Intelligent breeding monitoring method and system based on big data analysis

Abstract

This invention relates to the fields of smart aquaculture and big data analysis technology, and discloses a smart aquaculture monitoring method and system based on big data analysis. The method includes periodically acquiring raw data from the farm's sensor network, and constructing a multi-dimensional state image through spatiotemporal synchronization and normalization processing. This image is input into a pre-trained behavior-environment decoupling analysis network, and independent behavioral feature tensors and independent environmental background field tensors are extracted in parallel. Based on these two, a coupling degree index is calculated through a cross-influence evaluation matrix to generate a comprehensive anomaly score map. According to this map, a strategy generation engine dynamically synthesizes a set of control instructions including directional airflow guidance, refined feeding, and coordinated temperature and humidity regulation. This method achieves deep separation and correlation analysis between animal behavior and aquaculture environmental factors, accurately tracing the root cause of anomalies and realizing early warning and precise systemic control.

Description

Technical Field

[0001] This invention relates to the field of smart farming and big data analysis technology, specifically to a smart farming monitoring method and system based on big data analysis. Background Technology

[0002] Modern large-scale farms widely deploy sensor networks to monitor environmental parameters and animal behavior. Existing solutions primarily rely on setting fixed thresholds for various data types to trigger independent alarms, or using statistical models to analyze trends in single data streams. These methods typically treat environmental data such as temperature and humidity, and gas concentrations, and animal behavioral data such as activity levels and feeding frequency as parallel but separate monitoring indicators, with their correlations analyzed only through simple correlation analysis. The resulting control commands often trigger pre-set, standardized equipment actions in response to an anomaly in a single, isolated parameter.

[0003] These technical solutions have shortcomings. They cannot effectively distinguish whether abnormal behavior in animal groups stems from internal health problems or external environmental stress. When environmental factors and animal behavior are intertwined, the system struggles to trace the root cause of the abnormality, easily leading to misjudgments. This results in delayed and ineffective management measures, often resorting to passive regulation only after the problem becomes apparent, failing to provide early warning and lacking precision in intervention.

[0004] Current technology requires an analytical method capable of deeply separating animal behavioral characteristics from their breeding environment context. This method must not only independently characterize the inherent patterns of both, but also precisely quantify their dynamic and complex coupling relationships and the intensity of their interactions. Overcoming the bottleneck of unclear interplay between behavioral and environmental factors and ambiguous correlation mechanisms is key to moving from passive response to proactive, coordinated regulation. Summary of the Invention

[0005] The purpose of this invention is to provide a smart aquaculture monitoring method and system based on big data analysis to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides a smart aquaculture monitoring method based on big data analysis, the method comprising:

[0007] Raw monitoring data is periodically acquired from a sensor network deployed within the farm.

[0008] The original monitoring data is subjected to spatiotemporal field synchronization and normalization processing to construct a multidimensional state image of the farm with a unified spatiotemporal reference.

[0009] The multidimensional state image of the farm is input into a behavior-environment decoupling analysis network pre-trained with massive historical farming data for deep pattern mining. The behavior-environment decoupling analysis network performs animal behavior pattern stripping and environmental background field extraction in parallel, and outputs independent behavioral feature tensors and independent environmental background field tensors.

[0010] Based on the behavioral feature tensor and the environmental background field tensor, a coupling degree index between the behavioral feature tensor and the environmental background field tensor is calculated through a preset cross-influence evaluation matrix, and a comprehensive anomaly score map is generated based on the coupling degree index.

[0011] Based on the comprehensive anomaly score map, a set of control instructions is dynamically synthesized through a strategy generation engine. The set of control instructions includes at least directional airflow guidance parameters for the ventilation system, refined feeding formulas and schedules for the feeding device, and temperature and humidity coordinated adjustment instructions for the environmental control unit.

[0012] Preferably, the step of performing spatiotemporal field synchronization and normalization processing on the original monitoring data to construct a multidimensional state image of the farm with a unified spatiotemporal reference specifically includes:

[0013] The raw monitoring data includes temperature profiles generated by a thermal imaging sensor array, feed consumption rate curves acquired by a weight sensor, animal sound signal waveforms captured by a microphone array, and gas concentration distribution maps fed back in real time by an ammonia nitrogen sensor network.

[0014] A unified timestamp sequence and three-dimensional spatial coordinate information are added to the temperature profile, the feed consumption rate curve, the animal sound signal waveform, and the gas concentration distribution map, respectively.

[0015] The temperature profile image with added spatiotemporal information is subjected to hot spot region segmentation and contour extraction to generate a standardized temperature field layer;

[0016] Consumption events are detected and cumulative amounts are integrated on the feed consumption rate curve with added spatiotemporal information to generate a standardized feed consumption field layer.

[0017] The animal sound signal waveform with added spatiotemporal information is subjected to frequency domain transformation and feature event labeling to generate a standardized sound event field layer;

[0018] Concentration gradient calculation and high-concentration region clustering are performed on the gas concentration distribution map with added spatiotemporal information to generate a standardized gas concentration field layer;

[0019] The standardized temperature field layer, feed consumption field layer, sound event field layer, and gas concentration field layer are aligned and fused pixel-level according to their timestamp sequence and three-dimensional spatial coordinate information to generate a multi-dimensional state image of the farm with the unified spatiotemporal reference.

[0020] The multidimensional state image of the farm is a digital twin model that integrates temperature distribution, feed consumption dynamics, sound event markers, and gas concentration gradients.

[0021] Preferably, the step of performing frequency domain transformation and feature event labeling on the animal sound signal waveform with added spatiotemporal information to generate a standardized sound event field layer includes:

[0022] Perform a short-time Fourier transform on the waveform of the animal sound signal to obtain a time-spectrum diagram;

[0023] The temporal spectrogram is template-matched using a predefined vocal template library to identify a variety of predefined acoustic events, including feeding calls, stress calls, and pathological coughs.

[0024] Assign an event type label, confidence score, and sound source intensity estimate to each identified acoustic event;

[0025] Based on the three-dimensional spatial coordinate information corresponding to the acoustic event, the event type label, confidence score and sound source intensity estimate are mapped to a three-dimensional spatial grid to form a discrete sound event point cloud;

[0026] Spatial interpolation and smoothing are performed on the sound event point cloud to generate a continuously distributed standardized sound event field layer, wherein each pixel value in the layer represents a composite index of the intensity and frequency of a predefined acoustic event occurring at a spatial location within a specific time window.

[0027] Preferably, the step of inputting the multidimensional state image of the farm into a behavior-environment decoupling analysis network pre-trained with massive historical farming data for deep pattern mining, wherein the behavior-environment decoupling analysis network performs animal behavior pattern stripping and environmental background field extraction in parallel, and outputs independent behavioral feature tensors and independent environmental background field tensors, including:

[0028] The behavior-environment decoupling analysis network includes a shared feature encoder, a behavior feature decoding branch, and an environment background field decoding branch;

[0029] The shared feature encoder performs multi-layer convolutional downsampling on the input multi-dimensional state image of the aquaculture farm to extract a shared feature map containing multi-scale information;

[0030] The behavioral feature decoding branch receives the shared feature map and reconstructs the behavioral feature tensor, which mainly includes the animal group's movement trajectory, the frequency of interaction between individuals, and the periodicity of feeding and drinking behavior, by focusing on dynamic change patterns directly related to animal activities.

[0031] The environmental background field decoding branch receives the shared feature map and, by focusing on the relatively static physical field distribution and slow changing trend, reconstructs the environmental background field tensor, which mainly includes the basic temperature distribution, the reference humidity level, the background noise level, and the initial gas concentration.

[0032] During the training phase, an adversarial loss function is introduced to force the behavioral feature tensor to exclude environmental background information, and at the same time, to force the environmental background field tensor to exclude animal behavior information.

[0033] Preferably, the step of calculating the coupling index between the behavioral feature tensor and the environmental background field tensor using a preset cross-influence evaluation matrix, based on the behavioral feature tensor and the environmental background field tensor, includes:

[0034] The cross-influence evaluation matrix defines the theoretical correlation strength weights between different categories of behavioral patterns and different types of environmental background factors;

[0035] Projecting the behavior feature tensor onto the behavior pattern category dimension yields a behavior pattern probability distribution vector.

[0036] Projecting the environmental background field tensor onto the environmental background factor type dimension yields the environmental factor intensity distribution vector.

[0037] The dot product of the probability distribution vector of the behavior pattern and the intensity distribution vector of the environmental factors under the cross-influence evaluation matrix is ​​calculated to obtain the coupling index, which quantifies the extent to which the currently observed animal behavior can be explained by the current environmental background conditions.

[0038] When the coupling degree index is lower than the preset coupling degree threshold, it is determined that there is an anomaly. The anomaly indicates that the animal’s behavior deviates from the expected pattern in the current environment.

[0039] Preferably, generating a comprehensive anomaly scoring map based on the coupling degree index includes:

[0040] Obtain a reconstruction error map generated by the behavior-environment decoupling analysis network. The reconstruction error map characterizes the difference between the original multidimensional state image of the farm and the image reconstructed from the behavior feature tensor and the environmental background field tensor.

[0041] The coupling index is spatially extended to generate a coupling distribution map with the same spatial resolution as the reconstruction error map;

[0042] The reconstruction error map and the coupling degree distribution map are multiplied element by element to obtain a preliminary heatmap of abnormal attention.

[0043] A spatial prior weight map is introduced, which is pre-set based on the historical anomaly records of the farm, and is used to weight and correct the preliminary anomaly attention heat map.

[0044] The weighted and corrected heatmap of abnormal attention is normalized and nonlinearly stretched to generate the final comprehensive anomaly score map, where the score value of each pixel represents the probability of an anomaly occurring at the spatial location.

[0045] Preferably, based on the comprehensive anomaly scoring map, a set of control instructions is dynamically synthesized through a strategy generation engine, including:

[0046] The comprehensive anomaly scoring map marks the degree of anomaly at each spatial location within the farm in the time dimension;

[0047] Connectivity analysis is performed on the comprehensive anomaly score map to identify anomaly regions whose scores exceed the activation threshold;

[0048] For each of the aforementioned abnormal regions, extract its geometric center coordinates, spatial extent, average abnormal score, and score trend over time.

[0049] The feature information of the abnormal region block is input into a case-based reasoning control strategy library for matching and retrieval. The control strategy library stores the features of historical successful control cases and the corresponding control parameters.

[0050] Based on the matched historical cases, preliminary candidate control instructions for the abnormal area blocks are generated, including candidate ventilation parameters, candidate feeding adjustment amounts, and candidate temperature and humidity setpoints.

[0051] Based on the overall operating status and resource constraints of the current farm, the candidate control instructions are optimized and conflict resolved through multi-objective optimization to generate the final executable set of control instructions.

[0052] Preferably, the multi-objective optimization and conflict resolution of the candidate control commands includes:

[0053] Establish a multi-objective optimization function that includes animal welfare indicators, energy consumption indicators, feed conversion rate indicators, and equipment wear indicators;

[0054] Obtain the real-time operating status and capacity limits of all environmental control and feeding equipment in the current farm, and use them as constraints for the optimization problem.

[0055] A multi-objective evolutionary algorithm is used to adjust the candidate control commands, and under the premise of satisfying the constraints, the combination of control parameters that makes the multi-objective optimization function Pareto optimal is found.

[0056] The optimized control commands for different abnormal regions are coordinated globally to avoid conflicts in resource allocation or environmental field control between different regions.

[0057] The coordinated optimized control instructions for each abnormal region block, along with the global coordination instructions, are packaged together into the final executable control instruction set.

[0058] Preferably, the method further includes an online self-evolution step for the model:

[0059] After executing the set of control instructions, the changes in the original monitoring data are continuously monitored to form a multi-dimensional state image sequence of the farm after control.

[0060] The multidimensional images of the farm before and after regulation, the set of regulatory instructions executed, and the actual results of animal health and production performance are used as a new training sample pair.

[0061] Using new training sample pairs, the parameters of the behavior-environment decoupling analysis network are updated in an incremental learning manner, and the cross-influence evaluation matrix and the regulation policy library in the policy generation engine are optimized.

[0062] Through periodic or triggered model updates, the decision-making capabilities of the entire monitoring and control system can continuously evolve as the farm operates over time.

[0063] Preferably, when the processor executes the computer program, it implements the steps of the smart aquaculture monitoring method based on big data analysis as described in any of the above-mentioned embodiments.

[0064] Compared with the prior art, the beneficial effects of the present invention are:

[0065] A behavior-environment decoupling analysis network, pre-trained with massive historical data, is used to process multi-dimensional state images of a farm in parallel. It outputs behavioral feature tensors that purely represent animal activity patterns and environmental background field tensors that exclude transient animal interference. This separation of the animal's internal state from the external physical environment allows the system to trace the root cause of abnormal signals and accurately determine whether the problem stems from behavioral changes in the animal population or from environmental stressors, overcoming the shortcomings of traditional monitoring methods that often involve a mixture of both, leading to unclear root causes.

[0066] Using a pre-defined cross-influence evaluation matrix, the coupling degree index of the decoupled behavioral feature tensor and the environmental background field tensor is calculated. This quantifies the dynamic interaction strength between behavioral patterns and environmental parameters, and reveals abnormal deviations in the correlation through the generated comprehensive anomaly score map. It can identify complex and hidden problems that traditional threshold alarms cannot detect, such as the absence of expected behavior under specific microclimates, thereby achieving early comprehensive warning of potential risks.

[0067] Based on the nature and spatial distribution of associated anomalies revealed by the comprehensive anomaly scoring map, the strategy generation engine dynamically synthesizes customized combinations of regulatory instructions. These instructions are no longer fixed-pattern responses, but rather combined strategies designed to address the root causes of specific behavioral environmental dysfunctions. Acting on the identified key links, this achieves coordinated intervention at the holistic system level, improving the accuracy of management measures and optimizing the stability of the breeding environment and animal welfare. Attached Figure Description

[0068] Figure 1 This is a schematic diagram illustrating the working principle of the intelligent aquaculture monitoring method based on big data analysis described in this invention.

[0069] Figure 2 A flowchart for generating standardized sound event field layers;

[0070] Figure 3 A flowchart for deep pattern mining of networks used in behavior-environment decoupling analysis;

[0071] Figure 4 A heatmap showing the spatial distribution of reconstruction errors in a smart farming scenario;

[0072] Figure 5 A bar chart comparing multiple indicators of intelligent aquaculture control strategies. Detailed Implementation

[0073] 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.

[0074] Please see Figure 1This invention provides a smart aquaculture monitoring method based on big data analysis. The method includes: periodically acquiring raw monitoring data from a sensor network deployed within the farm, including but not limited to temperature, weight, sound, and gas concentration information. The raw monitoring data undergoes spatiotemporal field synchronization and normalization processing. This processing adds a unified timestamp sequence and three-dimensional spatial coordinate information to the temperature profile, feed consumption rate curve, animal sound signal waveform, and gas concentration distribution map, respectively. A standardized temperature field layer is generated through hot spot region segmentation and contour extraction; a standardized feed consumption field layer is generated through consumption event detection and cumulative integration; a standardized sound event field layer is generated through frequency domain transformation and feature event labeling; and a standardized gas concentration field layer is generated through concentration gradient calculation and high-concentration region clustering. These standardized layers are then pixel-level aligned and fused according to the timestamp sequence and three-dimensional spatial coordinate information to construct a multi-dimensional state image of the farm with a unified spatiotemporal reference. This image is a digital twin model that integrates temperature distribution, feed consumption dynamics, sound event labeling, and gas concentration gradient.

[0075] Multidimensional state images of the farm are input into a behavior-environment decoupling analysis network pre-trained with massive amounts of historical farming data for deep pattern mining. This network performs animal behavior pattern stripping and environmental background field extraction in parallel, outputting independent behavioral feature tensors and independent environmental background field tensors. Based on the behavioral feature tensors and environmental background field tensors, a coupling degree index between the two is calculated using a pre-defined cross-influence evaluation matrix. The coupling degree index quantifies the extent to which currently observed animal behavior can be explained by the current environmental background conditions, and a comprehensive anomaly score map is generated based on the coupling degree index. Based on the comprehensive anomaly score map, a set of control instructions is dynamically synthesized through a strategy generation engine. The control instruction set includes at least directional airflow guidance parameters for the ventilation system, refined feeding formulas and schedules for the feeding device, and temperature and humidity coordinated adjustment instructions for the environmental control unit.

[0076] Example 1: See Figure 2 In the process of constructing a multi-dimensional state image of a livestock farm through spatiotemporal field synchronization and normalization, the original monitoring data includes temperature profiles generated by a thermal imaging sensor array, feed consumption rate curves acquired by a weight sensor, animal sound signal waveforms captured by a microphone array, and gas concentration distribution maps fed back in real time by an ammonia nitrogen sensor network. Uniform timestamp sequences and three-dimensional spatial coordinate information are added to the temperature profiles, feed consumption rate curves, animal sound signal waveforms, and gas concentration distribution maps, respectively. Hotspot region segmentation and contour extraction are performed on the temperature profiles with added spatiotemporal information to generate a standardized temperature field layer. Consumption event detection and cumulative integration are performed on the feed consumption rate curves with added spatiotemporal information to generate a standardized feed consumption field layer.

[0077] Frequency domain transformation and feature event labeling are performed on animal sound signal waveforms with added spatiotemporal information to generate a standardized sound event field layer. The frequency domain transformation and feature event labeling include performing a short-time Fourier transform on the animal sound signal waveform to obtain a time-spectrum image. Template matching is performed on the time-spectrum image using a predefined vocal template library to identify various predefined acoustic events, including feeding calls, stress calls, and pathological coughs. Each identified acoustic event is assigned an event type label, confidence score, and sound source intensity estimate. Based on the three-dimensional spatial coordinates of the acoustic event, the event type label, confidence score, and sound source intensity estimate are mapped onto a three-dimensional spatial grid to form a discrete sound event point cloud. Spatial interpolation and smoothing are performed on the sound event point cloud to generate a continuously distributed standardized sound event field layer. Each pixel value in the layer represents a composite index of the intensity and frequency of a predefined acoustic event occurring at a specific spatial location within a specific time window. Concentration gradient calculation and high-concentration region clustering are performed on a gas concentration distribution map with added spatiotemporal information to generate a standardized gas concentration field layer. Standardized temperature field layers, feed consumption field layers, sound event field layers, and gas concentration field layers are aligned and fused pixel-level according to their timestamp sequences and three-dimensional spatial coordinate information to generate a multi-dimensional state image of the farm with a unified spatiotemporal reference.

[0078] In practice, raw monitoring data is periodically acquired from a sensor network deployed within the farm. This raw data includes temperature profiles generated by a thermal imaging sensor array, feed consumption rate curves collected by weight sensors, animal sound signal waveforms captured by a microphone array, and gas concentration distribution maps fed back in real time by an ammonia nitrogen sensor network. A unified timestamp sequence and three-dimensional spatial coordinate information are added to each of the temperature profile, feed consumption rate curve, animal sound signal waveform, and gas concentration distribution map. The timestamp sequence originates from a unified time synchronization server within the farm, and the three-dimensional spatial coordinate information is determined based on the fixed installation locations of the sensors within the farm and a pre-defined coordinate system mapping.

[0079] Hot spot region segmentation and contour extraction were performed on the temperature profile map with added spatiotemporal information to generate a standardized temperature field layer. Hot spot region segmentation adopted an image segmentation algorithm based on adaptive threshold, and contour extraction used an edge detection operator to identify the boundaries of each connected temperature region. Consumption event detection and cumulative integral were performed on the feed consumption rate curve with added spatiotemporal information to generate a standardized feed consumption field layer. Consumption event detection was achieved by identifying abrupt drops in feed weight over time, and cumulative integral was used to calculate the total consumption at each feeding point within a specified time window. The animal sound signal waveforms with added spatiotemporal information are subjected to frequency domain transformation and feature event labeling to generate a standardized sound event field layer. The frequency domain transformation uses short-time Fourier transform to convert the time-domain waveform into a time-spectrum graph. Feature event labeling uses a predefined vocal template library to perform template matching on the time-spectrum graph to identify various predefined acoustic events, including feeding calls, stress calls, and pathological coughs. Each identified acoustic event is assigned an event type label, confidence score, and sound source intensity estimate. Based on the three-dimensional spatial coordinate information corresponding to the acoustic event, the event type label, confidence score, and sound source intensity estimate are mapped to a three-dimensional spatial grid to form a discrete sound event point cloud. Spatial interpolation and smoothing are performed on the sound event point cloud. The spatial interpolation uses the Kriging interpolation algorithm, and the smoothing uses Gaussian filtering to generate a continuously distributed standardized sound event field layer.

[0080] It can be understood that each pixel value in the standardized sound event field layer represents a composite index of the intensity and frequency of a predefined acoustic event occurring at a spatial location within a specific time window. The composite index is calculated as follows:

[0081]

[0082] in: Represents spatial coordinates and time The pixel value of the layer at that location, This represents the total number of acoustic events identified within a temporal and spatial proximity range. Indicates the first Confidence score of an acoustic event. Indicates the first Estimation of the sound source intensity of an acoustic event. Indicates the first The time of occurrence of an acoustic event This is the time window width parameter that controls the rate of time decay. Concentration gradient calculation and high-concentration region clustering are performed on the gas concentration distribution map with added spatiotemporal information to generate a standardized gas concentration field layer. The concentration gradient calculation uses the Sobel operator, and the high-concentration region clustering uses a density-based spatial clustering algorithm to identify continuous spatial regions where the gas concentration is significantly higher than the background value.

[0083] In some embodiments, standardized temperature field layers, feed consumption field layers, sound event field layers, and gas concentration field layers are pixel-level aligned and fused according to their timestamp sequences and three-dimensional spatial coordinate information to generate a multi-dimensional state image of the farm with a unified spatiotemporal reference. Pixel-level alignment uses the three-dimensional digital model of the farm as the spatial reference, projecting each data point in each layer onto the corresponding spatial grid of the digital model according to its three-dimensional spatial coordinate information. The timestamp sequence is used to ensure that the data of different layers correspond to the same time slice or time window. The fusion operation adopts a channel overlay method, treating the temperature field layer, feed consumption field layer, sound event field layer, and gas concentration field layer as different data channels to jointly constitute a multi-channel tensor data. The multi-dimensional state image of the farm is a digital twin model that integrates temperature distribution, feed consumption dynamics, sound event markers, and gas concentration gradients.

[0084] Example 2: See Figure 3 The behavior-environment decoupling analysis network comprises a shared feature encoder, a behavior feature decoding branch, and an environmental background field decoding branch. The shared feature encoder performs multi-layer convolutional downsampling on the input multi-dimensional state image of the farm, extracting a shared feature map containing multi-scale information. The behavior feature decoding branch receives the shared feature map and, by focusing on dynamic change patterns directly related to animal activity, reconstructs a behavior feature tensor mainly containing animal group movement trajectories, inter-individual interaction frequency, and periodicity of feeding and drinking behaviors. The environmental background field decoding branch receives the shared feature map and, by focusing on the relatively static physical field distribution and slow changing trends, reconstructs an environmental background field tensor mainly containing baseline temperature distribution, baseline humidity level, background noise level, and initial gas concentration. During the training phase, an adversarial loss function is introduced to force the behavior feature tensor to exclude environmental background information, and simultaneously forces the environmental background field tensor to exclude animal behavior information.

[0085] In its implementation, the behavior-environment decoupling analysis network comprises a shared feature encoder, a behavior feature decoding branch, and an environmental background field decoding branch. The shared feature encoder performs multi-layer convolutional downsampling on the input multi-dimensional state image of the aquaculture farm to extract shared feature maps containing multi-scale information. The shared feature encoder consists of five concatenated convolutional modules, each containing a convolutional layer, a batch normalization layer, and an activation function. By using convolutional operations with a stride greater than one, it achieves gradual compression of spatial dimensions and an increase in the number of feature channels.

[0086] The behavioral feature decoding branch receives shared feature maps and reconstructs a behavioral feature tensor, primarily containing animal group movement trajectories, inter-individual interaction frequencies, and the periodicity of feeding and drinking behaviors, by focusing on dynamic change patterns directly related to animal activity. The behavioral feature decoding branch consists of a series of transposed convolutional layers and skip connections. Skip connections concatenate intermediate feature maps of the corresponding scale from the shared feature encoder with intermediate layer features from the behavioral feature decoding branch, fusing low-level spatial information with high-level semantic information. The environmental background field decoding branch receives shared feature maps and reconstructs an environmental background field tensor, primarily containing baseline temperature distribution, baseline humidity level, background noise level, and initial gas concentration, by focusing on the relatively static physical field distribution and slow changing trends. The structure of the environmental background field decoding branch is symmetrical to that of the behavioral feature decoding branch but has independent network parameters. Skip connections also pass intermediate features from the shared feature encoder to the environmental background field decoding branch.

[0087] In some embodiments, during the training phase, an adversarial loss function is introduced to force the behavior feature tensor to exclude environmental background information, and simultaneously force the environmental background field tensor to exclude animal behavior information. Adversarial training is implemented through two auxiliary discriminator networks. The first discriminator network takes the behavior feature tensor as input, and its training objective is to distinguish whether the input behavior feature tensor comes from the real animal behavior data distribution or from the network's reconstructed output. One of the training objectives of the behavior feature decoding branch is to generate a behavior feature tensor capable of deceiving the first discriminator network. The second discriminator network takes the environmental background field tensor as input, and its training objective is to distinguish whether the input environmental background field tensor comes from the real environmental background data distribution or from the network's reconstructed output. One of the training objectives of the environmental background field decoding branch is to generate an environmental background field tensor capable of deceiving the second discriminator network.

[0088] It can be understood that the total loss function of the behavior-environment decoupling analysis network is composed of reconstruction loss and adversarial loss. Reconstruction loss measures the difference between the sum of the reconstructed behavior feature tensor and the environmental background field tensor and the original multidimensional state image of the farm. Adversarial loss, on the other hand, causes the outputs of the two decoding branches to be mutually exclusive in terms of information content. The overall objective function for network training can be expressed as:

[0089]

[0090] in: It is the total loss. This represents the input multidimensional status image of the farm. The tensor representing the reconstructed behavioral features. This represents the reconstructed environmental background field tensor. It is a reconstruction loss function. and These are discriminator networks targeting behavioral characteristics and environmental context, respectively. and It is the corresponding adversarial loss function. It is a weighting coefficient used to balance reconstruction loss and adversarial loss.

[0091] Example 3: The cross-influence evaluation matrix defines the theoretical correlation strength weights between different categories of behavioral patterns and different types of environmental background factors. The behavioral feature tensor is projected onto the behavioral pattern category dimension to obtain the behavioral pattern probability distribution vector. The environmental background field tensor is projected onto the environmental background factor type dimension to obtain the environmental factor intensity distribution vector. The dot product of the behavioral pattern probability distribution vector and the environmental factor intensity distribution vector under the cross-influence evaluation matrix is ​​calculated to obtain the coupling degree index. When the coupling degree index is lower than a preset coupling degree threshold, an anomaly is identified. A reconstruction error map generated by the behavior-environment decoupling analysis network is obtained. The reconstruction error map characterizes the difference between the original multidimensional state image of the farm and the image reconstructed from the behavioral feature tensor and the environmental background field tensor. The coupling degree index is spatially expanded to generate a coupling degree distribution map with the same spatial resolution as the reconstruction error map. The reconstruction error map and the coupling degree distribution map are multiplied element-wise to obtain a preliminary anomaly attention heatmap. A spatial prior weight map is introduced. The spatial prior weight map is pre-set based on the farm's historical anomaly records to weight and correct the preliminary anomaly attention heatmap. The weighted and corrected heatmap of abnormal attention is normalized and nonlinearly stretched to generate the final comprehensive anomaly score map. The score value of each pixel in the map represents the probability of an anomaly occurring in the spatial location.

[0092] In practice, the cross-impact evaluation matrix defines the theoretical correlation strength weights between different categories of behavioral patterns and different types of environmental background factors. The cross-impact evaluation matrix is ​​a two-dimensional matrix. The rows of the matrix correspond to a predefined set of animal behavioral pattern categories, and the columns of the matrix correspond to a predefined set of environmental background factor types. The value of each element in the matrix represents the theoretical correlation strength between the corresponding behavioral pattern and the corresponding environmental factor. The correlation strength value is pre-set based on statistical regularities and domain knowledge in historical breeding data.

[0093] Projecting the behavioral feature tensor onto the behavioral pattern category dimension yields a behavioral pattern probability distribution vector. This projection operation is implemented using a fully connected layer or a combination of a global pooling layer and a Softmax activation function. Each component of the behavioral pattern probability distribution vector represents a probability estimate of a specific behavioral pattern exhibited by the animal group under the current observation. Projecting the environmental background field tensor onto the environmental background factor type dimension yields an environmental factor intensity distribution vector. This projection operation is implemented using another fully connected layer or a global pooling layer. Each component of the environmental factor intensity distribution vector represents the intensity or level of a specific background factor in the current environment. Calculating the dot product of the behavioral pattern probability distribution vector and the environmental factor intensity distribution vector under the cross-influence evaluation matrix yields the coupling degree index. The formula for calculating the coupling degree index is expressed as:

[0094]

[0095] in: This represents the calculated coupling degree scalar index. Represents the probability distribution vector of behavioral patterns. This indicates that vector b is transposed, that is, it is converted from a column vector to a row vector. This represents the pre-defined cross-impact evaluation matrix. This represents the distribution vector of environmental factor intensity. When the coupling degree index is lower than the preset coupling degree threshold, it is determined that there is an anomaly. The coupling degree threshold is determined by analyzing the distribution of the coupling degree index in historical normal data.

[0096] A reconstruction error map generated by a behavior-environment decoupling analysis network is obtained. This map characterizes the difference between the original multidimensional state image of the farm and the image reconstructed from the behavioral feature tensor and the environmental background field tensor. The reconstruction error map is generated by calculating the difference norm between the original and reconstructed images at each spatial pixel location. The coupling degree index is spatially expanded to generate a coupling degree distribution map with the same spatial resolution as the reconstruction error map. The expansion method assigns a scalar coupling degree index to all pixels in the spatially expanded map. Element-wise multiplication of the reconstruction error map and the coupling degree distribution map yields a preliminary heatmap of abnormal attention. This element-wise multiplication amplifies the signals in spatial regions that simultaneously exhibit high reconstruction errors and low coupling degrees.

[0097] In some embodiments, a spatial prior weight map is introduced. This spatial prior weight map is pre-defined based on the historical anomaly records of the farm. It is a two-dimensional matrix with the same spatial size as the comprehensive anomaly scoring map. The value of each location in the matrix is ​​assigned based on the frequency with which that location has historically been marked as an anomaly area; locations with higher historical anomaly frequencies have greater weight values. The spatial prior weight map is used to weight and correct the initial anomaly attention heatmap. This weighting correction is achieved by element-wise multiplying the initial anomaly attention heatmap with the spatial prior weight map.

[0098] It is understandable that the weighted and corrected heatmap of abnormal attention is normalized and nonlinearly stretched to generate the final comprehensive anomaly score map. Normalization linearly scales the values ​​of all pixels in the heatmap to the range of 0 to 1. Nonlinear stretching uses, for example, gamma correction or a sigmoid function to transform the normalized values ​​to enhance the difference in low-value areas or suppress saturation in high-value areas. The score value of each pixel in the image represents the probability of an anomaly occurring at that spatial location; a score value closer to 1 indicates a higher probability of an anomaly, and a score value closer to 0 indicates a lower probability of an anomaly.

[0099] See Figure 4 This is a heatmap showing the spatial distribution of reconstruction errors in a smart farming scenario. It's a spatial heatmap used to visualize the "reconstruction error values" (pixel-level difference indicators) at different spatial locations (divided by X and Y coordinates) within a farm, reflecting the degree of difference between the original monitoring image and the model-reconstructed image. It's commonly used in the anomaly monitoring phase of smart farming: areas with high reconstruction errors are usually potential anomaly areas, and the anomaly risk will be further analyzed in conjunction with coupling degree indicators. The core value of this type of chart is locating anomalous spatial points within the farm, providing a spatial basis for subsequent precise control.

[0100] Example 4: A comprehensive anomaly score map marks the degree of anomaly at various spatial locations within the farm over time. Connectivity analysis is performed on the comprehensive anomaly score map to identify anomalous regions whose scores exceed activation thresholds. For each anomalous region, its geometric center coordinates, spatial extent, average anomaly score, and score trend over time are extracted. The feature information of the anomalous regions is input into a case-based reasoning-based control strategy library for matching and retrieval. The control strategy library stores the features and corresponding control parameters of historically successful control cases. Based on the matched historical cases, preliminary candidate control instructions for the anomalous regions are generated. These candidate control instructions include candidate ventilation parameters, candidate feed adjustment amounts, and candidate temperature and humidity setpoints. Based on the current overall operational status and resource constraints of the farm, candidate control commands are optimized and conflict resolved through multi-objective optimization. This multi-objective optimization and conflict resolution involves establishing a multi-objective optimization function that includes animal welfare indicators, energy consumption indicators, feed conversion rate indicators, and equipment wear indicators. The real-time operating status and capacity limits of all environmental control and feeding equipment in the farm are obtained as constraints for the optimization problem. A multi-objective evolutionary algorithm is used to adjust the candidate control commands. Under the premise of satisfying the constraints, the combination of control parameters that makes the multi-objective optimization function Pareto optimal is found. The optimized control commands for different abnormal regions are globally coordinated to avoid contradictions in resource allocation or environmental field control between different regions. The coordinated optimized control commands for each abnormal region and the globally coordinated commands are packaged together into a final executable set of control commands.

[0101] In practice, the comprehensive anomaly score map marks the degree of anomaly at various spatial locations within the farm over time. Connectivity analysis is performed on the comprehensive anomaly score map to identify anomalous regions whose scores exceed an activation threshold. This connectivity analysis employs an image connected component labeling algorithm based on eight or four neighborhoods, with the activation threshold being a pre-defined value between 0 and 1. For each anomalous region, its geometric center coordinates, spatial extent, average anomaly score, and scoring trend over time are extracted. The geometric center coordinates are obtained by calculating the average of all pixel coordinates within the region. The spatial extent is described by recording the region's bounding box or minimum bounding rectangle. The average anomaly score is the arithmetic mean of all pixel scores within the region. The scoring trend over time is obtained by analyzing the score changes of the current region in the comprehensive anomaly score map sequence over the most recent time points.

[0102] Referring to Table 1, the feature information of the abnormal area block is input into a case-based reasoning control strategy library for matching and retrieval. The control strategy library stores the features and corresponding control parameters of historical successful control cases. Each case record in the control strategy library contains a problem feature field and a solution field. The problem feature field describes the characteristics of historical abnormal events and is consistent with the feature format of the currently extracted abnormal area block. The solution field records the set of control instructions that were executed at that time and produced positive effects. Matching and retrieval is accomplished by calculating the similarity between the features of the current abnormal area block and the problem feature fields of each case in the control strategy library. The similarity calculation can use methods such as Euclidean distance or cosine similarity, and the top K historical cases with the highest similarity are selected as references. Based on the matched historical cases, candidate control instructions for the abnormal area block are initially generated. Candidate control instructions include candidate ventilation parameters, candidate feeding adjustment amounts, and candidate temperature and humidity setpoints. The initial generation process involves performing a weighted average or cluster analysis on the control parameters in the solution fields of the K similar historical cases, with the weights determined by the matching similarity.

[0103] Table 1: Case Structure Table of Regulation Strategy Library

[0104] Case ID Problem characteristics (geometric center coordinates, spatial extent, average anomaly score, trend) Solutions (ventilation parameters, feed adjustment amount, temperature and humidity setpoints) C001 (x1, y1), area A1, score S1, increase Wind speed V1, wind direction D1, feed increment F1, temperature T1, humidity H1 C002 (x2, y2), area A2, score S2, decrease Wind speed V2, wind direction D2, feed reduction F2, temperature T2, humidity H2 ... ... ...

[0105] Based on the current overall operational status and resource constraints of the farm, multi-objective optimization and conflict resolution are performed on candidate control commands. Multi-objective optimization and conflict resolution involve establishing a multi-objective optimization function that includes animal welfare indicators, energy consumption indicators, feed conversion rate indicators, and equipment wear and tear indicators. Animal welfare indicators can be calculated comprehensively from predicted animal thermal comfort index, air quality index, etc.; energy consumption indicators are the estimated energy consumption for executing control commands; feed conversion rate indicators are the estimated feed input to animal weight gain ratio; and equipment wear and tear indicators are the estimated cumulative load on relevant equipment caused by executing control commands. The real-time operating status and capacity limits of all environmental control equipment and feeding equipment in the farm are obtained as constraints for the optimization problem. These constraints include, but are not limited to, the maximum airflow of the ventilation system, the maximum feeding rate of the feeding device, and the temperature and humidity adjustment range and power limit of the environmental control unit.

[0106] A multi-objective evolutionary algorithm is used to adjust candidate control commands, seeking the Pareto optimal combination of control parameters while satisfying constraints. The multi-objective optimization function can be expressed as:

[0107]

[0108] in: It is a multi-objective vector. This represents a vector of control parameters to be optimized, consisting of candidate ventilation parameters, candidate feed adjustment amounts, and candidate temperature and humidity setpoints. It is an animal welfare indicator function. It is an energy consumption index function. It is a feed conversion ratio indicator function. This is a function representing equipment wear and tear; the negative sign indicates that the original index needs to be negated to unify it into a minimization problem. Multi-objective evolutionary algorithms, such as NSGA-II, are used to search for a set of non-dominated solutions that constitute a Pareto front. In practical implementations, the application of multi-objective evolutionary algorithms aims to efficiently search the optimization parameter space of candidate regulatory instructions to achieve a Pareto optimal solution set for the multi-objective optimization function. This algorithm initializes a population containing multiple combinations of regulatory parameters, each combination representing a complete set of ventilation, feeding, and temperature / humidity instruction schemes, by simulating the natural evolutionary process. During iteration, the algorithm evaluates individuals in the population based on the multi-objective optimization function, which covers animal welfare indicators, energy consumption indicators, feed conversion rate indicators, and equipment wear and tear indicators, while strictly adhering to real-time equipment status and capacity constraints. NSGA-II, as a typical implementation, employs a fast non-dominated sorting and crowding calculation mechanism, prioritizing the retention of non-dominated solutions that achieve a good balance among multiple objectives and are evenly distributed, gradually evolving the population until it converges to the Pareto front.

[0109] In some embodiments, optimized control commands for different anomalous region blocks are globally coordinated to avoid conflicts in resource allocation or environmental field control between different regions. The global coordination checks whether the total demand for shared resources from optimized control commands from different anomalous region blocks exceeds the system's total supply capacity, and checks whether there are spatially conflicting settings for control commands targeting environmental fields such as temperature and humidity. If conflicts or exceedances exist, the parameters of the relevant control commands are iteratively adjusted and negotiated based on the average anomalous score or the urgency of the problem in each anomalous region block until all commands can be executed simultaneously by the system without conflict.

[0110] It is understandable that the coordinated optimized control instructions for each abnormal area block, along with the global coordination instructions, are packaged together into a final executable control instruction set. The control instruction set includes at least directional airflow guidance parameters for the ventilation system, refined feeding formulas and schedules for the feeding devices, and temperature and humidity coordinated adjustment instructions for the environmental control unit. The instruction set is output to the corresponding actuator system in the farm in a structured data format or standard control protocol message format.

[0111] Example 5: The online self-evolutionary model step includes continuously monitoring changes in the original monitoring data after executing the control instruction set, forming a multi-dimensional state image sequence of the farm after control. The multi-dimensional state images of the farm before and after control, the executed control instruction set, and the actual animal health and production performance results are used as a new training sample pair. Using this new training sample pair, the parameters of the behavior-environment decoupling analysis network are updated incrementally, and the cross-influence evaluation matrix and the control strategy library in the strategy generation engine are optimized. Through periodic or triggered model updates, the decision-making ability of the entire monitoring and control system continuously evolves as the farm operates.

[0112] In its implementation, the online self-evolutionary steps of the model include continuously monitoring changes in the original monitoring data after the execution of the control command set to form a sequence of multidimensional state images of the farm after control. This continuous monitoring is conducted with the same period and sensor network as the acquisition of the original monitoring data. The sequence of multidimensional state images of the farm after control records multidimensional state images of the farm at several time points from the start of the control command execution. The multidimensional state images of the farm before and after control, the executed control command set, and the actual animal health and production performance results are used as a new training sample pair. The multidimensional state images of the farm before and after control refer to an image at a time point before the execution of the control command and an image at a time point after the control effect has stabilized, respectively. The executed control command set is a complete set of commands output from the strategy generation engine. The actual animal health and production performance results are obtained through auxiliary records in the farm management system, including but not limited to animal morbidity records, mortality records, average daily weight gain data, and feed conversion rate data in subsequent periods.

[0113] Using new training sample pairs, the parameters of the behavior-environment decoupling analysis network are updated incrementally, and the cross-influence evaluation matrix and the regulation strategy library in the policy generation engine are optimized. Incremental learning employs an elastic weight consolidation algorithm or a playback memory-based method. While retaining key statistical features or some samples from historical data, new training sample pairs are added to the training set to fine-tune the behavior-environment decoupling analysis network. The fine-tuning process controls the magnitude of changes in network parameters to prevent catastrophic forgetting of learned knowledge. The cross-influence evaluation matrix is ​​optimized by analyzing the actual correlation strength between behavioral patterns and environmental factors in the new training sample pairs to adjust the weights of theoretical correlation strength in the matrix. The adjustment method is a weighted average between historical statistical values ​​and new observations. The regulation strategy library in the policy generation engine is optimized by adding new training sample pairs as successful or unsuccessful regulation cases after feature extraction to the case set of the regulation strategy library. When adding new cases, deduplication and similar case fusion operations are performed to control the size of the strategy library.

[0114] In some embodiments, a dynamic learning rate strategy is used when updating the network parameters for behavior-environment decoupling analysis. The dynamic learning rate decays as the number of online self-evolutionary steps of the model increases, ensuring that the model gradually stabilizes after accumulating a large number of new samples. The formula for calculating the dynamic learning rate is as follows:

[0115]

[0116] in: Indicates the first The learning rate used for updating network parameters in the online self-evolution step of the sub-model. This represents the base learning rate, which is a pre-defined positive number. This represents the decay coefficient, a positive number that controls the rate at which the learning rate decreases. Represents the natural exponential function. This represents the count of the number of times the online self-evolutionary steps of the model have been executed.

[0117] It is understandable that periodic or triggered model updates allow the decision-making capabilities of the entire monitoring and control system to continuously evolve as the farm operates. Periodic updates automatically execute a complete online self-evolution step of the model at fixed time intervals, while triggered updates are initiated when specific conditions are met. These conditions include the system detecting multiple consecutive prediction errors, significant fluctuations in farm production performance indicators, or a user manually initiating a model update command. The model update process is completed on the background computing node of the farm management system. The updated behavior-environment decoupling analysis network parameters, cross-influence evaluation matrix, and control strategy library are loaded into the online service system via hot replacement, ensuring uninterrupted monitoring and control operations.

[0118] See Figure 5 This is a bar chart comparing multiple indicators of smart farming control strategies. The original strategy had relatively low levels across all indicators, with animal welfare and feed conversion rate below 70%, and equipment durability at only 60%. After optimization, all indicators showed an upward trend, with animal welfare showing the most significant improvement (from approximately 75 to 90+); feed conversion rate and energy efficiency also gradually approached 90, while equipment durability stabilized at around 75. Optimized strategy 4 demonstrated the best overall performance, with animal welfare, feed conversion rate, and energy efficiency all at high levels. This type of chart is used for evaluating smart farming control strategies. By comparing multiple indicators, it helps to select the optimal strategy that balances animal welfare, production efficiency, and cost control, serving as a key reference tool for optimizing farming system decisions.

[0119] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0120] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A smart aquaculture monitoring method based on big data analysis, characterized in that, The method includes: Raw monitoring data is periodically acquired from a sensor network deployed within the farm. The original monitoring data is subjected to spatiotemporal field synchronization and normalization processing to construct a multidimensional state image of the farm with a unified spatiotemporal reference. The multidimensional state image of the farm is input into a behavior-environment decoupling analysis network pre-trained with massive historical farming data for deep pattern mining. The behavior-environment decoupling analysis network performs animal behavior pattern stripping and environmental background field extraction in parallel, outputting independent behavioral feature tensors and independent environmental background field tensors, including: The behavior-environment decoupling analysis network includes a shared feature encoder, a behavior feature decoding branch, and an environment background field decoding branch; The shared feature encoder performs multi-layer convolutional downsampling on the input multi-dimensional state image of the aquaculture farm to extract a shared feature map containing multi-scale information; The behavioral feature decoding branch receives the shared feature map and reconstructs the behavioral feature tensor, which mainly includes the animal group's movement trajectory, the frequency of interaction between individuals, and the periodicity of feeding and drinking behavior, by focusing on dynamic change patterns directly related to animal activities. The environmental background field decoding branch receives the shared feature map and, by focusing on the relatively static physical field distribution and slow changing trend, reconstructs the environmental background field tensor, which mainly includes the basic temperature distribution, the reference humidity level, the background noise level, and the initial gas concentration. During the training phase, an adversarial loss function is introduced to force the behavioral feature tensor to exclude environmental background information, and at the same time, to force the environmental background field tensor to exclude animal behavior information. Based on the behavioral feature tensor and the environmental background field tensor, a coupling degree index between the behavioral feature tensor and the environmental background field tensor is calculated using a preset cross-influence evaluation matrix. A comprehensive anomaly scoring map is then generated based on the coupling degree index, including: The cross-influence evaluation matrix defines the theoretical correlation strength weights between different categories of behavioral patterns and different types of environmental background factors; Projecting the behavior feature tensor onto the behavior pattern category dimension yields a behavior pattern probability distribution vector. Projecting the environmental background field tensor onto the environmental background factor type dimension yields the environmental factor intensity distribution vector. The dot product of the probability distribution vector of the behavior pattern and the intensity distribution vector of the environmental factors under the cross-influence evaluation matrix is ​​calculated to obtain the coupling index, which quantifies the extent to which the currently observed animal behavior can be explained by the current environmental background conditions. When the coupling degree index is lower than the preset coupling degree threshold, it is determined that there is an anomaly. The anomaly indicates that the animal’s behavior deviates from the expected pattern in the current environment. Obtain a reconstruction error map generated by the behavior-environment decoupling analysis network. The reconstruction error map characterizes the difference between the original multidimensional state image of the farm and the image reconstructed from the behavior feature tensor and the environmental background field tensor. The coupling index is spatially extended to generate a coupling distribution map with the same spatial resolution as the reconstruction error map; The reconstruction error map and the coupling degree distribution map are multiplied element by element to obtain a preliminary heatmap of abnormal attention. A spatial prior weight map is introduced, which is pre-set based on the historical anomaly records of the farm, and is used to weight and correct the preliminary anomaly attention heat map. The weighted and corrected heatmap of abnormal attention is normalized and nonlinearly stretched to generate the final comprehensive anomaly score map, where the score value of each pixel represents the probability of an anomaly occurring at the spatial location. Based on the comprehensive anomaly score map, a set of control instructions is dynamically synthesized through a strategy generation engine. The set of control instructions includes at least directional airflow guidance parameters for the ventilation system, refined feeding formulas and schedules for the feeding device, and temperature and humidity coordinated adjustment instructions for the environmental control unit.

2. The intelligent aquaculture monitoring method based on big data analysis according to claim 1, characterized in that, The process of performing spatiotemporal field synchronization and normalization on the original monitoring data to construct a multidimensional state image of the farm with a unified spatiotemporal reference specifically includes: The raw monitoring data includes temperature profiles generated by a thermal imaging sensor array, feed consumption rate curves collected by a weight sensor, animal sound signal waveforms captured by a microphone array, and gas concentration distribution maps fed back in real time by an ammonia nitrogen sensor network. A unified timestamp sequence and three-dimensional spatial coordinate information are added to the temperature profile, the feed consumption rate curve, the animal sound signal waveform, and the gas concentration distribution map, respectively. The temperature profile image with added spatiotemporal information is subjected to hot spot region segmentation and contour extraction to generate a standardized temperature field layer; Consumption events are detected and cumulative amounts are integrated on the feed consumption rate curve with added spatiotemporal information to generate a standardized feed consumption field layer. The animal sound signal waveform with added spatiotemporal information is subjected to frequency domain transformation and feature event labeling to generate a standardized sound event field layer; Concentration gradient calculation and high-concentration region clustering are performed on the gas concentration distribution map with added spatiotemporal information to generate a standardized gas concentration field layer; The standardized temperature field layer, feed consumption field layer, sound event field layer, and gas concentration field layer are aligned and fused pixel-level according to their timestamp sequence and three-dimensional spatial coordinate information to generate a multi-dimensional state image of the farm with the unified spatiotemporal reference. The multidimensional state image of the farm is a digital twin model that integrates temperature distribution, feed consumption dynamics, sound event markers, and gas concentration gradients.

3. The intelligent aquaculture monitoring method based on big data analysis according to claim 2, characterized in that, The step of performing frequency domain transformation and feature event labeling on the animal sound signal waveform with added spatiotemporal information to generate a standardized sound event field layer includes: Perform a short-time Fourier transform on the waveform of the animal sound signal to obtain a time-spectrum diagram; The temporal spectrogram is template-matched using a predefined vocal template library to identify a variety of predefined acoustic events, including feeding calls, stress calls, and pathological coughs. Assign an event type label, confidence score, and sound source intensity estimate to each identified acoustic event; Based on the three-dimensional spatial coordinate information corresponding to the acoustic event, the event type label, confidence score and sound source intensity estimate are mapped onto a three-dimensional spatial grid to form a discrete sound event point cloud; Spatial interpolation and smoothing are performed on the sound event point cloud to generate a continuously distributed standardized sound event field layer, wherein each pixel value in the layer represents a composite index of the intensity and frequency of a predefined acoustic event occurring at a spatial location within a specific time window.

4. The intelligent aquaculture monitoring method based on big data analysis according to claim 3, characterized in that, Based on the comprehensive anomaly scoring map, a set of control instructions is dynamically synthesized through a strategy generation engine, including: The comprehensive anomaly scoring map marks the degree of anomaly at each spatial location within the farm in the time dimension; Connectivity analysis is performed on the comprehensive anomaly score map to identify anomaly regions whose scores exceed the activation threshold; For each of the aforementioned abnormal regions, extract its geometric center coordinates, spatial extent, average abnormal score, and score trend over time. The feature information of the abnormal region block is input into a case-based reasoning control strategy library for matching and retrieval. The control strategy library stores the features of historical successful control cases and the corresponding control parameters. Based on the matched historical cases, preliminary candidate control instructions for the abnormal area blocks are generated, including candidate ventilation parameters, candidate feeding adjustment amounts, and candidate temperature and humidity setpoints. Based on the overall operating status and resource constraints of the current farm, the candidate control instructions are optimized and conflict resolved through multi-objective optimization to generate the final executable set of control instructions.

5. The intelligent aquaculture monitoring method based on big data analysis according to claim 4, characterized in that, The multi-objective optimization and conflict resolution of the candidate control instructions includes: Establish a multi-objective optimization function that includes animal welfare indicators, energy consumption indicators, feed conversion rate indicators, and equipment wear indicators; Obtain the real-time operating status and capacity limits of all environmental control and feeding equipment in the current farm, and use them as constraints for the optimization problem. A multi-objective evolutionary algorithm is used to adjust the candidate control commands, and under the premise of satisfying the constraints, the combination of control parameters that makes the multi-objective optimization function Pareto optimal is found. The optimized control commands for different abnormal regions are coordinated globally to avoid conflicts in resource allocation or environmental field control between different regions. The coordinated optimized control instructions for each abnormal region block, along with the global coordination instructions, are packaged together into the final executable control instruction set.

6. The intelligent aquaculture monitoring method based on big data analysis according to claim 1, characterized in that, The method also includes an online self-evolutionary model step: After executing the set of control instructions, the changes in the original monitoring data are continuously monitored to form a multidimensional state image sequence of the farm after control. The multidimensional images of the farm before and after regulation, the set of regulatory instructions executed, and the actual results of animal health and production performance are used as a new training sample pair. Using new training sample pairs, the parameters of the behavior-environment decoupling analysis network are updated in an incremental learning manner, and the cross-influence evaluation matrix and the regulation policy library in the policy generation engine are optimized. Through periodic or triggered model updates, the decision-making capabilities of the entire monitoring and control system can continuously evolve as the farm operates over time.

7. A smart aquaculture monitoring system based on big data analysis, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the smart aquaculture monitoring method based on big data analysis as described in any one of claims 1 to 6.

Images