Feeder optimization control method based on intelligent automatic feeding system

Abstract

The invention discloses a feeder optimization control method based on an intelligent automatic feeding system, and relates to the technical field of intelligent agricultural equipment and automatic control, and the method comprises the steps: constructing a space-time topological graph comprising feeder nodes, environment nodes and animal nodes based on preprocessed multi-source data; inputting the space-time topological graph into a space-time graph convolutional network, performing analysis through space neighborhood aggregation and a time evolution rule, predicting an abnormal propagation path of each node, and generating a risk diffusion graph according to the abnormal propagation path of each node; and constructing a multi-agent state vector for each feeder according to the risk diffusion map, defining a discrete action space based on the multi-agent state vectors, and carrying out collaborative optimization solution on the discrete action space by adopting a game equilibrium algorithm to generate an optimal control strategy. According to the invention, collaborative optimization control of multiple feeders is realized, and equipment conflicts and cascading failures are effectively avoided while the feeding efficiency is ensured.

Description

Technical Field

[0001] The present invention relates to the field of intelligent agricultural equipment and automated control technology, and in particular to a feeder optimization control method based on an intelligent automated feeding system. Background Art

[0002] Currently, intelligent feeding systems primarily utilize two technical approaches: PLC-based timing control solutions and machine learning-based single-device optimization methods. PLC control solutions, such as Germany's Big Dutchman's Feeding Control System, use preset timing logic to control the start and stop of feeders and integrate level sensors for basic automation. Machine learning solutions, such as Lely's T4C system, employ CNN algorithms to analyze vibration data from individual feeders for fault warning. Agritek's Enviflex system, a US company, further integrates temperature and humidity sensors, adjusting feeding strategies by establishing a linear regression model between environmental parameters and feed amount. Regarding animal behavior analysis, Israel's SCR's Heatime system tracks cow activity using RFID ear tags and uses logistic regression to predict feed intake.

[0003] Existing technologies have two key limitations: First, data fusion is insufficient. PLC solutions only achieve simple linkage between device status and environmental parameters. While machine learning solutions improve the accuracy of single-device analysis, they fail to establish a spatiotemporal correlation model between feeder vibration, environmental data, and animal behavior. Second, control strategies lack coordination. Existing feeding systems often use a master-slave control architecture or independent optimization algorithms. This non-coordinated control can amplify the negative impact of faulty nodes, resulting in a loss of feeding efficiency. Summary of the Invention

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

[0005] Therefore, the present invention provides a feeder optimization control method based on an intelligent automated feeding system to solve the problems of low accuracy in abnormal propagation prediction and poor efficiency in multi-device collaborative control caused by insufficient multi-source data fusion in the prior art.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] In a first aspect, the present invention provides a feeder optimization control method based on an intelligent automated feeding system, which includes collecting and preprocessing multi-source data, wherein the multi-source data includes vibration signals, environmental data, and animal position data; constructing a spatiotemporal topological graph including feeder nodes, environmental nodes, and animal nodes based on the preprocessed multi-source data; inputting the spatiotemporal topological graph into a spatiotemporal graph convolutional network, analyzing it through spatial neighborhood aggregation and time evolution rules, predicting the abnormal propagation path of each node, and generating a risk diffusion map based on the abnormal propagation path of each node; constructing a multi-agent state vector for each feeder according to the risk diffusion map, defining a discrete action space based on the multi-agent state vector, and using a game equilibrium algorithm to collaboratively optimize and solve the discrete action space to generate an optimal control strategy.

[0008] As a preferred embodiment of the feeder optimization control method based on the intelligent automatic feeding system of the present invention, the preprocessing comprises extracting vibration energy entropy from the vibration signal through multi-level wavelet packet decomposition, and arranging the vibration energy entropy in ascending order according to frequency bands to generate a vibration feature vector;

[0009] The timestamps of environmental data and animal location data are aligned with the timestamps of vibration feature vectors through the GPS clock synchronization mechanism.

[0010] As a preferred embodiment of the feeder optimization control method based on the intelligent automatic feeding system of the present invention, the spatiotemporal topological graph including feeder nodes, environment nodes and animal nodes is constructed based on the preprocessed multi-source data. The specific steps are as follows:

[0011] Based on the pre-processed multi-source data, feeder nodes, environment nodes, and animal nodes are created by assigning unique identifiers;

[0012] Calculate the Euclidean distance between feeder nodes, generate a symmetric distance matrix, combine the distance decay coefficient, calculate the spatial correlation weight between feeder nodes, and obtain the spatial correlation weight matrix between feeder nodes;

[0013] Calculate the Euclidean distance between the environment node and the animal node, combine the temperature and humidity gradient of the animal node, calculate the association weight between the environment node and the animal node, and obtain the association weight matrix between the environment node and the animal node;

[0014] Based on the spatial association weight matrix between feeder nodes and the association weight matrix between environment nodes and animal nodes, combined with a dynamic weight decay mechanism, a spatiotemporal topological map is constructed.

[0015] As a preferred solution of the feeder optimization control method based on the intelligent automatic feeding system of the present invention, the spatiotemporal topological graph is input into the spatiotemporal graph convolutional network and analyzed by spatial neighborhood aggregation and time evolution law. The specific steps are as follows:

[0016] Normalize the spatiotemporal topology graph to generate a standardized node feature matrix and edge weight matrix;

[0017] The standardized node feature matrix and edge weight matrix are input into the spatiotemporal graph convolutional network to calculate the weighted degree of each node. The global influence coefficient is obtained based on the weighted degree, and the edge weight is adjusted based on the global influence coefficient.

[0018] Based on the adjusted edge weights, the node basis function is constructed, and based on the node basis function, the node neighborhood aggregation is performed to obtain the node aggregation features.

[0019] As a preferred solution of the feeder optimization control method based on the intelligent automatic feeding system of the present invention, wherein: the abnormal propagation path of each node is predicted, and a risk diffusion map is generated according to the abnormal propagation path of each node. The specific steps are as follows:

[0020] Based on the node aggregation characteristics and global influence coefficient, the feeder node, environment node and animal node are detected for anomalies through weighted deviation detection to obtain the node anomaly probability and generate a node anomaly probability set;

[0021] Based on the node anomaly probability set, dynamic risk propagation modeling is performed to generate the risk value of each node and the risk propagation intensity between nodes, and thus generate a risk diffusion map.

[0022] As a preferred solution of the feeder optimization control method based on the intelligent automatic feeding system of the present invention, wherein: the multi-agent state vector is constructed for each feeder according to the risk diffusion map, and the specific steps are as follows:

[0023] Real-time collection of feeding queue length, motor speed and fault codes;

[0024] Discretize the risk level based on the risk value of the feeder node in the risk diffusion map;

[0025] Based on the feeding queue length, the queue pressure is obtained and discretized;

[0026] Based on all nodes around the feeder, obtain the neighborhood risk mean and discretize the neighborhood risk mean;

[0027] The discretized risk level, queue pressure and neighborhood risk mean are encode into feeder state through enumeration method to generate multi-agent state vector.

[0028] As a preferred solution of the feeder optimization control method based on the intelligent automatic feeding system of the present invention, wherein: the discrete action space is defined based on the multi-agent state vector, and the specific steps are as follows:

[0029] Based on the multi-agent state vector, by defining the speed gear and priority mode, the basic action combination is formed by combining the speed gear and the priority mode;

[0030] The basic action combinations are mapped through action-state mapping rules to obtain a discrete action space including high-risk mandatory constraints, neighborhood risk linkage constraints and default allowed action sets.

[0031] As a preferred solution of the feeder optimization control method based on the intelligent automatic feeding system of the present invention, wherein: the game equilibrium algorithm is used to collaboratively optimize and solve the discrete action space to generate the optimal control strategy. The specific steps are as follows:

[0032] Based on the multi-agent state vector and discrete action space, a three-dimensional Q table is constructed, and an initial uniform random strategy table is generated through initialization;

[0033] Select the action of each feeder from the initial uniform random strategy table according to the ε-greedy strategy, and record the joint action of all feeders;

[0034] Based on the joint actions of all feeders, the immediate reward and long-term reward are calculated, and the Nash-Q update rule is adopted to solve the Nash equilibrium strategy through the Lemke-Howson algorithm to generate the optimal control strategy.

[0035] In a second aspect, the present invention provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, any step of the feeder optimization control method based on the intelligent automatic feeding system as described in the first aspect of the present invention is implemented.

[0036] In a third aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, any step of the feeder optimization control method based on the intelligent automatic feeding system as described in the first aspect of the present invention is implemented.

[0037] The beneficial effects of the present invention are as follows: the space-time graph convolutional network is used for spatial neighborhood aggregation and time evolution analysis, which can accurately predict the abnormal propagation path and generate a risk diffusion map, providing a decision-making basis for risk prevention and control; based on the risk diffusion map, a multi-agent state vector is constructed and a discrete action space is defined, and the optimal control strategy is generated through a game equilibrium algorithm, realizing the collaborative optimization control of multiple feeders, effectively avoiding equipment conflicts and chain failures while ensuring feeding efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0039] Figure 1 Flowchart of the feeder optimization control method based on the intelligent automated feeding system.

[0040] Figure 2 Flowchart constructed for a space-time topology map.

[0041] Figure 3 Flowchart for generating a risk diffusion map.

[0042] Figure 4 Flowchart for generating the optimal control strategy. DETAILED DESCRIPTION

[0043] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0044] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0045] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0046] Reference Figures 1 to 4 , is an embodiment of the present invention, which provides a feeder optimization control method based on an intelligent automatic feeding system, comprising the following steps:

[0047] S1: Collecting and preprocessing multi-source data, wherein the multi-source data includes vibration signals, environmental data, and animal position data;

[0048] S1.1: Synchronously collect multi-source data including vibration signals, environmental data, and animal position data;

[0049] A vibration sensor is used at the feeder outlet to obtain a vibration signal;

[0050] Collect environmental data through temperature and humidity sensors in the breeding house;

[0051] The positioning signal transmission function is enabled in the positioning device worn by the animal to obtain the animal's location data and animal ID.

[0052] S1.2: Preprocess multi-source data;

[0053] Vibration energy entropy is extracted from the vibration signal through multi-level wavelet packet decomposition, and the vibration energy entropy is arranged in ascending order according to the frequency band to generate a vibration feature vector;

[0054] Furthermore, the original waveform of the vibration signal is decomposed into four layers of wavelet packets using the Daubechies 4 wavelet basis function to generate decomposed subbands. The 8th to 12th subbands (corresponding to frequencies 8-12 kHz) are selected from the decomposed subbands, and the energy entropy of each subband is calculated to generate the vibration energy entropy. The vibration energy entropies are arranged in ascending order according to the frequency band to generate a vibration eigenvector with a dimension of 5.

[0055] The timestamps of environmental data and animal location data are aligned with the timestamps of vibration feature vectors through the GPS clock synchronization mechanism.

[0056] S2: Based on the preprocessed multi-source data, a spatiotemporal topological map including feeder nodes, environment nodes, and animal nodes is constructed;

[0057] S2.1: Based on the pre-processed multi-source data, define and initialize the node types of the spatiotemporal topology graph;

[0058] A feeder node is generated by assigning a unique identifier to the feeder, associating the feeder coordinate position and the vibration characteristic vector, and generating a feeder node set, wherein the node attributes include the feeder coordinate position and the vibration characteristic vector;

[0059] By assigning a unique identifier to the temperature and humidity sensor, associating the temperature and humidity with the coordinate position of the temperature and humidity sensor, generating an environment node, and generating an environment node set, the node attributes include the temperature and humidity and the coordinate position of the temperature and humidity sensor;

[0060] By assigning a unique identifier to a positioning device worn by an animal, associating an animal ID with animal location data, generating an animal node, and generating an animal node set, the node attributes of which include the animal ID and animal location data.

[0061] S2.2: Based on the feeder node set, calculate the spatial correlation weights between the feeder nodes and obtain the spatial correlation weight matrix between the feeder nodes;

[0062] Calculate the Euclidean distance between each pair of feeder nodes in the feeder node set. The expression is:

[0063]

[0064] Among them, d ij is the Euclidean distance between feeder node i and feeder node j, x i is the coordinate position of feeder node i on the x-axis, y i is the coordinate position of feeder node i on the y-axis, z i is the coordinate position of feeder node i on the z axis, x j is the coordinate position of feeder node j on the x-axis, y j is the coordinate position of feeder node j on the y-axis, z j is the coordinate position of feeder node j on the z-axis;

[0065] Generate a symmetric distance matrix based on the Euclidean distance of the feeder nodes;

[0066] The minimum distance threshold is determined based on vibration propagation, and the spatial correlation weight between feeder nodes is calculated in combination with the distance attenuation coefficient. The expression is:

[0067]

[0068] Among them, w ij is the spatial association weight between feeder node i and feeder node j, δ is the distance attenuation coefficient, d min is the minimum distance threshold;

[0069] When d ij ≤d min When w ij =1 means that feeder node i and feeder node j have strong spatial correlation weights, and as the Euclidean distance increases, the spatial correlation weights decay nonlinearly;

[0070] It should be noted that the minimum distance threshold between feeders is determined by vibration propagation: in a standard farm environment, the correlation coefficients of the feeder vibration signals at different spacings are measured. When the spacing is ≥5 meters, the correlation coefficient decays to below 0.3, indicating that the mechanical vibration coupling effect is negligible; when the spacing is ≥5 meters, the correlation coefficient decays to below 0.3, while meeting the maximum distance at which the vibration energy of the steel structure is attenuated to within 10% (obtained by fitting an exponential decay curve to the measured data of the acceleration sensor), and its universality is confirmed through cross-validation in three groups of different farms.

[0071] According to the spatial correlation weights between feeder nodes, a spatial correlation weight matrix between feeder nodes is constructed;

[0072] It should be noted that the specific steps for constructing the spatial correlation weight matrix between feeder nodes are as follows: first, the Euclidean distance of each pair of feeder nodes is calculated to generate a symmetric distance matrix; based on the vibration propagation characteristics, the minimum distance threshold is determined, and combined with the distance attenuation coefficient, the spatial correlation weight between feeder nodes is calculated by the formula. When the Euclidean distance does not exceed the minimum distance threshold, the spatial correlation weight is set to 1. When it exceeds the minimum distance threshold, the spatial correlation weight decays nonlinearly with the distance; the calculation results are filled into the matrix according to the feeder node pairs, and finally a symmetric spatial correlation weight matrix is ​​formed. The matrix elements reflect the mechanical vibration coupling intensity between the feeder nodes.

[0073] S2.3: Based on the environment node set and the animal node set, obtain the association weight matrix between the environment nodes and the animal nodes;

[0074] Find the three environment nodes closest to the animal node and calculate the Euclidean distance between the environment node and the animal node;

[0075] Calculate the Euclidean distance between the environmental node and the animal node, take the inverse of the Euclidean distance to obtain the distance inverse weight, normalize the distance inverse weight, and then obtain the temperature and humidity gradient of the animal node through weighting;

[0076] According to the temperature and humidity gradient of the animal node, the association weight between the environment node and the animal node is calculated;

[0077]

[0078] Among them, w ca is the association weight between the environment node c and the animal node a, is the inverse distance weight, T a is the temperature gradient of animal node a, T c is the temperature of the ambient node c, H a is the humidity gradient of animal node a, H c is the humidity of the environmental node c, σ is the scaling factor of the temperature and humidity difference;

[0079] According to the association weights of the environment nodes and the animal nodes, the association weight matrix between the environment nodes and the animal nodes is constructed;

[0080] It should be noted that the specific steps of constructing the association weight matrix between environmental nodes and animal nodes are as follows: first, obtain the coordinate positions of the three environmental nodes closest to each animal node, and calculate the Euclidean distance between the animal node and each environmental node; take the inverse of each distance to obtain the initial distance inverse weight; normalize the initial distance inverse weights of all environmental nodes associated with the same animal node so that their sum is 1, and obtain the normalized distance inverse weight; combine the temperature and humidity differences between environmental nodes and animal nodes, and calculate the association weights between environmental nodes and animal nodes through the formula; fill the calculation results into the matrix according to the environmental node and animal node pairs, and finally form the association weight matrix between environmental nodes and animal nodes, and the matrix elements reflect the intensity of the influence of environmental nodes on the temperature and humidity of animal nodes.

[0081] S2.4: Construct a spatiotemporal topology map based on the spatial association weight matrix between feeder nodes and the association weight matrix between environment nodes and animal nodes;

[0082] Based on the feeder node set, the environment node set and the animal node set, the spatiotemporal topological graph is initialized to obtain an initial graph structure including a node set of feeder nodes, an environment node and an animal node, and an edge set of an empty set;

[0083] Traverse the spatial correlation weight matrix between feeder nodes, add undirected edges to the edge set for feeder node pairs whose spatial correlation weight between feeder nodes is greater than the spatial correlation weight threshold, and set the undirected edge attributes, including the spatial correlation weight between feeder nodes (as the initial weight of the edge), timestamp, and type tag (undirected edge connecting feeders);

[0084] It should be noted that the effective range of feeder vibration propagation is analyzed through historical data, and the spatial correlation weight threshold between feeder nodes is determined in combination with the mechanical vibration attenuation law. Specifically, the spatial correlation weight distribution of all feeder node pairs is counted, and the trend of spatial correlation weight changing with Euclidean distance is observed. The spatial correlation weight quantile that can cover 90% of the effective vibration propagation scenarios is selected as the spatial correlation weight threshold.

[0085] Traverse the association weight matrix between the environment node and the animal node, add the directed edge to the edge set for the environment node and the animal node whose association weight is greater than the association weight threshold, and set the directed edge attributes including the association weight between the environment node and the animal node (as the initial weight of the edge), timestamp and direction mark (from environment node to animal node);

[0086] It should be noted that the residence time or feeding frequency of animals under different temperature and humidity gradients was measured, a correlation model between association weights and animal behavioral responses was established, the impact of environmental data on animal behavior was quantified, the range of association weights that significantly affected behavior (referring to the sensitive response of animal behavior to changes in temperature and humidity) was screened out, and the association weight that can cover 80% of the significantly affected scenarios (referring to actual cases where behavior changes are caused by temperature and humidity differences) was selected as the association weight threshold.

[0087] Combined with the dynamic weight decay mechanism, a spatiotemporal topological graph is generated, which includes a set of nodes and a set of edges;

[0088] Furthermore, based on the initial weight of the edge and the attenuation coefficient, the attenuated weight of the edge is calculated, and the expression is:

[0089]

[0090] Where w(t) is the decayed weight of the edge at time t, w0 is the initial weight of the edge, λ is the decay coefficient, t-t0 is the length of time the edge exists, t is the current time, t0 is the time when the edge appears, and e is the base of the natural logarithm;

[0091] The removal threshold is set according to the attenuation law of edge weight. When the weight of the edge after attenuation is less than the removal threshold, it is deleted from the spatiotemporal topology graph. When the weight of the edge after attenuation is greater than or equal to the removal threshold, it is used as the edge weight.

[0092] It should be noted that the effective association duration of undirected edges between feeder nodes and directed edges from environment nodes to animal nodes was analyzed, and the initial removal threshold was calculated in combination with the attenuation coefficient. The removal threshold for undirected edges from feeders was set to 0.02 (reflecting the rapid attenuation characteristics of vibration propagation), and the removal threshold for directed edges from the environment to animals was set to 0.01 (taking into account the persistence of the influence of temperature and humidity). The final removal threshold needs to be optimized and confirmed on the validation set through grid search.

[0093] S3: Input the spatiotemporal topology graph into the spatiotemporal graph convolutional network, analyze it through spatial neighborhood aggregation and time evolution rules, predict the abnormal propagation path of each node, and generate a risk diffusion map based on the abnormal propagation path of each node;

[0094] S3.1: Normalize the spatiotemporal topology graph to generate a standardized node feature matrix and edge weight matrix;

[0095] The vibration eigenvectors of the feeder nodes were normalized using the Z-score, the temperature and humidity of the environment nodes were normalized using the maximum and minimum values, and the animal position data of the animal nodes were mapped to relative positions. The standardized coordinates were calculated with the center of the breeding house as the origin to generate a standardized node feature matrix.

[0096] The edge weights are normalized using the quantile normalization method to generate the edge weight matrix.

[0097] S3.2: Input the normalized node feature matrix and edge weight matrix into the spatiotemporal graph convolutional network and analyze them through spatial neighborhood aggregation and temporal evolution.

[0098] Phase 1: Adjust edge weights through neighborhood influence analysis;

[0099] Calculate the weighted degrees of feeder nodes, environment nodes, and animal nodes respectively;

[0100] Furthermore, the feeder nodes are traversed, the spatial association weights of all adjacent feeder nodes connected by undirected edges are read, and the spatial association weights of all adjacent feeders are accumulated to obtain the weighted degree of the feeder node.

[0101] Traverse the environmental nodes, read the associated weights corresponding to all animal nodes connected by directed edges, and accumulate all associated weights as the direct influence term; obtain the vibration energy entropy of the three feeder nodes closest to the environmental node, and obtain the vibration energy influence term based on the Euclidean distance from the feeder node to the environmental node, combined with the vibration-environment coupling coefficient, and weighted calculation based on vibration energy entropy and distance attenuation. Add the vibration energy influence terms of the three feeders to obtain the vibration indirect influence term, and add the direct influence term and the vibration indirect influence term to obtain the weighted degree of the environmental node;

[0102] It should be noted that the vibration-environment coupling coefficient is determined by controlling the variables: in a closed test environment, the feeder speed is fixed and the vibration energy entropy and temperature and humidity data are collected. The Pearson correlation coefficient of the vibration energy entropy and the ambient temperature change gradient at different speeds is analyzed, and the vibration-environment coupling coefficient is obtained by combining the linear regression equation fitting. The historical vibration energy entropy and temperature and humidity change data verify that the prediction error is <8%, and grid search cross-validation optimization confirms that the vibration-environment coupling coefficient has the smallest error and efficient calculation in temperature and humidity prediction.

[0103] Traverse the animal node, read the associated weights of all environment nodes connected by directed edges, accumulate all associated weights, and get the weighted degree of the animal node.

[0104] Calculate the weighted degree mean of the global nodes as the global benchmark, and normalize the weighted degree of each node to obtain the global influence coefficient;

[0105] It should be noted that all feeder nodes, environment nodes and animal nodes are traversed, and the sum of the spatial association weights of the undirected edges of the feeder nodes is accumulated as the weighted degree of the feeder node, the sum of the directed edge association weights of the environment nodes and the vibration indirect influence term is accumulated as the weighted degree of the environment node, and the sum of the directed edge association weights of the animal nodes is accumulated as the weighted degree of the animal node; the weighted degrees of all feeder nodes, environment nodes and animal nodes are added and divided by the total number of nodes to obtain the global weighted degree mean; the weighted degree of each feeder node is divided by the global weighted degree mean to obtain the global influence coefficient of the feeder node, the weighted degree of each environment node is divided by the global weighted degree mean to obtain the global influence coefficient of the environment node, and the weighted degree of each animal node is divided by the global weighted degree mean to obtain the global influence coefficient of the animal node.

[0106] Traverse all undirected edges and impose upper bound constraints on undirected edges with excessive spatial association weights. The expression is:

[0107]

[0108] in, is the spatial association weight of feeder node p and feeder node q after upper limit constraint, μ is the global weighted degree mean, w pq is the spatial association weight between feeder node r and feeder node q;

[0109] Traverse all directed edges and impose upper bound constraints on directed edges with excessively large associated weights. The expression is:

[0110]

[0111] in, is the association weight between the environment node r and the animal node s after the upper limit constraint, v rs is the association weight between the environment node r and the animal node s.

[0112] Phase 2: Construct node basis functions based on the adjusted edge weights;

[0113] Construct feeder node basis functions;

[0114] Furthermore, the feeder nodes are traversed to obtain the spatial correlation weights between all adjacent feeder nodes connected by undirected edges and the current feeder node. Based on the spatial correlation weights and the difference in vibration energy entropy, the feeder node basis function is calculated, which is expressed as:

[0115]

[0116] Among them, Φ p is the basis function of the feeder node p, E pis the vibration energy entropy of the feeder node p.

[0117] Construct environment node basis function;

[0118] Furthermore, the environmental nodes were traversed to obtain the arithmetic mean of the association weights of all animal nodes connected to the environmental nodes through directed edges, which was used as the direct impact intensity of the environmental node on animal behavior. The indirect impact term of vibration of the three feeder nodes closest to the environmental node was obtained. The environmental node basis function was obtained by weighting the influence of temperature and humidity gradients and the indirect influence of vibration in combination with the linear regression coefficient.

[0119] It should be noted that through the feeder start-stop control experiment, the temperature and humidity change data of the environmental nodes under different vibration energies were collected, and a linear regression model of the temperature and humidity change gradient, the energy entropy of the feeder vibration signal, and the Euclidean distance from the feeder node to the environmental node was established. The regression coefficient was fitted using the least squares method, and the weight coefficient that minimized the prediction error was screened through cross-validation and used as the fixed weight coefficient.

[0120] Construct animal node basis functions;

[0121] Furthermore, the animal nodes are traversed to obtain the behavioral activity of the animal nodes and the temperature and humidity of the environmental nodes connected by directed edges;

[0122] It should be noted that based on the animal location data, the sliding window mean of the animal's movement speed, the proportion of residence time, and the frequency of visits to the feeding area were calculated. The product of the movement speed and the visit frequency was divided by the proportion of residence time plus one to generate the original activity. The Z-score was then normalized using the statistical distribution of historical data to obtain the behavioral activity that obeys the normal distribution.

[0123] For environmental nodes connected by directed edges, the absolute difference between temperature and humidity and the animal's optimal temperature is calculated and mapped into a temperature and humidity comfort score using an exponential function. The temperature and humidity comfort score, behavioral activity, and feeder surrounding density are combined to obtain the animal node basis function through weighting.

[0124] It should be noted that the density around the feeder refers to the spatial distribution density index obtained by counting the number of feeder nodes within a preset radius with the animal node as the center, combining the Euclidean distance from each feeder node to the animal node, and using the weighted summation of the inverse of the distance to reflect the degree of feeder aggregation at the animal's location.

[0125] Phase 3: Based on the node basis function, node neighborhood aggregation is performed to obtain node aggregation features;

[0126] Furthermore, based on the vibration characteristics and spatial relationships of adjacent feeder nodes, feeder node features are aggregated to generate feeder node aggregation features; the environmental node aggregation features are calculated by integrating the influence of animal behavior, indirect influence of vibration and environmental basis functions, as the environmental node aggregation features; based on environmental comfort and behavioral activity, animal node features are calculated, as the animal node aggregation features.

[0127] S3.3: Predict the abnormal propagation path of each node and generate a risk diffusion map;

[0128] S3.3.1: Perform anomaly detection on the feeder node, environment node, and animal node respectively, obtain the node anomaly probability, and generate a node anomaly probability set;

[0129] Furthermore, the vibration energy entropy of the feeder node in the past 10 minutes is extracted, and the sliding window mean and standard deviation of the vibration energy entropy are calculated; when the vibration energy entropy exceeds 3 times the standard deviation of the historical mean, it is marked as an abnormal candidate node; the degree of abnormal deviation (the difference between the vibration energy entropy and the mean divided by the standard deviation) is multiplied by the global influence coefficient of the feeder node to generate the feeder node abnormal probability in the range of 0-1.

[0130] Compare the distribution difference of the current temperature and humidity gradient with that of the same period in history and calculate the KL divergence. When the KL divergence exceeds 2.5, it is determined to be an environmental anomaly. The KL divergence value is proportionally mapped to the node's global influence coefficient to generate the probability of environmental node anomaly.

[0131] Monitor the Z-score standardized value of animal behavioral activity. When it exceeds 3 times the standard deviation for 5 consecutive sampling periods, the behavior is determined to be abnormal; multiply the behavioral activity offset by the global influence coefficient to generate the animal node abnormality probability.

[0132] S3.3.2: Based on the abnormal probability of each node, perform risk propagation dynamic modeling to generate the risk value of each node and the risk propagation intensity between nodes;

[0133] Furthermore, nodes with a node abnormality probability ≥ 0.8 are marked as initial risk sources, and the initial risk value is equal to the node abnormality probability. Nodes with a node abnormality probability less than 0.8 have their initial risk value initialized to 0.

[0134] Traverse the feeder nodes and collect the initial risk values ​​of all adjacent feeder nodes connected by undirected edges; calculate the risk propagation strength based on the initial risk values, edge weights, and Euclidean distances of the adjacent feeder nodes, and update the maximum value of the feeder node's initial risk value and the risk propagation strength as the feeder node's risk value;

[0135] For the environmental node, obtain the risk value of the animal node and the vibration risk value of the adjacent feeder node;

[0136] The risk value of the animal node is: the weight of the directed edge multiplied by the initial risk value of the animal node;

[0137] Vibration risk value of the feeder node: the risk value of the feeder node multiplied by the vibration-environment coupling coefficient, and then divided by the square of the distance plus one;

[0138] After adding the risk value of the animal node and the vibration risk value of the feeder node, compare them with the initial risk value of the environment node and take the maximum value as the risk value of the environment node.

[0139] Traverse the animal nodes and collect the risk values ​​of the environmental nodes; multiply the risk value of the environmental node by the edge weight, and then multiply it by the temperature and humidity comfort score; take the maximum value of the initial risk value of the animal node and the risk propagation intensity as the risk value of the animal node.

[0140] Through iterative processing, when 10 iterations are performed or the change in the risk value of all nodes is less than 1%, the process stops, the final risk value of each node is generated, and the risk propagation intensity between nodes is recorded.

[0141] Based on the final risk value of each node and the risk propagation intensity between nodes, critical path backtracking is performed to obtain the abnormal propagation path of each node;

[0142] Furthermore, nodes with risk values ​​≥ 0.7 are extracted as the starting points of the critical path, and the risk values ​​are sorted from high to low, with high-risk nodes being processed first. Starting from each critical path starting point, the risk source is traced back in reverse to find the predecessor node that contributes the most to the node risk value (contribution = predecessor node risk value × edge weight). When the contribution is ≥ 30% of the node risk value, the predecessor node is added to the critical path; the reverse tracing of the risk source is repeated until the critical path length reaches 6 hops or the contribution is < 10%, generating the abnormal propagation path of each node.

[0143] S3.3.3: Generate a risk diffusion map based on the risk value of each node and the risk propagation intensity between nodes;

[0144] Based on the abnormal propagation path and node risk value, the node shape (circle for feeder nodes, square for environment nodes, triangle for animal nodes), node color (red for risk value ≥0.8, orange for risk value 0.5-0.8, yellow for risk value 0.3-0.5, green for risk value <0.3) and node size (proportional to the node influence coefficient) are defined. The edge width (proportional to the propagation intensity), edge type (solid line for feeder-feeder edge, dotted line for environment-animal edge) and edge arrow (direction arrow is added only on the environment-animal edge) are also defined. A path drilling function is designed to display the downstream path by clicking on the node, a risk tracing function to display the source path by right-clicking the node, and a real-time update function to refresh data every 5 minutes are designed to generate a vector format risk diffusion map and embed the node ID, coordinates, node risk value and path intensity metadata.

[0145] S4: Construct a multi-agent state vector for each feeder based on the risk diffusion map, define a discrete action space based on the multi-agent state vector, and use a game equilibrium algorithm to collaboratively optimize the discrete action space to generate the optimal control strategy;

[0146] S4.1: Construct a multi-agent state vector for each feeder based on the risk diffusion map;

[0147] Real-time collection of feeding queue length and feeder operation status;

[0148] Discretize the risk level based on the node risk value of the feeder in the risk diffusion map;

[0149] Furthermore, the node risk values ​​of all feeders in the risk diffusion map are extracted. The risk values ​​are continuous values ​​ranging from 0 to 1. The risk value range of 0 to 1 is divided into equally spaced intervals, each of which corresponds to a risk level number. Based on the interval in which the feeder node risk value falls, the continuous risk value is mapped to a corresponding discrete risk level number. The risk level number corresponding to each feeder node is generated, forming a discrete risk level classification result. The risk level numbers are arranged in ascending order according to the risk value.

[0150] Based on the feeding queue length, the queue pressure is obtained and discretized;

[0151] Furthermore, the number of animals queuing in front of each feeder is obtained in real time as the feeding queue length; the feeding queue length is divided by the maximum service capacity of the feeder to obtain the queue pressure ratio value; the queue pressure ratio range of 0% to 100% is divided into several equal-width intervals, each interval corresponding to a queue pressure level; according to the interval in which the queue pressure ratio value is located, the continuous ratio value is mapped to the corresponding discrete queue pressure level number; and a discrete queue pressure level reflecting the current feeding load level is generated.

[0152] Based on all nodes around the feeder, obtain the neighborhood risk mean and discretize the neighborhood risk mean;

[0153] Furthermore, the spatial neighborhood range centered on the feeder node is determined, and the risk values ​​of all feeder nodes, environmental nodes, and animal nodes within the range are collected; the Euclidean distance from each node to the central feeder node is calculated, and the node risk value is weighted using the inverse distance weighting method; the arithmetic mean of all weighted node risk values ​​is calculated to obtain the neighborhood risk mean of the feeder node; the neighborhood risk mean is divided into several levels according to the preset interval, and each level corresponds to a discrete identifier; and finally, the discretized neighborhood risk level corresponding to the feeder node is output.

[0154] The discretized risk level, queue pressure and neighborhood risk mean are encoded into the feeder state through enumeration method as the multi-agent state vector;

[0155] For example, the real-time collection of the feeding queue length is used to obtain the current number of animals in the queue through a counting device, and the motor speed and fault code are read in combination with the feeder operation status register; based on the feeder node risk value in the risk diffusion map, the 0.0-1.0 continuous risk value is divided into 10 risk levels from Lv1 to Lv10 at intervals of 0.1; the queue pressure ratio is obtained by dividing the feeding queue length by the maximum capacity of the feeder, and is divided into four pressure levels of Low (<25%), Medium (25-50%), High (50-75%) and Critical (≥75%) at intervals of 25%; based on all the animals within a 5-meter radius around the feeder, the risk level of the feeder is calculated. The risk values ​​of nodes (including feeder nodes, environmental nodes, and animal nodes) are calculated, and the weighted average of the inverse of the distance is calculated as the neighborhood risk mean, which is divided into five levels at intervals of 0.2: VL (0-0.2), L (0.2-0.4), M (0.4-0.6), H (0.6-0.8), and VH (0.8-1.0); the discretized risk level (10 levels), queue pressure (4 levels), and neighborhood risk mean (5 levels) are encoded through three-dimensional combination to generate 200 possible state vectors. Each state vector is represented as a triplet of (risk level number, queue pressure number, neighborhood risk number), completing the construction of the multi-agent state vector.

[0156] S4.2: Define discrete action space based on multi-agent state vectors;

[0157] Furthermore, based on the feeder status code (risk level number, queue pressure number, neighborhood risk number), the speed gears are defined as pause (0%), low speed (30%), medium speed (60%), and full speed (100%), and the priority modes are balanced distribution, neighboring high risk priority, and queue pressure priority. The speed gears and priority modes are combined to form 7 basic action combinations (pause + no mode, low speed + balance, low speed + high risk priority, medium speed + balance, medium speed + queue priority, full speed + balance, full speed + queue priority). Actions are established according to the feeder status code. -State mapping rules: When the risk level number is ≥8 and the queue pressure number is 4 (Critical), the high-risk mandatory constraint is triggered, and only the pause action is allowed; when the neighborhood risk number is ≥4 (H) and the queue pressure number is ≥3 (High), the neighborhood risk linkage constraint is triggered, and the mandatory action combination must include the high-risk priority mode; the remaining states match the allowed action set according to the default rules, and finally generate a complete discrete action space including the high-risk mandatory constraint action set, the neighborhood risk linkage constraint action set and the default allowed action set. Each feeder state code corresponds to 1-4 allowed action combinations.

[0158] S4.3: Obtain the optimal control strategy for the feeder using the Nash-Q learning algorithm;

[0159] Based on the multi-agent state vector and discrete action space, a three-dimensional Q table is constructed and initialized to generate a strategy table;

[0160] For each feeder, select an action according to the ε-greedy strategy and record the joint action of all feeders;

[0161] Based on the joint actions of all feeders, the immediate reward and long-term reward are calculated, and the Nash-Q update rule is adopted to solve the Nash equilibrium strategy through the Lemke-Howson algorithm to generate the optimal control strategy;

[0162] Furthermore, based on the multi-agent state vector (risk level number, queue pressure number, neighborhood risk number) and discrete action space (7 action combinations), a three-dimensional Q table (200 states × 7 actions × N feeders) is constructed. During initialization, the Q value of the "pause" action is set to 5.0 for states with risk level ≥ Lv8, and the Q values ​​of other actions are set to 0.0, and an initial uniform random strategy table is generated; for each feeder, an ε-greedy strategy with ε = 0.1 is used to select actions, with a 90% probability of selecting the action corresponding to the maximum Q value in the current Q table and a 10% probability of randomly selecting an allowed action, and the joint action combination of all feeders is recorded; based on the feeder state vector and In the discrete action space, the immediate reward including the risk control term, efficiency reward term and competition penalty term is first calculated, and the long-term reward is obtained by summing the rewards of the next three steps with a discount factor of 0.9; the Nash-Q algorithm is used to update the Q value, where the learning rate is initially 0.2 and increases to 0.5 when the global high-risk nodes exceed 15%. The Lemke-Howson algorithm is used to solve the Nash equilibrium strategy of each feeder's payoff matrix. The calculation is terminated when the strategy change amplitude is less than 0.01 for 20 iterations. Finally, the highest probability action is selected from the converged strategy as the optimal control instruction. At the same time, the feeder with continuously increasing neighborhood risk is forced to pause.

[0163] This embodiment also provides a computer device suitable for the feeder optimization control method based on the intelligent automated feeding system, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the feeder optimization control method based on the intelligent automated feeding system proposed in the above embodiment.

[0164] The computer device may be a terminal, comprising a processor, a memory, a communication interface, a display screen and an input device connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device comprises 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 the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner may be achieved through WIFI, an operator network, NFC (near field communication) or other technologies. The display screen of the computer device may be a liquid crystal display or an electronic ink display screen, and the input device of the computer device may be a touch layer covering the display screen, or a button, trackball or touchpad provided on the housing of the computer device, or an external keyboard, touchpad or mouse.

[0165] This embodiment also provides a storage medium having a computer program stored thereon, which, when executed by a processor, implements the feeder optimization control method based on the intelligent automatic feeding system proposed 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 static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk.

[0166] In summary, the present invention adopts the following methods: a spatiotemporal graph convolutional network for spatial neighborhood aggregation and temporal evolution analysis, which can accurately predict the abnormal propagation path and generate a risk diffusion map, providing a decision-making basis for risk prevention and control; a multi-agent state vector is constructed based on the risk diffusion map and a discrete action space is defined, and an optimal control strategy is generated through a game equilibrium algorithm, thereby realizing the collaborative optimization control of multiple feeders, effectively avoiding equipment conflicts and cascading failures while ensuring feeding efficiency.

[0167] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A feeder optimization control method based on an intelligent automatic feeding system, characterized in that: include, Collecting and preprocessing multi-source data, wherein the multi-source data includes vibration signals, environmental data, and animal position data; Based on the pre-processed multi-source data, a spatiotemporal topological map including feeder nodes, environment nodes and animal nodes is constructed; The spatiotemporal topology graph is input into the spatiotemporal graph convolutional network. Through spatial neighborhood aggregation and time evolution law analysis, the abnormal propagation path of each node is predicted. Based on the abnormal propagation path of each node, a risk diffusion map is generated. According to the risk diffusion map, a multi-agent state vector is constructed for each feeder. A discrete action space is defined based on the multi-agent state vector, and a game equilibrium algorithm is used to collaboratively optimize and solve the discrete action space to generate the optimal control strategy.

2. The feeder optimization control method based on the intelligent automatic feeding system according to claim 1, characterized in that: The preprocessing is to extract vibration energy entropy from the vibration signal through multi-level wavelet packet decomposition, and arrange the vibration energy entropy in ascending order according to frequency bands to generate a vibration feature vector; The timestamps of environmental data and animal location data are aligned with the timestamps of vibration feature vectors through the GPS clock synchronization mechanism.

3. The feeder optimization control method based on the intelligent automatic feeding system according to claim 2, characterized in that: The spatiotemporal topological graph including feeder nodes, environment nodes and animal nodes is constructed based on the pre-processed multi-source data. The specific steps are as follows: Based on the pre-processed multi-source data, feeder nodes, environment nodes, and animal nodes are created by assigning unique identifiers; Calculate the Euclidean distance between feeder nodes, generate a symmetric distance matrix, combine the distance decay coefficient, calculate the spatial correlation weight between feeder nodes, and obtain the spatial correlation weight matrix between feeder nodes; Calculate the Euclidean distance between the environment node and the animal node, combine the temperature and humidity gradient of the animal node, calculate the association weight between the environment node and the animal node, and obtain the association weight matrix between the environment node and the animal node; Based on the spatial association weight matrix between feeder nodes and the association weight matrix between environment nodes and animal nodes, combined with a dynamic weight decay mechanism, a spatiotemporal topological map is constructed.

4. The feeder optimization control method based on the intelligent automatic feeding system according to claim 3, characterized in that: The spatiotemporal topology graph is input into the spatiotemporal graph convolutional network and analyzed through spatial neighborhood aggregation and time evolution law. The specific steps are as follows: Normalize the spatiotemporal topology graph to generate a standardized node feature matrix and edge weight matrix; The standardized node feature matrix and edge weight matrix are input into the spatiotemporal graph convolutional network to calculate the weighted degree of each node. The global influence coefficient is obtained based on the weighted degree, and the edge weight is adjusted based on the global influence coefficient. Based on the adjusted edge weights, the node basis function is constructed, and based on the node basis function, the node neighborhood aggregation is performed to obtain the node aggregation features.

5. The feeder optimization control method based on the intelligent automatic feeding system according to claim 4, characterized in that: The specific steps of predicting the abnormal propagation path of each node and generating a risk diffusion map according to the abnormal propagation path of each node are as follows: Based on the node aggregation characteristics and global influence coefficient, the feeder node, environment node and animal node are detected for anomalies through weighted deviation detection to obtain the node anomaly probability and generate a node anomaly probability set; Based on the node anomaly probability set, dynamic risk propagation modeling is performed to generate the risk value of each node and the risk propagation intensity between nodes, and thus generate a risk diffusion map.

6. The feeder optimization control method based on the intelligent automatic feeding system according to claim 5, characterized in that: The multi-agent state vector is constructed for each feeder according to the risk diffusion map. The specific steps are as follows: Real-time collection of feeding queue length, motor speed and fault codes; Discretize the risk level based on the risk value of the feeder node in the risk diffusion map; Based on the feeding queue length, the queue pressure is obtained and discretized; Based on all nodes around the feeder, obtain the neighborhood risk mean and discretize the neighborhood risk mean; The discretized risk level, queue pressure and neighborhood risk mean are encode into feeder state through enumeration method to generate multi-agent state vector.

7. The feeder optimization control method based on the intelligent automatic feeding system according to claim 6, characterized in that: The discrete action space is defined based on the multi-agent state vector. The specific steps are as follows: Based on the multi-agent state vector, by defining the speed gear and priority mode, the basic action combination is formed by combining the speed gear and the priority mode; The basic action combinations are mapped through action-state mapping rules to obtain a discrete action space including high-risk mandatory constraints, neighborhood risk linkage constraints and default allowed action sets.

8. The feeder optimization control method based on the intelligent automatic feeding system according to claim 7, characterized in that: The game equilibrium algorithm is used to collaboratively optimize the discrete action space and generate the optimal control strategy. The specific steps are as follows: Based on the multi-agent state vector and discrete action space, a three-dimensional Q table is constructed, and an initial uniform random strategy table is generated through initialization; Select the action of each feeder from the initial uniform random strategy table according to the ε-greedy strategy, and record the joint action of all feeders; Based on the joint actions of all feeders, the immediate reward and long-term reward are calculated, and the Nash-Q update rule is adopted to solve the Nash equilibrium strategy through the Lemke-Howson algorithm to generate the optimal control strategy.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the feeder optimization control method based on the intelligent automatic feeding system according to any one of claims 1 to 8 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the feeder optimization control method based on an intelligent automatic feeding system according to any one of claims 1 to 8 are implemented.

Images