Deep-sea aquaculture net cage model design method and device based on machine learning

Abstract

The invention relates to a deep-sea aquaculture net cage model design method based on machine learning, and the method comprises the steps: firstly providing an agent model based on multi-neural network integration: constructing an integrated agent model comprising MLP, PINN and RBFNN, and can efficiently solve the problems of multi-input, multi-output, high-dimensional nonlinearity and multi-target optimization in deep-sea aquaculture net cage design; and secondly, providing a dynamic weight optimization mechanism: introducing a PPO algorithm, converting the distribution problem of the sub-model integration weight into a sequence decision problem in reinforcement learning, and through interaction of an intelligent agent and a simulation environment, realizing dynamic optimization and scene self-adaption of weight distribution, and significantly improving prediction precision and generalization ability of the integration model. And finally, providing an adaptive parallel training and hyper-parameter optimization framework: adopting a multi-model parallel training framework, and combining Bayesian optimization and a LOOCV strategy to realize synchronous training and hyper-parameter automatic optimization of a plurality of neural networks, so that the model development efficiency and the intelligent level are greatly improved, and the dependence on artificial experience is reduced.

Description

Technical Field

[0001] This invention relates to the technical field of deep-sea aquaculture cage model design, and particularly to a method and apparatus for designing deep-sea aquaculture cage models based on machine learning. Background Technology

[0002] When designing deep-sea aquaculture cages, ensuring the safety of the cage structure requires special attention to design optimization to prevent structural damage and maintain good performance under complex loads such as waves and currents. However, with the increasing complexity and intelligence of functions, deep-sea aquaculture cages have evolved into complex large-scale systems involving multiple disciplines. They have numerous design variables, complex operating conditions, and diverse performance targets, representing a typical high-dimensional nonlinear problem with multiple inputs and outputs. This leads to challenges in the design process, including high difficulty, long cycles, and high costs. In practical applications, the multiple stringent performance requirements and the extremely complex factors to consider in structural design make simulation analysis during the design optimization process time-consuming and costly.

[0003] Existing research on net cage design optimization commonly employs simulation analysis, physical model experiments, and numerical simulations; however, these methods fail to significantly address the aforementioned problems. Against this backdrop, surrogate model-assisted design optimization has become a widely adopted technical approach in industry. Current methods can be broadly categorized into two types: single surrogate models and ensemble surrogate models. Single surrogate models, such as the Kriging model, focus on representing relationships between few variables and single performance responses, making them difficult to adapt to the design environment of aquaculture net cages. Furthermore, in the absence of prior knowledge, the characteristics of the selected surrogate model may not match the features of the actual problem well, potentially leading to high prediction variance and insufficient generalization ability. Deep neural network surrogate models, due to their powerful ability to represent multi-input, multi-output nonlinear features and good transfer adaptability, can effectively compensate for these shortcomings. However, current research on deep neural network surrogate models often focuses on single models, which may not effectively improve the generalization ability of deep surrogate models. Ensemble surrogate models, by combining the advantages of different surrogate models, effectively reduce prediction variance and improve the generalization ability of the final model. However, the selection of model subsets and the method of integrating sub-models remain key factors affecting the computational efficiency and generalization ability of integrated proxy models.

[0004] While existing ensemble methods each have their advantages, they also have significant limitations. Weight-optimized surrogate model ensemble strategies are currently the most widely used approach. Global weight ensemble stands out for its computational simplicity and ease of implementation, making it suitable for scenarios where the global performance of the base model is stable. However, because it ignores local accuracy differences in the design space, it often struggles to guarantee sufficient prediction accuracy in complex nonlinear problems. In contrast, pointwise weight ensemble, by dynamically adjusting the weights of different design points, can fully adapt to local characteristics and significantly improve local prediction accuracy. However, it requires calculating the weights individually for each design point, which greatly increases computational complexity and cost in high-dimensional problems, limiting its practicality. Stacked ensemble methods construct an ensemble framework through multi-level learning mechanisms, enabling it to capture complex relationships between base models and thus enhance generalization ability. However, this method requires a large amount of data, and the design of secondary models and the implementation of cross-validation strategies increase the difficulty of implementation, making it prone to overfitting when data is insufficient. Surrogate ensemble methods based on component model optimization can effectively reduce redundancy and computational costs by filtering out poorly performing models, avoiding the negative impact of poor models on the ensemble results. However, the selection criteria of this method are highly dependent on error metrics, which may mistakenly remove models with excellent local performance but high global error, leading to a loss of base model diversity. In addition, existing methods have significant application limitations: on the one hand, insufficient attention is paid to the ensemble of neural network surrogate models, failing to fully utilize the potential of neural networks in handling nonlinear relationships; on the other hand, when facing high-dimensional nonlinear problems with multiple inputs and outputs, existing strategies often struggle to efficiently construct ensemble surrogate models with strong generalization capabilities.

[0005] Therefore, improving the multi-objective optimization capability and prediction accuracy of integrated agent models for aquaculture cage systems is an urgent problem to be solved. Summary of the Invention

[0006] The purpose of this invention is to at least address one of the shortcomings of the prior art by providing a machine learning-based method for designing deep-sea aquaculture cage models.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: Specifically, a machine learning-based method for designing deep-sea aquaculture cage models is proposed, including the following: Step 110: Determine the fatigue hazard points of the deep-sea aquaculture cage and construct the finite element model of the deep-sea aquaculture cage; Step 120: Based on the actual working conditions and simulation analysis, determine the objective function, design variables, value range, and optimization constraints; Step 130: Sample the design variables and constraints of the aquaculture cage by using Latin hypercube to obtain samples, substitute the samples into the finite element model to solve, obtain the corresponding objective function values, and divide the training set and test set. Then, train the surrogate model based on the divided training set. The surrogate model includes three sub-models: multilayer perceptron (MLP), physical information neural network (PINN), and radial basis function neural network (RBFNN). Step 140: Learn the optimal ensemble weights of the three sub-models by optimizing the PPO algorithm through the near-end strategy to minimize the ensemble prediction error; finally, construct an ensemble surrogate model based on the optimal ensemble weights, use the ensemble surrogate model to predict the test set of the finite element model to obtain the predicted value, and evaluate the predicted value to obtain the evaluation result. Step 150: Provide guidance on the design of deep-sea aquaculture cage models based on the evaluation results.

[0008] Furthermore, specifically, in step 2, the design variables of the deep-sea aquaculture cage model are expressed as follows: ,in, The diameter of the wire mesh cage, For the height of the cage, For the cone angle of the mesh, For the frame pipe wall thickness, The outer diameter of the frame pipe, For the density of the frame material, The areal density of the mesh material. Due to the resistance of ocean currents on the net cages, The density of the seawater where the net cage is located. The stress exerted on the cage by the ocean current. The stress exerted on the fish cage by sea waves, The stress exerted on the fish cage by the sea wind; Different value ranges are pre-defined for the design variables based on material properties and engineering practice; The objective function is set as follows: Output 1: Minimize the peak stress at the fatigue hazard point of the wire mesh cage. : Calculation method: ; Output 2: Minimize the material cost of the wire mesh cage. : Calculation method: ; in, This is the unit price of the frame material. This refers to the unit price of the mesh material. Output 3: Maximize the effective aquaculture volume of the net cage. : Calculation method: :; in, This is the space utilization coefficient; Output 4: Minimize the current resistance coefficient of the net cage. : Calculation method: ; The multi-objective function is expressed as .

[0009] Furthermore, specifically, step 130 includes the following sub-steps: Step 131: Use Latin hypercube sampling to generate a sample size of [number missing]. The initial sample set is expressed as follows: ; in, Number of variables; Step 132: Substitute each valid sample from the initial sample set into the finite element model of the deep-sea aquaculture cage, and conduct simulation calculations under multiple working conditions to obtain the objective function value corresponding to each sample; Step 133: Execute the initial sample set In each iteration of the second iteration, one sample is randomly selected as the test set, and the rest... One sample is used as the training set; Step 134: Based on the objective function value, construct the Multilayer Perceptron (MLP), Physical Information Neural Network (PINN), and Radial Basis Function Neural Network (RBFNN) respectively; Step 135: On different GPUs and computing nodes, LOOCV is used to train MLP, PINN, and RBFNN in parallel. The training of each sub-model is completed by optimizing its weights and bias parameters and minimizing the loss function through the backpropagation algorithm. The hyperparameters of learning rate, batch size, number of hidden layers, number of neurons, center point RBFNN, L2 regularization coefficient, physical loss function weight PINN, optimizer, and activation function are optimized using the Bayesian optimization algorithm. Finally, the root mean square error (RMSE) is used as the metric to measure the prediction accuracy of the model.

[0010] Furthermore, specifically, the formula for calculating RMSE is as follows: ; In the formula, Indicates the first The actual observed value or true value of a sample Indicates the first The predicted value for each sample, This represents the number of samples.

[0011] Furthermore, specifically, the process of learning the optimal ensemble weights of the three sub-models through the proximal strategy optimization PPO algorithm includes, If the output formula of the integrated prediction is: ; The error with the true value is minimized, that is ; in, These are the prediction results of the three trained sub-models MLP, PINN, and RBFNN, respectively; X represents the input features. This is the i-th ensemble prediction; This is the i-th true value; Based on the above objectives, the learning process for the optimal ensemble weights is as follows: Step 141: Initialize the agent's actor network to output the probability distribution of actions. , Representing state Take action below The conditional probability, Network parameters; initialize the critic network to output state value estimates. , Representing state The expected cumulative reward, Establish network parameters; create an experience pool Used to store trajectory data generated by the interaction between the agent and the environment; configures the learning rate of the actor network. Critical network learning rate Maximum number of training rounds Discount Factor and GAE parameters ; Step 142: Load the trained MLP, PINN, and RBFNN models into memory, fix their parameters, and integrate them as functions into the PPO environment; Step 143: The agent interacts with the environment and collects information about the current state. Information, namely the historical errors of the design parameters (design variables) of each cage and each sub-model, based on the current state. Utilizing old policy networks Acquire action That is, the weight allocation vector After obtaining the caching strategy, the agent executes the generated actions and receives immediate rewards based on the actions performed. The agent then transitions the environment to the next state, performs an action, and observes the reward. and obtain the state at the next moment. The interactive data is stored in the experience pool to form a trajectory sequence. The parameters for the old strategy network, i.e., the actor network; Step 144: When the trajectory reaches the cutoff length At that time, using the critic network The termination state is evaluated to obtain an estimated reward, and then the advantage function is calculated based on the trajectory sequence. The calculation formula is: ; in, The timing difference error is calculated using the following formula: ; It is a state-value function; Step 145: Extract complete trajectory data from the experience pool to train the critic network and actor network; The critic network is updated during training by minimizing the prediction error using the loss function, as shown in the following formula: ; in, For the critic network to evaluate the current state Value prediction The formula for calculating cumulative returns based on discounts is as follows: ; Update the critic network parameters The method is as follows: ; The actor network is updated during training by maximizing the pruning objective function, as shown in the following formula: ; in, Importance sampling ratio, This is a hyperparameter for the cropping range; ; Update actor network parameters The method is as follows: ; Continue iterating through the above steps until the strategy is optimal or the maximum number of training rounds is reached. ; Step 146: Based on the optimal weights obtained after training We will integrate proxy models and predict the objective function. Step 147: Evaluate the prediction results of Step 4.6 using the root mean square error (RMSE) and compare the obtained prediction results with the prediction results of a single surrogate model. The smaller the RMSE value, the better the prediction effect.

[0012] Furthermore, the method also includes adding two MLP hidden layer structures before the output layer of the RBFNN model to obtain a DRBFNN model. The activation function of the first hidden layer neuron uses the base function of RBFNN, while the activation functions of the other layers are consistent with MLP. The DRBFNN model is used as a sub-model to replace the RBFNN model for subsequent operations.

[0013] This invention also proposes a machine learning-based deep-sea aquaculture cage model design device, comprising the following: The finite element model construction module is used to determine the fatigue hazard points of the deep-sea aquaculture cage and construct the finite element model of the deep-sea aquaculture cage. The objective function determination module is used to determine the objective function, design variables, value range, and optimization constraints based on actual working conditions and simulation analysis. The surrogate model construction module is used to sample the design variables and constraints of the aquaculture cage through Latin hypercube to obtain samples, substitute the samples into the finite element model to solve, obtain the corresponding objective function values, and divide the training set and test set. Then, the surrogate model is trained based on the divided training set. The surrogate model includes three sub-models: multilayer perceptron (MLP), physical information neural network (PINN), and radial basis function neural network (RBFNN). The prediction and evaluation module is used to learn the optimal ensemble weights of the three sub-models through the near-end strategy optimization PPO algorithm to minimize the ensemble prediction error; finally, an ensemble surrogate model is constructed based on the optimal ensemble weights, and the ensemble surrogate model is used to predict the test set of the finite element model to obtain the predicted value, and the predicted value is evaluated to obtain the evaluation result. The cage model design module is used to guide the design of deep-sea aquaculture cage models based on the evaluation results.

[0014] The beneficial effects of this invention are as follows: This invention proposes a method and apparatus for designing deep-sea aquaculture cage models based on machine learning. 1. An integrated surrogate model architecture combining a multilayer perceptron (MLP), a physical information neural network (PINN), and a radial basis function neural network (RBFNN) was constructed, fully leveraging the advantages of each model in nonlinear fitting, embedding physical laws, and capturing local features. This integrated strategy can simultaneously handle multiple design variables and performance indicators, effectively addressing the complex system modeling requirements of high-dimensionality, strong coupling, and multiple constraints in deep-sea aquaculture cage design. It avoids the trade-off between diversity and accuracy inherent in single models, significantly improving the coordination of multi-objective optimization and overall prediction accuracy.

[0015] 2. This invention introduces the PPO algorithm to achieve intelligent model weight allocation and scene adaptation, improving the generalization ability and engineering applicability of the ensemble model. It innovatively transforms the ensemble weight optimization problem into a sequence decision-making task in reinforcement learning. Through interaction with the simulation environment via the PPO algorithm, the model autonomously learns the optimal weight combination for each sub-model under different working conditions. This method can dynamically adjust the weights according to the actual engineering state, maintaining the global stability of the ensemble surrogate model while significantly enhancing its prediction accuracy and environmental adaptability in local areas, thus effectively supporting multi-objective collaborative optimization of deep-sea aquaculture cages.

[0016] 3. Adaptive parallel training of the model significantly improves modeling efficiency and system intelligence. This invention employs a multi-model parallel training architecture, combining Bayesian hyperparameter optimization and the LOOCV strategy to achieve synchronous training and rapid hyperparameter tuning of different neural network models. This mechanism not only significantly shortens the model development cycle and reduces computational resource requirements, but also reduces reliance on manual hyperparameter tuning through an automated learning process, thereby enhancing the overall system intelligence. Attached Figure Description

[0017] The above and other features of this disclosure will become more apparent from the detailed description of the embodiments illustrated in conjunction with the accompanying drawings. In the accompanying drawings, the same reference numerals denote the same or similar elements. Obviously, the drawings described below are merely some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained from these drawings without any creative effort. In the drawings: Figure 1 The flowchart shown is a process for designing a deep-sea aquaculture cage model based on machine learning, according to the present invention. Figure 2 The diagram shows the flowchart of the optimal weight calculation using the PPO algorithm in this invention. Figure 3 The diagram shown is a schematic diagram of the MLP structure in a preferred embodiment of the present invention. Figure 4 The diagram shown is a schematic diagram of a PINN structure in a preferred embodiment of the present invention. Figure 5 The diagram shown is a schematic diagram of the RBFNN structure in a preferred embodiment of the present invention. Figure 6 The diagram shown is a schematic diagram of the DRBFNN structure in a preferred embodiment of the present invention. Figure 7 The diagram shown is a structural principle diagram of the integrated proxy model for deep-sea aquaculture cages proposed in this invention. Detailed Implementation

[0018] The following will provide a clear and complete description of the concept, specific structure, and technical effects of the present invention in conjunction with embodiments and accompanying drawings, so as to fully understand the purpose, solution, and effects of the present invention. It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The same reference numerals used throughout the accompanying drawings indicate the same or similar parts.

[0019] Example 1, referring to Figure 1 as well as Figure 7 This invention proposes a machine learning-based method for designing deep-sea aquaculture cage models, including the following: Step 110: Determine the fatigue hazard points of the deep-sea aquaculture cage and construct the finite element model of the deep-sea aquaculture cage; Step 120: Based on the actual working conditions and simulation analysis, determine the objective function, design variables, value range, and optimization constraints; Step 130: Sample the design variables and constraints of the aquaculture cage by using Latin hypercube to obtain samples, substitute the samples into the finite element model to solve, obtain the corresponding objective function values, and divide the training set and test set. Then, train the surrogate model based on the divided training set. The surrogate model includes three sub-models: multilayer perceptron (MLP), physical information neural network (PINN), and radial basis function neural network (RBFNN). Reference Figure 2 Step 140: Learn the optimal ensemble weights of the three sub-models by optimizing the PPO algorithm through the near-end strategy to minimize the ensemble prediction error; finally, construct an ensemble surrogate model based on the optimal ensemble weights, use the ensemble surrogate model to predict the test set of the finite element model to obtain the predicted value, and evaluate the predicted value to obtain the evaluation result. Step 150: Provide guidance on the design of deep-sea aquaculture cage models based on the evaluation results.

[0020] In this embodiment 1, refer to Figure 3 , Figure 4 as well as Figure 5 , 1. Construction of an surrogate model based on multi-neural network integration: An integrated surrogate model including MLP, PINN and RBFNN was constructed to comprehensively utilize the advantages of different neural network architectures in order to efficiently handle multi-input, multi-output, high-dimensional nonlinear and multi-objective optimization problems in deep-sea aquaculture cage design.

[0021] 2. Dynamic weight optimization mechanism based on reinforcement learning: The PPO algorithm is introduced to transform the problem of weight allocation in sub-model ensemble into a sequential decision problem in reinforcement learning. Through the interaction between the agent and the simulation environment, dynamic optimization of weight allocation and scene adaptation are achieved, which significantly improves the prediction accuracy and generalization ability of the ensemble model.

[0022] 3. Adaptive Parallel Training and Hyperparameter Optimization Framework: Adopting a multi-model parallel training architecture, combined with Bayesian optimization and LOOCV strategies, it realizes synchronous training of multiple neural networks and automatic hyperparameter optimization, which greatly improves the efficiency and intelligence of model development and reduces the dependence on human experience.

[0023] In a preferred embodiment of the present invention, specifically, in step 2, the design variables of the deep-sea aquaculture cage model are expressed as follows: ,in, The diameter of the wire mesh cage, For the height of the cage, For the cone angle of the mesh, For the frame pipe wall thickness, The outer diameter of the frame pipe, For the density of the frame material, The areal density of the mesh material. Due to the resistance of ocean currents on the net cages, The density of the seawater where the net cage is located. The stress exerted on the cage by the ocean current. The stress exerted on the fish cage by sea waves, The stress exerted on the fish cage by the sea wind; Different value ranges are pre-defined for the design variables based on material properties and engineering practice; The objective function is set as follows: Output 1: Minimize the peak stress at the fatigue hazard point of the wire mesh cage. : Calculation method: ; Output 2: Minimize the material cost of the wire mesh cage. : Calculation method: ; in, This is the unit price of the frame material. This refers to the unit price of the mesh material. Output 3: Maximize the effective aquaculture volume of the net cage. : Calculation method: ; in, This is the space utilization coefficient; Output 4: Minimize the current resistance coefficient of the net cage. : Calculation method: ; The multi-objective function is expressed as .

[0024] In a preferred embodiment of the present invention, step 130 specifically includes the following sub-steps. Step 131: Use Latin hypercube sampling to generate a sample size of [number missing]. The initial sample set is expressed as follows: ; in, Number of variables; Step 132: Substitute each valid sample from the initial sample set into the finite element model of the deep-sea aquaculture cage, and conduct simulation calculations under multiple working conditions to obtain the objective function value corresponding to each sample; Step 133: Execute the initial sample set In each iteration of the second iteration, one sample is randomly selected as the test set, and the rest... One sample is used as the training set; Step 134: Based on the objective function value, construct the Multilayer Perceptron (MLP), Physical Information Neural Network (PINN), and Radial Basis Function Neural Network (RBFNN) respectively; like Figure 3 As shown, the MLP is a classic feedforward neural network, mainly composed of three parts: an input layer, hidden layers, and an output layer. The number of hidden layers is typically 3 to 5, and each layer typically contains 32 to 128 neurons. Complex nonlinear mapping relationships are learned through densely connected multilayer neurons. The activation functions of the hidden layer neurons introduce nonlinear transformations, thereby learning the complex mapping relationships between multiple inputs and multiple outputs of data. The 12 design variables related to deep-sea aquaculture cages, after entering the hidden layer neurons through the input layer of the MLP, are first weighted, as follows: in, It is a linear combination result. Indicates the first One input variable, Indicates the first The weights of the input variables, These are the bias parameters. Then, activation functions are used to transform the data. Common activation functions include the Identity function, sigmoid function, tanh function, and ReLU function. Finally, a multi-output head simultaneously outputs the feature extraction results of four objective functions, sharing all hidden layers.

[0025] like Figure 4 As shown, the core of PINN is to use the residuals of the physical equations as part of the loss function, based on MLP, to constrain the network's output. After the 12 input variables pass through the hidden layer and output 4 output vectors, the dataset is formed. Perform data loss MSE calculation to obtain the data loss. In addition, outputs 1 and 4 also perform physical loss calculations. For output 1, the stress balance equation is embedded; for output 4, the continuity equation and momentum equation are embedded. Substituting these two outputs into the equations, partial differential calculations are performed to obtain the physical loss. The sum of the two yields the total loss function. Finally, the total loss function is fed into the optimizer for optimization. If the total loss is less than the set value, the data is output; if the total loss is greater than the set value, the data is returned to the neural network and optimized. Through the constraint of the physical loss term, the network's prediction results not only fit the data but also strictly follow known physical laws, ensuring the reasonableness of the predictions even in sparse data regions.

[0026] like Figure 5 As shown, RBFNN is a three-layer feedforward neural network. Its hidden layer activation function is the radial basis function, and the commonly used basis function is the Gaussian function, whose expression is: in, Indicates the first One input variable, and These are the centers of the basis functions and the points around which the function revolves. The width of each center point The distance between the input variable and the center vector is given. The output layer is a linear weighted sum of the hidden layer outputs. Due to the "local response" property of radial basis functions, RBFNN excels at capturing local features and abrupt changes in the input space, enabling it to more accurately predict local extrema problems such as the peak stress in output 1. RBFNN can also easily fit smooth and continuous functions like outputs 2, 3, and 4 with high accuracy. Furthermore, we add an MLP structure before the output layer of the RBFNN, such as... Figure 6The diagram shows a DRBFNN framework with two additional MLP hidden layers. The activation function of the first hidden layer neuron uses the base function commonly used in RBFNNs, while the activation functions of the other layers are consistent with the MLP. In this way, the outputs of the base functions are re-entered into the MLP for superposition mapping, improving the network's generalization ability.

[0027] Step 135: On different GPUs and computing nodes, LOOCV is used to train MLP, PINN, and RBFNN in parallel. The training of each sub-model is completed by optimizing its weights and bias parameters and minimizing the loss function through the backpropagation algorithm. Bayesian optimization algorithm is used to optimize hyperparameters such as learning rate, batch size, number of hidden layers, number of neurons, center point RBFNN, L2 regularization coefficient, physical loss function weight PINN, optimizer, and activation function to enhance the model's ability to adapt to the complex nonlinear relationship between the high-dimensional design variables of the net cage and the objective function. Finally, the root mean square error (RMSE) is used as the indicator to measure the model's prediction accuracy.

[0028] In a preferred embodiment of the present invention, the specific formula for calculating RMSE is as follows: ; In the formula, Indicates the first The actual observed value or true value of a sample Indicates the first The predicted value for each sample, This represents the number of samples.

[0029] In this preferred embodiment, the RMSE is calculated in the above manner to accurately measure the model's prediction accuracy.

[0030] In a preferred embodiment of the present invention, specifically, the process of learning the optimal ensemble weights of the three sub-models through the proximal strategy optimization PPO algorithm includes, If the output formula of the integrated prediction is: ; The error with the true value is minimized, that is ; in, These are pre-trained MLP, PINN, and RBFNN (or DRBFNN) sub-models; X represents the input features (12 design variables). For the integrated model predictions (4 performance targets); The actual value; Based on the above objectives, and referring to Figure 2Specifically, the process of learning the optimal ensemble weights of the three sub-models through the proximal strategy optimization PPO algorithm includes: If the output formula of the integrated prediction is: ; The error with the true value is minimized, that is ; in, These are the prediction results of the three trained sub-models MLP, PINN, and RBFNN, respectively; X represents the input features. This is the i-th ensemble prediction; This is the i-th true value; Based on the above objectives, the learning process for the optimal ensemble weights is as follows: Step 141: Initialize the agent's actor network to output the probability distribution of actions. , Representing state Take action below The conditional probability, Network parameters; initialize the critic network to output state value estimates. , Representing state The expected cumulative reward, Establish network parameters; create an experience pool Used to store trajectory data generated by the interaction between the agent and the environment; configures the learning rate of the actor network. Critical network learning rate Maximum number of training rounds Discount Factor and GAE parameters ; Step 142: Load the trained MLP, PINN, and RBFNN models into memory, fix their parameters, and integrate them as functions into the PPO environment; Step 143: The agent interacts with the environment and collects information about the current state. Information, namely the historical errors of the design parameters (design variables) of each cage and each sub-model, based on the current state. Utilizing old policy networks Acquire action That is, the weight allocation vector After obtaining the caching strategy, the agent executes the generated actions and receives immediate rewards based on the actions performed. The agent then transitions the environment to the next state, performs an action, and observes the reward. and obtain the state at the next moment. The interactive data is stored in the experience pool to form a trajectory sequence. The parameters for the old strategy network, i.e., the actor network; Step 144: When the trajectory reaches the cutoff length At that time, using the critic network The termination state is evaluated to obtain an estimated reward, and then the advantage function is calculated based on the trajectory sequence. The calculation formula is: ; in, The timing difference error is calculated using the following formula: ; It is a state-value function; Step 145: Extract complete trajectory data from the experience pool to train the critic network and actor network; The critic network is updated during training by minimizing the prediction error using the loss function, as shown in the following formula: ; in, For the critic network to evaluate the current state Value prediction The formula for calculating cumulative returns based on discounts is as follows: ; Update the critic network parameters The method is as follows: ; This is a gradient descent update step. New parameters. Equal to the old parameter Subtract its loss function gradient With critic network learning rate The product of, where For loss function Regarding parameters The gradient.

[0031] The actor network is updated during training by maximizing the pruning objective function, as shown in the following formula: ; in, Importance sampling ratio, This is a hyperparameter for the cropping range; ; Update actor network parameters The method is as follows: ; This is a gradient descent update step. New parameters. Equal to the old parameter Subtract its loss function gradient With critic network learning rate The product of, where For the pruning objective function Regarding parameters The gradient.

[0032] Continue iterating through the above steps until the strategy is optimal or the maximum number of training rounds is reached. ; Step 146: Based on the optimal weights obtained after training We will integrate proxy models and predict the objective function. Step 147: Evaluate the prediction results of Step 4.6 using the root mean square error (RMSE) and compare the obtained prediction results with the prediction results of a single surrogate model. The smaller the RMSE value, the better the prediction effect.

[0033] Example 2: This invention also proposes a machine learning-based deep-sea aquaculture cage model design device, comprising the following: The finite element model construction module is used to determine the fatigue hazard points of the deep-sea aquaculture cage and construct the finite element model of the deep-sea aquaculture cage. The objective function determination module is used to determine the objective function, design variables, value range, and optimization constraints based on actual working conditions and simulation analysis. The surrogate model construction module is used to sample the design variables and constraints of the aquaculture cage through Latin hypercube to obtain samples, substitute the samples into the finite element model to solve, obtain the corresponding objective function values, and divide the training set and test set. Then, the surrogate model is trained based on the divided training set. The surrogate model includes three sub-models: multilayer perceptron (MLP), physical information neural network (PINN), and radial basis function neural network (RBFNN). The prediction and evaluation module is used to learn the optimal ensemble weights of the three sub-models through the near-end strategy optimization PPO algorithm to minimize the ensemble prediction error; finally, an ensemble surrogate model is constructed based on the optimal ensemble weights, and the ensemble surrogate model is used to predict the test set of the finite element model to obtain the predicted value, and the predicted value is evaluated to obtain the evaluation result. The cage model design module is used to guide the design of deep-sea aquaculture cage models based on the evaluation results.

[0034] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated modules described above can be implemented in hardware or as software functional modules.

[0035] If the integrated module is implemented as a software functional module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or system capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0036] Although the description of the invention has been quite detailed and particularly of several described embodiments, it is not intended to limit it to any of these details or embodiments or any particular embodiment, but should be considered as providing a broad possible interpretation of the claims by referring to the appended claims and taking into account the prior art, thereby effectively covering the intended scope of the invention. Furthermore, the invention has been described above with respect to embodiments foreseeable by the inventors in order to provide a useful description, and non-substantial modifications to the invention that have not yet been foreseen may still represent equivalent modifications.

[0037] The above description is merely a preferred embodiment of the present invention. The present invention is not limited to the above-described embodiments. Any embodiment that achieves the technical effects of the present invention using the same means should fall within the protection scope of the present invention. Within the protection scope of the present invention, various modifications and variations can be made to the technical solutions and / or implementation methods.

Claims

1. A machine learning-based method for designing deep-sea aquaculture cage models, characterized in that, Including the following: Step 110: Determine the fatigue hazard points of the deep-sea aquaculture cage and construct the finite element model of the deep-sea aquaculture cage; Step 120: Based on the actual working conditions and simulation analysis, determine the objective function, design variables, value range, and optimization constraints; Step 130: Sample the design variables and constraints of the aquaculture cage by using Latin hypercube to obtain samples, substitute the samples into the finite element model to solve, obtain the corresponding objective function values, and divide the training set and test set. Then, train the surrogate model based on the divided training set. The surrogate model includes three sub-models: multilayer perceptron (MLP), physical information neural network (PINN), and radial basis function neural network (RBFNN). Step 140: Learn the optimal ensemble weights of the three sub-models by optimizing the PPO algorithm through the near-end strategy to minimize the ensemble prediction error; finally, construct an ensemble surrogate model based on the optimal ensemble weights, use the ensemble surrogate model to predict the test set of the finite element model to obtain the predicted value, and evaluate the predicted value to obtain the evaluation result. Step 150: Provide guidance on the design of deep-sea aquaculture cage models based on the evaluation results.

2. The method for designing deep-sea aquaculture cage models based on machine learning according to claim 1, characterized in that, Specifically, in step 2, the design variables of the deep-sea aquaculture cage model are expressed as follows: ,in, The diameter of the wire mesh cage, For the height of the cage, For the cone angle of the mesh, For the frame pipe wall thickness, The outer diameter of the frame pipe, For the density of the frame material, The areal density of the mesh material. Due to the resistance of ocean currents on the net cages, The density of the seawater where the net cage is located. The stress exerted on the cage by the ocean current. The stress exerted on the fish cage by sea waves, The stress exerted on the fish cage by the sea wind; Different value ranges are pre-defined for the design variables based on material properties and engineering practice; The objective function is set as follows: Output 1: Minimize the peak stress at the fatigue hazard point of the wire mesh cage. : Calculation method: ; Output 2: Minimize the material cost of the wire mesh cage. : Calculation method: ; in, This is the unit price of the frame material. This refers to the unit price of the mesh material. Output 3: Maximize the effective aquaculture volume of the net cage. : Calculation method: ; in, This is the space utilization coefficient; Output 4: Minimize the current resistance coefficient of the net cage. : Calculation method: ; The multi-objective function is expressed as .

3. The method for designing deep-sea aquaculture cage models based on machine learning according to claim 2, characterized in that, Specifically, step 130 includes the following sub-steps: Step 131: Use Latin hypercube sampling to generate a sample size of [number missing]. The initial sample set is expressed as follows: ； in, Number of variables; Step 132: Substitute each valid sample from the initial sample set into the finite element model of the deep-sea aquaculture cage, and conduct simulation calculations under multiple working conditions to obtain the objective function value corresponding to each sample; Step 133: Execute the initial sample set In each iteration of the second iteration, one sample is randomly selected as the test set, and the rest... One sample is used as the training set; Step 134: Based on the objective function value, construct the Multilayer Perceptron (MLP), Physical Information Neural Network (PINN), and Radial Basis Function Neural Network (RBFNN) respectively; Step 135: On different GPUs and computing nodes, LOOCV is used to train MLP, PINN, and RBFNN in parallel. The training of each sub-model is completed by optimizing its weights and bias parameters and minimizing the loss function through the backpropagation algorithm. The hyperparameters of learning rate, batch size, number of hidden layers, number of neurons, center point RBFNN, L2 regularization coefficient, physical loss function weight PINN, optimizer, and activation function are optimized using the Bayesian optimization algorithm. Finally, the root mean square error (RMSE) is used as the metric to measure the prediction accuracy of the model.

4. The method for designing deep-sea aquaculture cage models based on machine learning according to claim 3, characterized in that, Specifically, the formula for calculating RMSE is: ； In the formula, Indicates the first The actual observed value or true value of a sample Indicates the first The predicted value for each sample, This represents the number of samples.

5. The method for designing deep-sea aquaculture cage models based on machine learning according to claim 3, characterized in that, Specifically, the process of learning the optimal ensemble weights of the three sub-models through the proximal strategy optimization PPO algorithm includes, If the output formula of the integrated prediction is: ； The error with the true value is minimized, that is ; in, These are the prediction results of the three trained sub-models MLP, PINN, and RBFNN, respectively; X represents the input features. This is the i-th ensemble prediction; This is the i-th true value; Based on the above objectives, the learning process for the optimal ensemble weights is as follows: Step 141: Initialize the agent's actor network to output the probability distribution of actions. , Representing state Take action below The conditional probability, Network parameters; initialize the critic network to output state value estimates. , Representing state The expected cumulative reward, Establish network parameters; create an experience pool Used to store trajectory data generated by the interaction between the agent and the environment; configures the learning rate of the actor network. Critical network learning rate Maximum number of training rounds Discount Factor and GAE parameters ; Step 142: Load the trained MLP, PINN, and RBFNN models into memory, fix their parameters, and integrate them as functions into the PPO environment; Step 143: The agent interacts with the environment and collects information about the current state. Information, namely the historical errors of the design parameters (design variables) of each cage and each sub-model, based on the current state. Utilizing old policy networks Acquire action That is, the weight allocation vector After obtaining the caching strategy, the agent executes the generated actions and receives immediate rewards based on the actions performed. The agent then transitions the environment to the next state, performs an action, and observes the reward. and obtain the state at the next moment. The interactive data is stored in the experience pool to form a trajectory sequence. The parameters for the old strategy network, i.e., the actor network; Step 144: When the trajectory reaches the cutoff length At that time, using the critic network The termination state is evaluated to obtain an estimated reward, and then the advantage function is calculated based on the trajectory sequence. The calculation formula is: ； in, The timing difference error is calculated using the following formula: ; It is a state-value function; Step 145: Extract complete trajectory data from the experience pool to train the critic network and actor network; The critic network is updated during training by minimizing the prediction error using the loss function, as shown in the following formula: ； in, For the critic network to evaluate the current state Value prediction The formula for calculating cumulative returns based on discounts is as follows: ; Update the critic network parameters The method is as follows: ； The actor network is updated during training by maximizing the pruning objective function, as shown in the following formula: ； in, Importance sampling ratio, This is a hyperparameter for the cropping range; ; Update actor network parameters The method is as follows: ； Continue iterating through the above steps until the strategy is optimal or the maximum number of training rounds is reached. ; Step 146: Based on the optimal weights obtained after training We will integrate proxy models and predict the objective function. Step 147: Evaluate the prediction results of Step 4.6 using the root mean square error (RMSE) and compare the obtained prediction results with the prediction results of a single surrogate model. The smaller the RMSE value, the better the prediction effect.

6. The method for designing deep-sea aquaculture cage models based on machine learning according to claim 1, characterized in that, The method further includes adding two MLP hidden layer structures before the output layer of the RBFNN model to obtain a DRBFNN model. The activation function of the first hidden layer neuron uses the base function of RBFNN, while the activation functions of the other layers are consistent with MLP. The DRBFNN model is used as a sub-model to replace the RBFNN model for subsequent operations.

7. A machine learning-based deep-sea aquaculture cage model design device, characterized in that, Including the following: The finite element model construction module is used to determine the fatigue hazard points of the deep-sea aquaculture cage and construct the finite element model of the deep-sea aquaculture cage. The objective function determination module is used to determine the objective function, design variables, value range, and optimization constraints based on actual working conditions and simulation analysis. The surrogate model construction module is used to sample the design variables and constraints of the aquaculture cage through Latin hypercube to obtain samples, substitute the samples into the finite element model to solve, obtain the corresponding objective function values, and divide the training set and test set. Then, the surrogate model is trained based on the divided training set. The surrogate model includes three sub-models: multilayer perceptron (MLP), physical information neural network (PINN), and radial basis function neural network (RBFNN). The prediction and evaluation module is used to learn the optimal ensemble weights of the three sub-models through the near-end strategy optimization PPO algorithm to minimize the ensemble prediction error; finally, an ensemble surrogate model is constructed based on the optimal ensemble weights, and the ensemble surrogate model is used to predict the test set of the finite element model to obtain the predicted value, and the predicted value is evaluated to obtain the evaluation result. The cage model design module is used to guide the design of deep-sea aquaculture cage models based on the evaluation results.

Images