Intelligent pig house monitoring system

Abstract

The invention relates to the technical field of intelligent breeding, and discloses an intelligent hog house monitoring system, which comprises a plurality of agent nodes used for collecting physiological parameters, behavior characteristics and environment data of a hog; the edge calculation module is used for constructing an individual health state vector according to the collected data; the state atlas construction module is used for generating a health state atlas based on a time sequence state vector; the control feedback module is used for driving the environment adjusting equipment when the deviation from the optimal state is judged; the communication and cooperation module is used for exchanging prediction information and state feedback among different agent nodes; the federal learning coordination module carries out aggregation and synchronously updates to the agent node; and the system evolution module is used for adjusting a model structure or control parameters during system-level anomaly detection. A federal learning coordination mechanism is adopted to aggregate and synchronize model parameters uploaded by a plurality of agent nodes, and the technical effects of data local processing and model global optimization are achieved.

Description

Technical Field

[0001] This invention relates to the field of intelligent farming technology, specifically to an intelligent pigsty monitoring system. Background Technology

[0002] With the popularization and improvement of pig farming technology, the scale of pig farming is increasing day by day. The quality of the environment inside the pigsty directly affects the healthy growth of pigs. With the continuous increase in the density of pig farming, the environmental problems of pigsty are becoming more and more prominent. With the development of large-scale farming, modern pigsties have put forward higher requirements for environmental control, behavior monitoring and health management. In order to improve farming efficiency and reduce labor costs, pigsty monitoring systems are gradually evolving towards intelligence and automation.

[0003] Pig farm monitoring systems typically rely on a central server for centralized data processing and model training. However, this centralized approach has significant limitations. As farms expand and environmental conditions become more complex, data transmission bottlenecks become increasingly apparent. Frequent uploads of large amounts of sensitive data pose risks of privacy breaches and network latency, especially in remote areas with unstable networks. Therefore, existing technologies cannot effectively meet the demands of localized data processing and global model optimization, and cannot avoid the burden and privacy issues associated with data transmission.

[0004] Existing intelligent pigsty monitoring systems generally employ fixed model structures and preset rules for data processing and anomaly detection. However, as the farming environment and animal behavior change, fixed models struggle to adapt to constantly evolving realities, often leading to decreased prediction accuracy or system failure. This lack of dynamic adjustment capability fails to cope with real-time changes in the farming environment, resulting in compromised system stability and intelligence during long-term operation, and an inability to effectively address complex and ever-changing environmental challenges.

[0005] In traditional pigsty monitoring systems, model parameter updates between nodes typically employ a simple averaging aggregation method. This approach ignores the quality differences in data uploaded by each node and is susceptible to the influence of abnormal nodes or unreliable data. Simple averaging methods can lead to instability in training results and even model performance fluctuations, making it difficult to guarantee the system's reliability and accuracy. Furthermore, it has a weak ability to suppress low-quality information, failing to achieve accurate and reliable model updates. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides an intelligent pigsty monitoring system that solves the problems of data privacy leakage and transmission bottlenecks.

[0007] To achieve the above objectives, the present invention provides the following technical solution: an intelligent pigsty monitoring system, comprising: Multiple intelligent agent nodes, each configured in the monitoring area corresponding to at least one pig, are used to collect the pig's physiological parameters, behavioral characteristics and environmental data; An edge computing module, connected to the intelligent agent node, is used to construct an individual health status vector based on the collected data and to make predictions and control judgments locally. A state graph construction module, connected to the edge computing module, is used to generate a health state graph based on a time-series state vector. A control feedback module, connected to the edge computing module, is used to drive the environmental adjustment device when it is determined that the deviation from the optimal state is detected. The communication and coordination module is used to exchange prediction information and status feedback between different intelligent agent nodes, and to adjust the coordination priority based on the trust mechanism. The federated learning coordination module is used to receive model parameters uploaded by multiple agent nodes, aggregate them, and synchronously update them to the agent nodes. The system evolution module is used to adjust the model structure or control parameters during system-level anomaly detection to achieve adaptive system evolution.

[0008] Preferably, the intelligent agent node includes a temperature sensor, a heart rate sensor, a motion sensor, an ambient gas sensor, and a camera device, used to collect multimodal data of pigs in real time.

[0009] Preferably, the edge computing module is used to execute image processing functions to extract behavioral features and combine physiological and environmental data to form a state vector: ; In the formula: Indicates the An intelligent agent in time The multimodal state vector; Indicates body temperature value; This indicates the heart rate value; Indicates activity level; Indicates ambient temperature; Indicates ambient humidity; Indicates ammonia concentration; Represents a sequence of images Image processing function Extracted behavioral feature vectors.

[0010] Preferably, the state graph construction module constructs nodes based on the state vector and establishes graph edge weights using the following similarity function: ; In the formula: Indicates the The and the first The edge weights between state nodes; They represent the first An intelligent agent in time , The state vector; Represents the Euclidean distance norm; The bandwidth parameter controls the degree to which the weights are attenuated by the state differences.

[0011] Preferably, the edge computing module calculates the future state based on the prediction model: ; In the formula: Indicates the predicted time The state vector at any given time; Represents the current state vector; Indicates the prediction step size; This is a prediction function used to fit the trend of state evolution; For the first The parameter set of a local prediction model for an agent.

[0012] Preferably, the control feedback module controls at least one of a ventilation system, a sprinkler system, a drinking water system, or a lighting system, and activates the strategy based on state deviation control.

[0013] Preferably, the communication and coordination module constructs a trust matrix among intelligent agents using the following update formula: ; In the formula: Indicates the The first agent on the first An intelligent agent at time Trust value; Indicates the The first agent on the first An intelligent agent at time The trust value ranges from [0, 1]. The historical trust value retention coefficient has a value range of [0, 1]. Indicates the An intelligent agent at time The quality assessment value of the response to the suggestions or decisions provided.

[0014] Preferably, the federated learning coordination module aggregates the model parameters uploaded by the agent nodes using a weighted average method: ; In the formula: This represents the aggregated model parameters calculated by the coordinating node; Indicates the An intelligent agent at time Local model parameters; Indicates the The parameter contribution weights of each agent satisfy the following conditions. ; This represents the total number of agent nodes.

[0015] Preferably, when the system evolution module detects an abnormal synchronization of state deviations among multiple agent nodes, it performs anomaly detection, control threshold reset, or backup model activation operations.

[0016] Preferably, the system has a closed-loop operation mechanism of periodic learning, adaptive feedback, and self-evolutionary updating, enabling continuous optimization and collaborative evolution among intelligent agents.

[0017] This invention provides an intelligent pigsty monitoring system. It has the following beneficial effects: 1. This invention employs a federated learning coordination mechanism to aggregate and synchronize model parameters uploaded by multiple intelligent agent nodes, constructing a unified and shared intelligent behavior recognition model. This achieves the technical effect of local data processing and global model optimization. Compared to existing pig farm monitoring solutions that rely on centralized training on a central server, this invention solves the problems of data privacy leakage and transmission bottlenecks, making it particularly suitable for complex environments and unstable network environments in livestock farming.

[0018] 2. This invention introduces a system evolution module into the system, triggering model structure adjustments and control strategy reconstruction based on anomaly detection results. It possesses online learning and adaptive response capabilities, effectively improving the long-term stability and intelligence of the system. Unlike traditional pigsty monitoring systems that rely on fixed models and preset rules, this invention avoids the defect of models rapidly failing under changes in the feeding environment.

[0019] 3. This invention designs a parameter aggregation strategy based on node trust and data quality, making model updates more accurate and reliable, and suppressing the influence of low-quality information in multi-node collaborative learning. Compared with the simple average aggregation method commonly used in existing intelligent pigsty systems, this invention solves the problems of susceptibility to abnormal node interference and large fluctuations in training results.

[0020] 4. This invention incorporates a system evolution cooling mechanism and a model change recording mechanism, supporting the orderly management and tracking of the model's dynamic evolution process, and preventing performance fluctuations caused by frequent structural adjustments. Compared to traditional solutions that lack model management and evolution control, this overcomes the engineering problem of system instability due to erroneous adjustments, ensuring the safety and controllability of the intelligent pigsty monitoring system. Attached Figure Description

[0021] Figure 1 This is a system framework diagram of the present invention. Detailed Implementation

[0022] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Please see the appendix Figure 1 This invention provides an intelligent pigsty monitoring system, comprising:

[0024] Module 1: Agent Node In this embodiment, the intelligent pigsty monitoring system includes multiple intelligent agent nodes. Each intelligent agent node is preferably deployed at a corresponding location in each independent pigpen inside the pigsty, used to achieve accurate perception and status modeling of individual pigs or specific pig groups. The intelligent agent nodes constitute the system's data perception and preliminary processing unit, which is the prerequisite for realizing intelligent management and control of the pigsty.

[0025] The intelligent agent node includes various types of sensor modules for collecting physiological data, behavioral characteristics, and environmental status information of the target pig. Preferably, the sensor modules include, but are not limited to: a temperature sensor for acquiring the pig's body temperature, a non-contact electrocardiogram sensor for monitoring heart rate, an accelerometer for determining activity level, a temperature and humidity sensor and an ammonia concentration sensor for detecting environmental parameters, and a camera unit for acquiring behavioral images.

[0026] During operation, each sensor collects raw data related to the individual state of the target pig in real time. Through multimodal data fusion, the collected data of different types are structured and organized to define each moment. Individual state vector Its form is as follows: ; In the formula: Indicates the An intelligent agent in time The multimodal state vector; Indicates body temperature value; This indicates the heart rate value; Indicates activity level; Indicates ambient temperature; Indicates ambient humidity; Indicates ammonia concentration; Represents a sequence of images Image processing function Extracted behavioral feature vectors.

[0027] The state vector Characterizes the target pig individual at time [time]. The multidimensional health status is an important input basis for subsequent edge prediction, health map construction and system regulation.

[0028] Preferably, the intelligent agent node also integrates an embedded processing unit to perform preliminary feature transformation and state calculation after data collection, reducing the load on the upper-level system. This processing not only improves perception efficiency but also enables the system to possess edge intelligence capabilities, providing a structured data foundation for subsequent health prediction and collaborative learning.

[0029] Furthermore, to ensure cooperation among the intelligent agent nodes, the nodes also integrate low-power communication modules, which can exchange state vectors, parameter weights, or trust assessment data with neighboring nodes or a central coordinator. Dynamic topologies can be constructed between nodes to adapt to changes in pig herd behavior, enhancing the system's responsiveness.

[0030] To facilitate the construction of a continuous state graph, each node must have the ability to store historical state vectors and support state caching within a certain time window. During the system's periodic updates, nodes extract state data from multiple past moments to participate in the construction and evolution modeling of the graph structure.

[0031] Module 2: Edge Computing Module In this embodiment, the edge computing module is located in the local processing unit of the intelligent agent node. It is used to perform structured processing on the collected multimodal perception data and construct an individual health state vector accordingly. At the same time, it realizes the prediction of local health state and triggers the judgment of control decisions. This module constitutes the bridge between perception and cognition in this system, and is mainly responsible for local data conversion, feature summarization, state modeling, and execution of prediction logic.

[0032] In practical deployment, the edge computing module is connected to the sensor components in each intelligent agent node via wired or wireless means to receive multi-source data input collected from the front end in real time. Specifically, the edge computing module receives, but is not limited to, body temperature signals, heart rate signals, motion data, environmental parameters (temperature, humidity, ammonia concentration), and image sequence data.

[0033] First, the edge computing module cleans and preprocesses the raw data, including outlier removal, missing data completion, and timestamp alignment, to ensure semantic consistency and associativity of the multi-channel input data.

[0034] Subsequently, the edge computing module constructs individual state vectors. .

[0035] After constructing the state vector, the edge computing module is further used to achieve short-term prediction of the future state of the target individual. Specifically, a differential state modeling approach is used to construct the following prediction function: ; In the formula: Indicates the predicted time The state vector at any given time; Represents the current state vector; Indicates the prediction step size; This is a prediction function used to fit the trend of state evolution; For the first The parameter set of a local prediction model for an agent.

[0036] After the prediction is completed, the edge computing module uses the predicted values... With the set health reference status Perform deviation assessment: ; In the formula: This represents the vector norm, usually the Euclidean distance or the weighted norm.

[0037] If there is a deviation Exceeding the tolerance threshold Then the edge computing module generates a control trigger signal and transmits it to the control feedback module to activate the execution logic of control commands such as ventilation, spraying, heating, and drinking water.

[0038] To further enhance the accuracy of state judgment between individuals, the edge computing module can also call the graph construction module interface to participate in the construction of the health state graph. In this scenario, the edge module needs to provide state vectors from multiple time points for graph construction and participate in the preliminary calculation of edge weights. The edge weights are constructed as follows: ; In the formula: Indicates the The and the first The edge weights between state nodes; They represent the first An intelligent agent in time , The state vector; Represents the Euclidean distance norm; The bandwidth parameter controls the degree to which the weights are attenuated by the state differences.

[0039] In this embodiment, the edge computing module supports online operation, intermittent updates, and model adaptive adjustment mechanisms, and can periodically interact with the federated learning coordination module for parameter uploading and global model synchronization.

[0040] Module 3: State Graph Construction Module.

[0041] In this embodiment, the state graph construction module is used to construct a graph structure representation model that characterizes the health evolution process of pigs. It is connected to the edge computing module and is used to receive the state vector sequence from each agent node and generate a time-series health state graph based on it.

[0042] The graph construction module specifically includes sub-modules such as a graph structure generation unit, a graph node embedding unit, and an edge weight calculation unit. These three modules work together to achieve a visual representation of an individual's health status, model the relationships between states, and provide a structural foundation for subsequent prediction or classification tasks.

[0043] During system operation, the edge computing module periodically uploads the locally constructed state vectors to the graph construction module. Preferably, this upload is done at the time window level, and the state vectors within each time window are organized into a set of nodes, forming the initial nodes in the graph.

[0044] Each node represents the complete health status of a pig corresponding to a certain agent node at a certain moment in the system. This status includes multimodal physiological parameters, behavioral feature vectors, and environmental features, and is mapped to graph structure nodes through a unified embedding function to support subsequent graph neural network processing or graph-based clustering and prediction operations.

[0045] To characterize the semantic relationships between nodes, the graph construction module further calculates the edge weight coefficients between nodes. Specifically, based on the differences between state vectors, the system uses a function based on state similarity to calculate the edge connection strength between nodes. Such edges represent potential associations between different times or different individuals, and these associations may originate from similar states, consistent behavioral patterns, or similar environmental backgrounds.

[0046] Based on the foregoing description, calculate the edge weights between nodes. As shown below: ; In the formula: Indicates the The and the first The edge weights between state nodes; They represent the first An intelligent agent in time , The state vector; Represents the Euclidean distance norm; The bandwidth parameter controls the degree to which the weights are attenuated by the state differences.

[0047] Furthermore, to further enhance the flexibility of edge weight calculation, a dynamic adjustment factor based on time difference is introduced, making the correlation between states more accurate. This innovative formula is shown below: ; In the formula: Indicates the The state node and the first Dynamic weights between state nodes; To dynamically adjust the factor and control the degree of influence of the time difference on the weight; The time difference represents the distance between two points in time and reflects the frequency of changes in state.

[0048] By calculating the norm of the state vectors between node pairs and transforming them using similarity mapping functions such as Gaussian kernels, the resulting edge weights are used to construct and weight edges in the graph, thus forming a complete weighted directed or undirected graph. This graph not only has good scalability but is also suitable for incremental updates in dynamically changing scenarios.

[0049] Furthermore, to maintain the temporal structure of the graph, the system adds a timestamp to each node and introduces a sliding time window mechanism. Historical nodes that exceed the time window are either weakened or archived to control the graph size and improve computational efficiency. The graph update process can employ an incremental graph update algorithm, performing local calculations only on newly added nodes and edges to improve operational efficiency.

[0050] To further support the system's identification of group behavior and anomaly detection of pig herd health status, the graph construction module supports a fusion strategy for multi-head graph structures. Graph structures constructed by different agent nodes can be integrated into a unified global graph at the system level, where each subgraph corresponds to an individual. Cross edges are established between them through specific event triggers (such as spatial overlap or behavioral similarity overlap) to enhance the group representation capability of the graph.

[0051] During system operation, the graph construction module can also work in conjunction with the federated learning module to achieve graph structure learning and parameter sharing at the edge. Global feature encodings can be embedded in the graph structure for use as input to subsequent prediction models.

[0052] Ultimately, this graph can be used as input for the health scoring module, risk inference module, or graph neural network model execution platform, thereby enabling high-level intelligent judgment based on structured state evolution information.

[0053] Module 4: Control Feedback Module In this embodiment, the control feedback module is connected to the edge computing module and is used to drive the environmental adjustment equipment when it is determined that the pigsty deviates from the optimal state, so as to ensure that the environmental conditions in the pigsty meet the optimal health standards. This module receives individual health status data provided by the edge computing module, and makes real-time judgments based on the set health thresholds and control logic, and then issues control signals to adjust environmental parameters, such as temperature, humidity, air circulation, ammonia concentration, etc., so as to achieve optimized regulation of the pigs' growth environment.

[0054] The control feedback module includes core sub-modules such as a sensor interface unit, a deviation calculation unit, a control decision unit, and an execution interface unit. These sub-modules work closely together to ensure the system can provide effective real-time feedback and adjustment in dynamic environments.

[0055] The deviation between the current health status of the pigs and the preset optimal state. The health status vector includes multi-dimensional information such as physiological parameters, behavioral characteristics, and environmental data. Specifically, it is determined by comparing the health status vector with the target health status. The differences between them are calculated to obtain the individual's health deviation value. : ; In the formula: For individuals At any moment Health deviation values; For the target individual at any time The actual state vector; The optimal health state vector serves as a reference standard.

[0056] This deviation value reflects the degree of deviation between an individual's current health status and the ideal state. Based on this, the control decision unit generates adjustment instructions according to the deviation value, driving the environmental control equipment to make corresponding adjustments. These adjustment instructions are transmitted to the actual control equipment, such as ventilation systems, heating equipment, humidity regulators, and ammonia removal devices, through the execution interface unit.

[0057] Specifically, if the health deviation value Exceeding the preset threshold range The system then issues a control signal to activate the corresponding regulating device through the execution interface unit. The regulating method may include, but is not limited to, the following forms: Temperature regulation: When the system detects that an individual's body temperature deviates from the set optimal range, the regulating device will adjust the temperature of the pigsty according to the deviation value; Humidity regulation: When the humidity value deviates from the optimal range, the humidity control device will be activated to maintain a healthy standard by increasing or decreasing the air humidity; Ammonia concentration regulation: When the ammonia concentration is too high, the system will instruct the ventilation equipment to increase ventilation and reduce the concentration of harmful gases; Air circulation regulation: If air circulation is poor, the system will adjust the fan or exhaust system to ensure airflow and provide an adequate oxygen supply.

[0058] Furthermore, the control feedback module in this embodiment also supports adaptive adjustment based on intelligent learning. In the initial stage, the system achieves standardized adjustment by setting reasonable preset target states and threshold ranges. As the system operates further, the control feedback module, through the accumulation of historical data and combined with intelligent optimization algorithms (such as reinforcement learning or fuzzy control algorithms), can adaptively adjust the control strategy, thereby more accurately meeting the health needs of pigs and optimizing environmental regulation strategies.

[0059] This adaptive adjustment capability is based on the system's prediction of individual state change trends. Before each control decision, the system estimates the trend of health status changes based on historical data and predictive models, thereby selecting the most suitable control strategy. For example, if it is predicted that the pigs' body temperature may continue to rise in the future, the system can adjust the ambient temperature in advance to prevent the pigs from overheating.

[0060] Another key feature of the control feedback module is its real-time performance and response speed. To ensure that the environmental conditions inside the pigsty can be restored to their optimal state in the shortest possible time, the deviation calculation unit and control decision unit in the control feedback module employ efficient calculation methods to ensure the real-time nature of the feedback and adjustment process. Through close collaboration with environmental control equipment, this module can respond to changes in the system in an extremely short time, preventing pigs from being affected by excessive environmental changes.

[0061] Module 5: Communication and Collaboration Module In this embodiment, the communication and coordination module is used to realize information interaction and strategy coordination among multiple intelligent agent nodes. The module is connected between each intelligent agent node and the edge computing module, and its main functions include state information synchronization, prediction result transmission, feedback signal broadcasting, and coordination strategy adjustment based on node behavior credibility.

[0062] During system operation, each agent node establishes a point-to-point or multicast communication mechanism through this module to achieve cross-node sharing of local data. The shared data mainly includes, but is not limited to, the current health state vector, future state prediction results, control execution feedback information, individual behavioral feature summaries, and local environmental conditions.

[0063] Preferably, the communication and collaboration module integrates a lightweight communication protocol stack, supporting low-latency transmission and edge-level distributed synchronization mechanisms to ensure stable operation of the system in a multi-agent distributed environment.

[0064] To enhance the effectiveness of the collaborative strategy, this embodiment introduces a trust mechanism between nodes to dynamically evaluate the reliability of each node's prediction results and adjust the priority and decision reference weights in collaborative interactions accordingly. Specifically, each node maintains a trust value for other nodes. , indicating the current time Next node For nodes The level of trust. Trust value updates are based on the following strategy: ; In the formula: Indicates time node For nodes Trust value; Represents a node At any moment The degree of matching between the predicted results and the actual feedback, i.e., the inverse vector of the prediction error; The trust update coefficient is used to control the balance between historical memory weights and immediate feedback.

[0065] Prediction error matching degree This can be given by the following relationship: ; In the formula: Represents a node At any moment The predicted state; This indicates the actual state of the node; Represents the Euclidean distance norm; This is the error control bandwidth parameter, used to adjust the sensitivity of the matching degree function.

[0066] Through this mechanism, the system can assess the consistency and reliability of nodes in prediction, feedback, and other behaviors in real time. Nodes with higher trust values ​​have a greater weight in the group's collaborative judgment, while nodes with lower trust values ​​have their influence automatically reduced or temporarily removed to enhance the overall predictive stability of the system.

[0067] Based on the aforementioned trust model, the communication and coordination module can also construct a weighted coordination graph structure to schedule information transmission paths and synchronization update strategies among nodes. This coordination graph structure supports a topology adaptive adjustment mechanism. When the predicted reliability of a node continues to decline or data transmission anomalies occur, the system can automatically replan the coordination path to ensure network fault tolerance and decision robustness.

[0068] In practical applications, when a node detects that its own state deviates from its optimal health state, it not only drives the environmental adjustment equipment to respond through the control feedback module, but also broadcasts its predicted trend and state deviation signal through the communication and coordination module. Other nodes, based on the received coordination information and their local observations, complete the auxiliary judgment of the global state, forming an information loop.

[0069] In addition, this module also supports a cross-node state comparison mechanism, that is, a node can actively request the health status data of neighboring agent nodes under similar time windows and perform local comparative analysis to identify common anomalies, propagation trends or group effects caused by environmental factors, and provide cross-regional reference for control response strategies.

[0070] In terms of architecture, the communication and coordination module supports a coordination strategy that combines asynchronous broadcasting, priority queue processing, and event-triggered modes, thereby improving the system's responsiveness to scenarios with frequent state changes. Especially in large-scale deployment scenarios in livestock farms, this module effectively avoids network-wide information congestion and node synchronization conflicts through multicast routing and hierarchical scheduling mechanisms.

[0071] Module 6: Federated Learning Coordination Module In this embodiment, the federated learning coordination module is used to realize the convergence, aggregation and distribution of model parameters in a collaborative learning scenario of multiple agent nodes, thereby constructing a unified shared intelligent model and improving the overall cognitive ability and prediction accuracy of the system.

[0072] The federated learning coordination module connects to each agent node in the system and interacts with the edge computing module and the communication and coordination module. Its main functions include: receiving local model parameters uploaded by edge agent nodes; executing aggregation functions to construct a global model; distributing the updated model weights to all participating nodes; and dynamically adjusting the aggregation strategy or weight contribution value based on the trust score of node behavior.

[0073] During system deployment, each agent node independently receives data input from local sensors and state graphs, and performs model training tasks based on its local state vector. Training tasks typically employ a mini-batch update mechanism, optimizing local model parameters through gradient descent or its variants. After completing a certain number of training rounds, each node uploads its local model parameters to the federated learning coordination module.

[0074] In this embodiment, the federated learning coordination module uses the federated averaging algorithm as the basic parameter aggregation strategy. The specific form is as follows: ; In the formula: This represents the aggregated global model parameters; Indicates the Local model parameters uploaded by each node; This represents the parameter weight of the node, satisfying... It can be assigned based on the number of training samples or node trust level; This represents the number of agent nodes participating in the current round of aggregation.

[0075] Weighting factors The model can be configured using a multi-strategy fusion approach. In addition to the basic sample size weighting method, this embodiment further introduces a node trust adjustment mechanism to enhance the model's responsiveness to high-quality data sources. The trust score is provided by the communication and collaboration module and can be constructed based on prediction accuracy, feedback consistency, and historical reliability. ; In the formula: Represents a node The number of local training samples; Represents a node The trust score is a comprehensive evaluation based on the feedback behavior in the previous round; The number of agent nodes participating in the current round of aggregation; The denominator term is used to normalize the weighted contribution of all nodes.

[0076] After aggregation is complete, the federated learning coordination module will The parameters are synchronized by synchronously updating all agent nodes through the communication channel, and each node replaces its local old model weights. Based on this, each node can continue to carry out local training or prediction tasks, forming a new round of distributed learning loop.

[0077] To further improve the efficiency and stability of the federated learning process, the federated learning coordination module can also perform the following additional functions: Node filtering mechanism: When a node is removed from the current round of aggregation due to abnormal network status, excessive training accuracy deviation, or low trust level; Asynchronous aggregation support: When the training completion time of edge nodes is inconsistent, the system can adopt an asynchronous receiving mechanism and set buffer weights for delayed nodes to prevent aggregation offset; Differential privacy protection mechanism: To prevent individual privacy risks caused by model parameter leakage, perturbation terms or noise mechanisms can be introduced before uploading parameters to enhance the data security of the system.

[0078] In terms of deployment, the federated learning coordination module can be located at the system's central node, a local server, or at the edge of a cloud-edge collaborative architecture, preferably configured flexibly according to the system's distribution structure and bandwidth capabilities. All communication processes can be transmitted based on encrypted protocols to ensure the integrity and security of model parameters.

[0079] In practical applications, this module promotes the formation of a dynamically evolving shared cognitive ability throughout the system through a periodic training-aggregation-distribution approach. In the pigsty health monitoring system, the global model can undertake multiple tasks such as behavior recognition, anomaly prediction, and environmental control suggestions, while different nodes directly influence the direction of model evolution based on their participation and performance quality during the training phase.

[0080] Module 7: System Evolution Module In this embodiment, the system evolution module is used to automatically trigger adjustments to the system model structure or control strategy based on historical model performance, state feedback and global indicator dynamics when a system-level anomaly detection event occurs, so as to achieve adaptive evolution of system functions and maintenance of long-term performance.

[0081] The system evolution module is data-coupled with the federated learning coordination module, control feedback module, and communication and collaboration module. It can receive anomaly detection signals, predict deviation trends, and environmental interference indicators uploaded by nodes in real time, and determine whether to execute a structural evolution strategy based on a multi-factor triggering mechanism.

[0082] Preferably, the system evolution module integrates an anomaly state evaluation function to assess the deviation of the current system state from the optimal state. The function takes prediction error, invalidity ratio, and anomaly frequency as input variables and is defined as follows: ; In the formula: This indicates an overall abnormality index representing the current system status; This indicates that all agent nodes are at time [time]. The average prediction error; This indicates the ineffectiveness rate of control feedback per unit of time. Indicates the frequency of abnormal events reported by the node; , , , which is a weighting parameter used to balance the contribution ratio of each abnormal factor in the system evolution decision.

[0083] when Exceeding the preset threshold At this point, the system evolution module automatically enters the evolution response phase. The evolution response includes, but is not limited to, the following two types of strategies.

[0084] One type is the model structure evolution strategy. This strategy is applicable when the model fails or its generalization ability declines in the current environment. The system evolution module will call the historical model structure configuration records and, in conjunction with the current error distribution, node trust trends, and feature drift, trigger the model complexity adjustment or structural reconstruction process. Specific adjustments include the number of neural network layers, parameter sharing strategies, enabling and disabling the attention mechanism module, and selecting feature input channels.

[0085] During the evolutionary process, the following structural adjustment principles are preferred: If the current average error continues to rise and the consistency of prediction directions among trusted nodes is poor, then increase the model depth to improve the nonlinear fitting ability. If the abnormal regions are concentrated in a specific node, the input feature selection mechanism of that node is adjusted, for example, by introducing a convolutional submodule to enhance the ability to perceive local features. If the feature drift is significant, the feature transformation module structure can be reconstructed, and sliding window normalization or batch normalization operations can be introduced.

[0086] Another type is the control parameter adjustment strategy. When the overall predictive capability of the system is not ineffective but the control effect is unsatisfactory, the system evolution module can optimize the response by adjusting the weight parameters, trigger thresholds, or feedback cycle parameters in the control feedback. For example, adjusting the drive response time in the control feedback module. The following adjustments can be made: ; In the formula: , These represent the control delay parameters before and after adjustment, respectively. This indicates the changes in abnormal indicators between the current state and the previous stable state; This represents the average value of abnormal indicators during historical stable periods. The response adjustment coefficient is used to adjust the magnitude of change.

[0087] The above parameter adjustment process is executed within the control feedback module. The system evolution module is only responsible for parsing the global feedback state and generating response strategies. This mechanism ensures that the control strategy can adaptively evolve according to the actual operating conditions without significantly changing the model structure.

[0088] The system evolution module also incorporates an evolution cooling mechanism to prevent system oscillations caused by frequent adjustments. This mechanism, based on the minimum evolution interval, controls the shortest time interval between two consecutive structural evolutions and monitors the quality of the system response after adjustments to determine whether the evolution results meet expectations.

[0089] In addition, the system evolution module in this embodiment has a log recording function, which archives the context data of each model adjustment and control parameter change, including abnormal triggering factors, evolution strategy paths, and trends of changes in influencing indicators, for subsequent autonomous decision-making by the intelligent agent or manual auditing.

[0090] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. An intelligent pigsty monitoring system, characterized in that, include: Multiple intelligent agent nodes, each configured in the monitoring area corresponding to at least one pig, are used to collect the pig's physiological parameters, behavioral characteristics and environmental data; An edge computing module, connected to the intelligent agent node, is used to construct an individual health status vector based on the collected data and to make predictions and control judgments locally. A state graph construction module, connected to the edge computing module, is used to generate a health state graph based on a time-series state vector. A control feedback module, connected to the edge computing module, is used to drive the environmental adjustment device when it is determined that the deviation from the optimal state is detected. The communication and coordination module is used to exchange prediction information and status feedback between different intelligent agent nodes, and to adjust the coordination priority based on the trust mechanism. The federated learning coordination module is used to receive model parameters uploaded by multiple agent nodes, aggregate them, and synchronously update them to the agent nodes. The system evolution module is used to adjust the model structure or control parameters during system-level anomaly detection to achieve adaptive system evolution.

2. The intelligent pigsty monitoring system according to claim 1, characterized in that, The intelligent agent node includes a temperature sensor, a heart rate sensor, a motion sensor, an ambient gas sensor, and a camera device, used to collect multimodal data of pigs in real time.

3. The intelligent pigsty monitoring system according to claim 1, characterized in that, The edge computing module is used to execute image processing functions to extract behavioral features and combine physiological and environmental data to form a state vector. ； In the formula: Indicates the An intelligent agent in time The multimodal state vector; Indicates body temperature value; This indicates the heart rate value; Indicates activity level; Indicates ambient temperature; Indicates ambient humidity; Indicates ammonia concentration; Represents a sequence of images Image processing function Extracted behavioral feature vectors.

4. The intelligent pigsty monitoring system according to claim 1, characterized in that, The state graph construction module constructs nodes based on the state vectors and establishes graph edge weights using the following similarity function: ； In the formula: Indicates the The and the first The edge weights between state nodes; They represent the first An intelligent agent in time , The state vector; Represents the Euclidean distance norm; The bandwidth parameter controls the degree to which the weights are attenuated by the state differences.

5. The intelligent pigsty monitoring system according to claim 1, characterized in that, The edge computing module calculates the future state based on the prediction model: ； In the formula: Indicates the predicted time The state vector at any given time; Represents the current state vector; Indicates the prediction step size; This is a prediction function used to fit the trend of state evolution; For the first The parameter set of a local prediction model for an agent.

6. The intelligent pigsty monitoring system according to claim 1, characterized in that, The control feedback module controls at least one of the following: a ventilation system, a sprinkler system, a drinking water system, or a lighting system, and activates the control based on a state deviation control strategy.

7. The intelligent pigsty monitoring system according to claim 1, characterized in that, The communication and coordination module constructs a trust matrix among intelligent agents, using the following update formula: ； In the formula: Indicates the The first agent on the first An intelligent agent at time Trust value; Indicates the The first agent on the first An intelligent agent at time The trust value ranges from [0, 1]. The historical trust value retention coefficient has a value range of [0, 1]. Indicates the An intelligent agent at time The quality assessment value of the response to the suggestions or decisions provided.

8. The intelligent pigsty monitoring system according to claim 1, characterized in that, The federated learning coordination module aggregates the model parameters uploaded by agent nodes using a weighted average method: ； In the formula: This represents the aggregated model parameters calculated by the coordinating node; Indicates the An intelligent agent at time Local model parameters; Indicates the The parameter contribution weights of each agent satisfy the following conditions. ; This represents the total number of agent nodes.

9. The intelligent pigsty monitoring system according to claim 1, characterized in that, When the system evolution module detects an abnormal synchronization of state deviations among multiple agent nodes, it performs anomaly detection, control threshold reset, or backup model activation operations.

10. The intelligent pigsty monitoring system according to claim 1, characterized in that, The system possesses a closed-loop operation mechanism of periodic learning, adaptive feedback, and self-evolutionary updates, enabling continuous optimization and collaborative evolution among intelligent agents.

Images