CN-121997710-A - Deep sea aquaculture net cage model design method and device based on machine learning

CN121997710ACN 121997710 ACN121997710 ACN 121997710ACN-121997710-A

Abstract

The invention relates to a deep sea aquaculture net cage model design method based on machine learning, which firstly provides a proxy model based on multi-neural network integration, wherein an integrated proxy model comprising MLP, PINN and RBFNN is constructed, so that the problems of multi-input, multi-output, high-dimensional nonlinearity and multi-objective optimization in deep sea aquaculture net cage design can be effectively solved. And secondly, a dynamic weight optimization mechanism is provided, namely a PPO algorithm is introduced, the distribution problem of the sub-model integrated weight is converted into a sequence decision problem in reinforcement learning, and the dynamic optimization and scene self-adaption of weight distribution are realized through the interaction of an agent and a simulation environment, so that the prediction precision and generalization capability of the integrated model are remarkably improved. And finally, providing a self-adaptive parallel training and super-parameter optimizing framework, namely adopting a multi-model parallel training framework and combining Bayesian optimization and LOOCV strategy to realize synchronous training and super-parameter automatic optimizing of a plurality of neural networks, thereby greatly improving the development efficiency and the intelligence level of the model and reducing the dependence on the artificial experience.

Inventors

LIN JINGLIANG
SHAO YONGZHI
WU ZUODONG
LIU QIANG
XU XIAOMING

Assignees

广东海洋大学

Dates

Publication Date: 20260508
Application Date: 20251223

Claims (7)

1. The deep sea aquaculture net cage model design method based on machine learning is characterized by comprising the following steps of: Step 110, determining fatigue dangerous points of a deep sea culture net cage, and constructing a finite element model of the deep sea culture net cage; step 120, determining an objective function, design variables, a value range and optimization constraint conditions based on actual working conditions and simulation analysis conditions; 130, sampling design variables and constraints of a culture net cage through Latin hypercube to obtain a sample, substituting the sample into a finite element model to solve to obtain a corresponding objective function value, dividing a training set and a testing set, and then training a proxy model based on the divided training set, wherein the proxy model comprises three sub-models of a multi-layer perceptron MLP, a physical information neural network PINN and a radial basis function neural network RBFNN; Finally, constructing an integrated proxy model according to the optimal integration weights, predicting a test set of the finite element model by using the integrated proxy model to obtain a predicted value, and evaluating the predicted value to obtain an evaluation result; and step 150, guiding the design of the deep sea aquaculture net cage model based on the evaluation result.
2. The machine learning-based deep sea aquaculture net cage model design method according to claim 1, wherein, specifically, in said step 2, the design variables of the deep sea aquaculture net cage model are expressed as , wherein, The diameter of the net cage is the diameter of the net cage, The height of the net cage is the height of the net cage, For the cone angle of the net, For the wall thickness of the frame tube, Is the outer diameter of the frame tube, In order to achieve the density of the frame material, For the areal density of the material of the netting, Is the resistance of the sea current to the net cage, Is the density of the seawater in which the net cage is positioned, Is stressed by ocean currents applied to the net cage, Is stressed by sea waves of the net cage, The net cage is stressed by sea wind; different value ranges are preset for the design variables by combining material characteristics and engineering practices; the set objective function is as follows: output 1 peak stress to minimize cage fatigue hazard points : The calculation mode is as follows: ; output 2 minimizing cage Material cost : The calculation mode is as follows: ; Wherein, the Is a unit price of the frame material, The unit price of the net material is that; Output 3 maximizing effective farming volume of the cage : The calculation mode is as follows: ; Wherein, the Is a space utilization coefficient; Output 4 minimize the coefficient of resistance to ocean currents of the cage : The calculation mode is as follows: ; The multiple objective functions are expressed as 。
3. The method for designing a deep sea farming net cage model based on machine learning as defined in claim 2, wherein the step 130 comprises the sub-steps of, Step 131, sampling by Latin hypercube to generate sample number as The expression of which is as follows: ; Wherein, the Is the variable number; Step 132, substituting the effective samples in the initial sample set into a finite element model of a deep sea culture net cage one by one, and carrying out simulation calculation under multiple working conditions so as to obtain objective function values corresponding to the samples; Step 133, executing the initial sample set Dividing sub-loops, randomly selecting 1 sample in each loop as test set, and the rest The samples are used as training sets; step 134, respectively constructing a multi-layer perceptron MLP, a physical information neural network PINN and a radial basis function neural network RBFNN according to the objective function values; Step 135, performing parallel training on the MLP, the PINN and the RBFNN by adopting a leave-one-out cross-validation LOOCV on different GPUs and computing nodes, performing training on each sub-model by optimizing the weight and the bias parameters and minimizing the loss function through a back propagation algorithm, and performing optimization on the learning rate, the batch size, the hidden layer number, the neuron number, the central point RBFNN, the L2 regular coefficient, the physical loss function weight PINN, the optimizer and the super-parameters of the activation function by adopting a Bayesian optimization algorithm, wherein the root mean square error RMSE is finally used as an index for measuring the model prediction accuracy.
4. The machine learning-based deep sea aquaculture net cage model design method according to claim 3, wherein the specific calculation formula of RMSE is: ; in the formula, Represent the first Actual observations or true values of the individual samples, Represent the first The predicted value of the individual samples is calculated, Is the number of samples.
5. The method for designing a deep sea aquaculture net cage model based on machine learning according to claim 3, wherein the process of learning the optimal integration weights of three sub-models by optimizing the PPO algorithm through a near-end strategy comprises, If the output calculation formula of the integrated prediction is: ; with minimum error from true value, i.e. ; Wherein, the Respectively predicting results of the trained three sub-models MLP, PINN and RBFNN, wherein X is an input characteristic; an i-th integrated predictive value; is the ith true value; Based on the above objective, the learning process for performing the optimal integration weight is as follows: Step 141, initializing actor network of agents to output probability distribution of actions , Representing the status Take action downwards Is a function of the conditional probability of (1), Initializing critic a network to output state value estimates for network parameters , Representing the status Is a combination of the expected jackpot of (a), Establishing experience pools for network parameters For storing trajectory data generated by interaction of an agent with an environment, and for configuring actor a network learning rate Network learning rate critic Maximum training round Discount factor And GAE parameter ; Step 142, loading the trained MLP, PINN, RBFNN models into a memory, fixing parameters of the models, and integrating the models into a PPO environment as a function; step 143, the agent interacts with the environment to collect information about the current state Based on the current state, i.e., historical errors of various cage design parameters, i.e., design variables, and various sub-models Using old policy networks Obtaining an action I.e. weight distribution vector Obtaining a caching strategy, executing the generated action by the intelligent agent, and obtaining instant rewards according to the executed action And the environmental state is transferred to the next state, the intelligent agent executes the action and observes the rewards And obtain the state of the next moment Storing the interacted data into an experience pool to form a track sequence, Network parameters are the old policy network, namely actor; step 144 when the track reaches the cut-off length When utilizing critic network Evaluating the termination state to obtain an estimated reward, and then calculating a merit function based on the trajectory sequence The calculation formula is as follows: ; Wherein, the For the time sequence difference error, the calculation formula is as follows: ; As a function of state values; step 145, extracting complete track data from the experience pool for training critic networks and actor networks; updating critic the network in the training process, and realizing by minimizing the prediction error of the loss function, wherein the specific formula is as follows: ; Wherein, the For critic network pairs current state Is a value prediction of (1), For the discount cumulative return, the calculation formula is: ; updating critic network parameters The way of (2) is as follows: ; Updating actor the network in the training process, and realizing by maximizing the clipping objective function, wherein the specific formula is as follows: ; Wherein, the For the importance of the sampling rate, The cutting range is a super parameter; ; updating actor network parameters The way of (2) is as follows: ; continuing to iterate the steps until the strategy is optimal or the maximum training round is reached ; Step 146, according to the optimal weight obtained after training Carrying out agent model integration and predicting an objective function; And 147, evaluating the prediction result in the step 4.6 by using a Root Mean Square Error (RMSE), and comparing the obtained prediction result with the single agent model prediction result, wherein the smaller the RMSE value is, the better the prediction effect is.
6. The machine learning based deep sea aquaculture net cage model design method of claim 1, further comprising adding 2 MLP hidden layer structures before the output layer of the RBFNN model to obtain DRBFNN model, wherein the first hidden layer neuron activation function uses the RBFNN basis function, and the other layer activation functions are consistent with MLP, and using DRBFNN model as sub-model to replace the RBFNN model for subsequent operation.
7. Deep sea aquaculture net case model design device based on machine learning, its characterized in that includes following: The finite element model construction module is used for determining fatigue dangerous points of the deep sea culture net cage and constructing a finite element model of the deep sea culture net cage; The objective function determining module is used for determining an objective function, a design variable, a value range and an optimization constraint condition based on the actual working condition and the simulation analysis condition; The agent model construction module is used for sampling design variables and constraints of the culture net cage through Latin hypercube to obtain samples, substituting the samples into a finite element model to be solved to obtain corresponding objective function values, dividing a training set and a testing set, and then training an agent model based on the divided training set, wherein the agent model comprises three sub-models of a multi-layer perceptron MLP, a physical information neural network PINN and a radial basis function neural network RBFNN; The prediction evaluation module is used for learning the optimal integration weights of the three sub-models through a near-end strategy optimization PPO algorithm so as to minimize an integrated prediction error, constructing an integrated proxy model according to the optimal integration weights, predicting a test set of the finite element model by using the integrated proxy model to obtain a predicted value, and evaluating the predicted value to obtain an evaluation result; and the net cage model design module is used for guiding the design of the deep sea aquaculture net cage model based on the evaluation result.

Description

Deep sea aquaculture net cage model design method and device based on machine learning Technical Field The invention relates to the technical field related to deep sea aquaculture net cage model design, in particular to a method and a device for designing a deep sea aquaculture net cage model based on machine learning. Background In the design of the deep sea aquaculture net cage, special attention must be paid to the design optimization of the net cage in order to ensure the safety of the net cage structure, ensure that structural damage does not occur under the action of complex loads such as sea waves and ocean currents, and maintain good performance. However, with the improvement of the level of functional compounding and intelligence, the deep sea aquaculture net cage has evolved into a complex and large system with multiple disciplinary intersection, and has the defects of multiple design variables, complex operation conditions, multiple target performances to be considered, and high-dimensional nonlinearity problems of typical multiple inputs and multiple outputs, which lead to the difficulty of the design process, long period and high cost. In practical application, due to multiple severe requirements on performance, factors required to be considered in structural design are extremely complex, so that simulation analysis in the design optimization process is time-consuming and high in cost. In the existing research for the optimization of the net cage design, a simulation analysis method, a physical model experiment method, a numerical simulation method and the like are commonly used, however, the problems cannot be solved remarkably by the method. Under the background, the auxiliary design optimization of the agent model becomes a technical path commonly adopted in the industry, and the current method mainly comprises a single agent model and an integrated agent model. The single agent model, such as the kriging model, focuses on the representation of the relationship between few variables and single performance response, and is difficult to adapt to the design environment of the aquaculture net cage. Moreover, under the condition of lacking priori knowledge, the characteristics of the selected agent model and the characteristics of the actual problems are difficult to realize good matching, and the model prediction variance is possibly high, so that the generalization capability is insufficient. The deep neural network proxy model can better remedy the defects due to the characteristics of strong multi-input and multi-output nonlinear characteristic relation characterization capability, good migration adaptability and the like. However, the current research on the deep neural network proxy model focuses on a single model, and the generalization capability of the deep neural network proxy model may not be effectively improved. The integrated agent model effectively reduces the prediction variance and improves the generalization capability of the final model by fusing the advantages of different agent models. However, the selection of the subset of models and the way the sub-models are integrated are always key factors affecting the computational efficiency and generalization ability of the integrated proxy model. In the existing integration mode, various strategies have advantages, but have obvious limitations. The agent model integration strategy based on weight optimization is a widely applied method at present. The global weight integration is simple in calculation and easy to realize as a prominent advantage, and is suitable for scenes with stable global performance of the base model, but due to the fact that local precision differences in a design space are ignored, sufficient prediction precision is difficult to ensure in complex nonlinear problems. In contrast, the point-by-point weight integration can fully adapt to local characteristics by dynamically adjusting weights of different design points, so that local prediction accuracy is remarkably improved, but the weights are required to be calculated for each design point independently, calculation complexity and cost are greatly increased in a high-dimensional problem, and practicability is limited. The stacking integration method builds an integration framework through a multi-level learning mechanism, so that the integration framework can capture complex relations among base models, the generalization capability is enhanced, however, the method has higher data volume requirements, the design of a secondary model and the implementation of a cross verification strategy are both increased, the implementation difficulty is easy to fall into a dilemma of fitting when the data is insufficient, the agent integration method based on the optimization of the composition model can effectively reduce redundancy and reduce calculation cost by screening out models with poor performance, and avoid negative influence of a poor model on an integration res