Construction method and system of intelligent fishery platform based on federal learning

Abstract

The invention discloses an intelligent fishery platform construction method and system based on federated learning, and belongs to the technical field of federated learning, and the method comprises the steps: building an intelligent fishery platform to obtain fish school multi-modal data, data preprocessing, feature extraction, feature fusion, local fish school health assessment model construction, federated learning training and real-time early warning. The invention provides an efficient, reliable and safe intelligent fishery platform management solution, the problems of data islands, privacy security, model limitation and dynamic adjustment difficulty existing in the existing intelligent fishery platform are solved, and the scientificity and practicability of fish school health management are remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of federated learning, and in particular relates to a method and system for constructing a smart fishery platform based on federated learning. Background Art

[0002] Traditional fishery platforms mainly rely on basic farming techniques and experience-based operations, and these platforms usually lack advanced monitoring and management systems. These traditional methods face many limitations in data collection and processing, such as insufficient real-time data updates, lack of accurate monitoring equipment, and inefficient data utilization. In addition, traditional methods fail to introduce machine learning and deep learning technologies, making it difficult to accurately monitor and manage the complex and changing farming environment and the health status of fish stocks, thus affecting the scientific nature of decision-making and the improvement of farming efficiency.

[0003] In recent years, with the advent of the artificial intelligence era, some scholars have applied machine learning algorithms and Internet of Things technologies to fishery management. These methods attempt to improve the accuracy of fish health assessment and management efficiency by collecting and analyzing a large amount of aquaculture data. However, since the data mainly comes from a single farm and lacks a cross-platform data collaboration and sharing mechanism, the problem of data silos is serious, which limits the comprehensive utilization of data and the generalization ability of the model. In addition, existing technologies also have problems in data privacy protection, and it is difficult to ensure the security of data during transmission and sharing. Traditional fishery platforms also face challenges in data confidentiality and commercial value. For example, farmers may be reluctant to share data for fear of leaking commercial secrets, resulting in data silos and difficulty in achieving data collaboration and resource optimization within the industry. At the same time, traditional platforms can often only process single or simple data types and lack the ability to comprehensively capture multimodal data features, which limits the accuracy and reliability of health assessment models.

[0004] The shortcomings of the existing technology are the lack of cross-platform data sharing and collaboration mechanisms, making it difficult to fully utilize multimodal data for comprehensive analysis and real-time warning, and there are problems in data privacy protection. In order to overcome these problems, the present invention proposes a method and system for building a smart fishery platform based on federated learning, which can not only protect data privacy and achieve cross-platform data collaboration, but also improve the training effect and generalization ability of the model, thereby enhancing the scientificity and economic benefits of smart fishery management. This will help to achieve efficient use of data, improve the accuracy of fish health monitoring, and dynamically adjust management strategies, optimize resource allocation, and continuously improve breeding results, thereby promoting the development of the smart fishery industry. Summary of the invention

[0005] The purpose of the present invention is to provide a method and system for building a smart fishery platform based on federated learning to solve the above technical problems.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] The present invention discloses a method for constructing a smart fishery platform based on federated learning, the method comprising the following steps:

[0008] Step 1: Build a smart fishery platform to obtain multimodal data of fish schools: configure an integrated fish school monitoring module, health monitoring module, feed management module, and water quality monitoring module in the smart fishery platform to obtain multimodal data of fish schools;

[0009] Step 2, data preprocessing: The user terminal preprocesses the acquired multimodal data, and the preprocessing process includes applying anomaly detection and repair technology of deep learning algorithms to clean the data, adopting a standardization method based on a deep belief network, and using a generative adversarial network to enhance the data, so as to improve the diversity and quality of the data;

[0010] Step 3, feature extraction: extract features from the preprocessed data, and use different methods according to the data modality, including regression analysis, deep belief network, graph convolution network, variational autoencoder, long short-term memory network and multivariate statistical analysis to extract features from the data;

[0011] Step 4: Feature fusion: Use a multi-head self-attention mechanism combined with a graph convolutional network to perform multi-dimensional weighted concatenation and fusion on the extracted feature vectors. By introducing a dynamic weight adjustment mechanism, the feature weights are adaptively adjusted according to real-time data changes to form a more accurate comprehensive feature vector.

[0012] Step 5: Construction of a local fish health assessment model: A fish health assessment model is constructed by integrating feature vectors, and a convolutional neural network and a long short-term memory network are used for training to generate a fish health assessment model;

[0013] Step 6: Federated learning training: Adopt an improved federated learning algorithm, combined with adaptive optimization and reinforcement learning techniques, and conduct collaborative training through multiple smart fishery platforms to significantly improve the accuracy and generalization ability of the fish health assessment model. Homomorphic encryption technology and differential privacy protection algorithms are also applied to ensure data security and privacy.

[0014] Step 7: Real-time warning: Issue warnings based on the fish health assessment results obtained from the trained model.

[0015] Furthermore, the step 1 of establishing a smart fishery platform to collect aquaculture data includes: configuring a smart fishery platform that integrates a fish population monitoring module, a health monitoring module, a feed management module, and a water quality monitoring module;

[0016] Furthermore, multimodal data such as position data, behavioral data, body surface image data, physiological parameters, body length, weight, feed amount, eating behavior, water temperature, dissolved oxygen, pH value, and ammonia nitrogen and nitrite content of the fish school are collected from the module.

[0017] Furthermore, the preprocessing of the multimodal data in step 2 specifically includes: data cleaning, data standardization and data enhancement.

[0018] Furthermore, the data cleaning described in step 2 is to remove outliers, missing values ​​and duplicate data by applying the anomaly detection and repair technology of the deep learning algorithm, and fill the missing data by interpolation;

[0019] Furthermore, the data normalization in step 2 adopts a normalization method based on a deep belief network to normalize data of different scales for subsequent feature extraction and model training;

[0020] Furthermore, the data enhancement described in step 2 is to use a generative adversarial network to perform rotation, scaling and flipping data enhancement processing on the image data to increase the diversity and quality of the data.

[0021] Furthermore, the feature extraction of multimodal data described in step 3 is specifically as follows: for body length and weight data, regression analysis and statistical methods are used to extract features, including basic statistical features such as mean, variance, and maximum and minimum values; for health status data, a classification algorithm based on a deep belief network is used for feature extraction, and healthy, unhealthy and sub-healthy states are automatically detected and classified through multi-layer Bayesian learning; for behavioral pattern data, a hybrid model based on a graph convolutional network and a variational autoencoder is used for feature extraction to identify behavioral changes and abnormal behaviors of fish under different environmental conditions; for bait intake data, a time series analysis method based on a long short-term memory network is used for feature extraction to extract change trends, periodic characteristics and intake predictions; for water quality parameter data, multivariate statistical analysis methods and a model based on a Bayesian network are used for feature extraction to extract comprehensive characteristics of water quality parameters, and their impact on fish health is analyzed through causal inference.

[0022] Furthermore, the feature fusion method described in step 4 is specifically: using a multi-head self-attention mechanism combined with a graph convolutional network to perform multi-dimensional weighted splicing and fusion on the extracted feature data, and by introducing a dynamic weight adjustment mechanism, adaptively adjusting the feature weights according to real-time data changes to form a comprehensive feature vector as a training data set.

[0023] Furthermore, the local fish health assessment model is constructed in step 5, specifically: the comprehensive feature vector is input into the deep learning model for training to generate the fish health assessment model;

[0024] Furthermore, the health assessment model in step 5 comprehensively considers the body length, weight, health status and behavior pattern factors of the fish school, and is continuously updated through an adaptive optimization algorithm; in the health status, the health status of the fish school can be divided into: healthy, unhealthy and sub-healthy states.

[0025] Furthermore, the collaborative training of multiple smart fishery platforms using the federated learning algorithm described in step 6 includes:

[0026] (1) Each smart fishery platform independently trains a local health assessment model and aggregates the model parameters of each platform through homomorphic encryption technology;

[0027] (2) Integrate and optimize model parameters on the central server to generate a joint optimization model;

[0028] (3) After the joint optimization model is distributed to each smart fishery platform, it is dynamically adjusted using online learning technology.

[0029] The present invention also discloses a smart fishery platform system based on federated learning, the system comprising a multimodal data collection module, a data processing module, a fish health assessment model generation module, a federated learning collaborative optimization module, a real-time warning module interface module and a user terminal module;

[0030] The multimodal data collection module includes a fish monitoring module, a health monitoring module, a feed management module and a water quality monitoring module for collecting fish position data, behavior data, body surface image data, physiological parameters, body length, weight, feed quantity, feeding behavior, water temperature, dissolved oxygen, pH value and ammonia nitrogen and nitrite content;

[0031] The data processing module is used to perform preprocessing, feature extraction and feature fusion on the multimodal data;

[0032] The fish health assessment model generation module is used to generate a fish health assessment model based on the data, and train it using a method combining a convolutional neural network and a long short-term memory network;

[0033] The federated learning collaborative optimization module is used to optimize the health assessment model through collaborative training of multiple smart fishery platforms, and apply homomorphic encryption technology and differential privacy protection algorithm to ensure data security;

[0034] The real-time warning module issues warnings based on the fish health assessment results obtained from the trained model;

[0035] The interface module is used to realize communication between multiple smart fishery platforms and is responsible for uploading model parameters and real-time updating of local models during the federated learning training process;

[0036] The user terminal module is used to display the visual data to the user for easy monitoring and management.

[0037] The beneficial effects of the present invention are:

[0038] The method and system for constructing a smart fishery platform based on federated learning described in the present invention provide an efficient, reliable and secure smart fishery platform management solution, which solves the problems of data silos, privacy security, model limitations and dynamic adjustment difficulties existing in existing smart fishery platforms, and significantly improves the scientificity and practicality of fish health management. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 It is a flow chart of a method for constructing a smart fishery platform based on federated learning provided by the present invention;

[0040] Figure 2 It is a schematic diagram of a model training architecture based on federated learning provided by the present invention;

[0041] Figure 3 It is a system structure diagram of a smart fishery platform based on federated learning provided by the present invention; DETAILED DESCRIPTION

[0042] Embodiment 1

[0043] This embodiment discloses a method for constructing a smart fishery platform based on federated learning. Figure 1 As shown, the method comprises the following steps:

[0044] Step 1: Platform establishment and data acquisition: Establish a smart fishery platform to collect multimodal data of farmed fish.

[0045] In the method described in this embodiment, the smart ranch platform includes an integrated fish monitoring module, a health monitoring module, a feed management module and a water quality monitoring module, which are used to collect multimodal data of farmed fish so as to subsequently build a local fish health assessment model.

[0046] Among them, the data collected by the fish monitoring module includes: location data and behavior data.

[0047] Among them, the location data of fish schools include the real-time location and swimming trajectory of fish schools. Abnormal location distribution may be the response of fish schools to environmental changes, such as deteriorating water quality or external interference. By monitoring the location data, the breeding environment of fish schools can be optimized and the breeding efficiency can be improved. The behavioral data of fish schools include swimming speed and swimming direction. Abnormal swimming behavior may be an early sign of health problems of fish schools, such as infection with diseases or environmental stress. By monitoring behavioral data, health problems of fish schools can be detected early and timely measures can be taken.

[0048] Among them, the data collected by the health monitoring module include: body surface image data, physiological parameters, body length and weight.

[0049] Among them, the surface image data includes high-definition images of the fish surface, which are used to check for lesions or damage. Abnormal surface images may be a sign that the fish is infected with a disease or has been externally damaged. By monitoring the surface image data, health problems of the fish can be discovered early and dealt with in a timely manner. Physiological parameters include the respiratory rate and body temperature of the fish. Abnormal physiological parameters may be related to health problems of the fish, such as abnormal respiratory rate may be a sign of hypoxia or infection. By monitoring physiological parameters, the health status of the fish can be fully understood and the stability of the breeding environment can be ensured. Body length and weight are used to assess the growth status of the fish. Abnormal body length or weight growth rate may indicate a poor growth environment or health problems for the fish. By monitoring body length and weight data, breeding conditions can be optimized to ensure the normal growth of fish.

[0050] Among them, the data collected by the feed management module include: feed input amount and eating behavior.

[0051] Among them, the feed delivery volume records the daily feed delivery. Abnormal feed consumption may be a signal of feed waste or fish health problems. By monitoring the feed delivery volume, feed use can be optimized, costs can be reduced, and economic benefits can be improved. Feeding behavior includes the amount of food eaten by the fish and the speed of eating. Abnormal feeding behavior, such as reduced food intake or abnormal feeding speed, may be a sign of fish health problems. By monitoring feeding behavior, the health status of the fish can be discovered in time and the feeding strategy can be adjusted.

[0052] Among them, the data collected by the water quality monitoring module include: water temperature, dissolved oxygen, pH value, and ammonia nitrogen and nitrite content.

[0053] Among them, water temperature monitors the temperature of the water body. Abnormal water temperature may affect the growth and health of fish, such as too high or too low temperature. By monitoring the water temperature, we can ensure that the water temperature is within the appropriate range and ensure the normal growth of fish. Dissolved oxygen monitors the dissolved oxygen content in the water body. Abnormal dissolved oxygen levels may cause fish to be hypoxic, affecting their growth and health. By monitoring the dissolved oxygen content, we can maintain a good oxygen level in the water body and prevent fish from suffocating and dying. pH monitors the acidity and alkalinity of the water body. Abnormal pH values ​​may affect the physiological functions of fish, such as overly acidic or overly alkaline water quality. By monitoring the pH value, we can maintain the appropriate pH of the water body and ensure the healthy growth of fish. Ammonia nitrogen and nitrite content monitor the concentration of harmful substances in the water body. Abnormal ammonia nitrogen or nitrite content may be toxic to fish. By monitoring these harmful substances, timely measures can be taken to reduce their concentrations and maintain a good water quality environment.

[0054] Step 2: Data preprocessing: The acquired multimodal data is preprocessed, and the preprocessing process includes data cleaning, data standardization and data enhancement.

[0055] Among them, data cleaning is the application of deep learning algorithm's anomaly detection and repair technology to remove outliers, missing values ​​and duplicate data, and fill in missing data through interpolation. Data cleaning is the application of deep learning algorithm's anomaly detection and repair technology to remove outliers, missing values ​​and duplicate data, and fill in missing data through interpolation.

[0056] Among them, data standardization adopts a standardization method based on a deep belief network to normalize data of different scales to facilitate subsequent feature extraction and model training.

[0057] Among them, data enhancement is to use a generative adversarial network to rotate, scale, and flip image data to increase the diversity and quality of the data.

[0058] Step 3: Feature extraction: After preprocessing, feature extraction is performed on the data using a method suitable for each data modality.

[0059] The feature extraction method based on generative adversarial network is adopted. All preprocessed data is input into the generative adversarial network, and the features of the data are automatically extracted through the adversarial training process of the generator and the discriminator. The generator is responsible for generating realistic virtual data, while the discriminator is responsible for distinguishing between real data and generated data. In the process of multiple iterations, the generator and the discriminator compete with each other, and the generator is constantly improved, and finally it can generate virtual data that is very similar to real data. Through this adversarial training, the generative adversarial network can learn and extract complex features in the data.

[0060] Step 4: Feature fusion: Use a model based on multi-head self-attention mechanism and graph embedding technology to perform multi-dimensional weighted splicing and fusion of the extracted feature vectors, dynamically adjust the feature weights, and form a comprehensive feature vector.

[0061] In the present invention, through feature fusion, the information of different features can be comprehensively considered to provide a more comprehensive and accurate feature representation, thereby improving the performance and prediction ability of the model. Feature fusion can also reduce the redundancy between features, extract higher-level abstract features, and enhance the understanding and modeling capabilities of data. In addition, combining feature fusion with ambient temperature factors can make the feature fusion process more in line with actual scenarios and needs, improve the model's perception and response capabilities to environmental changes, and further optimize the performance and robustness of the system.

[0062] Step 5: Construction of local fish health assessment model: Determine the type, initialization parameters and hyperparameters of the local model, and train the local model using the training data set.

[0063] Based on the collected and preprocessed data and the extracted features, a local fish health assessment model is constructed. The specific steps are as follows:

[0064] (1) Determine the type of local model. Select the appropriate model structure according to different data types. For example, use a convolutional neural network to process surface image data to identify whether fish have lesions or injuries. Use a recurrent neural network or a long short-term memory network to analyze time series data such as swimming speed, breathing rate, and physiological data. Use a multilayer perceptron to process the fused feature data for comprehensive evaluation.

[0065] (2) Model initialization stage: Set the initial parameters of the model, including weights and biases. Use Xavier initialization to improve training stability. At the same time, set hyperparameters such as learning rate, batch size, and number of training rounds. The specific parameters can be adjusted by Bayesian optimization method to obtain the best performance. For example, the initial learning rate is set to 0.001, the batch size is set to 32, the number of training rounds is set to 100, and the L2 regularization parameter is set to 0.01 to prevent overfitting.

[0066] (3) Perform model training. First, the preprocessed data set is divided into a training set, a validation set, and a test set. The commonly used ratio is 70% training, 15% validation, and 15% test. During the training process, the back propagation algorithm and gradient descent method are used to optimize the model parameters. The early stopping strategy and regularization technology are used to prevent overfitting and ensure the generalization ability of the model. The loss functions used in the training process include the cross entropy loss function of image data and the mean square error (MSE) loss function of time series data. The optimization algorithm selects the Adam optimization algorithm to adaptively adjust the learning rate and accelerate the convergence speed.

[0067] Step 6: Federated learning training: A federated learning algorithm is used to coordinate training through multiple smart fishery platforms to optimize the accuracy and generalization ability of the fish health assessment model.

[0068] In this step, a federated learning algorithm is used, such as Figure 2 As shown in the figure, multiple smart fishery platforms are trained collaboratively to optimize the accuracy and generalization ability of the fish health assessment model. First, each smart fishery platform uses local data to independently train the local health assessment model. Each platform only uses local data for training to ensure data privacy and security. During the training process, homomorphic encryption technology is used to ensure the privacy and security of data during transmission.

[0069] Each platform encrypts the weight parameters obtained from local model training and sends them to the central server. The central server uses homomorphic encryption technology to aggregate the model parameters of each platform to ensure the privacy of the data of each platform. After receiving the model parameters of each platform, the central server uses the federated average algorithm (FedAvg) to integrate and optimize the model parameters and generate a joint optimization model. During the integration process, the parameters are weighted averaged considering the data volume and model quality of each platform to ensure the balance and generalization ability of the model.

[0070] After the generated joint optimization model is sent to each smart fishery platform, each platform uses online learning technology to dynamically adjust the model. Online learning technology can continuously fine-tune the model based on new data collected in real time to improve the adaptability and accuracy of the model. Regular parameter updates and model optimization are carried out between platforms, and the accuracy and generalization ability of the model are continuously improved through multiple rounds of federated learning iterations. Using real-time monitoring data, the model is dynamically adjusted and optimized so that it can respond to environmental changes in a timely manner and improve the accuracy and reliability of fish health assessment.

[0071] Through the above steps, this embodiment provides a method for building a smart fishery platform based on federated learning, which can effectively collect and analyze fish health data, build a high-precision and high-generalization fish health assessment model, and provide a scientific basis for smart fishery management.

[0072] Embodiment 2

[0073] This embodiment discloses a system of a smart fishery platform based on federated learning, such as Figure 3 As shown, the system includes a multimodal data collection module, a data processing module, a fish health assessment model generation module, a federated learning collaborative optimization module, a real-time warning module interface module and a user terminal module.

[0074] The multimodal data collection module includes a fish detection module, a health monitoring module, a feed management module and a water quality monitoring module for collecting fish location data, behavior data, body surface image data, physiological parameters, body length, weight, feed quantity, eating behavior, water temperature, dissolved oxygen, pH value and ammonia nitrogen and nitrite content.

[0075] The data processing module includes a data preprocessing module, a feature extraction module and a feature fusion module for preprocessing, feature extraction and feature fusion of the multimodal data.

[0076] The fish school health assessment model generation module is used to generate a fish school health assessment model based on the data, and train it using a method combining a convolutional neural network and a long short-term memory network.

[0077] The federated learning collaborative optimization module is used to optimize the health assessment model through collaborative training of multiple smart fishery platforms, and applies homomorphic encryption technology and differential privacy protection algorithm to ensure data security.

[0078] The real-time warning module issues warnings based on the fish health assessment results obtained from the trained model.

[0079] The interface module is used to realize communication between multiple smart fishery platforms and is responsible for uploading model parameters and real-time updating of local models during the federated learning training process.

[0080] The user terminal module is used to display the visual data to the user for easy monitoring and management.

Claims

1. A method for constructing a smart fishery platform based on federated learning, characterized in that: The method comprises: Step 1: Build a smart fishery platform to obtain multimodal data of fish schools: configure an integrated fish school monitoring module, health monitoring module, feed management module, and water quality monitoring module in the smart fishery platform to obtain multimodal data of fish schools; Step 2, data preprocessing: The user terminal preprocesses the acquired multimodal data, and the preprocessing process includes applying anomaly detection and repair technology of deep learning algorithms to clean the data, adopting a standardization method based on a deep belief network, and using a generative adversarial network to enhance the data, so as to improve the diversity and quality of the data; Step 3, feature extraction: extract features from the preprocessed data, and use different methods according to the data modality, including regression analysis, deep belief network, graph convolution network, variational autoencoder, long short-term memory network and multivariate statistical analysis to extract features from the data; Step 4: Feature fusion: Use a multi-head self-attention mechanism combined with a graph convolutional network to perform multi-dimensional weighted concatenation and fusion on the extracted feature vectors. By introducing a dynamic weight adjustment mechanism, the feature weights are adaptively adjusted according to real-time data changes to form a more accurate comprehensive feature vector. Step 5: Construction of a local fish health assessment model: A fish health assessment model is constructed by integrating feature vectors, and a convolutional neural network and a long short-term memory network are used for training to generate a fish health assessment model; Step 6: Federated learning training: Adopt an improved federated learning algorithm, combined with adaptive optimization and reinforcement learning techniques, and conduct collaborative training through multiple smart fishery platforms to significantly improve the accuracy and generalization ability of the fish health assessment model. Homomorphic encryption technology and differential privacy protection algorithms are also applied to ensure data security and privacy. Step 7: Real-time warning: Issue warnings based on the fish health assessment results obtained from the trained model.

2. The method for constructing a smart fishery platform based on federated learning according to claim 1, characterized in that: The step 1 of obtaining multimodal data of fish swarms from the fish swarm monitoring module, the health monitoring module, the feed management module and the water quality monitoring module specifically includes: Acquire the position data and behavior data of the school of fish from the school of fish monitoring module; Acquiring body surface image data, physiological parameters, body length and weight from the health monitoring module; Obtaining feed delivery amount and eating behavior from the feed management module; The water temperature, dissolved oxygen, pH value, and ammonia nitrogen and nitrite content are obtained from the water quality monitoring module.

3. The method for constructing a smart fishery platform based on federated learning according to claim 1, characterized in that: The preprocessing of multimodal data described in step 2 specifically includes: data cleaning, data standardization and data enhancement; The data cleaning is to remove outliers, missing values ​​and duplicate data by using the anomaly detection and repair technology of the deep learning algorithm, and fill in the missing data by interpolation; The data standardization adopts a standardization method based on a deep belief network to normalize data of different scales to facilitate subsequent feature extraction and model training; The data enhancement is to use a generative adversarial network to rotate, scale and flip the image data to increase the diversity and quality of the data.

4. The method for constructing a smart fishery platform based on federated learning according to claim 1, characterized in that: The feature extraction of multimodal data described in step 3 is specifically as follows: for body length and weight data, regression analysis and statistical methods are used to extract features, including basic statistical features such as mean, variance, and maximum and minimum values; for health status data, a classification algorithm based on a deep belief network is used for feature extraction, and healthy, unhealthy and sub-healthy states are automatically detected and classified through multi-layer Bayesian learning; for behavior pattern data, a hybrid model based on a graph convolutional network and a variational autoencoder is used for feature extraction to identify behavioral changes and abnormal behaviors of fish schools under different environmental conditions; for bait intake data, a time series analysis method based on a long short-term memory network is used for feature extraction to extract change trends, periodic characteristics and intake predictions; For water quality parameter data, multivariate statistical analysis methods and Bayesian network-based models were used for feature extraction to extract the comprehensive characteristics of water quality parameters, and their impact on fish health was analyzed through causal inference.

5. The method for constructing a smart fishery platform based on federated learning according to claim 1, characterized in that: The feature fusion method described in step 4 is specifically: using a multi-head self-attention mechanism combined with a graph convolutional network to perform multi-dimensional weighted splicing and fusion on the extracted feature data, and by introducing a dynamic weight adjustment mechanism, adaptively adjusting the feature weights according to real-time data changes to form a comprehensive feature vector as a training data set.

6. The method for constructing a smart fishery platform based on federated learning according to claim 1, characterized in that: The local fish health assessment model described in step 5 is constructed as follows: The comprehensive feature vector is input into a method combining convolutional neural network and long short-term memory network for training to generate a fish health assessment model; The health assessment model comprehensively considers the body length, weight, health status and behavior pattern factors of the fish school, and is continuously updated through an adaptive optimization algorithm; among the health status, the health status of the fish school can be divided into: healthy, unhealthy and sub-healthy status.

7. The method for constructing a smart fishery platform based on federated learning according to claim 1, characterized in that: The use of the federated learning algorithm described in step 6 to collaboratively train multiple smart fishery platforms includes: Each smart fishery platform independently trains a local health assessment model and aggregates the model parameters of each platform through homomorphic encryption technology; Integrate and optimize model parameters on the central server to generate a joint optimization model; After the joint optimization model is distributed to each smart fishery platform, it is dynamically adjusted using online learning technology.

8. A smart fishery platform system based on federated learning, characterized in that: include: Multimodal data collection module, data processing module, fish health assessment model generation module, federated learning collaborative optimization module, real-time warning module interface module and user terminal module; The multimodal data collection module includes a fish monitoring module, a health monitoring module, a feed management module and a water quality monitoring module for collecting fish position data, behavior data, body surface image data, physiological parameters, body length, weight, feed quantity, feeding behavior, water temperature, dissolved oxygen, pH value and ammonia nitrogen and nitrite content; The data processing module is used to perform preprocessing, feature extraction and feature fusion on the multimodal data; The fish health assessment model generation module is used to generate a fish health assessment model based on the data, and train it using a method combining a convolutional neural network and a long short-term memory network; The federated learning collaborative optimization module is used to optimize the health assessment model through collaborative training of multiple smart fishery platforms, and apply homomorphic encryption technology and differential privacy protection algorithm to ensure data security; The real-time warning module issues warnings based on the fish health assessment results obtained from the trained model; The interface module is used to realize communication between multiple smart fishery platforms and is responsible for uploading model parameters and real-time updating of local models during the federated learning training process; The user terminal module is used to display the visual data to the user for easy monitoring and management.

Images