An ai-based circulating water aquaculture environment adaptive regulation and decision method

Abstract

This invention discloses an AI-based adaptive regulation decision-making method for recirculating aquaculture systems (RAS), belonging to the fields of smart agriculture and deep learning. Addressing the problems of concept drift, lack of closed-loop parameter optimization capabilities, and high inference latency in existing RAS models, this invention trains a lightweight three-layer fully connected neural network model by collecting historical data of the aquaculture environment. Based on the physiological limits of the cultured species, it generates a massive number of virtual control parameter combinations. The neural network then uses parallel inference to select the optimal regulation scheme for growth prediction and executes it. The system continuously monitors the residual between actual growth data and predicted values. When the residual continuously exceeds a threshold, it automatically extracts incremental data to trigger online retraining, completing the model's self-evolution and iteration. This invention features low inference latency and low storage overhead, enabling rapid global optimization of high-dimensional aquaculture parameters, effectively combating environmental drift, and meeting the real-time adaptive regulation requirements of factory-scale RAS.

Description

Technical Field

[0001] This invention belongs to the fields of smart agriculture and computer deep learning technology, specifically involving a decision-making method for environmental regulation in recirculating aquaculture based on a multilayer fully connected neural network, a high-dimensional space random search optimization algorithm, and an online self-evolutionary retraining mechanism. Background Technology

[0002] Traditional factory-style recirculating aquaculture systems (RAS) rely heavily on human experience to adjust feeding rates and water quality parameters. Due to the extremely complex nonlinear coupling relationship between the growth of cultured organisms and environmental parameters (such as temperature, dissolved oxygen, pH, and concentration of toxic substances), it is difficult to establish accurate mathematical models to quantitatively guide daily aquaculture management.

[0003] While existing technologies have begun to incorporate machine learning models for fish growth prediction, they generally suffer from the following drawbacks:

[0004] Model staticization and environmental drift issues: During the aquaculture process, as the total amount of biomass increases, sensors age, and seasons change, the prediction accuracy of the initially trained model will rapidly decline due to "concept drift," making it impossible to maintain reliable prediction results in the long term.

[0005] Lack of closed-loop control capability: Most existing systems are in a "one-way prediction" mode, which can only output growth trend predictions. They cannot quickly search and lock the optimal combination of management parameters that can enable farmed organisms to achieve the maximum growth rate in the multi-dimensional parameter space, making it difficult to form a complete closed loop of prediction and control.

[0006] High computational energy consumption and latency: Some complex deep learning models (such as large recurrent neural networks or temporal networks) have extremely slow inference speeds when deployed at the edge, which cannot meet the industrial application requirements of real-time high-frequency, second-level adaptive control in recirculating aquaculture. Summary of the Invention

[0007] The purpose of this invention is to address the shortcomings of existing technologies by providing an AI-based adaptive regulation and decision-making method for recirculating aquaculture environments. By constructing a streamlined and efficient deep learning prediction network and a stochastic search space optimization algorithm, coupled with low-latency model management and an automatic triggered retraining evolution mechanism, it achieves closed-loop dynamic optimization of multi-dimensional aquaculture parameters and second-level control response.

[0008] To address the aforementioned technical problems, this invention provides the following technical solution: an AI-based adaptive regulation and decision-making method for recirculating aquaculture environments, comprising the following steps:

[0009] S1. Data collection and initialization training of aquaculture environment: Historical operation records of recirculating aquaculture ponds are collected as supervised learning samples. After standardizing and normalizing the input features, a fish growth prediction neural network model is trained and solidified into a lightweight model file.

[0010] S2. Batch generation of aquaculture control combinations: Determine the value range of each controllable environmental parameter based on the safety physiological limits of the target aquaculture species, and generate multiple sets of virtual control parameter combinations by uniformly distributed random sampling within the value range;

[0011] S3. Parallel optimization of optimal control scheme: After converting all virtual control parameter combinations into batch processing tensors, they are input into the solidified neural network model. The predicted growth index corresponding to each set of parameters is obtained through parallel inference, and the control parameter combination with the best prediction effect is selected from them.

[0012] S4. Optimal solution issuance and effect verification: Convert the optimal combination of control parameters into automated equipment control commands for execution, continuously collect actual growth data, and calculate the residual between actual growth indicators and predicted growth indicators.

[0013] S5. Online self-evolution update of the model: When the residual continuously exceeds the preset threshold, the newly added actual aquaculture feedback data is extracted as incremental training samples, and the neural network model is retrained online. After updating the model weights, the process returns to step S1 to enter the next round of regulation loop.

[0014] Further, in step S1, at least 100 historical aquaculture operation records are collected as supervised learning samples. The input features of each sample are 6-dimensional, specifically including water temperature, dissolved oxygen, feed amount, pH value, ammonia nitrogen concentration, and nitrite concentration. The output label is the daily weight gain of the corresponding fish. The 6-dimensional input features are dynamically normalized using a standardized formula to eliminate the influence of different physical dimensions on the convergence of the neural network gradient. A three-layer fully connected feedforward neural network is used for initial training. The weight matrix and bias vector inside the network are converged through the backpropagation algorithm, and finally solidified into a lightweight model file.

[0015] Furthermore, the topology of the three-layer fully connected feedforward neural network is 6→32→16→1, consisting of an input layer, a first hidden layer, a second hidden layer, and an output layer. Both the first and second hidden layers use the Rectified Linear Unit (ReLU) activation function. The first hidden layer is configured with a random dropout operator with a dropout rate of 0.1 to suppress overfitting. The output layer outputs a 1-dimensional prediction of the daily weight gain.

[0016] Further, in step S2, based on the safety physiological limits of the target cultured species, the value ranges of six controllable parameters are determined respectively: water temperature 18.0℃~28.0℃, dissolved oxygen 5.0~8.0mg / L, feed amount as a percentage of total fish weight 3.0%~5.0%, pH value 7.0~8.0, ammonia nitrogen concentration 0.01~0.10mg / L, and nitrite concentration 0.01~0.05mg / L. Within the high-dimensional control space formed by the above six parameters, 50,000 different virtual control parameter combinations are generated at one time through a multi-threaded parallel uniform distribution sampling mechanism, forming a parameter matrix of shape (50000,6).

[0017] Further, in step S3, the generated 50,000 sets of virtual control parameter combination matrices are converted into batch processing tensors and input into the solidified neural network model. The neural network performs layer-by-layer matrix multiplication and nonlinear activation mapping in parallel on all parameter combinations through the internally converged weight matrix and bias vector, and outputs the predicted daily weight gain result corresponding to each set of parameters. The maximum value retrieval operator (argmax) is called to locate the predicted daily weight gain with the largest value among all prediction results, and the corresponding input parameter combination is matched in reverse to serve as the optimal control decision suggestion for the current period.

[0018] Furthermore, in step S4, the optimal combination of control parameters is converted into standard control commands for the corresponding automated actuators, including the operating power of the variable frequency heat pump, the speed of the variable frequency aerator, and the single start-up duration of the intelligent feeder, and sent to the field equipment for execution; with a fixed period of 24 hours, the mean squared error (MSE) residual between the actual daily weight gain of the fish and the daily weight gain predicted by the model is dynamically calculated, and the prediction accuracy of the model is continuously monitored.

[0019] Further, in step S5, the predicted residuals are continuously subjected to time series discrimination. If the residuals exceed the preset 0.15g judgment threshold for 5 consecutive days, concept drift is judged to have occurred, and the model self-evolution process is automatically activated. Real aquaculture feedback data from the most recent 10 days are automatically extracted as the incremental training sample set. The Adaptive Moment Estimation (Adam) optimizer is called, and the mean squared error is used as the loss function to perform online forced retraining on the original neural network on the local gateway, dynamically correcting the weight matrix inside the network. An early stopping mechanism is enabled during the retraining process. When the validation set loss no longer decreases for 5 consecutive training cycles, the training is automatically terminated to avoid overfitting. After the retraining is completed, a new lightweight model file is generated and overwrites the original model, completing the online self-evolution iteration of the model.

[0020] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages compared with the prior art:

[0021] 1. Extremely low inference latency and storage overhead: Through a streamlined, highly cohesive 6→32→16→1 network architecture, the model file is only about 6KB, a single high-dimensional matrix prediction inference takes only about 0.003 seconds, and the model takes only 0.01 seconds from loading to being ready, which fully meets the industrial-grade requirements of second-level control at the edge.

[0022] 2. Rapid global optimization in high-dimensional space: By combining random search algorithms with the parallel batch processing capabilities of deep learning, 50,000 control schemes can be evaluated globally in a very short time, completely breaking the parameter blind spots of traditional experience-based regulation and achieving the global optimal matching of multi-dimensional aquaculture parameters.

[0023] 3. Closed-loop self-evolution to prevent accuracy drift: The system has a built-in error monitoring and self-training update mechanism, which can actively sense changes in the breeding environment and the succession of biological growth stages. Through online incremental retraining, it continuously corrects the model weights, achieving a self-evolution effect where the model becomes more and more accurate with use, effectively combating the concept drift problem.

[0024] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0025] Figure 1 This is the overall architecture and five-step closed-loop control diagram of the adaptive environmental regulation and self-evolutionary decision-making system of this invention;

[0026] Figure 2 This is a schematic diagram of the topology of the multilayer fully connected neural network of the present invention;

[0027] Figure 3 This is a flowchart of the global optimization and adaptive control based on the random search algorithm of the present invention;

[0028] Figure 4 This is a closed-loop logic flowchart of the self-evolutionary online incremental retraining mechanism of the model in this invention. Detailed Implementation

[0029] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] It should be noted that the terms "vertical," "horizontal," "up," "down," "left," "right," and similar expressions used in this article are for illustrative purposes only and do not represent the only possible implementation.

[0031] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains; the terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to limit the invention; the term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0032] The following is in conjunction with the appendix Figure 1-4 The present invention will be further described in detail with reference to specific embodiments, but the scope of protection of the present invention is not limited thereto. This embodiment takes the actual control application of California bass in a high-density factory recirculating aquaculture system as an example to describe the method of the present invention in detail.

[0033] Figure 1 This is the overall architecture and five-step closed-loop control flowchart of the method of this invention. (See diagram below.) Figure 1 As shown, the control loop in this embodiment mainly includes a closed-loop operation of five core steps:

[0034] Step 01, Aquaculture Environment Data Acquisition and Initial Model Training: Through IoT sensors deployed at the aquaculture site, multi-source heterogeneous environmental characteristic data are acquired in real time, providing initial data support for the system;

[0035] Step 02, exhaustively search all control combinations based on physiological limits: establish a high-dimensional search space based on the physiological limits of California bass, and generate a large number of virtual control parameter combinations in parallel within this space;

[0036] Step 03, Matrix parallel reasoning to find the optimal breeding scheme: Using a pre-trained and ready fully connected feedforward reasoning network, high-performance parallel matrix reasoning is performed on all generated control schemes to lock in the optimal scheme;

[0037] Step 04, Optimization results are converted into action and sent for actual verification: The optimal control decision is output to the automated actuator and sent to the actual aquaculture pond for equipment action response and verification;

[0038] Step 05, Residual error exceeding the limit triggers online self-evolution of the AI ​​model: continuously monitor actual aquaculture feedback, trigger online incremental retraining when the residual error continuously exceeds the limit, dynamically update the network weights, then update the model coverage and return to step 01 to start a new round of control loop.

[0039] The inference model used in this invention is a simplified fully connected feedforward neural network deployed in a computing gateway, and its topology is as follows: Figure 2 As shown, the model is strictly designed with a 6→32→16→1 architecture. The input layer receives a standardized 6-dimensional feature vector, including: water temperature, dissolved oxygen, feed amount, pH value, ammonia nitrogen concentration, and nitrite concentration. This vector is mapped to a 32-dimensional first hidden layer via a weight matrix. This layer uses a linear rectified activation function and a random deactivation operator (dropout rate set to 0.1-0.2) to effectively cut off neuronal collinearity under small sample training and prevent overfitting. The model is then further reduced in dimension to a 16-dimensional second hidden layer (also using a linear rectified activation function). Finally, the output layer performs 1-dimensional linear inference, which is then denormalized and scaled to restore the model to a physically meaningful "predicted daily weight gain (g)" for California bass. After the initial training, the model is fixed into a serialized file of approximately 6KB by the self-evolutionary model manager, ensuring a single system readiness loading latency of no more than 0.01 seconds.

[0040] In the initial training phase, the system first acquires 100 historical real-world operation records of physical aquaculture ponds as supervised learning samples. Each record is a complete supervised learning sample, with a total of 6-dimensional input features and an output label representing the fish's actual daily weight gain. To eliminate the influence of different physical dimensions on the convergence of the neural network gradient, the system calls a feature scaler to perform online dynamic standardization processing on the 6-dimensional input features. The mathematical mapping formula is as follows:

[0041]

[0042] Where X represents the original 6-dimensional features, \(\mu\) represents the historical baseline mean, and \(\sigma\) represents the standard deviation. After processing, a standard tensor is obtained. The normalized real samples are fed into the above neural network for initial training. The weight matrix and bias vector inside the network are converged through the backpropagation algorithm, thereby transforming the neural network into a prediction model that can map 6-dimensional inputs to 1-dimensional weighted outputs.

[0043] Figure 3 This is a schematic diagram of the global optimization and adaptive control process based on a random search algorithm of the present invention. Figure 3 As shown, when the aquaculture site issues a request for parameter optimization and adaptive control, the following steps are executed:

[0044] Step 301: Read the safe physiological limits of California bass and establish the boundary of the high-dimensional control space: The system reads the preset safe physiological limits of California bass and uses them as the boundary of the high-dimensional control space.

[0045] Step 302: Within the boundary limits, N=50000 sets of parameter combinations are generated by high-frequency parallel sampling using uniform distribution: The multi-dimensional parameter space random search optimization module performs high-frequency multi-threaded parallel sampling using uniform distribution within the above-mentioned boundary limits, and randomly generates N=50000 sets of different virtual control parameter combinations at once.

[0046] Step 303: Convert 50,000 sets of parameters into batch processing tensors: Convert these 50,000 sets of control parameters into tensor form in batches to facilitate matrix-based parallel throughput of the neural network.

[0047] Step 304: Input the neural network inference module to predict 50,000 corresponding daily weight gain values ​​in parallel: The batch tensor is fed into the neural network inference module as a whole, and the model completes parallel matrix inference within 0.003 seconds, outputting 50,000 corresponding predicted daily weight gain scalars.

[0048] Step 305: Run the maximum value retrieval operator to lock the golden control parameter combination corresponding to the predicted maximum weight gain: The system runs the maximum value retrieval operation to directly retrieve and lock the control combination that predicts the maximum daily weight gain.

[0049] Step 306: The optimal control parameters are converted into action commands and sent to the circulating water frequency converter and automated feeding actuator. The control terminal module converts the optimal parameters into physical equipment action commands and sends them to the automated actuator. In this embodiment, the range of controllable variables is: water temperature 18.0℃~28.0℃, dissolved oxygen 5.0~8.0mg / L, feed amount 3.0%~5.0%, pH value 7.0~8.0, ammonia nitrogen 0.01~0.10mg / L, and nitrite 0.01~0.05mg / L. The optimal control parameter combination generated by the system is specifically: optimal water temperature 29.8℃, optimal dissolved oxygen 5.5mg / L, optimal feed amount (as a percentage of total fish weight) 5.0%, optimal pH value 7.8, and both ammonia nitrogen and nitrite concentrations are within the safe limit of 0.01mg / L. Under this combination, the system infers that the predicted daily weight gain of California bass is as high as 5.39g. To further verify the model's decision sensitivity and rationality, the system conducted a full-link inference comparison across four typical recirculating aquaculture scenarios for California bass. The results showed that the growth rate of California bass was highest when the feed intake was maintained at 4.5%-5.0% and the water temperature was controlled within the range of 26℃-30℃. However, once water quality deteriorated (e.g., increased ammonia nitrogen and nitrite concentrations in harsh scenarios), even with a certain feed intake, the system predicted a sharp drop in weight gain to 1.23g. At this point, the system automatically triggered a linkage control command to increase water circulation and improve dissolved oxygen, demonstrating a high degree of adaptive environmental regulation capability.

[0050] Figure 4This is a closed-loop logic flowchart of the online incremental retraining mechanism for model self-evolution in this invention. Due to factors such as insufficient early training data and changes in the organism's own growth stage, the model's predicted values ​​and actual values ​​may deviate significantly. This embodiment achieves online self-evolution of the model through the following closed-loop iterative steps:

[0051] Step 401: Continuously monitor and record the actual daily weight gain data of California bass after implementing the control recommendations: The system continuously monitors and records the actual biological characteristics data of the fish after implementing the control actions during daily operation.

[0052] Step 402, dynamically calculate the residual error (mean square error) between the actual daily weight gain and the network inference prediction: dynamically calculate the mean square error residual value between the actual daily weight gain and the network prediction value;

[0053] Step 403, whether the residual error exceeds the set threshold for M consecutive days: The system determines whether the residual error exceeds the preset threshold for M consecutive days (M=4 in this embodiment). If it does not exceed the threshold, the system returns to step 401 to continue normal monitoring; if it exceeds the threshold for 4 consecutive days, the system determines that a concept drift has occurred due to biological stage change or equipment wear and tear, and the system automatically executes step 404.

[0054] Step 404, the model manager automatically extracts the local incremental aquaculture dataset containing the latest execution feedback: the self-evolutionary model manager module is activated and automatically extracts the local incremental aquaculture records containing the latest execution feedback as new training samples;

[0055] Step 405: Invoke the adaptive moment estimation optimizer for online forced retraining, and dynamically adjust the network weight parameters using an early stopping mechanism: The system calls the adaptive moment estimation optimizer to perform online incremental forced retraining of the neural network. An early stopping mechanism is configured during the retraining process, that is, retraining is automatically interrupted when the test set loss no longer decreases for 5 consecutive training epochs to prevent overfitting;

[0056] Step 406: Generate a new lightweight 6KB fixed file, overwriting the original network weights, and complete online self-evolution: After retraining, the system generates a new lightweight 6KB fixed file and overwrites the original network weights, completing online self-evolution. After self-evolution, the prediction residuals under the same conditions subsequently decrease back to within 0.15g, successfully achieving model self-iteration.

[0057] Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Anyone skilled in the art can make various modifications and alterations without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention should be determined by the claims.

Claims

1. An AI-based adaptive regulation decision-making method for recirculating aquaculture environments, characterized in that, Includes the following steps: S1. Data collection and initialization training of aquaculture environment: Historical operation records of recirculating aquaculture ponds are collected as supervised learning samples. After standardizing and normalizing the input features, a fish growth prediction neural network model is trained and solidified into a lightweight model file. S2. Batch generation of aquaculture control combinations: Determine the value range of each controllable environmental parameter based on the safety physiological limits of the target aquaculture species, and generate multiple sets of virtual control parameter combinations by uniformly distributed random sampling within the value range; S3. Parallel optimization of optimal control scheme: After converting all virtual control parameter combinations into batch processing tensors, they are input into the solidified neural network model. The predicted growth index corresponding to each set of parameters is obtained through parallel inference, and the control parameter combination with the best prediction effect is selected from them. S4. Optimal solution issuance and effect verification: Convert the optimal combination of control parameters into automated equipment control commands for execution, continuously collect actual growth data, and calculate the residual between actual growth indicators and predicted growth indicators. S5. Online self-evolution update of the model: When the residual continuously exceeds the preset threshold, the newly added actual aquaculture feedback data is extracted as incremental training samples, and the neural network model is retrained online. After updating the model weights, the process returns to step S1 to enter the next round of regulation loop.

2. The AI-based adaptive regulation and decision-making method for recirculating aquaculture systems according to claim 1, characterized in that, Step S1 specifically includes: Collect no fewer than 100 historical aquaculture operation records as supervised learning samples. The input features of each sample are 6-dimensional, specifically including water temperature, dissolved oxygen, feed amount, pH value, ammonia nitrogen concentration, and nitrite concentration. The output label is the daily weight gain of the corresponding fish. A standardized formula is used to dynamically normalize the 6-dimensional input features to eliminate the influence of different physical dimensions on the convergence of neural network gradients. A three-layer fully connected feedforward neural network is used for initial training. The weight matrix and bias vector inside the network are converged through the backpropagation algorithm, and finally solidified into a lightweight model file.

3. The AI-based adaptive regulation and decision-making method for recirculating aquaculture systems according to claim 2, characterized in that, The topology of the three-layer fully connected feedforward neural network is 6→32→16→1, consisting of an input layer, a first hidden layer, a second hidden layer, and an output layer. Both the first and second hidden layers use linear rectified activation functions. The first hidden layer is configured with a random deactivation operator with a dropout rate of 0.1 to suppress overfitting. The output layer outputs a 1-dimensional prediction of the daily weight gain.

4. The AI-based adaptive regulation and decision-making method for recirculating aquaculture systems according to claim 1, characterized in that, Step S2 specifically includes: Based on the safety physiological limits of the target cultured species, the ranges of six controllable parameters were determined as follows: water temperature 18.0℃~28.0℃, dissolved oxygen 5.0~8.0mg / L, feed intake as a percentage of total fish weight 3.0%~5.0%, pH value 7.0~8.0, ammonia nitrogen concentration 0.01~0.10mg / L, and nitrite concentration 0.01~0.05mg / L. Within the high-dimensional control space comprised of the aforementioned six parameters, a multi-threaded parallel uniform distribution sampling mechanism is used to generate 50,000 different combinations of virtual control parameters at once, forming a parameter matrix of shape (50,000, 6).

5. The AI-based adaptive regulation and decision-making method for recirculating aquaculture systems according to claim 1, characterized in that, Step S3 specifically includes: The generated 50,000 sets of virtual control parameter combination matrices are converted into batch tensors and input into the solidified neural network model; The neural network performs layer-by-layer matrix multiplication and non-linear activation mapping in parallel on all parameter combinations using the internally converged weight matrix and bias vector, and outputs the predicted daily weight gain result for each set of parameters. The maximum value retrieval operator is invoked to locate the daily weight gain with the largest value among all prediction results. The corresponding input parameter combination is then matched in reverse to provide the optimal control decision suggestion for the current period.

6. The AI-based adaptive regulation and decision-making method for recirculating aquaculture systems according to claim 1, characterized in that, Step S4 specifically includes: The optimal combination of control parameters is converted into standard control commands for the corresponding automated actuators, including the operating power of the variable frequency heat pump, the speed of the variable frequency oxygenation pump, and the single start-up duration of the intelligent feeder, and then sent to the field equipment for execution. Using a fixed 24-hour period, the mean squared error residual between the actual daily weight gain of fish and the daily weight gain predicted by the model is dynamically calculated to continuously monitor the prediction accuracy of the model.

7. The AI-based adaptive regulation and decision-making method for recirculating aquaculture systems according to claim 1, characterized in that, Step S5 specifically includes: The predicted residuals are continuously analyzed for time series. If the residuals exceed the preset 0.15g threshold for 5 consecutive days, concept drift is determined to have occurred, and the model self-evolution process is automatically activated. The system automatically extracts real aquaculture feedback data from the past 10 days as an incremental training sample set, calls the adaptive moment estimation optimizer, uses mean squared error as the loss function, and performs online forced retraining on the original neural network on the local gateway to dynamically correct the weight matrix inside the network. The retraining process employs an early stopping mechanism, automatically terminating training when the validation set loss no longer decreases for five consecutive training epochs to avoid overfitting. After retraining is completed, a new lightweight model file is generated and overwrites the original model, completing the online self-evolutionary iteration of the model.

Images