Aquaculture comprehensive guarantee method and system based on multi-modal data acquisition

Abstract

The invention discloses an aquaculture comprehensive guarantee method and system based on multi-modal data acquisition, and relates to the technical field of intelligent aquaculture, and the method comprises the steps: collecting aquaculture multi-modal data and an aquaculture image of an aquaculture region, carrying out the aquaculture feature capture of the aquaculture image, and outputting a visual feature group; inputting the visual feature group into a MobileNet lightweight model, and outputting a health state label and a confidence score; based on the health state label and the confidence score, outputting disease type data through a residual network structure optimized by transfer learning; constructing a graph neural network according to the disease type data, and generating a dynamic regulation and control scheme; and the control center executes the dynamic regulation and control scheme, carries out regulation and control effect detection and target comparison on the aquaculture area of the dynamic regulation and control scheme, judges whether the regulation and control effect is effective or not, and generates a feedback regulation instruction. According to the invention, through multi-modal data fusion and intelligent decision closed loop, dynamic and accurate guarantee of aquatic product health management is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent aquaculture, and in particular to an aquaculture comprehensive guarantee method and system based on multi-modal data collection. BACKGROUND

[0002] Aquaculture life health monitoring technology is accelerating evolution towards multi-modal perception and intelligent decision fusion. Current mainstream solutions rely on high-resolution visual sensors to collect biological posture behavior features, combine multi-parameter water quality sensor arrays and metagenomic sequencing technology to build a multi-source data collection system, and realize preliminary fusion screening of multi-modal data based on lightweight MobileNet model. Graph neural networks are gradually applied to environmental factor correlation analysis. Edge computing terminal deployment significantly reduces response delay, and multi-modal feature fusion algorithms promote health assessment to evolve from single threshold judgment to dynamic probability prediction, providing a technical foundation for intelligent aquaculture management.

[0003] However, existing technologies have limitations in data fusion depth and decision real-time. Traditional solutions usually process each modal data independently, without establishing a dynamic weight distribution mechanism, resulting in weakened cross-modal correlation between visual features and water quality parameters, insufficient confidence in lightweight model screening results, and reliance on preset rules to trigger control actions at the decision level, lack of Monte Carlo path optimization based on biological state-environment factor coupling, and difficulty in adapting to non-linear disease scenarios such as vibrio outbreaks. SUMMARY

[0004] In view of the above existing problems, the present application is proposed.

[0005] Therefore, the present application provides an aquaculture comprehensive guarantee method based on multi-modal data collection to solve the problem of limitations in data fusion depth and decision real-time.

[0006] To solve the above technical problems, the present application provides the following technical solutions:

[0007] In a first aspect, the present application provides a multi-modal data acquisition comprehensive guarantee method for aquaculture, which comprises: acquiring aquaculture multi-modal data and aquaculture images of an aquaculture area; capturing aquaculture features of the aquaculture images to output a visual feature group; inputting the visual feature group into a MobileNet lightweight model, combining water quality parameters and metagenomic data for multi-modal screening to output a health status label and a confidence score; based on the health status label and the confidence score, identifying a diseased area of the aquaculture through a residual network structure optimized by transfer learning and integrating a channel-space dual-path attention mechanism to output disease type data; constructing a graph neural network according to the disease type data to quantize multi-dimensional dynamic correlations between biological states, environmental parameters and control actions, generating a disease control path, and optimizing the disease control path through Monte Carlo simulation to generate a dynamic control scheme; a control center executes the dynamic control scheme, and detects and compares the aquaculture area of the dynamic control scheme to determine whether the control effect is effective, and generates a feedback adjustment instruction.

[0008] As a preferred scheme of the multi-modal data acquisition comprehensive guarantee method for aquaculture, the aquaculture features of the aquaculture images are captured to output the visual feature group, and the specific steps are as follows:

[0009] YOLOv8 target detection is performed on the aquaculture images to output boundary box coordinates;

[0010] Instance segmentation is performed on the aquaculture images based on the boundary box coordinates of the aquaculture to generate aquaculture segmentation masks;

[0011] RGB color values of the hepatopancreas region and pixel distribution of the gill region of the aquaculture are extracted from the aquaculture segmentation masks, and the RGB color values and the pixel distribution are mapped to body features to generate biological body features;

[0012] The swimming speed fluctuation of the biological body is analyzed according to the fixed time period boundary box coordinates to generate biological behavior features;

[0013] The biological body features and the biological behavior features are compared with the previous adjacent biological body features and biological behavior features to determine the disease trend, and the visual feature group is generated.

[0014] As a preferred scheme of the multi-modal data acquisition comprehensive guarantee method for aquaculture, the visual feature group is input into the MobileNet lightweight model, the water quality parameters and the metagenomic data are combined for multi-modal screening to output the health status label and the confidence score, and the specific steps are as follows:

[0015] The visual feature group, the water quality parameters and the metagenomic data are spatio-temporally aligned to generate aligned data;

[0016] The aligned data is input into the MobileNet lightweight model, the visual feature groups and water quality parameters are dynamically correlated, and pathogen information is screened from metagenomic data to output a weighted reconstructed feature vector.

[0017] A depthwise separable convolution is performed on the weighted reconstructed feature vectors to generate a high-dimensional feature map.

[0018] The global lesion features of the high-dimensional feature map are aggregated by spatial average pooling, and the lesion dataset is output.

[0019] Perform a three-dimensional health mapping on the disease dataset and output health status labels and confidence scores.

[0020] As a preferred embodiment of the multimodal data acquisition method for comprehensive aquaculture protection described in this invention, the method involves: identifying diseased areas in aquatic organisms based on health status labels and confidence scores, using a residual network structure optimized through transfer learning, and integrating a channel-spatial dual-path attention mechanism to output disease type data. The specific steps are as follows:

[0021] The health status labels and confidence scores are concatenated with historical lesion features from historical lesion data in a time sequence to generate a diagnostic feature vector.

[0022] The diagnostic feature vector is input into the residual network structure optimized by transfer learning. The diagnostic feature vector is then convolved and skipped through multiple residual blocks to output a multi-scale initial feature map.

[0023] The lesion channel dimension in the initial scale feature map is weighted using a channel attention mechanism, and a channel-weighted feature map is output.

[0024] Microlesion coordinates are located using a spatial attention mechanism on the channel-weighted feature map, and the lesion feature map is output.

[0025] Perform fully connected classification on the lesion feature map to identify the disease status of aquatic organisms and output disease type data.

[0026] As a preferred embodiment of the multimodal data acquisition comprehensive aquaculture protection method of the present invention, the residual network structure optimized by transfer learning refers to a residual network structure adapted to the aquatic disease identification task by reconstructing the top-level classifier by adding a channel-space dual-path attention mechanism within the basic framework based on the ResNet-50 model.

[0027] As a preferred embodiment of the multimodal data acquisition method for comprehensive aquaculture protection described in this invention, the specific steps of constructing a graph neural network based on disease type data to quantify the multidimensional dynamic correlation between biological state, environmental parameters, and regulatory actions, and generating disease regulation pathways, are as follows:

[0028] extracting the disease type and disease probability of the aquatic product from the disease type data as the biological state of the aquatic product;

[0029] combining the water quality parameters and the metagenomic data to generate the environmental parameters;

[0030] extracting the control actions and control results corresponding to different disease types from the historical lesion data;

[0031] taking the biological state, the environmental parameters and the control actions as biological state nodes, environmental parameter nodes and control action nodes of the graph neural network, and connecting the biological state nodes, the environmental parameter nodes and the control action nodes to generate node association edges;

[0032] matching the biological state nodes and the environmental parameter nodes to generate biological-environmental nodes;

[0033] matching the control action nodes and the biological state nodes to generate control-biological nodes;

[0034] performing graph convolution on the biological-environmental nodes and the control-biological nodes to quantify the dynamic dependency relationship between the biological state nodes and the environmental parameter nodes, and the intervention effect of the control action nodes on the biological state nodes, and generate node weights;

[0035] starting from the current biological state node, traversing the graph neural network through breadth-first search to generate a disease control path.

[0036] As a preferred scheme of the multi-modal data acquisition-based comprehensive guarantee method for aquaculture, the disease control path is optimized through Monte Carlo simulation to generate a dynamic control scheme, and the specific steps are,

[0037] randomly perturbing the execution parameters of the control actions in the disease control path to generate perturbed execution parameters;

[0038] combining the perturbed execution parameters, the node weights and the control results in the disease type data to perform dynamic association simulation to generate a control path execution scenario;

[0039] performing health recovery probability evaluation on the control path execution scenario, and selecting the control path execution scenario with the highest health recovery probability evaluation as the dynamic control scheme.

[0040] As a preferred scheme of the multi-modal data acquisition-based comprehensive guarantee method for aquaculture, the control center executes the dynamic control scheme, and detects and compares the control effect of the dynamic control scheme on the aquaculture area to determine whether the control effect is effective, and generates a feedback adjustment instruction, and the specific steps are,

[0041] The control center executes the dynamic regulation scheme through the control equipment of the aquaculture area;

[0042] The water quality parameters, biological state characteristics and metagenome data after the dynamic regulation scheme is executed are collected by the sensor network at fixed time intervals;

[0043] The water quality parameters, biological state characteristics and metagenome data are compared with the regulation target to generate a health recovery probability;

[0044] If the health recovery probability exceeds a preset recovery probability threshold, it is determined that the regulation effect is effective;

[0045] If the health recovery probability does not exceed the preset recovery probability threshold, it is determined that the regulation effect is ineffective, and a feedback adjustment instruction is generated.

[0046] As a preferred scheme of the multi-modal data acquisition comprehensive guarantee method for aquaculture, the multi-modal data for aquaculture includes historical lesion data, water quality parameters and metagenome data.

[0047] In a second aspect, the application provides a multi-modal data acquisition comprehensive guarantee system for aquaculture, comprising,

[0048] The acquisition module is configured to acquire multi-modal data for aquaculture and aquaculture images of the aquaculture area, capture water features from the aquaculture images, and output a visual feature group;

[0049] The screening module is configured to input the visual feature group into a MobileNet lightweight model, combine the water quality parameters and the metagenome data for multi-modal screening, and output a health state label and a confidence score;

[0050] The recognition module is configured to recognize a lesion area of the aquatic product based on the health state label and the confidence score, through a residual network structure optimized by transfer learning, and integrate a channel-space dual-path attention mechanism to output disease type data;

[0051] The regulation module is configured to construct a graph neural network based on the disease type data, quantify a multi-dimensional dynamic correlation between biological states, environmental parameters and regulation actions, generate a disease regulation path, and optimize the disease regulation path through Monte Carlo simulation to generate a dynamic regulation scheme;

[0052] The feedback module is configured to control the control center to execute the dynamic regulation scheme, detect and compare the regulation effect of the dynamic regulation scheme on the aquaculture area, determine whether the regulation effect is effective, and generate a feedback adjustment instruction.

[0053] The present application has the beneficial effects that: through multi-modal data fusion and intelligent decision closed loop, dynamic and accurate guarantee of aquatic health management is realized. Through the MobileNet light model, dynamic correlation and pathogen screening are performed to generate a weighted reconstruction feature vector, support early abnormal screening, and the residual network integrated channel-space dual-channel attention mechanism optimized by transfer learning is used to locate and capture micro-lesion areas, realize accurate identification of subtle lesions and generation of disease classification atlas, break through the limitations of traditional artificial experience, construct a graph neural network based on disease type data, quantify the dynamic correlation of biological state, environmental parameters and control actions, solve the response lag problem of complex disease scenes, and the closed loop execution link compares and adjusts the control target in real time through multi-source feedback data, drives the incremental learning optimization model parameters, forms a cross-cycle self-calibration mechanism, and achieves the continuous evolution of aquatic health guarantee. BRIEF DESCRIPTION OF DRAWINGS

[0054] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0055] Fig. 1 The flow chart of the multi-modal data acquisition comprehensive guarantee method for aquatic culture.

[0056] Fig. 2 The schematic diagram of the multi-modal data acquisition comprehensive guarantee system for aquatic culture.

[0057] Fig. 3 The flow chart of multi-modal screening.

[0058] Fig. 4 The flow chart of the closed loop control feedback mechanism. DETAILED DESCRIPTION

[0059] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the specific embodiments of the present application will be described in detail below with reference to the drawings of the specification.

[0060] In the following description, many specific details are set forth in order to provide a thorough understanding of the present application, but the present application can also be implemented in other ways different from those described herein, and those skilled in the art can make similar generalizations without departing from the connotation of the present application, therefore the present application is not limited to the specific embodiments disclosed below.

[0061] Second, the "one embodiment" or "embodiments" referred to herein can include a specific feature, structure, or characteristic in at least one implementation of the application. "In one embodiment" appearing in various places in the specification does not all refer to the same embodiment, nor are separate or alternative embodiments mutually exclusive of other embodiments.

[0062] Referring to Figs. 1-4 For one embodiment of the application, the embodiment provides a comprehensive guarantee method for multi-modal data acquisition of aquaculture, comprising the following steps:

[0063] S1, collect the multi-modal data of aquaculture in the aquaculture area and the aquaculture images, capture the aquaculture features of the aquaculture images, and output the visual feature group.

[0064] Collect historical lesion data, water quality parameters, metagenomic data, and aquaculture images in the aquaculture area;

[0065] Specifically, the historical lesion data is mainly collected through long-term pathological observation, laboratory detection, and breeding records in the aquaculture area; Specifically, it includes: historical lesion characteristics such as disease type, lesion site, lesion area, and mortality rate, etc., which are obtained through on-site observation, microscope examination, and pathogen isolation and culture methods; The control action and control result when the historical lesion occurs, that is, each intervention measure (such as drug dosage, oxygenation machine starting time, and water change frequency, etc.) and the corresponding effect (such as dissolved oxygen change, pathogen abundance reduction ratio, and aquaculture survival rate improvement amplitude, etc.);

[0066] Water quality parameters (such as dissolved oxygen, pH value, ammonia nitrogen, and nitrite, etc.) are collected in real time by a sensor network deployed in the aquaculture area; The sensor network (such as a dissolved oxygen sensor, a pH sensor, and an ammonia nitrogen detector) is directly immersed in the water body, continuously monitors and transmits the water quality parameters to the data center; For example, the dissolved oxygen sensor measures the oxygen content in the water by electrochemical principle, and the pH sensor detects the hydrogen ion concentration by glass electrode; In addition, some water quality parameters (such as nitrite) need to collect water samples regularly, and perform offline detection by laboratory spectrophotometric method or electrochemical analyzer;

[0067] The metagenomic data is generated by detecting the water sample through a chemical laboratory; The metagenomic data includes information related to pathogenic bacteria, algal content, and microbial distribution;

[0068] The aquaculture image is collected by a underwater 4K camera (3840x2160@30fps) at regular intervals or in real time; The 4K camera is installed in the aquaculture pond or net cage to shoot the three-dimensional distribution image of the aquatic products.

[0069] YOLOv8 target detection is performed on the aquaculture image to output the bounding box coordinates;

[0070] Specifically, the collected aquaculture images are subjected to YOLOv8 target detection, the aquaculture image contour features are extracted through a convolutional neural network, and the bounding box coordinates of the aquaculture individuals are drawn in the aquaculture image, each bounding box corresponding to an aquaculture individual; the YOLOv8 target detection is based on an anchor box mechanism and a non-maximum suppression (NMS) algorithm, through filtering the overlapping or redundant bounding boxes, ensuring that each aquaculture individual corresponds to only accurate bounding box coordinates, and outputting the bounding box coordinates;

[0071] Based on the bounding box coordinates of the aquatic products, the aquaculture images are subjected to instance segmentation to generate aquatic product segmentation masks;

[0072] The bounding box coordinates are used as input to call an instance segmentation model (such as Mask R-CNN) to perform pixel-level segmentation on the aquaculture images; the instance segmentation model identifies the accurate contours of the aquatic individuals through the pixel distribution within the bounding box, and generates corresponding binary segmentation masks (0 represents background and 1 represents aquatic area); the segmentation mask is consistent with the size of the original image, and each pixel point is marked as belonging to an aquatic individual; after segmentation, the segmentation mask is superimposed with the original image to ensure that the segmentation area of each aquatic individual corresponds to the bounding box coordinates, providing accurate spatial positioning for subsequent liver and pancreas region and gill region extraction.

[0073] The RGB color values of the aquatic liver and pancreas region and the pixel distribution of the gill region are extracted from the aquatic product segmentation mask, and the RGB color values and the pixel distribution are subjected to body posture feature mapping to generate biological body posture features;

[0074] Specifically, based on the aquaculture individual contours in the segmentation mask, the liver and pancreas region and the gill region are located through an aquaculture organ coordinate system (such as a coordinate grid with the head of the aquatic product as the origin); the RGB color values of the liver and pancreas region, i.e. the mean and standard deviation of the red, green and blue channels, are extracted; for the gill region, the gray histogram and texture features (such as the local binary pattern LBP histogram) of the gill region are extracted; after extraction, the RGB color values, the gray histogram and the LBP histogram are input into a predefined body posture feature mapping template (such as the liver and pancreas color difference index and the gill texture abnormality degree) for body posture feature mapping to generate biological body posture features; the biological body posture features are output in the form of a numerical vector for subsequent disease trend analysis;

[0075] Further, the body state feature mapping template is set based on the correlation of aquatic pathology and historical lesion data; the specific setting process is as follows: first, the labeled healthy and lesion samples (such as liver and pancreas jaundice and cheek ulceration) are extracted from the historical lesion data, and the RGB color value, gray histogram distribution and LBP histogram feature corresponding to the lesion sample are counted; then, the contribution weight of each feature to the lesion diagnosis is determined by a machine learning method (such as principal component analysis PCA or random forest feature importance sorting) (for example, the red channel mean value accounts for 50% in the liver and pancreas color difference index, the green channel mean value accounts for 30%, and the blue channel mean value accounts for 20%); finally, the biological states are weighted and fused according to the contribution weight, and a pre-defined body state feature mapping template is generated.

[0076] According to the fixed time period boundary box coordinates, the swimming speed fluctuation of the biological body is analyzed, and the biological behavior feature is generated;

[0077] The boundary box coordinates of each aquatic individual are recorded at a fixed time interval, and the displacement distance of the center point of the boundary box in adjacent time periods is obtained by the difference method; the instantaneous swimming speed of the biological body is obtained by the displacement distance and the time interval; the instantaneous swimming speed at multiple time points is counted, and the average swimming speed, the maximum swimming speed, the swimming speed fluctuation variance and the motion direction change frequency (such as the number of turns per unit time) of the aquatic individual are output;

[0078] The biological state feature and the biological behavior feature are compared with the previous adjacent biological state feature and the biological behavior feature, the lesion trend is judged, and a visual feature group is generated;

[0079] The numerical difference comparison between the biological state feature at the current time and the biological state feature in the previous fixed period is performed, and the body state change rate is obtained; similarly, the difference comparison between the biological behavior feature at the current time and the biological behavior feature in the previous period is performed, and the behavior change rate is obtained; based on the body state change rate and the behavior change rate (such as liver and pancreas color difference ΔE>5, cheek lesion area>2mm2) and the biological behavior feature (swimming speed mutation>50%), the lesion trend of the aquatic product is judged; the judgment result is integrated with the current biological state feature and the biological behavior feature, and a visual feature group containing the lesion trend label such as normal, early lesion, acute lesion and corresponding feature value is generated, which provides input data for subsequent disease diagnosis.

[0080] S2, input the visual feature group into the MobileNet light model, combine the water quality parameters and the metagenomic data for multi-modal screening, and output the health status label and the confidence score.

[0081] The visual feature group, the water quality parameters and the metagenomic data are spatio-temporally aligned to generate aligned data;

[0082] Specifically, the time dimension matching is performed according to the timestamps in the visual feature group and the real-time collection time of the water quality parameter sensor and the sampling time of the metagenomic data, the time interval difference is filled by an interpolation method (such as linear interpolation or spline interpolation), and the visual feature group, the water quality parameter and the metagenomic data are synchronized at the same time point; in the spatial dimension, the coordinate information of the aquatic individual in the visual feature group is spatially mapped with the layout position of the water quality parameter sensor, and the average value of the water quality parameter of the region where each aquatic individual is located is obtained by distance weighted average method; finally, the microbial abundance information (such as pathogenic bacteria species and proportion) in the metagenomic data is bound with the visual feature group and the water quality parameter at the corresponding time point to form aligned data containing visual features, water quality parameters and microbial information.

[0083] The aligned data is input into the MobileNet lightweight model, the visual feature group and the water quality parameter are dynamically associated, and the pathogenic information is screened from the metagenomic data, and a weighted reconstruction feature vector is output;

[0084] Specifically, first, the visual feature group and the water quality parameter in the aligned data are input into the input layer of the MobileNet model, the MobileNet model extracts cross-modal correlation features through lightweight convolution layers (such as deep separable convolution); the dynamic association process adopts an attention mechanism to obtain the correlation weight between the visual feature group and the water quality parameter (for example, the correlation weight of the hepatopancreas color difference index and dissolved oxygen is 0.6, and the correlation weight of the swimming speed fluctuation rate and ammonia nitrogen is 0.4), and a dynamic correlation feature vector is generated by weighted fusion; secondly, the pathogenic information (such as the abundance of Vibrio and Aeromonas hydrophila) is screened from the metagenomic data, the high-risk pathogenic features are extracted by pathogenic threshold filtering, and are spliced with the dynamic correlation feature vector; finally, the spliced feature vector is input into the fully connected layer of the MobileNet, a weighted reconstruction feature vector is generated by normalization operation, and the weighted reconstruction feature vector represents the comprehensive health status of the aquatic individual in numerical form;

[0085] Further, the setting of the pathogenic threshold is based on aquatic pathology and historical lesion data, the critical value is determined by statistical analysis of the metagenomic data distribution range of healthy and lesion samples, and is dynamically optimized according to the actual aquatic farming environment (such as water temperature and dissolved oxygen) and the transmission characteristics of pathogens; for example, the aquatic farm finds through analysis that fish is prone to gill ulcer when Vibrio abundance exceeds 1%, and combines with the detection stability of the aquatic laboratory to set the pathogenic threshold to 1% as a high-risk early warning standard, and when Vibrio abundance exceeds 1%, an intervention measure is triggered.

[0086] The weighted reconstruction feature vector is subjected to deep separable convolution to generate a high-dimensional feature map;

[0087] Specifically, the weighted reconstruction feature vector is expanded into a two-dimensional tensor, which is input into the depth separable convolution layer of the MobileNet model; the depth separable convolution is divided into two steps: depth convolution: using a convolution kernel to independently extract features from each input channel of the weighted reconstruction feature vector, preserving the local spatial relationship of the original channel, and generating a convolution feature map; point convolution: mixing the channel of the convolution feature map after depth convolution through a convolution kernel to generate a high-dimensional feature map spectrum across modalities; during the convolution process, the MobileNet model enhances the non-linear expression ability through the activation function, and stabilizes the convolution process through batch normalization; the finally output high-dimensional feature map spectrum contains multiple channel lesion-related features, which are used for subsequent global feature aggregation;

[0088] The global lesion features of the high-dimensional feature map spectrum are aggregated by spatial average pooling to output the lesion data set;

[0089] The high-dimensional feature map spectrum is input into the spatial average pooling layer, and the mean value of each channel region is calculated to generate a global feature vector; for example, the mean value of the first channel region is calculated to obtain a global feature value; the mean value of the second channel region is calculated to obtain a global feature value, and the process is repeated until multiple global lesion feature values are obtained; the global lesion feature values are bound with the timestamp and the aquatic individual ID to form a structured lesion data set;

[0090] The lesion data set is mapped to three-dimensional health, and the health status label and confidence score are output;

[0091] The global lesion features of the lesion data set are reduced to a three-dimensional space, and the health, early lesion and acute lesion regions are divided in the three-dimensional space; at the same time, the confidence scores corresponding to the health, early lesion and acute lesion are calculated according to the global lesion feature values, and finally the results containing the health status label and the confidence score are output, which provides a decision basis for disease diagnosis;

[0092] The confidence score calculation formula is,

[0093] F = 1-(1+e -k(μ-θ) ) -1 ;

[0094] Where F represents the confidence score corresponding to the health, early lesion and acute lesion, e represents the natural constant, generally equal to about 2.71828; k represents the sensitivity coefficient of the confidence score to the lesion feature value change, μ represents the global lesion feature value in the high-dimensional feature map spectrum, and θ represents the health state boundary value (the boundary value of the health, early lesion and acute lesion regions divided according to the three-dimensional health mapping).

[0095] S3, based on the health status label and the confidence score, a residual network structure optimized by transfer learning is used to identify the diseased area of aquatic products by integrating channel-space dual-channel attention mechanism, and output disease type data.

[0096] The health status label and the confidence score are time-series spliced with the historical lesion features in the historical lesion data to generate a diagnosis feature vector.

[0097] Specifically, according to the health status of aquatic products at the current time and the corresponding confidence score, historical lesion features (such as liver and pancreas lesion area, gill ulceration degree and mortality rate, etc.) matching the health status of aquatic products at the current time are retrieved from historical lesion data. Then, the health status label, the confidence score and the historical lesion features are aligned by field in chronological order. For example, the health status label corresponds to the disease type field, the confidence score corresponds to the risk probability field, and the historical lesion features correspond to the lesion site and lesion area fields. A complete record containing time, health status, confidence score and historical lesion features, i.e. a diagnosis feature vector, is generated by horizontal splicing.

[0098] The diagnosis feature vector is input into the residual network structure optimized by transfer learning, and the diagnosis feature vector is convolved and connected by jump layer by layer through multiple residual blocks to output multi-scale initial feature maps.

[0099] The residual network structure optimized by transfer learning refers to a residual network structure adapted to the aquatic lesion recognition task, which is reconstructed by adding a channel-space dual-channel attention mechanism to the top classifier based on a ResNet-50 model as a basic framework. The diagnosis feature vector is expanded into a two-dimensional tensor and input into the residual network structure optimized by transfer learning based on the ResNet-50 model. The residual network structure initializes the bottom convolutional layer through the pre-trained ResNet-50 model parameters (such as ImageNet pre-training weights) to ensure that the ResNet-50 model has general feature extraction capability. Then, the diagnosis feature vector is processed layer by layer through multiple residual blocks: each residual block contains two convolutional layers (such as 3x3 convolutional kernels) and a shortcut connection (Shortcut Connection). The first layer of convolution extracts local features of aquatic products, the second layer of convolution enhances nonlinear representation of aquatic products, and the shortcut connection preserves the original input information of the diagnosis feature vector to alleviate the gradient vanishing problem. After multiple layers of convolution and jump connection, the residual network structure outputs multi-scale initial feature maps, which contain lesion-related features of different scales (such as liver and pancreas texture changes and gill ulcer edge), providing input for the subsequent channel attention mechanism.

[0100] The channel attention mechanism is used to weight the lesion channel dimension in the scale initial feature map to output a channel weighted feature map.

[0101] Specifically, the multi-scale initial feature map is input into the channel attention layer. The channel attention layer calculates the mean value of each channel region in the multi-scale initial feature map through global average pooling to generate a 64-dimensional channel weight vector. Then, the channel weight vector is nonlinearly transformed through two fully connected layers (such as the first layer with 64 neurons and the second layer with 64 neurons) to generate normalized channel attention weights. Finally, the channel attention weights are combined with the multi-scale initial feature map in the channel dimension to generate a channel weighted feature map. In the channel weighted feature map, high-weight channels (such as channels related to liver and pancreas lesions) are enhanced, and low-weight channels (such as channels related to normal tissues) are suppressed, providing input for the subsequent spatial attention mechanism.

[0102] The channel weighted feature map is subjected to micro-lesion coordinate positioning through the spatial attention mechanism to output a lesion feature map.

[0103] The channel weighted feature map is subjected to maximum branch and mean branch matching according to the spatial attention mechanism. The maximum branch: global maximum pooling is performed on the channel weighted feature map to generate a channel maximum value feature map. The mean branch: global average pooling is performed on the channel weighted feature map to generate a channel mean value feature map. Then, the channel maximum value feature map and the channel mean value feature map are concatenated in the channel dimension, and convolution operation (such as 7x7 convolution kernel) is applied for feature fusion to generate a spatial attention weight map. The spatial attention weight map is combined with the channel weighted feature map in the spatial dimension to generate a lesion feature map.

[0104] The lesion feature map is subjected to full connection classification to identify the disease state of aquatic products and output disease type data.

[0105] The lesion feature map is input into the full connection classification layer. First, global average pooling is used to compress the lesion feature map into a one-dimensional feature vector. Then, two fully connected layers (such as the first layer with 256 neurons and the second layer with disease category number) are used for classification processing. The first fully connected layer enhances the nonlinear expression ability through the activation function (such as ReLU), and the second fully connected layer generates a disease probability distribution (such as [0.1, 0.7, 0.2] representing the probabilities of Vibrio disease, parasitic disease, and other diseases, respectively) through the Softmax function. Finally, the disease type corresponding to the highest disease probability (such as Vibrio disease) is determined according to the disease probability distribution, and disease type data containing the disease type and disease probability (such as [Vibrio disease, 0.7]) is output, providing input for subsequent graph neural network construction.

[0106] The residual network structure optimized by transfer learning combined with the channel-spatial dual-channel attention mechanism significantly improves the recognition accuracy of aquatic pathological regions and the disease classification efficiency. Compared with traditional methods (such as microscope examination relying on artificial experience and PCR detection requiring complex experimental conditions), the residual network structure optimized by transfer learning combined with the channel-spatial dual-channel attention mechanism uses a pre-trained ResNet-50 model for parameter initialization, reduces the difficulty of small sample training, and preserves the global information of pathological features through multi-scale convolution and skip connection to avoid gradient disappearance. The channel attention mechanism strengthens the key pathological channels (such as liver and pancreas texture changes), and the spatial attention mechanism accurately locates the micro-pathological regions (such as the edge of the cheek ulcer), and the collaborative effect of the two mechanisms is better than that of a single attention mechanism in perceiving complex pathological features. Finally, the Softmax probability output of the fully connected classification layer realizes the automatic mapping from pathological features to disease types.

[0107] S4, constructing a graph neural network according to the disease type data, quantifying the multidimensional dynamic correlation between biological states, environmental parameters and regulation actions, generating a disease regulation path, and optimizing the disease regulation path through Monte Carlo simulation to generate a dynamic regulation scheme.

[0108] The disease type and disease probability of the aquatic product are extracted from the disease type data as the biological state of the aquatic product.

[0109] Specifically, the disease type data is used as the core information source, and the disease type data includes different disease types of aquatic products (such as bacterial diseases, viral diseases and parasitic diseases, etc.) and corresponding disease occurrence probabilities (the numerical range is 0 to 1, representing the possibility of disease occurrence). By analyzing the fields of the disease type data, the name and disease probability value of each disease type are obtained, and the name and disease probability value information of the disease type are combined into a representation vector of the biological state of the aquatic product.

[0110] The water quality parameters and metagenomic data are combined to generate environmental parameters.

[0111] The water quality parameters include physical and chemical indicators such as water temperature, pH value, dissolved oxygen concentration and ammonia nitrogen content, and the metagenomic data includes the species composition and abundance information of the microbial community in the water body. By integrating the water quality parameters and metagenomic data, the environmental parameters are generated by feature splicing: the water quality parameters are directly included in the form of numerical values, and the metagenomic data is converted into quantitative features through the mapping of microbial species-abundance. The environmental parameters comprehensively reflect the physicochemical properties and biological composition of the water body, and are used to represent the multidimensional state of the aquatic environment.

[0112] The regulation actions and regulation results corresponding to different disease types are extracted from the historical pathological data.

[0113] The historical lesion data records intervention measures taken in past aquaculture disease events, i.e., regulation actions, and final health recovery conditions, i.e., regulation results; the regulation actions include operations such as feeding specific drugs, adjusting water quality parameters (such as oxygenation and adjusting pH), and physical intervention (such as isolating sick fish); the regulation results are quantified by health recovery probability, representing the likelihood of disease relief after intervention; by extracting the regulation actions and regulation results corresponding to different disease types, a corresponding regulation relationship table of disease type, regulation action, and regulation result is established;

[0114] The biological state, the environmental parameter, and the regulation action are taken as the biological state node, the environmental parameter node, and the regulation action node of the graph neural network, and the biological state node, the environmental parameter node, and the regulation action node are connected to generate a node association edge;

[0115] Based on the biological state, the environmental parameter, and the regulation action, a node system of the graph neural network is constructed: each disease type-probability pair is taken as a biological state node, each environmental parameter vector corresponds to an environmental parameter node, and each regulation action-result pair corresponds to a regulation action node; the connection between the biological state node, the environmental parameter node, and the regulation action node is realized by a full connection mode: an association edge is established between the biological state node and the environmental parameter node, representing the dependence relationship of the biological state on the environmental parameter, and an association edge is established between the biological state node and the regulation action node, representing the intervention effect of the regulation action on the biological state; the initial weight of each association edge is obtained by the similarity of node features (such as cosine similarity), to generate graph structure data containing node and connection relationship.

[0116] The biological state node and the environmental parameter node are matched to generate a biological-environmental node;

[0117] The regulation action node and the biological state node are matched to generate a regulation-biological node;

[0118] The biological-environmental node and the regulation-biological node are subjected to graph convolution to quantify the dynamic dependence relationship between the biological state node and the environmental parameter node, and the intervention effect of the regulation action node on the biological state node, to generate node weights;

[0119] Specifically, taking the bio-environment node and the regulation-bio node as input, the graph convolution operation is performed: the feature information of the biological state node and the environmental parameter node is aggregated through the convolution layer to extract the dynamic dependency relationship between the biological state node and the environmental parameter node (for example, the probability change of a specific disease type under different environmental parameters) and the intervention effect of the regulation action node on the biological state node (for example, the recovery probability improvement range of different regulation actions on the same disease type); the dynamic dependency relationship and the intervention effect are quantitatively analyzed, that is, the higher the dynamic dependency relationship and the stronger the intervention effect, the greater the node weight; the node weight reflects the dependency strength of the biological state node and the environmental parameter node, and the intervention effect of the regulation action node on the biological state node, providing a quantitative basis for subsequent generation of disease regulation paths;

[0120] Starting from the current biological state node, the graph neural network is traversed through breadth-first search to generate a disease regulation path;

[0121] Starting from the current biological state node, the graph neural network is traversed through breadth-first search to generate a disease regulation path;

[0122] Starting from the current biological state node, the graph neural network is traversed through breadth-first search to generate a disease regulation path;

[0123] Starting from the current biological state node, the graph neural network is traversed through breadth-first search to generate a disease regulation path;

[0124] Starting from the current biological state node, the graph neural network is traversed through breadth-first search to generate a disease regulation path;

[0125] Specifically, the regulation action execution parameter after disturbance, the node weight, and the regulation result in the disease type data are taken as inputs to perform dynamic correlation simulation: first, the characteristics of the regulation action node are adjusted according to the disturbance execution parameter (such as the dose change affecting the intervention intensity); second, the influence of the environmental parameter node on the biological state node is obtained in combination with the node weight (such as the water temperature change adjusting the disease probability); finally, according to the intervention effect of the regulation action node on the biological state node and the regulation result in the disease type data, the health recovery probability distribution after the execution of the regulation path is deduced, and the regulation path execution scene is generated;

[0126] The health recovery probability of the regulation path execution scene is evaluated, and the regulation path execution scene with the highest health recovery probability evaluation is selected as the dynamic regulation scheme;

[0127] For the regulation path execution scene, the health recovery probability is taken as the core evaluation index: the health recovery probability values in each scene are sorted in ascending order; the scene with the highest health recovery probability is selected as the dynamic regulation scheme, and the dynamic regulation scheme includes a specific regulation action sequence, execution parameters, and an expected health recovery probability.

[0128] S5, the control center executes the dynamic regulation scheme, and detects and compares the regulation effect of the dynamic regulation scheme on the aquaculture area to determine whether the regulation effect is effective, and generates a feedback adjustment instruction.

[0129] The control center executes the dynamic regulation scheme through the control equipment of the aquaculture area;

[0130] Specifically, after receiving the dynamic regulation scheme, the control center sends instructions to the control equipment of the aquaculture area according to the specific regulation action sequence and execution parameters in the dynamic regulation scheme to ensure that each control equipment runs synchronously according to the specified parameters; the regulation actions included in the dynamic regulation scheme include but are not limited to: the intelligent AGV physically isolates the sick fish, the RAS adjusts the water environmental parameters (such as increasing the water exchange rate or adjusting the dissolved oxygen content), the UV sterilization equipment starts the timed sterilization, the automatic drug feeding equipment feeds the treatment drug according to the dose, and the food and medicine treatment device releases specific nutrient ingredients; the control center transmits the instructions to the corresponding control equipment through the Internet of Things communication protocol (such as MQTT or HTTP); the control center first sends a moving instruction to the intelligent AGV, positions to the breeding tank where the sick aquatic product is located, and completes the isolation operation, then sends a water exchange rate adjustment instruction to the RAS to increase the water exchange rate to 5 m 3 / h, at the same time, sends a start signal to the UV sterilization equipment to set the sterilization time, and finally sends the drug dose and feeding frequency parameters to the automatic drug feeding equipment to ensure that the drug is accurately delivered to the target area; during the execution process, the control center continuously monitors the running state of each regulation equipment to ensure the integrity and timeliness of the regulation action.

[0131] After the dynamic regulation scheme is executed, the water quality parameters, biological state characteristics and metagenome data after the execution of the dynamic regulation scheme are collected by the sensor network at fixed time intervals;

[0132] The water quality parameters, biological state characteristics and metagenome data are compared with the regulation target to generate a health recovery probability;

[0133] Specifically, the water quality parameters, biological state characteristics and metagenome data are compared with the regulation target set in the dynamic regulation scheme item by item; the regulation target includes the target value of the water quality parameter, the improvement index of the biological state characteristic and the pathogen concentration of the metagenome data; the target comparison process is divided into three stages: the first stage is water quality parameter comparison, the real-time collected parameters such as dissolved oxygen, pH value and ammonia nitrogen are compared with the target value one by one to obtain the water quality deviation rate; the second stage is biological state characteristic comparison, the biological state characteristics are compared with the health benchmark value to analyze the difference, and the health state change trend is evaluated combined with the biological behavior characteristics to generate a biological state recovery rate, for example, if the liver and pancreas color difference ΔE value decreases from 7 to 3.5 and the swimming speed fluctuation rate decreases from 50% to 10%, it is determined that the body state and behavior characteristics are significantly improved; the third stage is metagenome data comparison, the metagenome data are compared with the regulation target to output the metagenome regulation probability, for example, the pathogen abundance is compared with the pathogen abundance before adjustment to determine whether the pathogen is effectively inhibited; the comparison results are weighted and fused to generate a health recovery probability;

[0134] If the health recovery probability exceeds the preset recovery probability threshold, it is determined that the regulation effect is effective;

[0135] Specifically, the health recovery probability is compared with the preset recovery probability threshold, and the recovery probability threshold is determined according to the statistical analysis of historical regulation cases (for example, set to 0.8, that is, the health recovery probability ≥ 80% is considered effective); if the health recovery probability is higher than the recovery probability threshold, it is determined that the regulation effect is effective, indicating that the current dynamic regulation scheme has successfully improved the health state of the aquatic product, and the control center will generate a regulation effect effective mark, record the execution parameters and result data of this regulation, and upload them to the historical lesion data for data updating of the historical lesion data;

[0136] Further, the recovery probability threshold is determined by analyzing the health recovery probability distribution of successful and failed historical regulation data; for example, if the average recovery probability of historical successful cases is 78% and the standard deviation is 5%, the recovery probability threshold can be set to 83% (mean + 1 times standard deviation) to ensure that most effective regulations are identified;

[0137] If the health recovery probability does not exceed the preset recovery probability threshold, it is determined that the regulation effect is ineffective, and a feedback adjustment instruction is generated;

[0138] If the health recovery probability does not reach the preset threshold, it is determined that the current dynamic regulation scheme is insufficient, and a feedback adjustment mechanism needs to be started. The control center generates a feedback adjustment instruction based on the target comparison result. Specifically, after the dynamic regulation scheme is executed, the water quality parameters, biological state characteristics and metagenomic data collected by the sensor network at fixed time intervals are re-input into the graph neural network for iterative judgment until the improvement and treatment of the water product disease are achieved. The feedback adjustment instruction is issued to the control center and the execution equipment through the Internet of Things communication protocol to ensure rapid iteration of the regulation scheme. The closed-loop control mechanism dynamically optimizes the regulation strategy through the cycle of real-time data collection, comparison, feedback and adjustment, and finally realizes precise intervention of the water product health status.

[0139] The embodiment also provides a water product breeding comprehensive guarantee system for multi-modal data collection, comprising:

[0140] A collection module is configured to collect water product breeding multi-modal data and breeding images of a water product breeding area, capture water product features of the breeding images, and output a visual feature group.

[0141] A screening module is configured to input the visual feature group into a MobileNet lightweight model, perform multi-modal screening in combination with water quality parameters and metagenomic data, and output a health state label and a confidence score.

[0142] An identification module is configured to identify a disease area of the water product based on the health state label and the confidence score, through a residual network structure optimized by transfer learning, and integrate a channel-space dual-path attention mechanism to output disease type data.

[0143] A regulation module is configured to construct a graph neural network based on the disease type data, quantify multi-dimensional dynamic correlations between biological states, environmental parameters and regulation actions, generate a disease regulation path, and optimize the disease regulation path through Monte Carlo simulation to generate a dynamic regulation scheme.

[0144] A feedback module is configured to control the control center to execute the dynamic regulation scheme, detect and compare a regulation effect of the water product breeding area of the dynamic regulation scheme, determine whether the regulation effect is effective, and generate a feedback adjustment instruction.

[0145] The embodiment also provides a computer device suitable for the water product breeding comprehensive guarantee method for multi-modal data collection, comprising a memory and a processor. The memory is configured to store computer executable instructions, and the processor is configured to execute the computer executable instructions to implement the water product breeding comprehensive guarantee method for multi-modal data collection as described in the above embodiment.

[0146] The computer device can be a terminal, and the computer device includes a processor, a memory, a communication interface, a display screen and an input device connected by a system bus. The processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for running the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is configured to perform wired or wireless communication with an external terminal. The wireless communication can be achieved by WIFI, a carrier network, NFC (Near Field Communication) or other technologies. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer overlaid on the display screen, or a key, a trackball or a touchpad arranged on the shell of the computer device, or an external keyboard, a touchpad or a mouse.

[0147] The embodiment also provides a storage medium having a computer program stored thereon, and the computer program is executed by a processor to implement the method for realizing multi-modal data acquisition of aquaculture comprehensive guarantee provided in the above embodiment. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as a static random access memory (SRAM), an electrically erasable programmable read-only memory (EEPROM), an erasable programmable read-only memory (EPROM), a programmable read-only memory (PROM), a read-only memory (ROM), a magnetic storage, a flash memory, a magnetic disk or an optical disk.

[0148] To sum up, the present application realizes dynamic and accurate guarantee of aquatic health management through multi-modal data fusion and intelligent decision-making closed loop. Through the MobileNet light model, dynamic correlation and pathogen screening are performed to generate a weighted reconstruction feature vector, supporting early abnormality screening. The residual network integrated channel-space dual-channel attention mechanism optimized by transfer learning locates and captures micro lesion areas, realizes accurate identification of subtle lesions and generation of disease classification atlas, breaks through the limitations of traditional artificial experience, constructs a graph neural network based on disease type data, quantifies the dynamic correlation of biological state, environmental parameters and control actions, solves the response lag problem of complex disease scenes, and realizes real-time comparison of control targets through multi-source feedback data in the closed loop execution link, drives the incremental learning optimization model parameters, forms a cross-cycle self-calibration mechanism, and achieves the continuous evolution of aquatic health guarantee.

[0149] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and not to limit it. Although the present application has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solutions of the present application can be modified or replaced equivalently without departing from the spirit and scope of the technical solutions of the present application, which should be covered in the scope of the claims of the present application.

Claims

1. A comprehensive aquaculture support method for multimodal data acquisition, characterized in that: include, Collect multimodal data and images of aquaculture in aquaculture areas, capture aquatic features from the aquaculture images, and output visual feature groups; The visual feature set is input into the MobileNet lightweight model, and multimodal screening is performed by combining water quality parameters and metagenomic data to output health status labels and confidence scores. Based on health status labels and confidence scores, a residual network structure optimized through transfer learning is used to integrate a channel-spatial dual-path attention mechanism to identify diseased areas in aquatic organisms and output disease type data. A graph neural network is constructed based on disease type data to quantify the multidimensional dynamic relationship between biological state, environmental parameters and regulatory actions, generate disease regulation pathways, and optimize the disease regulation pathways through Monte Carlo simulation to generate dynamic regulation schemes. The control center executes the dynamic control plan and monitors the control effect of the plan in the aquaculture area, comparing it with the target to determine whether the control effect is effective and generating feedback adjustment instructions.

2. The comprehensive aquaculture support method for multimodal data acquisition as described in claim 1, characterized in that: The specific steps for capturing aquatic features from aquaculture images and outputting visual feature sets are as follows: Perform YOLOv8 object detection on aquaculture images and output bounding box coordinates; Based on the bounding box coordinates of aquatic products, instance segmentation is performed on the aquaculture image to generate an aquatic product segmentation mask; The RGB color values ​​of the hepatopancreas region and the pixel distribution of the gill region are extracted from the aquatic segmentation mask, and the RGB color values ​​and pixel distribution are mapped to the body features to generate biological body features. Analysis of organism swimming speed fluctuations based on bounding box coordinates over a fixed time period generates biological behavioral characteristics. By comparing the physical and behavioral characteristics of an organism with those of its immediate neighbors, the trend of disease can be determined, and a set of visual features can be generated.

3. The comprehensive aquaculture support method for multimodal data acquisition as described in claim 1, characterized in that: The steps involve inputting visual feature sets into the MobileNet lightweight model, combining water quality parameters and metagenomic data for multimodal screening, and outputting health status labels and confidence scores. The visual feature set, water quality parameters and metagenomic data are spatiotemporally aligned to generate aligned data. The aligned data is input into the MobileNet lightweight model, the visual feature groups and water quality parameters are dynamically correlated, and pathogen information is screened from metagenomic data to output a weighted reconstructed feature vector. A depthwise separable convolution is performed on the weighted reconstructed feature vectors to generate a high-dimensional feature map. The global lesion features of the high-dimensional feature map are aggregated by spatial average pooling, and the lesion dataset is output. Perform a three-dimensional health mapping on the disease dataset and output health status labels and confidence scores.

4. The comprehensive aquaculture support method for multimodal data acquisition as described in claim 1, characterized in that: The method involves identifying diseased areas in aquatic organisms based on health status labels and confidence scores, using a residual network structure optimized through transfer learning, and integrating a channel-spatial dual-path attention mechanism to output disease type data. The specific steps are as follows: The health status labels and confidence scores are concatenated with historical lesion features from historical lesion data in a time sequence to generate a diagnostic feature vector. The diagnostic feature vector is input into the residual network structure optimized by transfer learning. The diagnostic feature vector is then convolved and skipped through multiple residual blocks to output a multi-scale initial feature map. The lesion channel dimension in the initial scale feature map is weighted using a channel attention mechanism, and a channel-weighted feature map is output. Microlesion coordinates are located using a spatial attention mechanism on the channel-weighted feature map, and the lesion feature map is output. Perform fully connected classification on the lesion feature map to identify the disease status of aquatic organisms and output disease type data.

5. The comprehensive aquaculture support method for multimodal data acquisition as described in claim 4, characterized in that: The transfer learning-optimized residual network structure refers to a residual network structure adapted to aquatic disease identification tasks, which is based on the ResNet-50 model as the basic framework and reconstructs the top-level classifier by adding a channel-space dual-path attention mechanism within the basic framework.

6. The comprehensive aquaculture support method for multimodal data acquisition as described in claim 1, characterized in that: The process involves constructing a graph neural network based on disease type data to quantify the multidimensional dynamic relationships between biological states, environmental parameters, and regulatory actions, thereby generating disease regulation pathways. The specific steps are as follows: Extract the disease types and probabilities of aquatic organisms from disease type data as the biological status of aquatic organisms; Water quality parameters and metagenomic data are combined to generate environmental parameters; Extract regulatory actions and results corresponding to different disease types from historical lesion data; Biological state, environmental parameters, and regulatory actions are used as biological state nodes, environmental parameter nodes, and regulatory action nodes in a graph neural network. These nodes are then connected to generate node association edges. Match biological state nodes and environmental parameter nodes to generate biological-environment nodes; Match regulatory action nodes with biological state nodes to generate regulatory-biological nodes; Graph convolution is performed on biological-environment nodes and regulation-biological nodes to quantify the dynamic dependency between biological state nodes and environmental parameter nodes, as well as the intervention effect of regulation action nodes on biological state nodes, and to generate node weights. Starting from the current biological state node, a disease regulation path is generated by traversing the graph neural network through breadth-first search.

7. The comprehensive aquaculture support method for multimodal data acquisition as described in claim 1, characterized in that: The optimization of disease regulation pathways through Monte Carlo simulation to generate dynamic regulation schemes involves the following steps: Randomly perturb the execution parameters of the regulatory actions in the disease regulation pathway to generate perturbation execution parameters; By combining the perturbation execution parameters, node weights, and regulation results in disease type data, dynamic correlation simulation is performed to generate regulation path execution scenarios; A health recovery probability assessment is conducted for each control path execution scenario, and the control path execution scenario with the highest health recovery probability assessment is selected as the dynamic control scheme.

8. The comprehensive aquaculture support method for multimodal data acquisition as described in claim 1, characterized in that: The control center executes a dynamic control plan and monitors the control effect of the plan in the aquaculture area, comparing it with the target to determine its effectiveness. It then generates feedback adjustment instructions. The specific steps are as follows: The control center implements dynamic regulation and control schemes through the control equipment in the aquaculture area; Water quality parameters, biological state characteristics, and metagenomic data are collected at fixed time intervals using a sensor network after the implementation of a dynamic control scheme. By comparing water quality parameters, biological state characteristics, and metagenomic data with regulatory targets, a probability of health recovery is generated. If the probability of health recovery exceeds the preset recovery probability threshold, the regulation is deemed effective. If the probability of health recovery does not exceed the preset recovery probability threshold, the control effect is deemed ineffective, and a feedback adjustment instruction is generated.

9. The comprehensive aquaculture support method for multimodal data acquisition as described in claim 1, characterized in that: The aquaculture multimodal data includes historical lesion data, water quality parameters, and metagenomic data.

10. A multimodal data acquisition integrated aquaculture support system, based on the multimodal data acquisition integrated aquaculture support method according to any one of claims 1 to 9, characterized in that: include, The acquisition module is used to collect multimodal data and images of aquaculture in the aquaculture area, capture aquatic features from the aquaculture images, and output visual feature groups. The screening module is used to input visual feature sets into the MobileNet lightweight model, combine water quality parameters and metagenomic data to perform multimodal screening, and output health status labels and confidence scores. The identification module is used to identify diseased areas in aquatic organisms based on health status labels and confidence scores, through a residual network structure optimized by transfer learning, and integrates a channel-spatial dual-path attention mechanism, and outputs disease type data. The regulation module is used to construct a graph neural network based on disease type data, quantify the multidimensional dynamic relationship between biological state, environmental parameters and regulatory actions, generate disease regulation paths, and optimize the disease regulation paths through Monte Carlo simulation to generate dynamic regulation schemes. The feedback module is used by the control center to execute the dynamic control plan, detect the control effect of the dynamic control plan in the aquaculture area and compare it with the target, determine whether the control effect is effective, and generate feedback adjustment instructions.

Images