimizer is used to update the parameters. In S4, the parameters that affect the training performance of the multilayer perceptron model include the number of hidden layers, the number of neurons in each hidden layer, and the order. m And learning rate, order m That is, calculating the output of the current step. The number of historical steps required; calculating the regression values ​​of the test set under different parameters. To measure the predictive performance of a multilayer perceptron neural network model; regression value This represents the correlation between the predicted output and the theoretical output. The closer the value is to 1, the smaller the error between the predicted output and the theoretical output. The closer the value is to 0, the greater the error between the predicted output and the theoretical output. The calculation formula is: (13) In the above formula, Indicates the first test set i One theoretical output; Indicates the first test set i One predicted output; This represents the mean of the theoretical output sequence in the test set; In S4, the trained multilayer perceptron model is used to predict the data in the test set, and the corresponding two-dimensional active absorption wave-generating plate displacement is predicted. A deep learning wave direction angle prediction method using Bayesian optimization-convolution-bidirectional long short-term memory network, abbreviated as BO-CNN-BiLSTM model, is used to predict wave direction angle.

6. The three-dimensional active absorption method for a wave generator based on a neural network according to claim 5, characterized in that, In S5, the wave height meter data is preprocessed in batches. Specifically, for each angle, the wave height meter has… w One sampling point; Define the sliding window size as T Step size is h , h < T Then this w Each sampling point was divided into u The first sample consists of data from the first sampling point to the second sampling point. T The set consists of 1 sampling points; the second sample data is: the 1st h+ 1 sampling point to the T + h + A set consisting of 1 sampling point; the first i The sample data is: the ( ) i -1) h +1 sampling point to the T +( i -1) h A set consisting of +1 sampling points; a total of 91 angle values, that is, a total of 91 u 91 samples; each training iteration uses these 91 samples. u Randomly selected from the samples B Each sample is used as input data for a batch; In S5, the input data of a batch is normalized to obtain the pre-normalized input data. The specific expression is: (14) In the above formula, This represents a batch of input data; express The mean; express The variance; In step S5, Gaussian noise is added to the initially normalized input data to obtain the final normalized input data. This aims to improve the stability of the model in complex environments; In S5, the CNN model in the BO-CNN-BiLSTM model consists of convolutional layers, average pooling layers, and fully connected layers. The convolutional layers extract local features through convolution, specifically as follows: (15) In the above formula, P This represents the input to the convolutional layer, i.e., the input data after final normalization. Matrix form; W This indicates the weights of the convolutional layer; b This indicates the bias of the convolutional layer; f Indicates the activation function; M Indicates input P The output after passing through this convolutional layer; The activation function is the ReLU function, and its specific expression is: (16) After activation function processing, dimensionality reduction is performed using an average pooling layer to finally obtain the output of the fully connected layer. Z ; In S5, the BiLSTM model in the BO-CNN-BiLSTM model is a combination of a forward LSTM model and a backward LSTM model; For a forward LSTM model, it consists of a memory unit and three gating units, including a forget gate, an input gate, and an output gate. Each gating unit outputs a value between 0 and 1 through a sigmoid activation function, representing the degree to which information is allowed to pass through, with 0 indicating complete blocking and 1 indicating complete permission. Forget gate: Filters stored content and determines whether the stored data should be retained or discarded. The specific expression is as follows: (17) In the above formula, This represents the output of the fully connected layer of the CNN model at the current time step, and also serves as the input of the LSTM model at the current time step. This represents the output of the forget gate. It is the activation function, specifically the sigmoid function whose output range is 0-1, used to indicate the degree to which the door is open; This is the current time step input. The weight matrix between the forget gate and the forget gate It is a hidden state of historical time steps. The weight matrix between the forget gate and the forget gate It is the bias matrix of the forget gate; Input gate: Updates the cell state, determining whether the cell can remember new information. The specific expression is: (18) In the above formula, Indicates the output of the input gate. This is the current time step input. The weight matrix between the input gate and the input gate It is a hidden state of historical time steps. The weight matrix between the input gate and the input gate It is the bias matrix of the input gate; The expression for the candidate memory unit is: (19) In the above formula, It is a candidate memory unit. This is the current time step input. The weight matrix between candidate memory units, It is a hidden state of historical time steps. The weight matrix between candidate memory units, It is the bias matrix of candidate memory units; Current time step memory unit The update is performed using the output of the forget gate, the output of the input gate, and candidate memory units. The specific expression is as follows: (20) In the above formula, ⊙ represents the memory unit of the previous time step, and ⊙ represents the multiplication of corresponding elements; The output gate controls the output ratio of the memory unit, thereby obtaining the final hidden state output; the specific expression of the output gate is: (21) In the above formula, This indicates the output of the output gate. This is the current time step input. The weight matrix between the output gate and the output gate It is a hidden state of historical time steps. The weight matrix between the output gate and the output gate It is the bias matrix of the output gate; Current time step hidden state : (22) The hidden states at all time steps combined together constitute the output of the LSTM model after the input has passed through it. Equations (17)-(22) are the expressions for the forward LSTM model. Similarly, the expressions for the backward LSTM model are: (23) (24) (25) (26) (27) (28) In the above formula, , , , , , These represent the output of the forget gate, the output of the input gate, the candidate memory unit, the memory unit, the output of the output gate, and the hidden state of the backward LSTM model, respectively. , , , These represent the inputs in the backward LSTM model. The weights between the forget gate, input gate, candidate memory units, and output gate. , , , Indicates the corresponding bias; , , , These represent the hidden states at historical time steps in the feedforward LSTM model. The weights between the forget gate, input gate, candidate memory units, and output gate; Finally, the output vectors of the forward LSTM model and the backward LSTM model are concatenated to obtain the output after passing through the BiLSTM model: (29)； In S5, the Bayesian optimization algorithm is used to optimize the hyperparameters of the CNN-BiLSTM model. The optimized hyperparameters are: the number of convolutional kernels in the second convolutional layer of the CNN model, the kernel length in all convolutional layers of the CNN model, the number of neurons in a single LSTM model, the initial learning rate, and the dropout rate. S5.1 Within the preset hyperparameter space, a combination of hyperparameters is randomly selected to train the CNN-BiLSTM model; the parameters of the CNN-BiLSTM model are updated and iterated through the Adam optimizer and backpropagation algorithm to finally obtain the corresponding prediction target value, which constitutes the observation dataset. S5.2, based on the observed dataset, a surrogate model is used to fit the objective function. A Gaussian model is chosen as the surrogate model to generate a probability distribution for predicting the objective function value. The prediction formula for the Gaussian model is: (30) (31) In the above formula, The combination of hyperparameters to be predicted; Hyperparameter combination The corresponding predicted value of the objective function; Hyperparameter combination The corresponding covariance; For the current hyperparameter combination The vector formed by the covariances of the various hyperparameter combinations in the observed dataset. The matrix formed by the covariances between the various hyperparameter combinations in the observation dataset; Hyperparameter combination Its own covariance; This is the variance term for noise. n The number of samples in the observation dataset, It is the identity matrix. p This is a vector composed of the target values ​​corresponding to each hyperparameter combination in the observation dataset. To represent the transpose of a matrix, Represents the inverse of a matrix; S5.3 utilizes the Gaussian regression process to evaluate the sampled hyperparameter points by improving the acquisition function through expectation, and then selects the optimal sample points as the next hyperparameter combination. The desired improvement to the acquisition function expression is: (32) In the above formula, To improve the acquisition function, It is the predicted value of the objective function, given by the Gaussian process; The current known optimal value, E For expectations; S5.4, Update the surrogate model, repeating S4.1-S4.3 within a preset number of iterations until the lowest test set loss value is found. RMSE The corresponding hyperparameter combination is used as the final hyperparameter combination and output. Using mean absolute mean square error RMSE Mean absolute error MAE Correlation coefficient R 2 To evaluate the predictive performance of the model, the specific expression is: (33) (34) (35) In the above formula, Indicates sample u The corresponding theoretical value; This represents the average of the predicted values ​​for all samples in a batch. Indicates sample u The corresponding predicted value.

7. A three-dimensional active absorption method for a wave generator based on a neural network according to claim 6, characterized in that, In S6, the models trained in S4 and S5 are integrated together. The multilayer perceptron neural network model is responsible for solving the two-dimensional active absorption problem, while the BO-CNN-BiLSTM model provides it with orientation angle information, thereby realizing three-dimensional active absorption when the orientation angle is unknown.

8. A three-dimensional active absorption device for wave generators based on neural networks, characterized in that, include: At least one processor, and a memory communicatively connected to said at least one processor; The memory stores instructions executable by the processor, which are executed by the at least one processor to cause the at least one processor to perform a three-dimensional active absorption method for a wave generator based on a neural network as described in any one of claims 1-7.

Images