Multi-modal information fusion feeding decision-making system and method for breeding chicken behavior recognition

Abstract

The invention provides a multi-modal information fusion feeding decision-making system and method for chicken breeding behavior recognition, and the method comprises the steps: collecting a multi-source heterogeneous data set, and extracting a primary fusion feature vector; constructing a triple knowledge graph; mining implicit association rules of the ingestion frequency and the body temperature; performing secondary fusion on the primary fusion feature vector and an implicit association rule to generate an intermediate decision feature; and generating a feeding decision instruction through the pre-training decision model and the expert rule base. According to the method, the knowledge graph is constructed, GNN reasoning is utilized, and a manual preset rule static mode is replaced; performing secondary feature fusion, generating intermediate decision features by combining primary fusion features and implicit rules, and then combining a pre-training model and an expert rule base, ensuring decision real-time performance, integrating domain knowledge, outputting accurate adjustment parameters, realizing full-link intelligence, improving the accuracy and adaptability of breeding chicken feeding decisions, and improving the accuracy and adaptability of chicken feeding decisions. Therefore, dynamic requirements of complex breeding scenes are met, and chicken flock health and breeding efficiency improvement are promoted.

Description

Technical Field

[0001] This invention relates to the field of feeding decision-making technology for farmed chickens, and in particular to a multimodal information fusion feeding decision-making system and method for recognizing farmed chicken behavior. Background Technology

[0002] Current feeding decisions for farmed chickens are mostly based on a single data source (such as manual observation of feed intake or environmental temperature and humidity) or simply overlaying multi-source data without deep integration. Some studies have attempted to introduce visual monitoring (such as movement trajectories) or physiological indicators (such as heart rate), but lack efficient mechanisms for fusing cross-modal features. The application of knowledge graphs in the field of aquaculture is still in its early stages, and the mining of behavior-physiology-environment association rules largely relies on manual definition, making it difficult to dynamically adapt to complex aquaculture scenarios.

[0003] Existing technologies suffer from insufficient fusion of multi-source heterogeneous data and lack in-depth mining of semantic relationships between modalities, leading to biased decision-making and an inability to effectively address dynamic changes in flock behavior and fluctuations in physiological state. Traditional methods do not fully utilize the reasoning capabilities of knowledge graphs, and the implicit association rules between behavioral and physiological indicators require manual pre-setting, lacking adaptive learning mechanisms. In complex farming environments, single-modal data is susceptible to noise interference, and the timeliness and accuracy of cross-modal feature fusion are insufficient, resulting in delayed feeding decisions or deviations from the actual needs of the flock. Summary of the Invention

[0004] This invention aims to at least address the technical problem of low accuracy in feeding decisions for farmed chickens in the prior art, and innovatively proposes a multimodal information fusion feeding decision system and method for recognizing farmed chicken behavior.

[0005] To achieve the above-mentioned objectives of this invention, this invention provides a multimodal information fusion feeding decision-making method for recognizing the behavior of farmed chickens, the method comprising: S1. Collect multi-source heterogeneous datasets of chicken flocks, including: visual-thermal imaging data, acoustic-physiological data, and environmental-metabolic data; S2. Extract and fuse features from the multi-source heterogeneous dataset to obtain a primary fused feature vector; S3. Construct a behavior-physiology-environment triplet knowledge graph based on the aforementioned primary fusion feature vector; S4. Based on the triplet knowledge graph, use GNN to perform relational reasoning to mine the implicit association rules between feeding frequency and body temperature; S5. Perform secondary feature fusion between the primary fusion feature vector and the implicit association rule to generate intermediate decision features that include the probability of abnormal behavior, physiological stress index and environmental risk level. S6. Based on the intermediate decision features, a feeding decision instruction is generated through a pre-trained decision model and an expert rule base. The feeding decision instruction includes feeding time, feeding amount, and feed type adjustment parameters.

[0006] In another aspect, the present invention also provides a multimodal information fusion feeding decision system for recognizing the behavior of farmed chickens, the system comprising: processor; Memory used to store processor-executable instructions; The processor is configured to implement the multimodal information fusion feeding decision method for recognizing the behavior of farmed chickens when executing the executable instructions.

[0007] The beneficial effects of this invention are as follows: This invention employs a multi-head attention mechanism to achieve deep fusion of visual-thermal imaging, acoustic-physiological, and environmental-metabolic features, overcoming the limitations of traditional single data sources or simple superposition. It constructs a behavior-physiology-environment triplet knowledge graph and utilizes GNN for relational reasoning, dynamically mining implicit association rules between feeding frequency and body temperature, replacing the static mode of manually preset rules. Through secondary feature fusion, it combines primary fusion features with implicit rules to generate intermediate decision features including behavioral abnormality probability, physiological stress index, and environmental risk level. Combined with a pre-trained decision model and expert rule base, it ensures both real-time decision-making and incorporates domain knowledge, ultimately outputting precise adjustment parameters for feeding time, feeding amount, and feed type. This achieves end-to-end intelligentization from data collection to decision execution, significantly improving the accuracy and adaptability of feeding decisions for poultry, effectively addressing dynamic needs in complex farming scenarios, and promoting flock health and farming efficiency.

[0008] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0009] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which: Figure 1 This is a flowchart of the multimodal information fusion feeding decision-making method for recognizing the behavior of farmed chickens according to the present invention. Detailed Implementation

[0010] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0011] Example 1 like Figure 1 As shown, a multimodal information fusion feeding decision-making method for recognizing the behavior of farmed chickens includes: S1. Collect a multi-source heterogeneous dataset of chickens, including visual-thermal imaging data, acoustic-physiological data, and environmental-metabolic data. In step S1, visual-thermal imaging data is collected simultaneously using a high-definition camera and an infrared thermal imager, covering the chickens' activity area to obtain information on chicken behavior, activity range, and body surface temperature distribution. Acoustic-physiological data is collected using a directional microphone array and wearable physiological sensors to record chicken call characteristics, respiratory rate, and heart rate data. Environmental-metabolic data is collected using temperature and humidity sensors, an ammonia concentration detector, and a smart feeding device to monitor breeding environment parameters and feed consumption in real time. All data undergoes preliminary cleaning and time synchronization through edge computing nodes to form a structured multimodal dataset.

[0012] S2. Extract and fuse features from multi-source heterogeneous datasets to obtain a primary fused feature vector; S3. Construct a triplet knowledge graph of behavior-physiology-environment based on primary fusion feature vectors; S4. Based on the triplet knowledge graph, use GNN to perform relational reasoning to mine the implicit association rules between feeding frequency and body temperature. S5. Perform secondary feature fusion between the primary fusion feature vector and the implicit association rule to generate intermediate decision features that include the probability of abnormal behavior, physiological stress index and environmental risk level. S6. Based on intermediate decision features, feed decision instructions are generated through a pre-trained decision model and expert rule base. Feed decision instructions include feeding time, feeding amount and feed type adjustment parameters.

[0013] In this embodiment, the principle of a multimodal information fusion feeding decision-making method for recognizing the behavior of farmed chickens is as follows: First, by collecting multi-source heterogeneous datasets (S1), the visual behavior, thermal imaging temperature distribution, acoustic call characteristics, physiological indicators, and environmental parameters of the chicken flock are comprehensively acquired, forming a structured multimodal input. Next, in the feature extraction and fusion stage (S2), a multi-head attention mechanism is employed to achieve deep interaction among the visual-thermal imaging, acoustic-physiological, and environmental-metabolic modalities, overcoming data heterogeneity and generating robust primary fusion feature vectors. Based on this, a behavior-physiology-environment triplet knowledge graph is constructed (S3), where behavioral entities (e.g., feeding frequency), physiological entities (e.g., body temperature), and environmental entities (e.g., temperature and humidity) are associated through a graph structure. Subsequently, a graph neural network (GNN) is used for two-hop relation reasoning (S4), dynamically mining the implicit path weights and rule confidence from feeding frequency to body temperature, replacing the static pattern of manually preset rules and enhancing the adaptability of rule discovery. A gating mechanism is used to filter high-confidence rules and map them to a predefined template library (S405), generating a set of implicit association rules. In the secondary feature fusion stage (S5), the primary fused feature vector is aligned with the implicit rules through cross-modal attention, decoupling it into three independent components: behavioral abnormality probability, physiological stress index, and environmental risk level. After activation function and normalization constraints, these components form intermediate decision features. Finally, based on these features (S6), the pre-trained decision model, combined with an expert rule base, performs multi-dimensional parameter prediction and conflict resolution, outputting the feeding time execution point, precise feed amount, and feed type adjustment code. This ensures that the decision is optimized within the preset time window, flock size requirements, and nutritional standards, achieving a fully intelligent closed loop from data to decision. This principle integrates multimodal information fusion, knowledge graph reasoning, and rule-driven decision-making, significantly improving feeding accuracy and effectively addressing the dynamic complexity of farming scenarios.

[0014] As an optional embodiment of the present invention, optionally, obtaining the primary fusion feature vector in step S2 includes: S201. Based on visual-thermal imaging data, use a convolutional neural network to extract visual-thermal imaging features from the visual-thermal imaging data. The visual-thermal imaging features include motion trajectory and heat distribution features. The expression for extracting visual-thermal imaging features is: in, Indicates visual characteristics, This represents an instance segmentation model, and the output is a chicken flock mask plus the coordinates of the skeleton points. This represents a 4K / 8K visual input image. This indicates a pixel-by-pixel feature fusion operation. This represents a spatiotemporal graph convolutional network used to extract temporal motion features at 30fps. This represents time-series data on the movement trajectories of a flock of chickens, including the spatial location of individual chickens, changes in posture, and group movement patterns. Indicates thermal imaging features, This indicates that the 3×3 convolution kernel extracts the heat distribution characteristics. Indicates infrared thermal imaging input. This indicates a Gaussian blur filtering operation. This represents the linear smoothing filtering operation of a two-dimensional Gaussian function. Indicates the degree of smoothness. Indicates the characteristics of the motion trajectory. This describes a 3D reconstruction algorithm that outputs the spatial coordinates of the chicken flock. Represents the rotation matrix (for capturing attitude changes). This represents a translation vector (for locating the chickens). In step S201, when processing visual-thermal imaging data using a convolutional neural network, the 4K / 8K high-resolution image is first analyzed at the pixel level using an instance segmentation model to generate a chicken flock mask and skeletal point coordinates, accurately locating the individual spatial positions and posture changes. Subsequently, a pixel-wise feature fusion operation is used to deeply couple visual information and thermal imaging data at the feature level. The spatiotemporal graph convolutional network extracts temporal data of the chicken flock's movement trajectory at a frame rate of 30fps, capturing the spatial displacement, posture dynamics, and group movement patterns of individual chickens. The thermal imaging branch extracts heat distribution features from the infrared thermal imaging input using a 3×3 convolutional kernel, and combines this with a Gaussian blur filtering operation (two-dimensional Gaussian function linear smoothing filter, with adjustable smoothness) to suppress noise interference. Finally, a three-dimensional reconstruction algorithm outputs the chicken flock's spatial coordinates, and rotation matrices and translation vectors are used to accurately describe posture changes and group positioning. This process achieves multi-scale feature extraction of visual behavior and thermal imaging temperature.

[0015] S202. Extract acoustic-physiological features based on acoustic-physiological data. The acoustic-physiological features include Mel frequency cepstral coefficients, captured sound patterns, and physiological signals such as heart rate and respiratory rate. The expression for extracting acoustic-physiological features is: in, Indicates MFCC features, Represents the discrete cosine transform (compressed frequency domain information). This indicates the number of filters in the Mel filter bank. This represents the Mel filter bank (triangular filter matrix). Indicates the first One filter, Represents the FFT spectrum of an audio frame. Indicates heart rate characteristics, This indicates the calculation of the PPG crest interval. This represents the photoplethysmography (PPG) signal. Indicates the sampling rate. Indicates respiratory rate characteristics, This represents a counting function that counts the number of R-wave peaks in a respiratory waveform signal. This indicates the peak point in the respiratory waveform. Inside, This represents the respiratory waveform signal (acquired via a 9-axis IMU chest strap). Indicates the length of the time window; In step S202, when collecting acoustic-physiological data using a directional microphone array and wearable physiological sensors, the FFT spectrum of the audio frame is first decomposed into multiple frequency bands using a Mel filter bank. Discrete cosine transform is then used to compress the frequency domain information, generating Mel frequency cepstral coefficients (MFCC) to capture the acoustic pattern characteristics of the chicken's calls. Simultaneously, photoplethysmography (PPG) signals are acquired at a fixed sampling rate using a chest strap sensor, and heart rate features are extracted by calculating the interval between adjacent peaks. Respiratory waveform signals are monitored in real-time using a 9-axis IMU chest strap, and the number of R-wave peaks within a time window is counted using a counting function to determine the respiratory rate. This process combines the synchronous extraction of acoustic features and physiological signals.

[0016] S203. Extract environmental-metabolic features based on environmental-metabolic data. These features include normalized temperature parameters, humidity parameters, ammonia concentration parameters, and metabolic indicators. The expression for extracting environmental-metabolic features is: in, This indicates the temperature normalization result. This represents the real-time temperature sampling value. This represents the average of historical temperature data. The standard deviation of historical temperature data This represents the mean function. This represents the set of all temperature sample values ​​(such as real-time temperature data collected from multiple locations and time periods within a chicken coop). This represents the function for calculating the standard deviation. This represents the normalized ammonia concentration. This indicates the real-time reading of the ammonia sensor. Indicates the safety threshold. Indicates the standardized step size. This indicates the normalized results of metabolic indicators. Indicates fecal nitrogen content, Indicates the minimum nitrogen content in feces. This indicates the maximum nitrogen content in feces; In step S203, the collection and feature extraction of environmental-metabolic data are closely integrated with the real-time monitoring needs of the poultry farming scenario. Temperature normalization involves collecting real-time temperature data from multiple locations and time periods within the chicken house, calculating the historical temperature mean and standard deviation, and mapping the current temperature sample value to a standard normal distribution range, thus eliminating the impact of environmental fluctuations on the model. Ammonia concentration parameters are calculated by comparing the real-time readings of the ammonia sensor with a preset safety threshold, and adjusting the numerical range using a standardized step size to ensure the comparability of environmental risk indicators. Regarding metabolic indicators, fecal samples are collected periodically, and nitrogen content is measured and normalized to the minimum-maximum content range to quantify the metabolic health status of the flock. This process unifies the scale of multi-source environmental and metabolic data through normalization, preserving the physical meaning of the original data and facilitating parameter calibration and decision interpretation for the expert rule base.

[0017] S204. Based on the extracted visual-thermal imaging features, acoustic-physiological features, and environmental-metabolic features, feature-level fusion is performed using a multi-head attention mechanism to generate a primary fused feature vector.

[0018] The expression for feature-level fusion using a multi-head attention mechanism is: in, This represents the attention mechanism. This represents the query / key / value matrix (obtained by linear transformation of visual / acoustic / physiological features). Indicates normalization, The dimension of the key vector (controls the distribution of attention weights). Represents the primary fusion feature vector. This indicates a multi-head feature splicing operation (preserving the spatiotemporal correlation of each modality). This represents the output layer weight matrix. Indicates the first The learnable weight matrix of the head, Indicates the first Head attention mechanisms; In step S204, the multi-head attention mechanism dynamically captures the correlation between features of different modalities by computing multiple attention heads in parallel. Specifically, visual-thermal imaging features, acoustic-physiological features, and environmental-metabolic features are first mapped to query matrix, key matrix, and value matrix through linear transformation, respectively, where the dimension of the key vector is... This determines the sparsity of the attention weight distribution. Each attention head independently computes scaled dot product attention: by multiplying the query matrix by the transpose of the key matrix and dividing by... Scale normalization is performed, followed by the generation of intermodal correlation weights using the Softmax function. Finally, a weighted summation of the summation matrix yields the single-head output. The multi-head feature concatenation operation (Concat) concatenates the outputs of each attention head along the feature dimension, preserving the spatiotemporal correlation information between different modalities. A linear transformation is then applied to the output layer weight matrix to generate a primary fusion feature vector containing cross-modal interaction information. This mechanism adaptively adjusts feature importance through a learnable weight matrix, enabling the model to focus on key behavioral-physiological-environmental coupling features in the aquaculture scenario.

[0019] As an optional embodiment of the present invention, optionally, constructing a behavior-physiology-environment triplet knowledge graph based on the primary fusion feature vector in step S3 includes: S301. Based on the primary fusion feature vector, behavioral entities, physiological entities and environmental entities are extracted through the entity recognition module; among them, behavioral entities include flock movement patterns, feeding behavior and abnormal postures, physiological entities include heart rate variability, respiratory rate and body temperature distribution, and environmental entities include temperature fluctuations, humidity changes and ammonia concentration. In step S301, the entity recognition module employs a multi-level feature decoding architecture: First, a spatiotemporal graph convolutional network (ST-GCN) is used to parse the spatiotemporal correlation patterns in the primary fusion feature vector, where node features represent the individual behavior state of chickens, and edge weights encode the group interaction strength. This network dynamically models the flock topology through a learnable adjacency matrix, outputting behavioral entity labels (including pecking frequency, pacing trajectory clusters, abnormal wing flapping postures, etc.). Physiological entity recognition uses a bidirectional LSTM network to process the temporal physiological signals in the feature vector, capturing the temporal features of heart rate variability (such as RMSSD, LF / HF ratio) and respiratory waveform phase shift through a gating mechanism. Body temperature distribution entities generate the chicken surface temperature field through three-dimensional spatial interpolation of thermal imaging feature channels, and segment abnormally heated areas using existing region growing algorithms. Environmental entities are identified through sliding window statistics: temperature fluctuation entities calculate the standard deviation within adjacent 30-minute windows, humidity change entities extract the zero-crossing rate of the difference sequence, and ammonia concentration entities eliminate sensor drift errors through Kalman filtering. All entities are mapped to a predefined ontology library (such as AnimalBehaviorOntology), and the output is a triple data structure of <entity type, entity attribute, spatiotemporal stamp>.

[0020] S302. Use a relation extraction algorithm to define the types of relationships between behavioral entities, physiological entities, and environmental entities. The types of relationships include behavior influencing physiology, environment influencing behavior, and physiology influencing metabolism. In step S302, the relation extraction algorithm employs a semantic parsing framework based on graph neural networks. First, the triplet data structure output by the entity recognition module is mapped to a heterogeneous graph structure, where nodes represent behavioral, physiological, or environmental entities, and edges represent potential relation types between entities. Semantic relevance between nodes is dynamically calculated using a graph attention mechanism (GAT). Specifically, the feature vector of each node aggregates neighbor node information through a multi-head attention layer to generate a node embedding containing contextual semantics. For relation types where behavior influences physiology, the algorithm compares the spatiotemporal synchronization of chicken flock movement patterns (e.g., pecking frequency, pacing trajectory) with physiological entities (e.g., heart rate variability, respiratory rate). It uses the Dynamic Time Warping (DTW) algorithm to quantify the phase coupling strength between behavior and physiological signals. When the coupling coefficient exceeds a preset threshold, a relationship where behavior influences physiology is determined. For relationships where environmental factors influence behavior, the algorithm uses a sliding window to statistically analyze the abrupt changes in environmental entities (such as temperature fluctuations and ammonia concentration) and the response delays of behavioral entities (such as abnormal postures and feeding behavior). Granger causality tests are used to verify the predictive power of environmental changes on behavioral patterns. If the lagged terms of environmental variables significantly improve the goodness of fit of the behavioral sequence, the relationship is confirmed. For relationships where physiological factors influence metabolism, the algorithm combines regression analysis of physiological entities (such as body temperature distribution and respiratory rate) and metabolic indicators (such as fecal nitrogen content). Elastic mesh regularization is used to screen key physiological features. When the absolute value of the Pearson correlation coefficient between physiological parameters and metabolic indicators is greater than 0.6 and passes the significance test, a physiological-metabolic relationship is determined to exist. Finally, all relationship types are converted into structured knowledge representations using predefined rule templates (such as "behavioral entity → influence type → physiological entity") and stored in a triplet knowledge graph.

[0021] S303. Integrate behavioral entities, physiological entities, environmental entities, and relationship types to construct a triplet knowledge graph of behavior-physiology-environment.

[0022] In step S303, the construction process of the behavior-physiology-environment triplet knowledge graph adopts a hierarchical integration architecture. First, the triplets <entity type, entity attribute, spatiotemporal stamp> output by the entity recognition module are aligned and verified with the structured relations generated by the relation extraction module. Specifically, spatiotemporal stamp matching ensures the spatiotemporal synchronization of behavioral entities (such as pecking frequency) and corresponding physiological entities (such as heart rate variability). At the same time, semantic constraint rules in the ontology library (such as the "feeding behavior - stomach contraction frequency" association defined in AnimalBehaviorOntology) are used to filter out erroneous associations. In the graph initialization stage, Neo4j graph database is used as the storage engine. Behavioral entities, physiological entities, and environmental entities are modeled as different types of nodes, and directed edges are constructed through relation types (behavior affects physiology, environment affects behavior, physiology affects metabolism). To enhance the interpretability of the graph, each node is attached with an attribute dictionary to store quantitative indicators (such as the RMSSD value of the respiratory frequency node and the standard deviation of the temperature fluctuation node), and each edge is associated with a confidence score (calculated by DTW phase coupling strength or Granger causality test p-value). The knowledge graph update mechanism employs an incremental learning strategy. When newly collected aquaculture data triggers a preset threshold (e.g., the coefficient of variation of respiratory rate exceeds 15% for 30 consecutive minutes), the entity recognition and relation extraction process is dynamically activated, updating only the affected subgraph and avoiding the computational overhead of full graph reconstruction. The resulting knowledge graph supports three query modes: source tracing queries based on behavioral entities (e.g., tracing a sudden change in environmental ammonia concentration by inputting "abnormal posture"), prediction queries based on physiological indicators (e.g., predicting metabolic disease risk through body temperature distribution), and cross-modal association analysis (e.g., jointly analyzing the lagged correlation between humidity changes and feeding behavior).

[0023] As an optional embodiment of the present invention, the expression of the relation extraction algorithm is optionally: in, express, Indicates normalization, Represents the relation classification weight matrix. Represents the characteristics of a behavioral entity. Represents physiological physical characteristics, Represents the characteristics of environmental entities. Indicates the characteristics of metabolic entities. Indicates bias. This represents the standard triplet structure in a knowledge graph. , and All represent relation thresholds.

[0024] As an optional embodiment of the present invention, optionally, the implicit correlation rules between feeding frequency and body temperature in step S4 include: S401. Embed and initialize the behavioral entities, physiological entities, and environmental entities in the triplet knowledge graph, generate entity vectors, and encode the relationship types of behavioral entities, physiological entities, and environmental entities into a relation matrix; In step S401, when initializing the embedding of entities in the triplet knowledge graph, a pre-trained language model and graph structure feature fusion strategy are adopted. First, the text descriptions of behavioral entities (such as pecking frequency and pacing trajectory), physiological entities (such as heart rate variability and respiratory rate), and environmental entities (such as temperature fluctuation and ammonia concentration) are input into the BERT model to obtain context-related semantic vectors. At the same time, the adjacency matrix of the knowledge graph is feature extracted through a graph convolutional network (GCN) to generate graph embedding vectors containing topological structure information. The two types of vectors are weighted and concatenated (the weights are determined by cross-validation) to form the initial entity vector. Among them, behavioral entities focus on spatiotemporal dynamic features (such as the temporal variance of pecking frequency), physiological entities emphasize the periodic patterns of physiological signals (such as the frequency domain energy of respiratory rate), and environmental entities highlight spatial distribution characteristics (such as the three-dimensional gradient of the temperature field). The relationship matrix is ​​encoded using a bilinear transformation, mapping relationship types (behavior influencing physiology, environment influencing behavior, etc.) to learnable weight parameters. Relationship strength scores are calculated through interactions between entity vectors. For example, the relationship matrix elements between the behavioral entity "pecking frequency" and the physiological entity "stomach contraction frequency" are generated by the dot product of their vectors followed by sigmoid activation, with a threshold of 0.7 to filter weak associations. To preserve the physical meaning of the original data, all embedded vectors are scaled using L2 regularization and post-processed with ontology constraints (such as the "feeding behavior-digestive system" association rule defined in AnimalBehaviorOntology) to ensure that the generated entity vectors and relationship matrices possess both numerical representation capabilities and conform to the professional knowledge system of the aquaculture field.

[0025] S402. Based on entity vectors and relation matrices, cross-entity relation reasoning is performed using graph attention networks. Neighbor node information is aggregated through a multi-head attention mechanism to obtain aggregated node feature vectors and locate the association path between feeding frequency and body temperature. The expression for cross-entity relation reasoning is: in, Represents a node In the Updated features of the layer This represents the activation function. Represents a node The set of neighboring nodes, Indicates the first The weighting coefficients of head attention. Indicates the first The query weight matrix for head attention. Indicates the first Each neighbor entity feature vector Representation of relation matrix Embedded vectors of the corresponding relation type Indicates the number of heads of attention. This represents the output layer weight matrix. This represents the combined feature vector after concatenation and linear transformation by the multi-head attention mechanism. This indicates a multi-head feature splicing operation. Indicates the first Attention head to node Feature aggregation results This represents the scaling matrix of the multi-head splicing layer; In step S402, when the graph attention network performs cross-entity relation reasoning, it first performs feature aggregation for each node using its set of neighboring nodes. Specifically, for node i, its k-th attention head performs a dot product operation on the entity feature vector of neighboring node j and the embedding vector of the corresponding relation type based on the query weight matrix, thereby obtaining the original weight coefficients of the k-th attention head. Subsequently, the Softmax function is used to normalize these original weight coefficients to obtain the weight coefficients of the k-th attention head, which reflect the importance of neighboring node j to node i in the k-th attention dimension.

[0026] Next, the entity feature vector of each neighbor node j is weighted and summed with the embedding vector of the corresponding relation type, where the weights are the weight coefficients of the k-th attention head, resulting in the feature aggregation result of the k-th attention head for node i. This process is executed in parallel through a multi-head attention mechanism, with each attention head focusing on a different feature subspace, thereby capturing diverse relational patterns between nodes.

[0027] After aggregating features from all attention heads, a multi-head feature concatenation operation is applied to these aggregating results. A comprehensive feature vector is generated through concatenation and linear transformation. During this process, a scaling matrix is ​​used to adjust the dimension of the concatenated vector to ensure it matches the output layer weight matrix. Finally, the comprehensive feature vector is linearly transformed through the output layer weight matrix to obtain the feature vector at node i in the first position. Updated features of the layer.

[0028] When locating the association path between feeding frequency and body temperature, graph attention networks pay special attention to nodes directly connected to these two entities or indirectly connected via short paths. By analyzing the updated feature vectors of these nodes, potential implicit association rules between feeding frequency and body temperature can be identified. For example, if the updated feature vectors of the feeding frequency node and the body temperature node show high similarity across multiple attention heads, and this similarity corresponds to a strong association type (such as behavior influencing physiology) in the relationship matrix, then it can be preliminarily determined that there is some implicit association between feeding frequency and body temperature.

[0029] To further verify this association, visualization tools can be used to visualize the attention weights in the graph attention network, intuitively presenting the strength and direction of the association path between feeding frequency and body temperature. Simultaneously, by incorporating expertise from the aquaculture field, the identified association rules can be interpreted and evaluated to ensure they conform to the physiological and behavioral patterns observed in actual aquaculture scenarios.

[0030] S403. Based on node feature vectors and associated paths, use GNN two-hop inference to generate path weights from feeding frequency to body temperature, and combine cosine similarity to calculate rule confidence. in, This represents the path weight from feeding frequency to body temperature. Represents the behavioral entity. Represents a physiological entity. This represents the activation function. Represents the gate weight matrix. , This represents the feature vector of the feeding frequency behavior entity and the body temperature physiological entity after being updated by a two-hop GNN. It represents the Hadamah accumulation. Represents the cosine similarity function. Represents the embedding vector of the behavioral entity. Represents the physiological entity embedding vector. Indicates the relation type decay coefficient. Represents a node In the Updated features of the layer In step S403, the core of the GNN two-hop inference lies in capturing the indirect relationship between the feeding frequency behavior entity and the body temperature physiological entity through two propagation processes of the graph neural network, via intermediate nodes (such as environmental entities or other physiological entities). Specifically, firstly, a weighted combination of the feature vectors of the feeding frequency behavior entity and the body temperature physiological entity after the one-hop GNN update is performed using a gating weight matrix. This process uses the Hadamard product to perform element-wise multiplication of the feature vectors to preserve the interaction information between the two. Subsequently, the weighted feature vectors are input into an activation function (such as ReLU or Sigmoid) for nonlinear transformation to generate intermediate features containing two-hop relationship information.

[0031] When calculating path weights, a relation type decay coefficient is introduced to quantify the impact of intermediate nodes on the associated path. For example, if feeding frequency indirectly affects body temperature through the environmental ammonia concentration node, the relation type decay coefficient corresponding to the ammonia concentration node will reduce the overall path weight, reflecting the weakening effect of indirect association. Simultaneously, the updated features of nodes at layer l are compared with the embedding vectors of behavioral and physiological entities using a cosine similarity function to calculate rule confidence. Cosine similarity quantifies the reliability of the association rule between feeding frequency and body temperature by measuring the consistency of the directions of the two vectors. If the cosine similarity value is close to 1, it indicates that the directions of the two feature vectors are highly consistent, and the association rule has high confidence; conversely, if the value is close to 0, the association rule has low confidence.

[0032] Ultimately, the combined result of path weights and rule confidence is used to assess the strength of the implicit association between feeding frequency and body temperature. For example, if the path weight is high and the rule confidence exceeds a preset threshold (e.g., 0.8), a significant association can be identified, thus providing a basis for feeding decisions. This process ensures that the generated feeding strategy conforms to the laws of aquaculture biology and possesses data-driven precision by quantifying the physical meaning and statistical significance of the association rules.

[0033] S404. Based on path weight and rule confidence, information flow is controlled through a gating mechanism to extract high-confidence rules and map them to a predefined rule template library to obtain the mapping results. In step S404, the gating mechanism is designed based on a dynamic threshold comparison between path weights and rule confidence. Its core lies in filtering input information using a learnable parameter matrix. Specifically, the system first concatenates the path weights and rule confidence into a joint feature vector, and then maps it to the gating activation space through a fully connected layer. This space is generated by the Sigmoid function, producing continuous values ​​between 0 and 1, which serve as the throughput of the information flow. For example, when the path weight is higher than a preset dynamic threshold (which is adaptively adjusted based on historical data) and the rule confidence exceeds 0.85, the gating output is close to 1, allowing the corresponding rule to enter the subsequent processing flow; conversely, the output is close to 0, suppressing the transmission of low-quality rules.

[0034] During information flow control, the system employs a residual connection strategy to preserve original rule features and prevent gradient vanishing. High-confidence rules are extracted through thresholding: rules are marked as "high-confidence" only when the gate output is greater than 0.7, and then enter the rule template library mapping stage. The predefined rule template library contains 52 association patterns summarized by experts in the aquaculture field (such as "behavior-physiology" causal chains, "environment-behavior" triggering relationships, etc.). Each template defines a standardized format for input entity type, relationship type, and output decision. The mapping process is completed by calculating the semantic similarity between high-confidence rules and each pattern in the template library. Similarity calculation uses BERT-based sentence embedding comparison, combined with the Jaccard index to measure entity overlap. Finally, the system selects the template with the highest similarity as the mapping result; if the maximum similarity is less than 0.6, a manual review process is triggered. For example, when a rule chain of "phasing frequency ↑ → stomach contraction frequency ↑ → body temperature ↑" is identified, the system can automatically match it to the "behavior chain-physiological feedback" template, generating a decision suggestion of "increasing feeding frequency to maintain stable body temperature."

[0035] S405. Generate a set of implicit association rules between feeding frequency and body temperature based on high-confidence rules and mapping results.

[0036] In step S405, generating a set of implicit association rules between feeding frequency and body temperature based on high-confidence rules and mapping results includes: S4051, setting a confidence threshold, filtering high-confidence rules, and retaining high-confidence rules with confidence levels higher than the preset threshold; setting the confidence threshold in step S4051 is a key step to ensure the quality of implicit association rules. The system first dynamically determines a reasonable lower confidence limit based on the actual needs and data characteristics of the farming scenario. For example, in large-scale farms requiring high decision-making accuracy, this threshold is set to 0.85 in this embodiment to filter out rules with low reliability; while in the exploratory research phase, the threshold can be appropriately relaxed to 0.75 to retain more potential associations. In specific implementation, the system will statistically analyze the confidence distribution of effective rules in historical data, combine expert experience to set an initial threshold, and continuously optimize it through cross-validation. During the screening process, the system iterates through all high-confidence rules, retaining only those with confidence levels above a preset threshold. For example, it includes the rule "increased feeding frequency → increased stomach contraction frequency → increased body temperature" (confidence level 0.88) in the set, while removing the rule "changes in activity level → fluctuations in body temperature" (confidence level 0.72). This threshold-based filtering ensures the rigor of the subsequent rule set. Simultaneously, the system records the confidence level and reasons for removed rules, creating a screening log for later analysis. For instance, if a rule is removed because its confidence level (0.79) is below the threshold of 0.8, the log will indicate "insufficient confidence (0.79 < 0.8)" and possible influencing factors (such as insufficient sample size). Ultimately, the filtered rule set will only contain high-quality rules that pass the confidence level test.

[0037] S4052. Based on high-confidence rules, traverse the predefined rule template library, match the mapping results with the templates in the predefined rule template library, and select the most suitable rule template. In step S4052, the matching process of the predefined rule template library relies on the semantic and structural similarity between high-confidence rules and each template in the template library. The system first extracts entity types (e.g., behavioral entity "feeding frequency," physiological entity "body temperature"), relationship types (e.g., causal relationship, triggering relationship), and logical directions (e.g., positive influence, negative inhibition) from the high-confidence rules, and encodes them into structured feature vectors. Simultaneously, each template in the template library is also converted into the same feature representation, containing predefined entity type constraints, relationship type patterns, and output decision formats. During matching, the system employs a two-stage strategy: the first stage calculates the semantic similarity between the rule description and the template description using a BERT-based sentence embedding model, filtering out candidate templates with a similarity higher than 0.6; the second stage combines the Jaccard index to calculate the overlap rate between rule entities and template entities, further narrowing down the candidate range. For example, for the rule "feeding frequency ↑ → stomach contraction frequency ↑ → body temperature ↑", the system will prioritize matching templates containing a "behavioral-physiological" causal chain and whose output decision is "adjust feeding strategy". If multiple candidate templates with similarity exist (such as "behavioral chain - physiological feedback" and "intake behavior - body temperature regulation"), the system will introduce domain knowledge weights, prioritizing templates with higher expert annotation priority. Ultimately, the system selects the template with the highest overall similarity as the matching result and records the similarity score, candidate template list, and selection criteria during the matching process, forming a traceable matching log. If the similarity of all templates is below a threshold (e.g., 0.5), a manual review process is triggered, where domain experts intervene to adjust the template library or redefine the rules.

[0038] S4053, Instantiate the most suitable rule template to generate specific implicit association rules between feeding frequency and body temperature; In step S4053, the instantiation process is the core step in transforming the most suitable rule template into specific, executable implicit association rules. The system first extracts the standardized format of the selected template from the matching results, including input entity type constraints (e.g., it must include behavioral and physiological entities), relation type definitions (e.g., causal relationships require explicit antecedent and consequent), and output decision frameworks (e.g., the calculation method for feeding frequency adjustment). Subsequently, the system fills the corresponding positions in the template with specific entities (e.g., replacing "feeding frequency" with "pecking frequency" and "body temperature" with "rectal temperature") and relational parameters (e.g., replacing "↑" with a specific numerical change range "increase of 15%-20%) from the high-confidence rules. For example, for the matched "behavioral chain-physiological feedback" template, if the input rule is "feeding frequency ↑ → gastric contraction frequency ↑ → body temperature ↑", the system will transform it into a specific rule description: "When the number of pecking increases by 15-20 times per hour, the gastric contraction frequency increases by 3-5 times per minute, leading to an increase in rectal temperature of 0.3-0.5℃." Simultaneously, the system generates quantifiable decision suggestions based on the decision calculation logic in the template, combined with aquaculture biological parameters (such as the energy consumption increment corresponding to each 0.1℃ change in body temperature), for example, "It is recommended to increase the daily feed intake by 8%-12% to maintain stable body temperature." During instantiation, the system performs multi-level checks: the first level checks the consistency of entity types, ensuring that the filled entities conform to the categories defined in the template (e.g., behavioral entities must be observable actions); the second level checks the rationality of the relational logic, verifying whether the transformed rules conform to aquaculture knowledge (e.g., "increased feed intake" should not lead to "decreased body temperature"); the third level checks the validity of the numerical range, checking whether the parameter changes are within the physiologically acceptable range (e.g., body temperature changes do not exceed ±1℃). If a conflict is found during the checks (e.g., the rule logic contradicts the template definition), the system will trigger a rollback mechanism, reselecting a suboptimal template or marking it as a rule awaiting manual review. Ultimately, the instantiated implicit association rules will include a complete entity-relationship-decision chain, a confidence range (e.g., 0.82-0.88), an applicable scenario description (e.g., "applicable to adult chickens in an environment with an ambient temperature of 25-30℃"), and a version identifier. These rules will be stored in the rule base in a structured format (e.g., JSON), while the association index between the rules and the template will be updated.

[0039] S4054. Integrate all generated implicit association rules to form an implicit association rule set.

[0040] In step S4054, the integration process is a step of systematically summarizing and structuring the multiple specific implicit association rules generated in the previous steps. The system first classifies all instantiated rules according to entity type (behavior-physiology, environment-behavior, etc.) and relationship type (causality, triggering, synergy). For example, "increased pecking frequency → increased stomach contraction frequency → increased body temperature" and "increased drinking frequency → enhanced kidney metabolic rate → body temperature regulation" are grouped into a behavior-physiology causal chain, while "increased environmental temperature → decreased activity" is separately classified as an environment-behavior triggering category. After classification, the system performs secondary semantic and logical checks on each type of rule: semantic checks ensure consistency in the expression of rules of the same type by calculating the cosine similarity between rule descriptions (e.g., unifying the terminology of "feeding frequency" and "pecking frequency"); logical checks verify the biological rationality of the antecedents and consequents of the rules based on the ontology library of the aquaculture field (e.g., "increased feed intake" should correspond to "enhanced digestive system activity" rather than "sudden drop in body temperature").

[0041] As an optional embodiment of the present invention, optionally, the generation of intermediate decision features in step S5 includes: S501. Perform dimensional matching and semantic alignment between the primary fusion feature vector and the implicit association rules to obtain the aligned feature vector. In step S501, dimensional matching and semantic alignment are fundamental steps to ensure the effective fusion of the primary fusion feature vector and the implicit association rules. The system first analyzes the dimensional structure of the primary fusion feature vector to clarify the distribution of its included behavioral features (such as pecking frequency and activity level), physiological features (such as body temperature and heart rate), and environmental features (such as temperature and humidity). At the same time, the system parses the entity types (such as behavioral entities and physiological entities) and relation types (such as causal relationships and triggering relationships) defined in the implicit association rules and converts them into semantic representations compatible with the feature vector. For example, the rule "feeding frequency↑→stomach contraction frequency↑→body temperature↑" is converted into a semantic label containing a behavioral-physiological causal chain and associated with the corresponding dimensions "pecking frequency", "stomach contraction frequency", and "body temperature" in the feature vector. During semantic alignment, the system employs synonym expansion and concept mapping techniques based on an ontology library in the aquaculture field to resolve terminological differences between feature names and rule entities (e.g., unifying "feeding frequency" and "pecking count" into the same concept). Furthermore, the system checks the compatibility of feature vectors and rules across time scales (e.g., real-time data versus daily cumulative rules) and spatial scopes (e.g., individual data versus group rules), performing data aggregation or decomposition as necessary. For example, if the feature vector provides the number of pecking counts per minute, while the rule describes hourly feed intake changes, the system converts minute-level data to hourly granularity using time window statistics. After alignment, the system generates an aligned feature vector whose dimensions perfectly match the input requirements of the implicit association rule, and the semantics of each dimension precisely correspond to the entities and relationships in the rule. This process is efficiently handled by automated scripts, while simultaneously recording the mapping relationships and transformation parameters during alignment, forming a traceable alignment log.

[0042] S502. The feature patterns of implicit association rules are integrated into the aligned feature vector through a cross-modal attention mechanism to obtain fused features. In step S502, the cross-modal attention mechanism is the core technology connecting the primary fused feature vector and the implicit association rules. Its core lies in achieving deep interaction of feature patterns through dynamic weight allocation. The system first constructs a dual-branch attention network. One branch inputs the aligned feature vector (containing multi-dimensional features of behavior, physiology, and environment), and the other branch inputs the semantic encoding of implicit association rules (such as the weight matrix of the behavior-physiology causal chain). In the attention calculation stage, the system adopts a scaled dot product attention model, using each dimension in the feature vector as a query and the entity relationship in the rule semantic encoding as the key and value. For example, when processing the "pecking frequency" feature, the network will prioritize the "feeding behavior → digestive system" causal chain related to it in the rule base, and dynamically determine the contribution weight of each rule to the current feature by calculating the similarity score between the query and the key. To enhance domain adaptability, the system incorporates aquaculture biology constraints into attention calculation, adjusting the original attention score through a predefined rule priority matrix (e.g., physiological feedback rules have higher weights than environmental trigger rules). Specifically, the system maps the confidence range of a rule (e.g., 0.82-0.88) to an attention weight scaling factor, allowing high-confidence rules to achieve a larger feature fusion ratio. During fusion, the system employs a residual connection strategy to preserve original feature information, preventing feature distortion due to excessive rule intervention. For example, for the "body temperature" feature, the network retains the basic influence of environmental temperature on body temperature while incorporating the rule "increased feed intake → increased body temperature". The final fused feature not only includes the spatiotemporal characteristics of the original data but also embeds rule-driven causal reasoning capabilities. Each dimension is labeled with a rule source identifier (e.g., "R001: behavior-physiological causal chain") and a fusion weight value, forming an interpretable enhanced feature representation. This feature achieves full interaction of cross-modal information through a multi-head attention mechanism. Each attention head focuses on different rule types (such as causal relationship head and trigger relationship head), and finally generates a fusion feature vector with unified dimensions through concatenation and linear transformation.

[0043] S503. Decouple the fused features into three independent components: behavioral anomaly probability, physiological stress index, and environmental risk level. In step S503, the decoupling process is a crucial step in decomposing the fused feature vector into independent decision factors with clear business implications. The system first constructs a decoupling model based on a knowledge graph of the aquaculture field. This model contains three parallel processing sub-networks: a behavioral anomaly detection network, a physiological stress assessment network, and an environmental risk quantification network. Each sub-network employs an independent feature selection strategy. For example, the behavioral anomaly detection network focuses on abnormal behavioral patterns such as sudden changes in pecking frequency and a sharp decrease in activity; the physiological stress assessment network focuses on physiological indicators such as body temperature fluctuations and heart rate variability; and the environmental risk quantification network analyzes extreme values ​​of environmental parameters such as temperature and humidity. During decoupling, the system performs a nonlinear transformation on the fused features using a multilayer perceptron. The first layer maps the high-dimensional fused features to three subspaces, each corresponding to a decision component. The second layer dynamically adjusts the feature weights within each subspace using a gating mechanism; for example, when the pattern of "sudden increase in body temperature + increased respiratory rate" is detected, the gating value of the physiological stress subnetwork increases significantly. The third layer converts the subspace features into probability or exponential values ​​using a normalized exponential function (Softmax). In practice, the system incorporates domain prior knowledge to constrain the decoupling process. For example, a predefined "behavior-physiology correlation matrix" ensures the correlation between the probability of abnormal behavior and the physiological stress index (e.g., a high probability of abnormal behavior corresponds to a moderate or higher physiological stress index). Simultaneously, the system employs an adversarial training strategy to enhance the independence of the decoupled features. The discriminator network distinguishes whether the decoupled components contain information from other components. If cross-interference is detected (e.g., the abnormal behavior component contains ambient temperature information), the decoupling model parameters are adjusted through a gradient inversion layer. The three decoupled components have clearly defined scales: the probability of behavioral abnormality ranges from [0,1], representing the degree to which the current behavioral pattern deviates from the normal baseline; the physiological stress index uses standardized Z-scores to reflect the degree to which the physiological state deviates from the group mean; and the environmental risk level is divided into three levels: low (0-0.3), medium (0.3-0.7), and high (0.7-1), corresponding to different warning thresholds. The system generates a detailed explanatory report for each component, including the contribution of key features (e.g., "the contribution of elevated body temperature to the physiological stress index is 65%"), rule trigger records (e.g., "triggering rule R001 increases the probability of behavioral abnormality by 0.2"), and historical comparison data (e.g., "the current probability of behavioral abnormality is 2.3 standard deviations higher than the group mean for the same age"). Finally, the three decoupled components are output in a structured format.

[0044] S504, the activation function and normalization operation apply numerical constraints to the probability of abnormal behavior, the physiological stress index and the environmental risk level, generating intermediate decision features that include the probability of abnormal behavior, the physiological stress index and the environmental risk level.

[0045] In step S504, the activation function and normalization operation are key steps to ensure the stability and interpretability of intermediate decision features. The system first selects suitable activation functions for the characteristics of the three independent components (probability of abnormal behavior, physiological stress index, and environmental risk level): For the probability of abnormal behavior, the Sigmoid function is used to compress its output range to [0,1], ensuring the validity of the probability value; for the physiological stress index, the Tanh function is used to map it to the [-1,1] interval, and then a linear transformation is applied to convert it into a Z-score form, maintaining its relative relationship with the population mean; for the environmental risk level, a piecewise linear function discretizes the continuous value into three levels: low (0-0.3), medium (0.3-0.7), and high (0.7-1), matching the early warning threshold system. During the normalization phase, the system employs a dynamic range adjustment strategy: the abnormal behavior probability retains its original probability value, the physiological stress index is standardized using Z-score based on the population data distribution (e.g., the difference between the current value and the population mean divided by the standard deviation), and the environmental risk level linearly maps the original risk score to the [0,1] interval through Min-Max normalization. To prevent extreme values ​​from interfering with decision-making, the system introduces robust processing mechanisms, such as truncating samples with the physiological stress index exceeding ±3 standard deviations and smoothing the boundary values ​​of the environmental risk level (e.g., 0.3 and 0.7). Simultaneously, the system records parameters (e.g., population mean, standard deviation, mapping coefficients) and activation function types during the normalization process, forming a traceable numerical transformation log. The final generated intermediate decision features have clear business implications: the abnormal behavior probability directly reflects the degree to which an individual's behavior deviates from the normal pattern, the physiological stress index quantifies abnormal fluctuations in physiological state, and the environmental risk level provides an early warning level for environmental parameters. These features are stored in a structured format (e.g., JSON), with each component accompanied by metadata such as activation function type, normalization parameters, and calculation timestamp.

[0046] As an optional embodiment of the present invention, optionally, generating the feeding decision instruction in step S6 includes: S601. Based on intermediate decision features, a pre-trained decision model is used to predict multi-dimensional parameters and generate preliminary feeding decision outputs. The outputs include feeding time adjustment values, feeding amount adjustment values, and feed type adjustment parameters. In step S601, the pre-trained decision model is a deep neural network trained on a large amount of aquaculture data. This model takes intermediate decision features (probability of abnormal behavior, physiological stress index, and environmental risk level) as input and achieves accurate prediction of feeding parameters through multi-layer nonlinear transformation. The model structure includes three core modules: the feature embedding layer converts the structured input into a high-dimensional vector representation and uses a 1×1 convolutional kernel to extract the interaction features between components; the temporal prediction layer uses a bidirectional LSTM network to capture the dynamic impact of historical decisions and stores the feeding adjustment records of the past 7 days through a memory unit; the output decoding layer dynamically allocates the weights of each feature dimension using an attention mechanism. For example, when the physiological stress index increases significantly, the model will automatically increase the prediction weight of the feeding amount adjustment. During training, the model employs a multi-task learning framework to simultaneously optimize three prediction objectives. The loss function consists of a weighted average of mean squared error and classification cross-entropy, with regression loss used for feeding time adjustment and classification loss used for feed type adjustment. To enhance model robustness, 20% noisy samples and 15% cross-species farming data were introduced into the training data, and an adversarial training strategy was used to improve the model's predictive stability under abnormal conditions. The final preliminary feeding decision output includes adjustment values ​​in three dimensions: feeding time adjustment in minutes (range ±60 minutes), feeding amount adjustment in grams (range ±200 grams), and feed type adjustment parameters using one-hot encoding (e.g., the probability vector for switching from basal feed to high-protein feed). The system generates a confidence score (0-1 interval) and uncertainty interval for each prediction value (e.g., the 95% confidence interval for feeding amount adjustment is ±15 grams), and reveals key decision factors through SHAP value analysis (e.g., "the physiological stress index contributes 72% to feeding amount adjustment"). The output is stored in a structured format, and records the model version number, prediction timestamp, and snapshot of the input features.

[0047] S602. Combine the expert rule base to perform rule verification and conflict resolution on the preliminary decision output, and then integrate the decision model output and expert rules through a weighted fusion algorithm to generate an optimized feeding instruction. In step S602, the expert rule base is a knowledge system built by combining the experience of experts in the aquaculture field, containing hundreds of decision-making rules for different aquaculture scenarios. The system first performs pattern matching between the preliminary decision outputs (feeding time adjustment values, feed amount adjustment values, and feed type adjustment parameters) and the rules in the rule base. For example, when "physiological stress index > 0.8 and environmental risk level is high" is detected, the rule combination of "emergency feed increase + electrolyte addition" is triggered. During the rule verification phase, the system uses a forward chain reasoning mechanism, starting from the initial facts (intermediate decision features) and gradually deriving conclusions through rule condition judgments, while recording the trigger rule numbers (such as R203, R205) in the reasoning path. During conflict resolution, the system employs a priority ranking strategy for possible multi-rule matching situations (such as simultaneously triggering "feed increase" and "feed restriction" rules): weighted ranking based on rule source (e.g., veterinary advice > historical experience > default settings), confidence score (model prediction confidence > rule preset threshold), and timeliness (real-time data rules > historical statistical rules). For example, when the model's predicted feed adjustment value (+150g) conflicts with the "high-temperature feed restriction" rule (-100g) in the rule base, the system will prioritize adopting the model output (because its confidence level of 0.92 is higher than the rule threshold of 0.85), but will correct the final adjustment value to +80g through a weighted fusion algorithm (model weight 0.6 × 150g + rule weight 0.4 × (-100g)). During the weighted fusion stage, the system dynamically calculates the fusion weights of the model output and expert rules: for decisions related to the probability of abnormal behavior, the model weight is set to 0.7 (because it is based on big data analysis), and the rule weight is set to 0.3; for decisions related to physiological stress, the model and rule weights each account for 0.5 (balancing data and experience); for decisions related to environmental risks, the rule weight is increased to 0.6 (emphasizing the expert's judgment on environmental factors). The fused feeding instructions consist of three parts: basic adjustment values ​​(original model output), expert correction values ​​(adjustments triggered by rules), and final decision values ​​(weighted fusion results), such as "Feeding time: Model suggestion +30 minutes, rule trigger -15 minutes, final +18 minutes." The system generates an explanation report for each instruction, including a list of triggered rules, the basis for weight allocation, the conflict resolution process, and comparisons with similar historical cases (e.g., "Decision adjustments similar to those made during the high temperatures of June 2023"). The final optimized feeding instructions are output in a standardized format, including feeding time (accurate to minutes), feeding amount (accurate to grams), feed type (coded format), and adjustment basis (model version number + rule base version number).

[0048] S603. Normalize and constrain the optimized feeding instructions to ensure that the feeding time is within the preset time window, the feeding amount meets the needs of the flock size, and the feed type matches the nutritional standards. In step S603, normalization and range constraint processing are crucial steps to ensure the feasibility and safety of feeding instructions. The system first establishes independent constraint rules for three dimensions: feeding time, feeding amount, and feed type. Feeding time uses a time window constraint, limiting the adjusted values ​​output by the model to the range allowed by the flock's physiological rhythm (e.g., 4 AM to 8 PM). Values ​​exceeding this range are truncated to the most recent valid value. Feeding amount is calculated dynamically based on the flock's age, weight, and breed parameters, generating a basic requirement (e.g., 200 grams / bird / day). Adjusted values ​​are allowed to fluctuate within ±30% of the basic requirement, triggering a secondary verification mechanism when the threshold is exceeded. Feed type uses a coding mapping table constraint to ensure that the adjusted feed type meets the nutritional standards for the current growth stage (e.g., during the brooding period, it must contain more than 21% crude protein). During the normalization phase, the system normalizes feeding time adjustments at the minute level (e.g., mapping ±60 minutes to the [0,1] interval), and feed amounts are relatively normalized as a percentage of the baseline amount per gram (e.g., +50 grams is mapped to 0.25). Feed types are normalized by entropy value normalization using probability vectors converted from unique heat codes (ensuring the sum of probabilities for all feed types is 1). To enhance the robustness of constraints, the system introduces a dynamic adjustment mechanism: when the environmental risk level is high, the feeding amount adjustment range is automatically tightened to ±15%; when the probability of abnormal behavior exceeds 0.7, feed type adjustments are paused, and only time and amount optimization is performed. Regarding conflict handling, the system establishes a three-dimensional constraint matrix. When feeding time conflicts with feed type adjustments (e.g., changing feed type at night), the compliance of feeding time is prioritized; when feeding amounts conflict with nutritional standards (e.g., excessive feeding due to high-protein feed), the feeding amount is corrected according to nutritional standard priority. The final generated constraint instructions contain three validation fields: time validity flag (0 / 1), quantity compliance flag (0 / 1), and type matching flag (0 / 1). A manual review process is triggered when any flag is 0. The system records all constraint operation logs, including the original value, adjustment amount, constraint rule version number, and operator ID, forming a traceable decision-making chain. The constraint-based feeding instructions are output in enhanced JSON format, with a new "Constraint Type" field indicating the reason for the adjustment (e.g., "Time Window Limitation" or "Nutritional Standard Matching"), and include a comparison chart before and after constraint (e.g., an overlay of the feeding amount adjustment curve and the allowable range).

[0049] S604. Output the final feeding decision instruction, which includes the feeding time execution point, the precise value of the feeding amount, and the feed type adjustment code.

[0050] In step S604, the system uses a structured data protocol to encapsulate the instruction content, which includes three core fields: the feeding time execution point is stored in dual mode using an absolute timestamp (UTC+8 time zone, accurate to the second) and a relative time offset (the minute difference relative to the most recent standard feeding time), for example, "2024-03-15T06:30:00+08:00 (relative +15 minutes)", ensuring that execution can still be based on the relative offset even in the event of clock synchronization failure; the precise value of the feeding amount is recorded separately through the basic amount and the adjustment amount. The basic amount is dynamically calculated based on the flock's age-weight model (e.g., the basic amount for 20-day-old broilers is 180 grams / bird), and the adjustment amount includes... It includes triple data: model prediction, rule correction, and final fusion value. For example, "base amount 180 grams, model suggestion +25 grams, rule correction -10 grams, final execution 195 grams". The unit of measurement (grams / animal) and calculation accuracy (±1 gram) are also indicated. The feed type adjustment code adopts a three-level coding system. The first letter indicates the feed category (P-protein / E-energy / V-vitamin), the middle three digits represent the specific formula number (e.g., P001 represents a 21% crude protein formula), and the last letter indicates the adjustment direction (A-increase / R-decrease / M-maintain). For example, "P001A" means increasing the proportion of 21% protein feed.

[0051] When issuing commands, the system simultaneously generates decision traceability information, including model version number, rule base version number, input feature snapshot (abnormal behavior probability 0.62 / physiological stress index 1.23 / environmental risk level 0.8), and decision path graph (visually displaying the processing trajectory of each stage from input to output). To adapt to different aquaculture scenarios, the system provides three output modes: fully automatic mode directly pushes commands to feeding equipment for execution; semi-automatic mode triggers a manual confirmation process after command generation, with a 10-minute buffer period for the feeder to review; emergency mode automatically switches to the basic feeding plan when a system anomaly is detected, while simultaneously pushing alarm information to the management terminal. The transmission of output commands uses the encrypted MQTT protocol, transmitted via the farm's intranet or 4G / 5G dual-link redundancy, ensuring that commands can still be issued even in the event of a single network failure. Finally, the commands are stored in a blockchain distributed ledger, recording metadata such as decision time, executing device ID, and operator digital signature, forming an immutable decision audit trail.

[0052] As an optional embodiment of the present invention, the expression of the decision model may be: in, This represents the initial set of decision parameters. This represents the linear transformation matrix of the output layer. This represents the activation function. Represents the linear transformation matrix of the hidden layer. Indicates the characteristics of intermediate decisions. This represents the hidden layer offset. This represents the output layer offset. Indicates time adjustment. Indicates time adjustment. This refers to the Softplus function. Indicates time adjustment. This indicates normalization.

[0053] Example 2 A multimodal information fusion feeding decision system for recognizing the behavior of farmed chickens includes: processor; Memory used to store processor-executable instructions; The processor is configured to implement a multimodal information fusion feeding decision method for recognizing the behavior of farmed chickens when executing executable instructions.

[0054] It should be noted that the computer device includes a processor, a memory, and may also include one or more of a multimedia component, an input / output (I / O) interface, and a communication component.

[0055] The processor controls the overall operation of the computer device to complete all or part of the steps in the multimodal information fusion feeding decision method for the recognition of poultry behavior described above.

[0056] Memory is used to store various types of data to support the operation of the computer device. This data may include, for example, instructions for any application or method used to operate on the computer device, as well as application-related data. Memory 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 storage, flash memory, magnetic disk, or optical disk.

[0057] The multimedia component may include a screen and an audio component, wherein the screen may be, for example, a touch screen, and the audio component is used to output and / or input audio signals; for example, the audio component may include a microphone for receiving external audio signals, the received audio signals may be further stored in memory or transmitted via a communication component; the audio component may also include at least one speaker for outputting audio signals.

[0058] I / O interfaces provide interfaces between the processor and other interface modules, such as keyboards, mice, buttons, etc.; these buttons can be virtual buttons or physical buttons.

[0059] The communication component is used for wired or wireless communication between the computer device and other devices; wireless communication, such as Wi-Fi, Bluetooth, Near Field Communication (NFC), 2G, 3G, 4G or 5G, or one or more combinations thereof, and the corresponding communication component may include: Wi-Fi module, Bluetooth module, NFC module, mobile communication module.

[0060] As a preferred embodiment, the computer device may be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to execute the above-described multimodal information fusion feeding decision method for recognizing the behavior of farmed chickens.

[0061] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A multimodal information fusion feeding decision-making method for recognizing the behavior of farmed chickens, characterized in that, The method includes: S1. Collect multi-source heterogeneous datasets of chicken flocks, including: visual-thermal imaging data, acoustic-physiological data, and environmental-metabolic data; S2. Extract and fuse features from the multi-source heterogeneous dataset to obtain a primary fused feature vector; S3. Construct a behavior-physiology-environment triplet knowledge graph based on the aforementioned primary fusion feature vector; S4. Based on the triplet knowledge graph, use GNN to perform relational reasoning to mine the implicit association rules between feeding frequency and body temperature; S5. Perform secondary feature fusion between the primary fusion feature vector and the implicit association rule to generate intermediate decision features that include the probability of abnormal behavior, physiological stress index and environmental risk level. S6. Based on the intermediate decision features, a feeding decision instruction is generated through a pre-trained decision model and an expert rule base. The feeding decision instruction includes feeding time, feeding amount, and feed type adjustment parameters.

2. The multimodal information fusion feeding decision-making method for chicken behavior recognition as described in claim 1, characterized in that, The primary fusion feature vector obtained in step S2 includes: S201. Based on the visual-thermal imaging data, a convolutional neural network is used to extract visual-thermal imaging features from the visual-thermal imaging data. The visual-thermal imaging features include motion trajectory and heat distribution features. S202. Based on the acoustic-physiological data, extract acoustic-physiological features, including Mel frequency cepstral coefficients, captured sound patterns, and physiological signals such as heart rate and respiratory rate. S203. Extract environmental-metabolic features based on the environmental-metabolic data. The environmental-metabolic features include normalized temperature parameters, humidity parameters, ammonia concentration parameters, and metabolic indicators. S204. Based on the extracted visual-thermal imaging features, acoustic-physiological features, and environmental-metabolic features, feature-level fusion is performed using a multi-head attention mechanism to generate a primary fused feature vector.

3. The multimodal information fusion feeding decision-making method for chicken behavior recognition as described in claim 1, characterized in that, In step S3, constructing a behavior-physiology-environment triplet knowledge graph based on the primary fusion feature vector includes: S301. Based on the primary fusion feature vector, behavioral entities, physiological entities, and environmental entities are extracted through the entity recognition module; wherein, behavioral entities include flock movement patterns, feeding behavior, and abnormal postures; physiological entities include heart rate variability, respiratory rate, and body temperature distribution; and environmental entities include temperature fluctuations, humidity changes, and ammonia concentration. S302. Define the relationship types between the behavioral entities, physiological entities and environmental entities using a relation extraction algorithm. The relationship types include behavior influencing physiology, environment influencing behavior, and physiology influencing metabolism. S303. Integrate the behavioral entities, physiological entities, environmental entities, and relationship types to construct the behavioral-physiological-environment triplet knowledge graph.

4. The multimodal information fusion feeding decision-making method for chicken behavior recognition as described in claim 3, characterized in that, The expression for the relation extraction algorithm is: in, express, Indicates normalization, Represents the relation classification weight matrix. Represents the characteristics of a behavioral entity. Represents physiological physical characteristics, Represents the characteristics of environmental entities. Indicates the characteristics of metabolic entities. Indicates bias. This represents the standard triplet structure in a knowledge graph. , and All represent relation thresholds.

5. The multimodal information fusion feeding decision-making method for chicken behavior recognition as described in claim 1, characterized in that, The implicit association rules between feeding frequency and body temperature discovered in step S4 include: S401. Embedding and initializing the behavioral entities, physiological entities, and environmental entities in the triplet knowledge graph to generate entity vectors, and encoding the relationship types of the behavioral entities, physiological entities, and environmental entities as a relationship matrix; S402. Based on the entity vector and relation matrix, cross-entity relation reasoning is performed using a graph attention network. Neighbor node information is aggregated through a multi-head attention mechanism to obtain the aggregated node feature vector and locate the association path between feeding frequency and body temperature. S403. Based on the node feature vector and associated path, use GNN two-hop inference to generate the path weight from feeding frequency to body temperature, and combine cosine similarity to calculate rule confidence. S404. Based on the path weight and rule confidence, the information flow is controlled through a gating mechanism to extract high-confidence rules and map them to a predefined rule template library to obtain the mapping result; S405. Generate a set of implicit association rules between feeding frequency and body temperature based on the high confidence rules and mapping results.

6. The multimodal information fusion feeding decision-making method for chicken behavior recognition as described in claim 5, characterized in that, The set of implicit association rules between feeding frequency and body temperature generated in step S405 includes: S4051. Set a confidence threshold, filter the high confidence rules, and retain high confidence rules with a confidence level higher than the preset threshold; S4052. Based on the high confidence rule, traverse the predefined rule template library, match the mapping result with the templates in the predefined rule template library, and select the most suitable rule template. S4053, Instantiate the most suitable rule template to generate specific implicit association rules between feeding frequency and body temperature; S4054. Integrate all generated implicit association rules to form an implicit association rule set.

7. The multimodal information fusion feeding decision-making method for chicken behavior recognition as described in claim 1, characterized in that, The intermediate decision features generated in step S5 include: S501. Perform dimensional matching and semantic alignment between the primary fusion feature vector and the implicit association rule to obtain the aligned feature vector; S502. By using a cross-modal attention mechanism, the feature patterns of implicit association rules are incorporated into the aligned feature vector to obtain fused features. S503. Decouple the fused features into three independent components: behavioral abnormality probability, physiological stress index, and environmental risk level. S504. The activation function and normalization operation apply numerical constraints to the abnormal behavior probability, physiological stress index and environmental risk level, generating intermediate decision features that include the abnormal behavior probability, physiological stress index and environmental risk level.

8. The multimodal information fusion feeding decision-making method for chicken behavior recognition as described in claim 1, characterized in that, The feeding decision instruction generated in step S6 includes: S601. Based on the intermediate decision features, multi-dimensional parameter prediction is performed through a pre-trained decision model to generate preliminary feeding decision output, which includes feeding time adjustment value, feeding amount adjustment value and feed type adjustment parameter. S602. Combine the preliminary decision output with the expert rule base to perform rule verification and conflict resolution, and then integrate the decision model output and expert rules through a weighted fusion algorithm to generate an optimized feeding instruction. S603. Normalize and range-constrain the optimized feeding instructions to ensure that the feeding time is within the preset time window, the feeding amount meets the needs of the flock size, and the feed type matches the nutritional standards. S604. Output the final feeding decision instruction, which includes the feeding time execution point, the precise value of the feeding amount, and the feed type adjustment code.

9. The multimodal information fusion feeding decision-making method for chicken behavior recognition as described in claim 8, characterized in that, The expression for the decision model is: in, This represents the initial set of decision parameters. This represents the linear transformation matrix of the output layer. This represents the activation function. Represents the linear transformation matrix of the hidden layer. Indicates the characteristics of intermediate decisions. This represents the hidden layer offset. This represents the output layer offset. Indicates time adjustment. Indicates time adjustment. This refers to the Softplus function. Indicates time adjustment. This indicates normalization.

10. A multimodal information fusion feeding decision-making system for recognizing the behavior of farmed chickens, characterized in that, The system includes: processor; Memory used to store processor-executable instructions; The processor is configured to implement the multimodal information fusion feeding decision method for recognizing the behavior of farmed chickens as described in any one of claims 1 to 9 when executing the executable instructions.

Images