A method and system for monitoring oxygenation in aquaculture based on artificial intelligence

Abstract

This invention discloses an artificial intelligence-based method and system for monitoring oxygenation in aquaculture. It collects multimodal environmental data and inputs it into an ST-BiGRU spatiotemporal bidirectional gated recurrent unit. The time channel processes historical time-series data, and a dynamic weighted attention mechanism is introduced to adaptively adjust the contribution of different environmental factors, resulting in a multi-dimensional feature vector. A dissolved oxygen prediction model is established based on an LSTM (Long Short-Term Memory) network. An improved IPOS (In-Plane System) particle swarm optimization algorithm is used to optimize the model's hyperparameters. The multi-dimensional feature vector is then input into the target dissolved oxygen prediction model for prediction, outputting a dissolved oxygen concentration index for a preset future time period. Oxygenation resource management is then performed based on the dissolved oxygen concentration index, combined with a reinforcement learning algorithm and a preset dissolved oxygen safety threshold range. This effectively reduces power consumption and equipment wear, lowers aquaculture costs, and improves resource utilization efficiency in aquaculture.

Description

Technical Field

[0001] This invention relates to the field of aquaculture technology, and in particular to an artificial intelligence-based method and system for monitoring oxygenation in aquaculture. Background Technology

[0002] In aquaculture, dissolved oxygen (DO) is a critical water quality parameter that directly affects the growth, health, and survival rate of farmed organisms. Traditional aquaculture oxygenation monitoring methods mainly rely on setting fixed thresholds (activating aerators when DO falls below a certain value) or simple on / off control based on real-time DO sensor data. This approach is highly lagging, failing to provide early warning and intervention when DO levels begin to decline but have not yet reached the danger threshold, easily leading to oxygen deprivation stress and even death in fish and shrimp. Fixed thresholds or simple timed control can easily lead to over-operation of aerators, resulting in energy waste and low oxygenation efficiency. Summary of the Invention

[0003] The purpose of this invention is to solve the above problems by designing an artificial intelligence-based aquaculture oxygenation monitoring method and system.

[0004] To achieve the above objectives, the technical solution of the present invention further includes the following steps in the above-mentioned artificial intelligence-based aquaculture oxygenation monitoring method:

[0005] Multimodal environmental data is collected and input into the ST-BiGRU spatiotemporal bidirectional gated cyclic unit. The spatial channel is used to process the grid sensor data, and the temporal channel is used to process the historical time series data. A dynamic weight attention mechanism is introduced to adaptively adjust the contribution of different environmental factors to obtain a multidimensional feature vector.

[0006] A dissolved oxygen prediction model is established based on LSTM (Long Short-Term Memory) network, and the hyperparameters of the model are optimized using an improved IPOS (Integrated Population Swarm Optimization) algorithm to obtain the target dissolved oxygen prediction model.

[0007] The multi-dimensional feature vector is input into the target dissolved oxygen prediction model for prediction, and the dissolved oxygen concentration index of the water body is output within a preset time period in the future.

[0008] Oxygenation resource management is carried out based on the dissolved oxygen concentration index of the water body, combined with reinforcement learning algorithms and preset dissolved oxygen safety threshold ranges.

[0009] Furthermore, in the aforementioned artificial intelligence-based aquaculture oxygenation monitoring method, the collection of multimodal environmental data includes:

[0010] Data on dissolved oxygen concentration, environmental parameters, meteorological information, and fish images of the aquaculture water were collected. Outlier detection was performed on the collected data, and the Laida criterion was used to identify and remove outliers to obtain cleaned data.

[0011] Random noise in the sensor-acquired data is processed using wavelet transform denoising method to obtain denoised data;

[0012] The z-score standardization method is used to transform the data into standardized data with a mean of 0 and a standard deviation of 1, thus obtaining multimodal environment data.

[0013] Furthermore, in the aforementioned artificial intelligence-based aquaculture oxygenation monitoring method, the step of inputting the multimodal environmental data into the ST-BiGRU spatiotemporal bidirectional gated loop unit includes:

[0014] Multimodal environmental data is input into the spatial channel in the form of a three-dimensional matrix. The three-dimensional matrix is ​​then split into multiple two-dimensional grid layers according to the type of environmental factor. Each layer represents the spatial distribution of an environmental factor.

[0015] Multi-layer stacked convolution operations are used to process each grid layer. A 3×3 convolution kernel slides on the grid to calculate the correlation features of sensor data in local areas.

[0016] Global pooling is used to integrate local features into a feature vector that reflects the spatial pattern of the entire aquaculture area; cross-channel convolutional layers are used to calculate the correlation of the spatial distribution of different environmental factors, and intermediate results containing multi-factor spatial interaction features are output.

[0017] Furthermore, in the aforementioned artificial intelligence-based aquaculture oxygenation monitoring method, the step of inputting the multimodal environmental data into the ST-BiGRU spatiotemporal bidirectional gated loop unit further includes:

[0018] The system employs a BiGRU bidirectional gated cyclic unit structure. The forward GRU traverses the data from the historical starting point to the current moment to capture the upward trend of dissolved oxygen over time, while the backward GRU traverses the data from the current moment to the historical starting point to capture the downward trend.

[0019] By using the reset and update gates of the GRU unit, historical information is filtered out and irrelevant noise is forgotten; the feature weights at the moment of mutation are strengthened through the gating mechanism, and the time-series feature vector containing trend, periodic and mutation information is output.

[0020] Furthermore, in the aforementioned artificial intelligence-based aquaculture oxygenation monitoring method, the dissolved oxygen prediction model is established based on an LSTM (Long Short-Term Memory) network, and the hyperparameters of the model are optimized using an improved IPOS (Indexed Population Swarm Optimization) algorithm to obtain the target dissolved oxygen prediction model, including:

[0021] An initial particle swarm is randomly generated within a preset hyperparameter range. The hyperparameters corresponding to each particle are substituted into the LSTM model. After training the model with the training set, the prediction error of the validation set is calculated as the fitness of the particle.

[0022] Calculate the individual optimal position and global optimal position of each particle, and update the velocity and position of the particles by combining adaptive inertial weights and global optimal guidance. At the same time, new particles are generated through crossover mutation.

[0023] The iteration stops when the number of iterations reaches the preset maximum value. At this point, the combination of hyperparameters corresponding to the globally optimal particle is the optimal solution.

[0024] Furthermore, in the aforementioned artificial intelligence-based aquaculture oxygenation monitoring method, the step of inputting the multi-dimensional feature vector into the target dissolved oxygen prediction model for prediction, and outputting the dissolved oxygen concentration index of the water body within a preset future time period, includes:

[0025] The input layer of the model is used to initially encode the feature vector, and the key temporal information in the feature vector is extracted through the hidden layer. The calculation accuracy is dynamically adjusted according to the length of the preset time period, and the dissolved oxygen concentration index of the water body in the future preset time period is output.

[0026] Furthermore, in the aforementioned artificial intelligence-based aquaculture oxygenation monitoring method, the step of managing oxygenation resources based on the dissolved oxygen concentration index of the water body combined with a reinforcement learning algorithm and a preset dissolved oxygen safety threshold range includes:

[0027] If the dissolved oxygen concentration is predicted to remain above the safe limit in the future, the current oxygenation intensity will be maintained; if the dissolved oxygen concentration is predicted to remain within the safe range in the future, the current oxygenation strategy will be maintained.

[0028] If it is predicted that the dissolved oxygen concentration will drop to near the safe lower limit in the future, oxygenation measures will be initiated in advance; if it is predicted that the dissolved oxygen concentration will drop sharply in the future, the highest level of oxygenation response will be implemented immediately, and a high-level warning will be issued.

[0029] Furthermore, in an artificial intelligence-based aquaculture oxygenation monitoring system, the aquaculture oxygenation monitoring system includes the following modules:

[0030] The multi-dimensional vector extraction module is used to collect multimodal environmental data and input the multimodal environmental data into the ST-BiGRU spatiotemporal bidirectional gated cyclic unit. The spatial channel is used to process grid sensor data, and the temporal channel is used to process historical time series data. A dynamic weight attention mechanism is introduced to adaptively adjust the contribution of different environmental factors to obtain multi-dimensional feature vectors.

[0031] The prediction model building module is used to build a dissolved oxygen prediction model based on LSTM long short-term memory network, and optimize the hyperparameters of the model using an improved IPOS particle swarm optimization algorithm to obtain the target dissolved oxygen prediction model.

[0032] The dissolved oxygen concentration prediction module is used to input the multi-dimensional feature vector into the target dissolved oxygen prediction model for prediction and output the dissolved oxygen concentration index of the water body within a preset time period in the future.

[0033] The oxygenation resource management module is used to manage oxygenation resources based on the dissolved oxygen concentration index of the water body, combined with a reinforcement learning algorithm and a preset dissolved oxygen safety threshold range.

[0034] Furthermore, in an artificial intelligence-based aquaculture oxygenation monitoring system, the dissolved oxygen concentration prediction module includes the following sub-modules:

[0035] The prediction submodule is used to initially encode the feature vector using the input layer of the model, extract key temporal information from the feature vector through the hidden layer, dynamically adjust the calculation accuracy according to the length of the preset time period, and output the dissolved oxygen concentration index of the water body within the preset time period in the future.

[0036] Furthermore, in an artificial intelligence-based aquaculture oxygenation monitoring system, the dissolved oxygen concentration prediction module includes the following sub-modules:

[0037] The management submodule is used to maintain the current oxygenation intensity if it is predicted that the dissolved oxygen concentration will continue to be above the safe limit in the future; and to maintain the current oxygenation strategy if it is predicted that the dissolved oxygen concentration will continue to be within the safe range in the future.

[0038] The early warning submodule is used to initiate oxygenation measures in advance if it is predicted that the dissolved oxygen concentration will drop to near the safe lower limit in the future; if it is predicted that the dissolved oxygen concentration will drop sharply in the future, it will immediately execute the highest level of oxygenation response and issue a high-level early warning.

[0039] Its beneficial effects are as follows: 1. It can plan the activation time, quantity, and intensity of aeration equipment in advance based on future dissolved oxygen trends, ensuring the most rational use of aeration resources. Continuous optimization of resource management effectively reduces power consumption and equipment wear, lowers aquaculture costs, and improves resource utilization efficiency. 2. Aquaculture personnel only need to view relevant information through monitoring equipment, eliminating the need for complex operations and decisions, thus simplifying the aquaculture management process. This intelligent management model improves the efficiency and timeliness of aquaculture management. 3. By introducing a dynamic weighted attention mechanism, an improved IPOS particle swarm optimization algorithm, and reinforcement learning algorithms, a precise dissolved oxygen prediction and intelligent aeration management system is constructed, significantly improving the resource utilization efficiency of aquaculture. This results in a 30% reduction in power consumption, a 40% reduction in equipment wear, an 80% reduction in labor costs, and a 66.7% reduction in fish mortality losses during the aquaculture process. Simultaneously, oxygen utilization rate increases by 20%-25%, feed conversion rate improves by 10%-15%, and overall, aquaculture costs are reduced by more than 40%, providing efficient and sustainable technical support for large-scale, green aquaculture. Attached Figure Description

[0040] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention.

[0041] Figure 1 This is a schematic diagram of the first embodiment of an artificial intelligence-based aquaculture oxygenation monitoring method according to the present invention;

[0042] Figure 2 This is a schematic diagram of a second embodiment of an artificial intelligence-based aquaculture oxygenation monitoring method according to the present invention;

[0043] Figure 3 This is a schematic diagram of the first embodiment of an artificial intelligence-based aquaculture oxygenation monitoring system according to the present invention. Detailed Implementation

[0044] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0045] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0046] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 As shown, an artificial intelligence-based method for monitoring oxygenation in aquaculture includes the following steps:

[0047] Step 101: Collect multimodal environmental data and input the multimodal environmental data into the ST-BiGRU spatiotemporal bidirectional gated cyclic unit. Use the spatial channel to process the grid sensor data and the temporal channel to process the historical time series data. Introduce a dynamic weight attention mechanism to adaptively adjust the contribution of different environmental factors and obtain a multidimensional feature vector.

[0048] Specifically, in this embodiment, dissolved oxygen concentration data, environmental parameters, meteorological information data, and fish image data of the aquaculture water are collected. Outlier detection is performed on the collected data, and the Raida criterion is used to identify and remove outliers to obtain cleaned data.

[0049] Random noise in the sensor-acquired data is processed using wavelet transform denoising method to obtain denoised data;

[0050] The z-score standardization method is used to transform the data into standardized data with a mean of 0 and a standard deviation of 1, thus obtaining multimodal environment data.

[0051] Multimodal environmental data is input into the spatial channel in the form of a three-dimensional matrix. The three-dimensional matrix is ​​then split into multiple two-dimensional grid layers according to the type of environmental factor. Each layer represents the spatial distribution of an environmental factor.

[0052] Multi-layer stacked convolution operations are used to process each grid layer. A 3×3 convolution kernel slides on the grid to calculate the correlation features of sensor data in local areas.

[0053] Global pooling is used to integrate local features into a feature vector that reflects the spatial pattern of the entire aquaculture area; cross-channel convolutional layers are used to calculate the correlation of the spatial distribution of different environmental factors, and intermediate results containing multi-factor spatial interaction features are output.

[0054] The system employs a BiGRU bidirectional gated cyclic unit structure. The forward GRU traverses the data from the historical starting point to the current moment to capture the upward trend of dissolved oxygen over time, while the backward GRU traverses the data from the current moment to the historical starting point to capture the downward trend.

[0055] By using the reset and update gates of the GRU unit, historical information is filtered out and irrelevant noise is forgotten; the feature weights at the moment of mutation are strengthened through the gating mechanism, and the time-series feature vector containing trend, periodic and mutation information is output.

[0056] Specifically:

[0057] (a) Data collection methods;

[0058] Dissolved oxygen concentration data: High-precision dissolved oxygen sensors are used, deployed at different depths and areas within the aquaculture water body to ensure a comprehensive reflection of dissolved oxygen distribution. The sensors collect data every 5 minutes and transmit it in real-time to the data processing center via a wireless transmission module.

[0059] Environmental parameters include water temperature, pH value, salinity, and ammonia nitrogen content. Water temperature is collected by a temperature sensor, pH value is obtained through a pH sensor, and salinity and ammonia nitrogen content are collected by salinity sensors and ammonia nitrogen sensors, respectively. These sensors are also deployed at different locations in the water body, collecting and transmitting data every 10 minutes.

[0060] Meteorological data: Data is collected through weather stations set up around the farm, including temperature, humidity, light intensity, wind speed, and precipitation. The weather stations record data every 30 minutes and upload it to the data processing center in real time.

[0061] Fish school image data: High-definition cameras installed underwater capture images of fish activity. The camera's shooting angle and position are carefully adjusted to ensure clear capture of the fish's condition in different areas. The camera records a 1-minute video every 15 minutes, and the video data is compressed before being transmitted to the data processing center.

[0062] (ii) Data preprocessing;

[0063] Data cleaning: Outlier detection was performed on all types of collected data. The Raida criterion (3σ criterion) was used to identify and remove outliers. For missing values, if the missing percentage was less than 5%, linear interpolation was used to impute them; if the missing percentage was greater than or equal to 5%, all data corresponding to that time period were discarded.

[0064] Data denoising: To address potential random noise in sensor-acquired data, wavelet transform denoising is employed to preserve the main features of the data. For fish school image data, Gaussian filtering is used to remove noise from the image.

[0065] Data standardization: To enable data of different magnitudes to be processed in the same model, all data are standardized. For numerical data, the z-score standardization method is used to convert the data into standardized data with a mean of 0 and a standard deviation of 1; for fish image data, pixel values ​​are normalized to the range [0,1].

[0066] Data fusion: The dissolved oxygen concentration data, environmental parameters, meteorological information data and fish school image data processed above are fused to form multimodal environmental data.

[0067] I. ST-BiGRU Overall Architecture:

[0068] ST-BiGRU (Spatiotemporal Bidirectional Gated Recurrent Unit) is a processing framework specifically designed for the spatiotemporal characteristics of multimodal environmental data. It consists of three parts: a spatial channel module, a temporal channel module, and a dynamic weighted attention fusion module. The multimodal environmental data is first split into spatially distributed data (grid data such as dissolved oxygen and water temperature from different sensor nodes) and temporally varied data (historical dissolved oxygen time series and meteorological data sequences from a single node). These are input into two channels for feature extraction, and finally fused into a unified multidimensional feature vector through an attention mechanism.

[0069] II. Spatial channel processing of grid sensor data:

[0070] (a) Data input format;

[0071] Multimodal environmental data is input into the spatial channel in the form of a three-dimensional matrix, with the matrix dimensions being (number of sensor grid rows × number of sensor grid columns × number of environmental factors). The number of sensor grid rows and columns corresponds to the actual sensor deployment grid in the aquaculture area (a 10×10 grid covers the entire aquaculture pond). The number of environmental factors includes preprocessed numerical parameters such as dissolved oxygen, water temperature, and pH, as well as key indicators (fish aggregation density, activity intensity, etc.) extracted from the fish school images.

[0072] (II) Spatial Feature Extraction Process;

[0073] Grid data hierarchical mapping: The three-dimensional matrix is ​​split into multiple two-dimensional grid layers according to the type of environmental factor (separate dissolved oxygen grid layer, water temperature grid layer, etc.), and each layer represents the spatial distribution of an environmental factor.

[0074] Local spatial correlation capture: Multi-layer stacked convolution operations are used to process each grid layer. By sliding 3×3 or 5×5 convolution kernels on the grid, correlation features of sensor data within local regions are calculated (such as dissolved oxygen gradients and the continuity of water temperature distribution among three adjacent sensors). Spatial location information is preserved after each round of convolution while compressing the data dimensionality.

[0075] Global spatial pattern integration: After multiple convolutions, global pooling is used (taking the maximum or average value of each feature map) to integrate local features into a feature vector that reflects the spatial pattern of the entire aquaculture water area (the overall uniformity of dissolved oxygen distribution, the spatial proportion of high temperature areas, etc.).

[0076] Cross-factor spatial correlation mining: Through cross-channel convolutional layers, the correlation of spatial distribution of different environmental factors is calculated (the spatial correspondence between high water temperature and low dissolved oxygen in a certain area), and intermediate results containing multi-factor spatial interaction features are output.

[0077] III. Time Channel Processing of Historical Time-Series Data:

[0078] (a) Data input format;

[0079] The historical time series data received by the time channel is a multivariate time series. Each variable corresponds to a historical monitoring sequence of an environmental factor (dissolved oxygen data recorded every 10 minutes in the past 24 hours, daily precipitation sequence in the past 7 days, etc.). The data format is a two-dimensional matrix of (time steps × number of environmental factors). The number of time steps is determined according to the historical data span (usually 24-168 steps).

[0080] (II) Temporal Feature Extraction Process;

[0081] Bidirectional temporal traversal: A bidirectional gated cyclic unit (BiGRU) structure is adopted. The forward GRU traverses the data from the historical starting point to the current time to capture the upward trend of dissolved oxygen over time (the increase in dissolved oxygen due to enhanced photosynthesis after sunrise); the backward GRU traverses the data from the current time to the historical starting point in reverse to capture the downward trend (the decrease in dissolved oxygen due to nighttime respiration).

[0082] Capture of both short-term and long-term dependence: Through the reset and update gates of the GRU unit, important historical information (dissolved oxygen change patterns under similar weather conditions the previous day) is automatically filtered and irrelevant noise (instantaneous fluctuations of the sensor) is forgotten. It is particularly good at capturing periodic features across days and weeks (the diurnal rhythm of dissolved oxygen and the continuous impact of rainy weather).

[0083] Temporal mutation feature extraction: For sudden drops in dissolved oxygen (surge in oxygen consumption caused by a sudden increase in fish density), a gating mechanism is used to strengthen the feature weights at the mutation moment, and output a temporal feature vector containing trend, periodic and mutation information.

[0084] IV. Operation of the Dynamic Weighted Attention Mechanism:

[0085] (a) Basis for weight calculation;

[0086] The dynamic weighted attention mechanism focuses on the "real-time impact of environmental factors on dissolved oxygen" and dynamically allocates weights based on the following criteria:

[0087] Factor correlation: Calculate the real-time Pearson correlation coefficient between each environmental factor and dissolved oxygen (water temperature has a stronger negative correlation with dissolved oxygen in summer).

[0088] Time sensitivity: assign higher weight to factors that have changed drastically in the near future (such as precipitation during sudden downpours).

[0089] Spatial saliency: Increase the weight of sensor data in the region with the largest dissolved oxygen gradient (the difference between the edge and center of the aquaculture pond).

[0090] Fish school correlation: Combined with fish school image features (when fish rise to breathe, the weight of fish school activity intensity increases significantly).

[0091] V. Output Results:

[0092] The final output multi-dimensional feature vector includes both the spatial distribution characteristics of each area of ​​the aquaculture water (which corner has low dissolved oxygen) and the trend of change in the time dimension (whether it will fall below the threshold in the next few hours). At the same time, dynamic weights highlight the most critical influencing factors (whether the current dominant factor is light or fish density), providing high-value input for subsequent dissolved oxygen prediction models.

[0093] Step 102: Establish a dissolved oxygen prediction model based on LSTM long short-term memory network, and optimize the hyperparameters of the model using the improved IPOS particle swarm optimization algorithm to obtain the target dissolved oxygen prediction model.

[0094] Specifically, in this embodiment, an initial particle swarm is randomly generated within a preset hyperparameter range. The hyperparameters corresponding to each particle are substituted into the LSTM model. After training the model with the training set, the prediction error of the validation set is calculated as the fitness of the particle.

[0095] Calculate the individual optimal position and global optimal position of each particle, and update the velocity and position of the particles by combining adaptive inertial weights and global optimal guidance. At the same time, new particles are generated through crossover mutation.

[0096] The iteration stops when the number of iterations reaches the preset maximum value. At this point, the combination of hyperparameters corresponding to the globally optimal particle is the optimal solution.

[0097] I. Construction of LSTM Long Short-Term Memory Network Model:

[0098] (a) Overall model architecture design;

[0099] The LSTM dissolved oxygen prediction model adopts a three-layer architecture of "input layer-hidden layer-output layer", which is specifically designed for the long-term dependency characteristics of time series data. It can effectively avoid the "gradient vanishing" problem of traditional recurrent neural networks and accurately capture the dynamic change pattern of dissolved oxygen concentration over time.

[0100] (II) Structure and function of each layer;

[0101] Input layer: Receives the multi-dimensional feature vector output from step 2. This vector includes spatial distribution features (environmental differences in different aquaculture areas), temporal trend features (diurnal variation of dissolved oxygen), and dynamic weights of key influencing factors (real-time impact of water temperature and fish density). The input layer transforms the feature vector into a format suitable for LSTM processing, ensuring that the data dimensionality matches the number of hidden layer units.

[0102] Hidden Layer: Composed of multiple LSTM units, the number of which is dynamically adjusted according to the complexity of the aquaculture scenario (typically 32-128). Each LSTM unit filters and stores information through memory gates, input gates, and forget gates: the memory gate retains key historical information (dissolved oxygen change patterns at the same time point in the previous 3 days), the input gate controls the inclusion of new information (abrupt changes in current weather conditions), and the forget gate filters out irrelevant noise (instantaneous fluctuation data from sensors). Through the synergistic effect of multiple hidden units, the model can deeply mine the complex temporal correlations contained in the feature vectors, such as the cumulative decline trend of dissolved oxygen under continuous rainy weather.

[0103] Output layer: This layer transforms the features processed by the hidden layer into specific predicted dissolved oxygen concentration values, outputting a numerical sequence for a preset time period (the predicted value for each hour within the next 6 hours). The output layer uses a linear activation function to ensure that the prediction results are within a reasonable numerical range and consistent with the magnitude of the actual dissolved oxygen concentration.

[0104] (III) Model training process;

[0105] Data partitioning: The preprocessed historical data is divided into training set, validation set and test set in a ratio of 7:2:1.

[0106] Loss function: The mean square error between the predicted value and the actual dissolved oxygen concentration is used as the loss function to measure the prediction accuracy and guide the model to optimize in the direction of minimizing the error.

[0107] II. Hyperparameter Optimization of the Improved IPOS Particle Swarm Optimization Algorithm:

[0108] (a) Optimization objective and hyperparameter selection;

[0109] The optimization objective is to find the hyperparameter combination that minimizes the prediction error of the LSTM model. The hyperparameters to be optimized include:

[0110] Structural parameters: number of LSTM hidden layers (1-3 layers), number of cells per layer (32-128);

[0111] Training parameters: learning rate (0.001-0.1), training batch size (32-128), maximum training epochs (50-200);

[0112] Regularization parameter: L2 regularization coefficient (0.0001-0.01) to prevent overfitting.

[0113] (ii) Advantages of the improved IPOS particle swarm algorithm;

[0114] Compared to the traditional particle swarm optimization algorithm, the improved IPOS algorithm has been optimized in the following aspects:

[0115] Adaptive inertia weights: The inertia weights are dynamically adjusted according to the algorithm iteration stages. In the early stages, larger weights are used to enhance the global search capability (exploring a wider hyperparameter space), and in the later stages, the weights are reduced to enhance the local search capability (fine-tuning the current high-quality parameters), thus balancing search efficiency and accuracy.

[0116] Crossover and mutation mechanism: The crossover operation of the genetic algorithm is introduced to combine the parameters of two better particles to generate new particles, increasing parameter diversity; at the same time, a small probability mutation is set to randomly change a certain parameter of the particle to avoid the algorithm getting stuck in a local optimum (suboptimal result caused by improper selection of initial parameters).

[0117] Elite retention strategy: After each iteration, the top 5% of particles are retained and directly enter the next iteration to ensure that high-quality parameters are not discarded and accelerate the convergence process.

[0118] (iii) Optimize processes;

[0119] Initialize the particle swarm: Treat each hyperparameter combination as a "particle", and randomly generate 50-100 particles within the preset hyperparameter range to form the initial particle swarm.

[0120] Fitness evaluation: Substitute the hyperparameters corresponding to each particle into the LSTM model, train the model with the training set, and calculate the prediction error (mean squared error) of the validation set. This error is used as the fitness of the particle (the smaller the error, the higher the fitness).

[0121] Particle Update: Based on the particle's fitness, the particle position (i.e., the hyperparameter combination) is updated through the following steps:

[0122] Calculate the individual optimal position (the parameter corresponding to its own historical best fitness) and the global optimal position (the parameter corresponding to the historical best fitness of the entire particle swarm) for each particle.

[0123] By combining adaptive inertia weights and guidance from individual and global optima, the velocity and position of particles are updated, while new particles are generated through cross-mutation.

[0124] Iteration Termination: When the number of iterations reaches the preset maximum value (usually 30-50 rounds), or when the global optimal fitness does not improve significantly after 5 consecutive iterations, the iteration stops. The hyperparameter combination corresponding to the global optimal particle at this time is the optimal solution.

[0125] Step 103: Input the multi-dimensional feature vector into the target dissolved oxygen prediction model for prediction, and output the dissolved oxygen concentration index of the water body in the future preset time period;

[0126] Specifically, in this embodiment, the input layer of the model is used to initially encode the feature vector, and the key temporal information in the feature vector is extracted through the hidden layer. The calculation accuracy is dynamically adjusted according to the length of the preset time period, and the dissolved oxygen concentration index of the water body in the future preset time period is output.

[0127] Specifically, after receiving the multi-dimensional feature vectors input, the target dissolved oxygen prediction model initiates a hierarchical parsing mechanism. First, the model's input layer performs preliminary encoding on the feature vectors, separating spatial distribution features (dissolved oxygen gradients in different areas of the aquaculture pond), temporal trend features (dissolved oxygen fluctuation patterns over the past 24 hours), and dynamic weight information (the influence weight of current water temperature on dissolved oxygen), and directing them to the corresponding processing modules.

[0128] During internal model computation, the optimized LSTM hidden layer prioritizes extracting key temporal information from the feature vectors. In the prediction process, the model dynamically adjusts its computational precision based on the length of the preset time period. If the preset time period is 1 hour (short-term prediction), it focuses on minute-level subtle changes, such as dissolved oxygen fluctuations every 10 minutes; if it is 24 hours (long-term prediction), it emphasizes capturing hourly-level trend changes.

[0129] III. Output Format and Content:

[0130] After the prediction is completed, the system outputs the dissolved oxygen concentration index of the water body for the preset time period in the future. The results are presented in two forms: a time-concentration series table and a trend curve. The series table is accurate to the hourly predicted dissolved oxygen concentration value (in mg / L). For example, the output for the next 6 hours may be: 6.2 mg / L after 1 hour, 5.8 mg / L after 2 hours, 5.5 mg / L after 3 hours, etc., which makes it convenient for aquaculture personnel to quickly obtain specific values. The trend curve visually shows the overall trend of dissolved oxygen change and marks possible key nodes.

[0131] Step 104: Manage oxygenation resources based on the dissolved oxygen concentration index of the water body, combined with reinforcement learning algorithms and preset dissolved oxygen safety threshold range.

[0132] Specifically, in this embodiment, if it is predicted that the dissolved oxygen concentration will continue to be higher than the safe limit in the future, the current oxygenation intensity will be maintained; if it is predicted that the dissolved oxygen concentration will continue to be within the safe range in the future, the current oxygenation strategy will be maintained.

[0133] If it is predicted that the dissolved oxygen concentration will drop to near the safe lower limit in the future, oxygenation measures will be initiated in advance; if it is predicted that the dissolved oxygen concentration will drop sharply in the future, the highest level of oxygenation response will be implemented immediately, and a high-level warning will be issued.

[0134] Specifically, if it is predicted that the dissolved oxygen concentration will continue to be higher than the safe limit in the future, the current oxygenation intensity should be maintained or reduced (the oxygenator power should be turned off or reduced) to avoid excessive oxygenation and wasting energy.

[0135] If the dissolved oxygen concentration is predicted to remain within a safe range in the future, the current oxygenation strategy will be maintained.

[0136] If it is predicted that the dissolved oxygen concentration will drop to near or below the safe limit in the future (even if the current DO value is still within the safe range), then start or enhance oxygenation measures in advance (turn on the aerator in advance, increase the power of the aerator).

[0137] If a sharp drop in dissolved oxygen concentration is predicted in the future (a steep drop trend is predicted), the highest level of oxygenation response will be immediately executed (all available oxygenation equipment will be turned on at full power) and a high-level warning will be issued.

[0138] Application of Reinforcement Learning Algorithms: Reinforcement learning algorithms learn optimal decision-making strategies through interaction with the environment. Using aeration as the action and maintaining the dissolved oxygen concentration (DOC) within a safe threshold range as the reward, a reinforcement learning model is constructed. Based on the DOC prediction model's output for a predetermined future time period, and combined with the reinforcement learning algorithm, an aeration resource management strategy is formulated. This strategy includes the activation time, number of aeration devices, and aeration intensity, aiming to achieve precise and efficient aeration management, reduce aquaculture costs, and improve aquaculture efficiency.

[0139] Its beneficial effects are as follows: 1. It can plan the activation time, quantity, and intensity of aeration equipment in advance based on future dissolved oxygen trends, ensuring the most rational use of aeration resources. Continuous optimization of resource management effectively reduces power consumption and equipment wear, lowers aquaculture costs, and improves resource utilization efficiency. 2. Aquaculture personnel only need to view relevant information through monitoring equipment, eliminating the need for complex operations and decisions, thus simplifying the aquaculture management process. This intelligent management model improves the efficiency and timeliness of aquaculture management. 3. By introducing a dynamic weighted attention mechanism, an improved IPOS particle swarm optimization algorithm, and reinforcement learning algorithms, a precise dissolved oxygen prediction and intelligent aeration management system is constructed, significantly improving the resource utilization efficiency of aquaculture. This results in a 30% reduction in power consumption, a 40% reduction in equipment wear, an 80% reduction in labor costs, and a 66.7% reduction in fish mortality losses during the aquaculture process. Simultaneously, oxygen utilization rate increases by 20%-25%, feed conversion rate improves by 10%-15%, and overall, aquaculture costs are reduced by more than 40%, providing efficient and sustainable technical support for large-scale, green aquaculture.

[0140] Please see Figure 2 In an AI-based aquaculture oxygenation monitoring method, inputting multimodal environmental data into the ST-BiGRU spatiotemporal bidirectional gated loop unit includes the following steps:

[0141] Step 201: Input the multimodal environmental data into the spatial channel in the form of a three-dimensional matrix, and split the three-dimensional matrix into multiple two-dimensional grid layers according to the type of environmental factor. Each layer represents the spatial distribution of an environmental factor.

[0142] Step 202: Process each grid layer using multi-layer stacked convolution operations. Calculate the correlation features of sensor data in local areas by sliding a 3×3 convolution kernel across the grid.

[0143] Step 203: Use global pooling to integrate local features into a feature vector that reflects the spatial pattern of the entire aquaculture water area; calculate the correlation of spatial distribution of different environmental factors through cross-channel convolutional layers, and output intermediate results containing multi-factor spatial interaction features.

[0144] The above describes an embodiment of an artificial intelligence-based aquaculture oxygenation monitoring method of the present invention. Please refer to [link / reference]. Figure 3 In an artificial intelligence-based aquaculture oxygenation monitoring system, the aquaculture oxygenation monitoring system includes the following modules:

[0145] The multi-dimensional vector extraction module is used to collect multimodal environmental data. The multimodal environmental data is input into the ST-BiGRU spatiotemporal bidirectional gated cyclic unit. The spatial channel is used to process grid sensor data, and the temporal channel is used to process historical time series data. A dynamic weight attention mechanism is introduced to adaptively adjust the contribution of different environmental factors to obtain multi-dimensional feature vectors.

[0146] The prediction model building module is used to build a dissolved oxygen prediction model based on LSTM long short-term memory network, and optimize the hyperparameters of the model using an improved IPOS particle swarm optimization algorithm to obtain the target dissolved oxygen prediction model.

[0147] The dissolved oxygen concentration prediction module is used to input multi-dimensional feature vectors into the target dissolved oxygen prediction model for prediction and output the dissolved oxygen concentration index of the water body within a preset time period in the future.

[0148] The oxygenation resource management module is used to manage oxygenation resources based on the dissolved oxygen concentration index of the water body, combined with reinforcement learning algorithms and preset dissolved oxygen safety threshold ranges.

[0149] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for monitoring oxygenation in aquaculture based on artificial intelligence, characterized in that, The aquaculture oxygenation monitoring method includes the following steps: Multimodal environmental data is collected and input into the spatial channel in the form of a three-dimensional matrix. The three-dimensional matrix is ​​then split into multiple two-dimensional grid layers according to the type of environmental factor, with each layer representing the spatial distribution of one environmental factor. Multi-layered stacked convolutional operations are used to process each grid layer. By sliding a 3×3 convolutional kernel across the grid, the correlation features of sensor data within a local area are calculated. Global pooling is used to integrate the local features into a feature vector reflecting the spatial pattern of the entire aquaculture area. The correlation between the spatial distributions of different environmental factors is calculated through cross-channel convolutional layers, outputting intermediate results containing multi-factor spatial interaction features. A Bi-channel convolutional algorithm is employed. The GRU bidirectional gated cyclic cell structure uses a forward GRU to traverse data from the historical starting point to the current time to capture the upward trend of dissolved oxygen over time, and a backward GRU to traverse data from the current time to the historical starting point to capture the downward trend. Through the reset and update gates of the GRU cells, historical information is filtered and irrelevant noise is discarded. The gating mechanism strengthens the feature weights at abrupt changes, outputting a temporal feature vector containing trend, periodic, and abrupt change information. The spatial channel processes grid sensor data, and the temporal channel processes historical time-series data. A dynamic weighted attention mechanism is introduced to adaptively adjust the contribution of different environmental factors, resulting in a multi-dimensional feature vector. A dissolved oxygen prediction model is established based on LSTM (Long Short-Term Memory) network, and the hyperparameters of the model are optimized using an improved IPOS (Integrated Population Swarm Optimization) algorithm to obtain the target dissolved oxygen prediction model. The multi-dimensional feature vector is input into the target dissolved oxygen prediction model for prediction, and the dissolved oxygen concentration index of the water body is output within a preset time period in the future. Oxygenation resource management is carried out based on the dissolved oxygen concentration index of the water body, combined with reinforcement learning algorithm and preset dissolved oxygen safety threshold range. If it is predicted that the dissolved oxygen concentration will continue to be higher than the safety upper limit in the future, the current oxygenation intensity will be maintained; if it is predicted that the dissolved oxygen concentration will continue to be within the safety range in the future, the current oxygenation strategy will be maintained; if it is predicted that the dissolved oxygen concentration will drop to close to the safety lower limit in the future, oxygenation measures will be initiated in advance; if it is predicted that the dissolved oxygen concentration will drop sharply in the future, the highest level of oxygenation response will be executed immediately and a high-level warning will be issued.

2. The aquaculture oxygenation monitoring method based on artificial intelligence as described in claim 1, characterized in that, The collection of multimodal environmental data includes: Data on dissolved oxygen concentration, environmental parameters, meteorological information, and fish images of the aquaculture water were collected. Outlier detection was performed on the collected data, and the Laida criterion was used to identify and remove outliers to obtain cleaned data. Random noise in the sensor-acquired data is processed using wavelet transform denoising method to obtain denoised data; The z-score standardization method is used to transform the data into standardized data with a mean of 0 and a standard deviation of 1, thus obtaining multimodal environment data.

3. The aquaculture oxygenation monitoring method based on artificial intelligence as described in claim 1, characterized in that, The dissolved oxygen prediction model is established based on an LSTM (Long Short-Term Memory) network. The hyperparameters of the model are optimized using an improved IPOS (Index-Pool Swarm Optimization) algorithm to obtain the target dissolved oxygen prediction model, which includes: An initial particle swarm is randomly generated within a preset hyperparameter range. The hyperparameters corresponding to each particle are substituted into the LSTM model. After training the model with the training set, the prediction error of the validation set is calculated as the fitness of the particle. Calculate the individual optimal position and global optimal position of each particle, and update the velocity and position of the particles by combining adaptive inertial weights and global optimal guidance. At the same time, new particles are generated through crossover mutation. The iteration stops when the number of iterations reaches the preset maximum value. At this point, the combination of hyperparameters corresponding to the globally optimal particle is the optimal solution.

4. The aquaculture oxygenation monitoring method based on artificial intelligence as described in claim 1, characterized in that, The step of inputting the multi-dimensional feature vector into the target dissolved oxygen prediction model for prediction and outputting the dissolved oxygen concentration index of the water body within a preset future time period includes: The input layer of the model is used to initially encode the feature vector, and the key temporal information in the feature vector is extracted through the hidden layer. The calculation accuracy is dynamically adjusted according to the length of the preset time period, and the dissolved oxygen concentration index of the water body in the future preset time period is output.

5. An artificial intelligence-based aquaculture oxygenation monitoring system, characterized in that, The aquaculture oxygenation monitoring system includes the following steps: A multi-dimensional vector extraction module is used to collect multimodal environmental data. This data is input into the spatial channel as a three-dimensional matrix. The three-dimensional matrix is ​​then split into multiple two-dimensional grid layers according to environmental factor types, with each layer representing the spatial distribution of a particular environmental factor. Multi-layered stacked convolutional operations are used to process each grid layer. A 3×3 convolution kernel slides across the grid to calculate the correlation features of sensor data within local areas. Global pooling is used to integrate these local features into a feature vector reflecting the spatial patterns of the entire aquaculture area. Finally, cross-channel convolutional layers are used to calculate the correlations between the spatial distributions of different environmental factors, outputting an intermediate structure containing multi-factor spatial interaction features. Results: A BiGRU bidirectional gated cyclic unit structure is adopted. The forward GRU traverses the data from the historical starting point to the current time to capture the upward trend of dissolved oxygen over time, while the backward GRU traverses the data from the current time to the historical starting point to capture the downward trend. Through the reset gate and update gate of the GRU unit, historical information is filtered and irrelevant noise is forgotten. The feature weights at abrupt change times are strengthened through the gating mechanism, and the output time-series feature vector contains trend, periodic and abrupt change information. The spatial channel is used to process grid sensor data, and the temporal channel is used to process historical time-series data. A dynamic weight attention mechanism is introduced to adaptively adjust the contribution of different environmental factors to obtain a multi-dimensional feature vector. The prediction model building module is used to build a dissolved oxygen prediction model based on LSTM long short-term memory network, and optimize the hyperparameters of the model using an improved IPOS particle swarm optimization algorithm to obtain the target dissolved oxygen prediction model. The dissolved oxygen concentration prediction module is used to input the multi-dimensional feature vector into the target dissolved oxygen prediction model for prediction and output the dissolved oxygen concentration index of the water body within a preset time period in the future. The oxygenation resource management module is used to manage oxygenation resources based on the dissolved oxygen concentration index of the water body, combined with a reinforcement learning algorithm and a preset dissolved oxygen safety threshold range. If it is predicted that the dissolved oxygen concentration will continue to be higher than the safety upper limit in the future, the current oxygenation intensity will be maintained; if it is predicted that the dissolved oxygen concentration will continue to be within the safety range in the future, the current oxygenation strategy will be maintained; if it is predicted that the dissolved oxygen concentration will drop to near the safety lower limit in the future, oxygenation measures will be initiated in advance; if it is predicted that the dissolved oxygen concentration will drop sharply in the future, the highest level of oxygenation response will be executed immediately, and a high-level warning will be issued.

6. The aquaculture oxygenation monitoring system based on artificial intelligence as described in claim 5, characterized in that, The dissolved oxygen concentration prediction module includes the following sub-modules: The prediction submodule is used to initially encode the feature vector using the input layer of the model, extract key temporal information from the feature vector through the hidden layer, dynamically adjust the calculation accuracy according to the length of the preset time period, and output the dissolved oxygen concentration index of the water body within the preset time period in the future.

7. The aquaculture oxygenation monitoring system based on artificial intelligence as described in claim 5, characterized in that, The dissolved oxygen concentration prediction module includes the following sub-modules: The management submodule is used to maintain the current oxygenation intensity if it is predicted that the dissolved oxygen concentration will continue to be above the safe limit in the future; and to maintain the current oxygenation strategy if it is predicted that the dissolved oxygen concentration will continue to be within the safe range in the future. The early warning submodule is used to initiate oxygenation measures in advance if it is predicted that the dissolved oxygen concentration will drop to near the safe lower limit in the future; if it is predicted that the dissolved oxygen concentration will drop sharply in the future, it will immediately execute the highest level of oxygenation response and issue a high-level early warning.

Images