A whole cycle growth track monitoring management system for meat ducks

Abstract

The present application relates to livestock and poultry breeding intelligent monitoring technical field, specifically to a kind of meat duck whole cycle growth trajectory monitoring management system, comprising: through multi-source perception module, original perception stream is formed by using multi-source sensing network to duck group full-time state perception;Characteristic decoupling module extracts individual meat duck independent perception segment, and static shape feature and dynamic behavior feature are separated by depth feature decoupling.The state modeling module fusion generates continuum state evolution sequence, quantifies behavior mode spectrum, and constructs individual growth state diagram by coupling the two space-time alignment.Growth trajectory generation module iteratively optimizes growth state diagram according to growth consistency intrinsic coordination index, and outputs meat duck individual whole cycle growth trajectory after time dimension integration and smoothing processing.The system can realize individual perception data accurate split, shape and behavior feature independent analysis, construct multi-dimensional coupled growth model, generate continuous and complete, actual meat duck whole cycle growth trajectory.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring technology for livestock and poultry farming, and in particular to a monitoring and management system for the entire growth trajectory of meat ducks. Background Technology

[0002] Current methods for monitoring the growth of meat ducks employ a group-based sensing acquisition model. This model uses multi-source sensing devices to acquire overall flock-wide sensory data, forming a raw sensory data stream. This data exists in a mixed, group-wide form, without individual duck-specific data segmentation, making it impossible to extract independent sensory fragments for each duck. Current data processing only performs basic feature extraction, failing to perform refined decoupling of relevant duck features. Static physical characteristics and dynamic behavioral characteristics are mixed, making it impossible to generate independent representations of these two types of features.

[0003] Current analyses of duck growth status only perform temporal or spatial processing on single features, failing to align and couple cross-timestamp fused physical feature data with quantified behavioral feature data in a spatiotemporal manner. This makes it difficult to construct an individual growth status model that integrates multi-dimensional information. Existing growth trajectory generation methods do not include growth consistency indicators for feedback optimization of growth status data, nor do they perform integration and smoothing of growth status data over time, resulting in discontinuous and distorted trajectory data.

[0004] This invention requires the extraction of individual independent perceptual fragments and decoupling of deep features from the original perceptual flow of the group, the spatiotemporal coupling modeling of the body evolution sequence and the behavioral pattern spectrum, the optimization of the growth state diagram based on the internal coordination index and the completion of the integration processing of the time dimension, thereby solving the problems of missing individual data, inability to separate features, single modeling dimension and incomplete growth trajectory in the existing technology. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing a full-cycle growth trajectory monitoring and management system for meat ducks.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a monitoring and management system for the entire growth trajectory of meat ducks, comprising: The multi-source sensing module performs real-time sensing of the duck flock based on a multi-source sensor network, forming an original sensing stream. The feature decoupling module parses the original sensory stream, extracts independent sensory segments related to individual ducks, performs deep feature decoupling on each independent sensory segment, and separates static body features from dynamic behavior features. The state modeling module performs cross-timestamp association and fusion on the separated static body features to generate a continuous body evolution sequence, performs pattern recognition and quantification on the separated dynamic behavior features to generate a quantized behavior pattern spectrum, and spatiotemporally aligns and couples the continuous body evolution sequence with the quantized behavior pattern spectrum to construct an individual growth state diagram. The trajectory generation module calculates an internal coordination index reflecting growth consistency based on the individual growth state map, performs feedback iterative optimization on the individual growth state map based on the internal coordination index, forms a final individual growth state map, integrates and smooths the final individual growth state map in the time dimension, and outputs the full-cycle growth trajectory of the meat duck individual. The resource allocation module is used to predict the slaughter time of individual meat ducks based on the full-cycle growth trajectory, and dynamically adjust the allocation strategy of breeding resources and the slaughter logistics plan according to the prediction results.

[0007] As a further aspect of the present invention, the step of performing real-time perception of the duck flock based on a multi-source sensor network to form an original perception stream includes: Deploy a heterogeneous sensor network that includes a non-contact weight sensor array, a distributed acoustic sensor network, and a millimeter-wave radar scanning device; The non-contact weight sensor array is used to acquire time-series data of the pressure distribution generated by the duck flock in the ground activity area; The distributed acoustic sensor network synchronously collects sound spectrum information within the aquaculture environment to form acoustic fingerprint time-series data; The millimeter-wave radar scanning device periodically transmits and receives modulated signals to extract point cloud time-series data containing the duck flock's location, speed, and direction of movement. The pressure distribution time-series data, the voiceprint time-series data, and the point cloud time-series data are spliced ​​and fused together using a unified time reference to generate the original perception stream.

[0008] As a further aspect of the present invention, the original sensory stream is analyzed to extract independent sensory fragments related to individual meat ducks, including: The original sensory stream is subjected to joint time-frequency domain analysis to identify signal segments that match preset individual activity primitives; For each marked signal segment, the pressure centroid position and pressure distribution profile are calculated from the pressure distribution time-series data channel of the original sensing stream; By combining the pressure center of gravity position with the pressure distribution contour, motion trajectory segments belonging to the same individual are separated from the point cloud time series data channel of the original sensing stream. From the acoustic feature data channel of the original sensing stream, extract the acoustic feature segment that is time-synchronized with the motion trajectory segment; The pressure distribution profile, the motion trajectory segment, and the acoustic feature segment are packaged together to form an independent sensing segment.

[0009] As a further aspect of the present invention, deep feature decoupling is performed on each of the independent perceptual segments to separate static shape features and dynamic behavior features, including: The pressure distribution contour in the independent sensing segment is fitted with a contour curve, and the contour curvature change sequence is extracted as a preliminary shape feature. Kinematic differential analysis is performed on the motion trajectory segments in the independent sensing segments to calculate the temporal values ​​of velocity, acceleration, and changes in motion direction, forming a motion feature sequence; Mel frequency cepstral coefficients are extracted from the acoustic feature segments in the independent sensing segments to obtain acoustic feature vectors; The contour curvature change sequence, the motion feature sequence, and the acoustic feature vector are input into a pre-trained feature decoupling network; The feature decoupling network outputs two mutually orthogonal latent variables. One latent variable is decoded as the static body features representing the individual's body shape and posture at the current moment, and the other latent variable is decoded as the dynamic behavioral features representing the individual's actions and gait at the current moment.

[0010] As a further aspect of the present invention, the step of performing cross-timestamp correlation and fusion on the separated static shape features to generate a continuous body evolution sequence includes: Map the static shape features from different timestamps to a high-dimensional feature space, and calculate the feature distance between any two static shape features; Using time sequence as nodes and the feature distance as edge weights, a feature spatiotemporal correlation graph is constructed; On the aforementioned spatiotemporal correlation graph, a path search algorithm based on minimum spanning tree is used to find an optimal correlation path that runs through all time nodes; According to the node order of the optimal association path, nonlinear smooth interpolation is performed on all the static body features on the path to generate the continuous body evolution sequence that is continuous and smooth in time.

[0011] As a further aspect of the present invention, the step of performing pattern recognition and quantization on the separated dynamic behavioral features to generate a quantized behavioral pattern spectrum includes: Cluster analysis is performed on the dynamic behavioral features within multiple time windows to identify basic behavioral patterns; Define a unique pattern code for each of the identified basic behavioral patterns; Statistics were compiled on the frequency, average duration, and intensity distribution of each basic behavioral pattern within a preset evaluation period. The pattern encoding, frequency of occurrence, average duration, and intensity distribution are integrated to construct a multi-dimensional behavior quantification matrix; Singular value decomposition is performed on the behavior quantization matrix to extract principal components, forming a low-dimensional spectrum of the quantized behavior patterns.

[0012] As a further aspect of the present invention, the continuous body evolution sequence is spatiotemporally aligned and coupled with the quantified behavioral pattern spectrum to construct an individual growth state diagram, including: Establish a unified spatiotemporal coordinate system with time as the horizontal axis, the feature dimension in the continuous body evolution sequence as the first vertical axis, and the pattern intensity in the quantified behavior pattern spectrum as the second vertical axis; In the unified spatiotemporal coordinate system, the continuous volume evolution sequence is represented as a multidimensional feature evolution trajectory line; In the unified spatiotemporal coordinate system, the quantized behavior pattern spectrum is represented as a set of time-varying multidimensional pattern intensity surfaces; Calculate the correlation strength between the feature evolution trajectory and the multidimensional mode intensity surface at each time point to form a correlation strength time series; Using the points on the feature evolution trajectory as master nodes, the corresponding points on the multidimensional pattern intensity surface as attribute nodes, and the association strength as edge weights, a dynamic heterogeneous graph network is constructed, which is the individual growth state graph.

[0013] As a further aspect of the present invention, based on the individual growth state diagram, an intrinsic coordination index reflecting growth consistency is calculated, including: In the individual growth state diagram, the mutual information value between the set of master nodes and the set of attribute nodes is calculated; In the individual growth state diagram, traverse all connection paths between master nodes and attribute nodes, and calculate the average path length and clustering coefficient of all paths; Extract the topological structure of the individual growth state map on multiple consecutive time slices, calculate the similarity of the topological structure between adjacent time slices, and form a topological evolution consistency sequence. The mutual information value, the average path length, the clustering coefficient, and the topological evolution consistency sequence are weighted and fused to calculate the intrinsic coordination index, which characterizes the degree of synergy between physical changes and behavioral patterns during individual growth and development.

[0014] As a further aspect of the present invention, the individual growth state diagram is subjected to feedback iterative optimization based on the inherent coordination index to form a final individual growth state diagram, including: Set an optimization target threshold for the intrinsic coordination index; Compare the intrinsic coordination index corresponding to the current individual growth state diagram with the optimization target threshold; If the intrinsic coordination index does not reach the optimization target threshold, then the edges with association strength lower than the preset threshold in the individual growth state graph are pruned, and the main node and attribute node connected by the edge with the highest association strength are enhanced. Based on the graph structure after pruning and feature enhancement, the intrinsic coordination index is recalculated. Repeat the steps of comparing indicators, pruning and enhancing the graph structure, and recalculating the indicators until the intrinsic coordination indicator reaches or exceeds the optimization target threshold. The resulting graph structure is the final individual growth state graph.

[0015] As a further aspect of the present invention, the final individual growth state map is integrated and smoothed over time to output the full-cycle growth trajectory of the meat duck individual, including: In the final individual growth state diagram, piecewise linear integration is performed on the change curves of all feature dimensions of the master node set along the time axis to calculate the cumulative change of each feature dimension within the observation period. The pattern intensity change curve of each attribute node is processed by time window moving average to generate a smoothed behavior pattern evolution curve. The cumulative change and the smoothed behavior pattern evolution curve are aligned in time and combined into a multidimensional, time-continuous vector sequence. Principal component analysis is performed on the vector sequence to extract its first few principal components, forming a low-dimensional trajectory curve that represents the overall growth and development process of a meat duck individual from the initial stage to the end of the observation. The trajectory curve is the full-cycle growth trajectory.

[0016] As a further aspect of the present invention, the resource allocation module is used to perform the following operations: Extract the cumulative changes in each shape feature dimension in the full-cycle growth trajectory, and the intensity trends of basic feeding and drinking behavior patterns in the smoothed behavior pattern evolution curve; The cumulative change is matched with a preset standard slaughter size threshold, and combined with the intensity trend of the behavior pattern, the expected slaughter time window of a single meat duck is predicted. Based on the expected slaughter time window, the feed ratio, pen space allocation and environmental control parameters of the corresponding duck flock are dynamically adjusted to generate a breeding resource allocation strategy. The estimated slaughter time window for all individuals is aggregated, and a slaughter logistics scheduling plan is generated by combining cold chain logistics capacity and slaughter production line schedules.

[0017] Compared with the prior art, the advantages and positive effects of the present invention are as follows: The raw sensor stream collected by the multi-source sensor network is parsed and processed to extract independent sensor segments corresponding to individual ducks. Deep feature decoupling is performed on each independent sensor segment to directly separate static physical features and dynamic behavioral features. The sensor data can accurately match the individual information of a single duck, eliminating information interference between different individuals in the group sensor data. Static physical features and dynamic behavioral features are presented in independent data form, and the information boundaries of the two types of features are clearly defined. The physical parameters and behavioral information of individual ducks can be directly distinguished and obtained, and the exclusivity and purity of the feature data remain stable.

[0018] Static physical features are fused across timestamps to form a continuous physical evolution sequence. Dynamic behavioral features are processed by pattern recognition and quantification to generate a quantified behavioral pattern spectrum. After spatiotemporal alignment and coupling of the continuous physical evolution sequence and the quantified behavioral pattern spectrum, an individual growth state map is constructed. Based on the internal coordination index reflecting growth consistency, the individual growth state map is subjected to feedback iterative optimization. The optimized final individual growth state map is integrated and smoothed in the time dimension, outputting the full-cycle growth trajectory of the duck. The continuous physical evolution sequence can completely present the change process of the duck's body shape over time, and the quantified behavioral pattern spectrum can intuitively reflect the quantitative characteristics of the duck's behavior. Spatiotemporal alignment and coupling allow physical and behavioral information to form a correlation and match in the corresponding spatiotemporal dimensions. Feedback iterative optimization corrects the data deviation in the growth state map, and the integration and smoothing in the time dimension keeps the growth trajectory continuous and complete, outputting full-cycle trajectory data that fits the actual growth state of the duck. Attached Figure Description

[0019] Figure 1 This is a time-series diagram of a meat duck full-cycle growth trajectory monitoring and management system according to the present invention; Figure 2 A flowchart for the formation of the original sensory flow; Figure 3 This is a flowchart for separating static shape features from dynamic behavior features. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0021] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0022] See Figure 1 This invention provides a monitoring and management system for the entire growth trajectory of meat ducks, specifically comprising: The multi-source sensing module continuously senses the duck flock based on a deployed heterogeneous sensor network, simultaneously collecting pressure distribution, acoustic information, and millimeter-wave radar point cloud data, and fusing them to form a raw sensing stream. The feature decoupling module analyzes this raw sensing stream, extracting independent sensing segments belonging to specific duck individuals from the stream data through time-frequency analysis and spatial correlation. Then, a pre-trained feature decoupling network is used to deeply process these segments, separating static physical features representing individual body shape and dynamic behavioral features representing individual movements. The state modeling module models the separated features in the time dimension. On one hand, it performs cross-timestamp correlation and smoothing of static physical features to generate a continuous body evolution sequence; on the other hand, it performs pattern recognition and statistical analysis of dynamic behavioral features to generate a quantified behavioral pattern spectrum. Finally, these two sequences are aligned and coupled in time and space to construct a dynamic individual growth state map. Based on the constructed individual growth state map, the trajectory generation module calculates the intrinsic coordination index reflecting the synergy between body shape and behavior during the growth process. Using this index as feedback, the growth state map is iteratively optimized to obtain the final individual growth state map. Finally, the map is integrated and smoothed in the time dimension to output the full-cycle growth trajectory representing the complete growth and development process of the meat duck.

[0023] In one embodiment of the invention, within a breeding shed, see [reference] Figure 2A heterogeneous sensor network was deployed, comprising a non-contact weight sensor array, a distributed acoustic sensor network, and a millimeter-wave radar scanning device. The non-contact weight sensor array was deployed in the main activity area of ​​the duck flock to acquire time-series data of pressure distribution generated when the ducks walked and stood. The distributed acoustic sensor network was deployed at multiple points in the environment to synchronously collect sound spectrum information, forming acoustic signature time-series data. The millimeter-wave radar scanning device periodically transmitted and received modulated signals, and obtained point cloud time-series data containing the position, speed, and direction of movement of individual ducks through signal analysis. The pressure distribution time-series data, acoustic signature time-series data, and point cloud time-series data were spliced ​​and fused using a unified time reference to generate a raw sensing stream. To extract data related to specific individuals from the raw sensing stream, a joint time-frequency domain analysis was performed on the raw sensing stream to mark signal segments that matched preset individual activity primitives. For each marked signal segment, the pressure center of gravity position and pressure distribution contour at that moment were calculated from the pressure distribution time-series data channel. By combining the pressure center of gravity location and pressure distribution contour, motion trajectory segments belonging to the same individual with corresponding spatial locations are separated from the point cloud temporal data channel. Simultaneously, acoustic feature segments that are completely synchronized in time with the motion trajectory segments are extracted from the acoustic signature temporal data channel. Finally, the calculated pressure distribution contour, the separated motion trajectory segments, and the extracted acoustic feature segments are packaged together to form an independent sensing segment.

[0024] In a specific implementation, an embodiment of a full-cycle growth trajectory monitoring and management system for meat ducks involves the formation of the original sensing flow and the extraction of independent sensing segments. In an exemplary meat duck breeding house scenario, the system's multi-source sensing module deploys a heterogeneous sensing network consisting of a non-contact weight sensor array, a distributed acoustic sensor network, and a millimeter-wave radar scanning device. The non-contact weight sensor array is laid out in a grid pattern on the ground area where the ducks are active, and its output pressure distribution time-series data records the pressure values ​​at different locations over time. The distributed acoustic sensor network consists of multiple sound pickup nodes, which are evenly distributed on the top of the breeding environment, synchronously collecting mixed sound signals in the environment and processing them into acoustic fingerprint time-series data. The millimeter-wave radar scanning device is installed at a specific height in the breeding house, periodically emitting frequency-modulated continuous waves and receiving reflected signals. Through signal processing, point cloud time-series data reflecting the position, speed, and direction of movement of the ducks are extracted. In practice, the aforementioned pressure distribution time-series data, acoustic signature time-series data, and point cloud time-series data are transmitted to the central processing unit. The central processing unit assigns a unified timestamp to all data streams based on a high-precision clock source, and splices and fuses the multi-channel data according to the time reference to generate a raw perception stream containing time-synchronized pressure, acoustic, and spatial motion information.

[0025] In practical implementation, the feature decoupling module analyzes the original sensing stream to extract independent sensing segments. Through joint time-frequency domain analysis of the original sensing stream, the system identifies signal segments that match preset individual activity primitives, including continuous walking gait signals and brief pecking sound pulse signals. For each identified signal segment, the processing begins with the pressure distribution time-series data channel of the original sensing stream, calculating the centroid coordinates of the pressure distribution within the corresponding time window and extracting the pressure distribution profile at that moment. The formula for calculating the pressure centroid position can be understood as:

[0026] in: The centroid coordinates represent the pressure distribution profile. Indicates the first The pressure value of each sensing unit Indicates the first The planar coordinates of each sensing unit This represents the total number of sensing units. In some embodiments, combining the calculated pressure center of gravity position and the shape information of the pressure distribution contour, the system filters out point cloud sequences from the point cloud time-series data channel of the original sensing stream that spatially coincide with the pressure center of gravity position and match the pressure distribution contour in shape, thereby separating motion trajectory segments belonging to the same individual. From the acoustic signature time-series data channel of the original sensing stream, the system precisely extracts audio data segments that are completely synchronized with the separated motion trajectory segments in start and end time, as acoustic feature segments. Finally, the calculated pressure distribution contour, the separated motion trajectory segments, and the extracted acoustic feature segments are packaged into a packaged independent sensing segment data packet. It can be understood that each independent sensing segment data packet contains a timestamp, pressure contour data, motion trajectory coordinate sequence, and acoustic spectrum data.

[0027] In some embodiments, data comparison is reflected in the correlation between different data channels. For example, changes in local pressure peaks appearing in the pressure distribution time-series data channel correspond to the movement of point cloud clusters in the point cloud time-series data channel, while in the acoustic signature time-series data channel, they may correspond to the calls or activity sounds of ducks. Time-frequency domain joint analysis identifies this cross-channel synchronization pattern to mark individual activities. In specific implementations, for overlapping or intersecting individual signals, the system distinguishes them by analyzing the separation degree of the pressure distribution contour and the spatial independence of the point cloud motion trajectory. When two pressure center positions are too close and their contours overlap, the system combines historical motion trajectory segments for tracking correlation to determine which continuously tracked individual the current segment should belong to. Optionally, for individuals that are briefly lost and then reappear, the system compares the similarity of the reappearing individual's pressure distribution contour, typical motion pattern, and acoustic feature segments with the independent sensing segments in the historical record to decide whether to continue tracking the original individual or initialize it as a new tracking target. Ultimately, the system outputs a series of independent sensory fragments that may be discontinuous in time, but each corresponds to the activity of a specific individual duck within a specific time period, for subsequent feature decoupling module processing.

[0028] In one embodiment of the present invention, see [reference] Figure 3 The process involves: fitting contour curves to the pressure distribution contours within the independent perception segments, extracting the curvature of each point on the contour, and forming a contour curvature change sequence as preliminary shape features; performing kinematic differential analysis on the motion trajectory segments within the independent perception segments, calculating the temporal values ​​of their velocity, acceleration, and direction of motion changes, and forming a motion feature sequence; and extracting Mel-frequency cepstral coefficients from the acoustic feature segments within the independent perception segments to obtain a set of acoustic feature vectors representing sound characteristics. Subsequently, the extracted contour curvature change sequence, motion feature sequence, and acoustic feature vectors are input into a pre-trained feature decoupling network. This feature decoupling network, after training, can decompose the mixed representation of the input features into two independent latent spaces. The network outputs two mutually orthogonal latent variables. One latent variable is reconstructed by the decoder, outputting static shape features representing the individual's body shape and posture at the current moment; the other latent variable is reconstructed by the decoder, outputting dynamic behavioral features representing the individual's actions and gait at the current moment.

[0029] In practical implementation, the system receives independent sensing segment data packets from the feature decoupling module. These packets contain pressure distribution contours, motion trajectory segments, and acoustic feature segments. The pressure distribution contour in the independent sensing segments is processed by fitting elliptical or spline curves to the boundary points using the least squares method, resulting in a smooth closed contour curve. Subsequently, the curvature values ​​of a series of sampling points along the closed contour curve are calculated, forming a sequence of contour curvature changes as preliminary shape features. The motion trajectory segments in the independent sensing segments are processed. These segments consist of a series of three-dimensional spatial coordinate points arranged in chronological order. Instantaneous velocities are calculated by performing first-order differences on these coordinate points, instantaneous accelerations are calculated by performing first-order differences on the velocity sequence, and the change in the direction angle of the line connecting consecutive coordinate points is calculated. This yields the temporal values ​​of velocity, acceleration, and motion direction changes, which together form the motion feature sequence. The acoustic feature segments in the independent sensing segments are processed. The acoustic feature segments are a time-domain audio signal. The system pre-emphasizes, frames, and windows the audio signal. Then, a fast Fourier transform is performed on each frame of the signal to obtain the spectrum. The spectrum is passed through a set of Mel-scale triangular filter banks, and the logarithmic energy of the output of each filter is calculated. Finally, a discrete cosine transform is performed on the logarithmic energy, and the first few coefficients are extracted to obtain the Mel-frequency cepstral coefficient acoustic feature vector that characterizes the sound characteristics.

[0030] In practice, the initial body features, motion feature sequences, and Mel-frequency cepstral coefficient acoustic feature vectors are integrated into a single multimodal feature vector, which is then input into a pre-trained feature decoupling network. This feature decoupling network is a deep neural network structure containing a shared encoder, two independent latent variable spaces, and corresponding decoders. The shared encoder encodes the input multimodal feature vector into an intermediate representation, which is decomposed into two mutually orthogonal latent variables. One latent variable is labeled as a body latent variable, and the other as a behavior latent variable. In practice, the body latent variable is input into the body feature decoder, which outputs a decoded feature vector representing the static body features of the duck at the current moment, such as body size, trunk posture, and overall outline. The behavior latent variable is input into the behavior feature decoder, which outputs a decoded feature vector representing the dynamic behavior features of the duck at the current moment, such as movement type, gait cycle, and activity intensity. In some embodiments, the training objective of the pre-trained feature decoupling network is to minimize the mutual information between static shape features and dynamic behavior features, while ensuring high-fidelity reconstruction of the corresponding original feature subsets from their respective latent variables. The training loss function includes reconstruction loss, orthogonality constraint loss, and mutual information estimation loss. After the feature decoupling network is trained, during the inference phase, inputting multimodal features extracted from an independent perceptual segment will yield separated static shape feature vectors and dynamic behavior feature vectors at the output.

[0031] In some embodiments, the data comparison reflects the contribution of features from different sources to the final decoupling result. For example, low-frequency components in the contour curvature change sequence are more associated with static shape features, while high-frequency change components in the motion feature sequence and specific frequency band energy in the acoustic feature vector are more associated with dynamic behavior features. Optionally, the weight distribution learned by the feature decoupling network during training shows that the neural network path processing the contour curvature change sequence is more strongly connected to the decoder outputting static shape features, while the path processing the motion feature sequence and acoustic feature vector is more strongly connected to the decoder outputting dynamic behavior features. It can be understood that static shape features have high stability in a short time, and their feature vectors have a small Euclidean distance between consecutive time stamps, while dynamic behavior features exhibit high time-varying characteristics, and their feature vectors change drastically. By introducing an evaluation metric to measure the feature decoupling separation degree, the formula for calculating this metric is:

[0032] in: This represents the degree of decoupling. This indicates the number of independent sensory segments in a batch. Indicates the first The static shape feature vectors decoupled from each segment This represents the mean vector of the static shape feature vectors of this batch. Indicates the first Dynamic behavior feature vectors decoupled from each segment This represents the mean vector of the dynamic behavior feature vectors of this batch. It is a very small positive number used to prevent division by zero errors. Higher The value indicates that the consistency of static shape features within a batch is higher than that of dynamic behavior features, thus indirectly reflecting the effectiveness of decoupling. Optionally, in actual operation, the system will periodically calculate the decoupling separation degree index of the current batch of data to monitor the running status of the feature decoupling network. Finally, the system outputs the decoupled static shape features and dynamic behavior features corresponding to the timestamp of each independent sensing segment for use by the state modeling module.

[0033] In one embodiment of the present invention, the separated static body features are correlated and fused across time stamps, mapping static body features from different time stamps to a high-dimensional feature space, and calculating the feature distance between any two features. A feature spatiotemporal correlation graph is constructed using time sequence as nodes and feature distance as edge weights. On this feature spatiotemporal correlation graph, a path search algorithm based on minimum spanning tree is used to find an optimal correlation path that runs through all time nodes. According to the node order of the optimal correlation path, nonlinear smooth interpolation is performed on all static body features on the path to generate a temporally continuous and smooth continuous body evolution sequence. Pattern recognition and quantification are performed on the separated dynamic behavior features. Cluster analysis is performed on dynamic behavior features within multiple time windows to identify several basic behavior patterns. A unique pattern code is defined for each identified basic behavior pattern. The frequency, average duration, and intensity distribution of each basic behavior pattern are statistically analyzed within a preset evaluation period. The pattern codes, frequency, average duration, and intensity distribution are integrated to construct a multi-dimensional behavior quantification matrix. Singular value decomposition is performed on the behavior quantization matrix to extract its principal components, forming a low-dimensional quantization behavior pattern spectrum.

[0034] In practical implementation, the system receives a set of static body features and a set of dynamic behavior features, arranged by timestamps, from the feature decoupling module. For the static body feature set, the system performs cross-timestamp correlation fusion to generate a continuous body evolution sequence. The processing maps each static body feature from different timestamps to a preset high-dimensional feature space. In this high-dimensional feature space, the feature distance between any two static body features is calculated. The formula for calculating the feature distance is:

[0035] in: Represents the static shape feature vector and The characteristic distance between them It is a positive semi-definite weight matrix used to adjust the contributions of different feature dimensions. A small positive adjustment parameter is used to ensure computational stability. In practice, each timestamp is used as a node, and the calculated feature distance is used as the edge weight connecting the corresponding two nodes to construct a fully connected weighted undirected graph, namely the feature spatiotemporal correlation graph. On the feature spatiotemporal correlation graph, a path search algorithm based on minimum spanning tree is used to find the optimal correlation path that runs through all time nodes and has the minimum sum of edge weights, with time order as a constraint. It can be understood that this optimal correlation path reflects the smoothest and most coherent sequence of shape changes in the feature space. According to the node order of the optimal correlation path, nonlinear smooth interpolation is performed on all static shape feature vectors on the path along the time axis. The interpolation method adopts the radial basis function-based interpolation method, which finally generates a continuous and smooth continuous body evolution sequence in time. The continuous body evolution sequence can be represented as a vector function that changes with time. .

[0036] In some embodiments, for a set of dynamic behavioral features, the system performs pattern recognition and quantification to generate a quantified behavioral pattern spectrum. The process first performs cluster analysis on the dynamic behavioral feature vectors within multiple consecutive time windows, using a density clustering algorithm to identify representative basic behavioral patterns. A unique pattern code is defined for each identified basic behavioral pattern, using numeric identifiers such as "B001" and "B002". The frequency, average duration, and intensity distribution of each basic behavioral pattern are statistically analyzed within a preset evaluation period, such as a complete feeding cycle. The intensity distribution is described by calculating the mean and variance of the magnitudes of all dynamic behavioral feature vectors corresponding to that behavioral pattern. It can be understood that the identified basic behavioral patterns may include standing still, walking, running, feeding, drinking, and preening. Refer to Table 1, which shows a quantitative statistical analysis of a basic behavioral pattern. Table 1: Quantitative Statistics of Basic Behavioral Patterns

[0037] In practice, multiple statistics, including pattern encoding, frequency of occurrence, average duration, and intensity distribution, are integrated to construct a multidimensional behavior quantification matrix consisting of behavioral patterns (rows) multiplied by statistical dimensions (columns). Optionally, the intensity distribution can be described using multiple statistics, such as the mean in the table above, or extended to a high-dimensional vector including variance. Singular value decomposition is performed on this behavior quantification matrix to extract its k largest singular values ​​and their corresponding left and right singular vectors, forming a low-dimensional quantified behavior pattern spectrum. The quantified behavior pattern spectrum can be represented as a set of low-dimensional vectors that centrally characterize the main components of an individual's behavior patterns and their intensity relationships within the evaluation period. In some embodiments, data comparison is reflected in the results of cluster analysis. For example, dynamic behavior features from different time windows but with highly similar feature vectors are clustered into the same basic behavior pattern and assigned the same pattern encoding, while dynamic behavior features with significant feature differences are classified into different patterns. Optionally, principal component analysis following singular value decomposition can help compare the intrinsic correlations between different behavioral patterns. For example, the first principal component might represent the "activity" dimension, explaining most of the variance for both "running" and "walking" patterns, while the second principal component might represent the "feeding-related" dimension. Ultimately, the system outputs a continuous sequence of body posture evolution. And quantify the behavioral pattern spectrum for coupling with subsequent modules.

[0038] In one embodiment of the present invention, a continuous body evolution sequence and a quantized behavioral pattern spectrum are spatiotemporally aligned and coupled to establish a unified spatiotemporal coordinate system with time as the horizontal axis, the feature dimension in the continuous body evolution sequence as the first vertical axis, and the pattern intensity in the quantized behavioral pattern spectrum as the second vertical axis. In this unified spatiotemporal coordinate system, the continuous body evolution sequence is represented as a multidimensional feature evolution trajectory line, and the quantized behavioral pattern spectrum is represented as a set of time-varying multidimensional pattern intensity surfaces. The correlation strength between the feature evolution trajectory line and the multidimensional pattern intensity surfaces at each time point is calculated to form a correlation strength time series. Using points on the feature evolution trajectory line as master nodes, corresponding points on the multidimensional pattern intensity surfaces as attribute nodes, and correlation strength as edge weights, a dynamic heterogeneous graph network is constructed; this dynamic heterogeneous graph network is the individual growth state graph. Based on the constructed individual growth state graph, an intrinsic coordination index reflecting growth consistency is calculated. In the individual growth state graph, the mutual information value between the master node set and the attribute node set is calculated. Traverse all connection paths between master nodes and attribute nodes in the graph, and calculate the average path length and clustering coefficient of all paths. Extract the topological structure of the individual growth state graph across multiple consecutive time slices, calculate the similarity of the topological structure between adjacent time slices, and form a topological evolution consistency sequence. Perform weighted fusion of the calculated mutual information values, average path lengths, clustering coefficients, and topological evolution consistency sequences to calculate an intrinsic coordination index characterizing the degree of synergy between physical changes and behavioral patterns during individual growth and development.

[0039] In its implementation, the system receives a continuous body evolution sequence and a quantized behavioral pattern spectrum from the state modeling module. The continuous body evolution sequence is a static body feature vector function that varies over time, while the quantized behavioral pattern spectrum is a low-dimensional representation set containing multiple behavioral patterns and their intensity information. To fuse the two, the system first establishes a unified spatiotemporal coordinate system. The horizontal axis of the unified spatiotemporal coordinate system is the time axis, the first vertical axis corresponds to each feature dimension in the continuous body evolution sequence, such as body length index, chest circumference index, and body curvature, and the second vertical axis corresponds to the intensity value of each basic behavioral pattern in the quantized behavioral pattern spectrum. In the unified spatiotemporal coordinate system, the continuous body evolution sequence is represented as a multidimensional feature evolution trajectory line, where each point is determined by the time coordinate and the static body feature vector at that moment. In the unified spatiotemporal coordinate system, the quantized behavioral pattern spectrum is represented as a set of multidimensional pattern intensity surfaces that vary over time. Each pattern intensity surface represents the change in intensity of a basic behavioral pattern over time, and its height is determined by the intensity value of that pattern.

[0040] In practical implementation, the correlation strength between the feature evolution trajectory line and the multidimensional pattern intensity surface is calculated at each time point. For each discrete time point, the system extracts the static shape feature vector on the feature evolution trajectory line at that moment, and simultaneously extracts the intensity values ​​of each basic behavioral pattern at that moment to form a behavioral pattern intensity vector. The formula for calculating the correlation strength is:

[0041] in: Indicates at a point in time The strength of the association, Indicates a point in time The static shape feature vector, Indicates a point in time Behavioral pattern intensity vector This is a minimal positive number to prevent the denominator from being zero. This calculation is performed at each time point, thus forming a time series of association strengths. The system uses each point on the feature evolution trajectory as a master node and the corresponding point on the multidimensional mode intensity surface at the same time point as an attribute node, using the calculated association strength... Using the edge weights connecting the corresponding master nodes and attribute nodes, a dynamic heterogeneous graph network is constructed, which is the individual growth state graph. It can be understood that the individual growth state graph is a graph that changes over time; the set of master nodes represents the physical state, the set of attribute nodes represents the behavioral state, and the edge weights represent the coupling relationship between the two at this point.

[0042] In some embodiments, based on the constructed individual growth state graph, the system calculates an intrinsic coordination index reflecting growth consistency. The calculation process first involves calculating the mutual information value between the set of master nodes and the set of attribute nodes in the individual growth state graph. This mutual information value measures the probability of predicting behavioral pattern changes from the evolution of physical features. In the individual growth state graph, all connection paths between master nodes and attribute nodes are traversed, and the average path length and clustering coefficient of all paths are calculated. The average path length reflects the average separation between nodes in the graph, and the clustering coefficient reflects the trend of nodes clustering together. The topological structure of the individual growth state graph is extracted across multiple consecutive time slices, and the similarity of the topological structure between adjacent time slices is calculated to form a topological evolution consistency sequence. See Table 2, which shows the results of partial metric calculations for an individual growth state graph. Table 2: Topology and Mutual Information Measurement Table of Individual Growth State Diagram

[0043] In practical implementation, the calculated mutual information value Average path length Clustering coefficient and topological evolution consistency sequence Weighted fusion and internal coordination indicators The calculation method is as follows:

[0044] in: , , , These are preset weighting coefficients. This represents the average value of the topological evolution consistency sequence. The final calculated intrinsic coordination index. This is used to characterize the degree of coordination between physical changes and behavioral patterns during an individual's growth and development. In some embodiments, data comparisons are reflected in the differences in indicators between different individuals or at different growth stages of the same individual; for example, the internal coordination indicators of a healthy developing individual. Individuals whose physical and behavioral development is not coordinated typically maintain relative stability or exhibit specific trends of change, while those with internal coordination indicators... It may exhibit unusually low or high values, or its average path length. With clustering coefficient The combination pattern will deviate from the norm. Optional, weighting coefficients. , , , It can be obtained by analyzing and learning from data of historically normal-growing individuals. It can be understood that the intrinsic coordination index is a comprehensive scalar that integrates the characteristics of dynamic heterogeneous graph networks in multiple aspects, such as information association, connection efficiency, clustering, and evolutionary smoothness.

[0045] In one embodiment of the present invention, the individual growth state graph is iteratively optimized based on the calculated intrinsic coordination index, and an optimization target threshold for the intrinsic coordination index is set. The intrinsic coordination index corresponding to the current individual growth state graph is compared with the optimization target threshold. If the intrinsic coordination index does not reach the optimization target threshold, edges in the individual growth state graph with association strength lower than the preset threshold are pruned, and the master node and attribute node connected by the edge with the highest association strength are enhanced. Based on the graph structure after pruning and feature enhancement, the intrinsic coordination index is recalculated. The steps of index comparison, graph structure pruning and enhancement, and index recalculation are repeated until the intrinsic coordination index reaches or exceeds the optimization target threshold. The resulting graph structure is the final individual growth state graph. The final individual growth state graph is integrated and smoothed over time. In the final individual growth state graph, piecewise linear integration is performed along the time axis on the feature dimension change curves of all feature dimensions of the master node set to calculate the cumulative change of each feature dimension within the observation period. The pattern intensity change curve of each attribute node is processed by time window moving average to generate a smoothed behavior pattern evolution curve. The cumulative changes and the smoothed behavioral pattern evolution curves are aligned over time and combined into a multidimensional, time-continuous vector sequence. Principal component analysis is performed on this vector sequence, and the first few principal components are extracted to form a low-dimensional trajectory curve representing the overall growth and development process of a meat duck individual from the initial stage to the end of the observation. This trajectory curve is the full-cycle growth trajectory.

[0046] In practical implementation, the system receives the individual growth state graph output by the state modeling module and the calculated intrinsic coordination index. Based on the intrinsic coordination index, it performs feedback iterative optimization of the individual growth state graph. The system presets an optimization target threshold for the intrinsic coordination index, which is the lower limit of the reference value range obtained from statistical data of the intrinsic coordination index of historically growing individuals. The system compares the intrinsic coordination index corresponding to the current individual growth state graph with the optimization target threshold. If the current intrinsic coordination index does not reach the optimization target threshold, the optimization process is initiated. Edges in the current individual growth state graph with association strength below the preset threshold are pruned and removed. The preset threshold is determined based on the quantile of the distribution of all edge weights. Simultaneously, feature enhancement is performed on the master node and attribute node connected by the edge with the highest association strength in the graph. Feature enhancement is achieved by increasing the norm of the node's feature vector or introducing attention weights. Based on the new graph structure after pruning and feature enhancement, the system recalculates the intrinsic coordination index and compares it again with the optimization target threshold corresponding to the new graph structure. This process of index comparison, graph structure pruning and enhancement, and index recalculation is repeated until the intrinsic coordination index reaches or exceeds the optimization target threshold. The resulting stable graph structure is defined as the final individual growth state graph.

[0047] In some embodiments, the system integrates and smooths the final individual growth state map over time to output a full-cycle growth trajectory. The processing involves performing piecewise linear integration along the time axis on the feature dimension change curves of the master node set within the final individual growth state map. The cumulative change of each feature dimension over the observation period is calculated, reflecting the total change in physical characteristics such as body length and weight estimation index during the growth cycle. A time-window moving average is applied to the pattern intensity change curve of each attribute node to generate a smoothed behavior pattern evolution curve. The window width of the moving average is set according to the typical cycle of behavior pattern changes. In a specific implementation, the calculated cumulative changes of each feature dimension and the smoothed behavior pattern evolution curves are aligned along a unified time axis and combined into a multidimensional, time-continuous vector sequence. This vector sequence integrates information from cumulative physical changes and behavior pattern intensity evolution. It can be understood that the dimension of this multidimensional vector sequence is equal to the sum of the number of physical feature dimensions and the number of behavior patterns. Principal component analysis was performed on this multidimensional vector sequence to extract its first k principal components, forming a low-dimensional trajectory curve. This trajectory curve represents the overall growth and development process of a meat duck individual from the initial observation to the end of the observation, which is the full-cycle growth trajectory. The full-cycle growth trajectory Φ can be expressed as:

[0048] in: It is the full-cycle growth trajectory over time. The value (position in principal component space). It is time The corresponding original multidimensional vector, It is the first Principal component direction vectors It represents the number of principal components retained.

[0049] In practical implementation, data comparison is reflected in the graph structure before and after optimization iterations, as well as the final generated trajectory. For example, after pruning and feature enhancement, the average path length and clustering coefficient of the individual growth state graph may change, and the intrinsic coordination index can be improved. Optionally, the final individual growth state graphs of different individuals or different individuals within the same group will have different network topologies, and these differences will be transmitted and reflected in the shape of their full-cycle growth trajectory curves. It can be understood that the first principal component of the full-cycle growth trajectory curve obtained after principal component analysis may mainly explain the growth trend strongly correlated with age, while the second principal component may explain the fluctuations related to health or behavioral status. In some embodiments, the output form of the full-cycle growth trajectory can be a two-dimensional or three-dimensional spatial curve, which is convenient for visualization and comparison. The starting point of the curve corresponds to the initial growth state of the individual, the ending point of the curve corresponds to the growth state at the end of the observation, and the shape of the intermediate path of the curve represents the dynamics of the growth process. Finally, the full-cycle growth trajectory output by the system provides a digital and visualized basis for individualized growth assessment and group comparison.

[0050] In one embodiment of the present invention, the resource allocation module extracts the cumulative changes in each physical feature dimension of the full-cycle growth trajectory, as well as the intensity trends of basic behavioral patterns such as feeding and drinking in the smoothed behavioral pattern evolution curve. Taking a large-scale meat duck farm as an example, the farm raises 1,000 white-feathered meat ducks. The full-cycle growth trajectory includes three physical feature dimensions: weight, breast muscle thickness, and leg length, with corresponding cumulative changes of 2.8 kg, 1.2 cm, and 3.5 cm, respectively. The intensity trend of the basic behavioral pattern of feeding is an average daily feeding time of 12 hours, and the intensity trend of the basic behavioral pattern of drinking is an average daily water consumption of 1.5 L.

[0051] In some embodiments, the resource allocation module matches the cumulative changes with preset standard slaughter size thresholds and, combined with behavioral pattern intensity trends, predicts the expected slaughter time window for a single duck. The preset standard slaughter size thresholds are: weight ≥ 3 kg, breast muscle thickness ≥ 1.5 cm, and leg length ≥ 3.8 cm. The basic behavioral pattern intensity trend thresholds for feeding are: average daily feeding time ≥ 10 hours, and the basic behavioral pattern intensity trend thresholds for drinking are: average daily water intake ≥ 1.2 L. For a single duck A, its cumulative changes are: weight 2.9 kg, breast muscle thickness 1.3 cm, leg length 3.6 cm, feeding time 11 hours, and water intake 1.4 L. After matching, the expected slaughter time window is calculated to be from day 38 to day 40 of rearing.

[0052] Optionally, the resource allocation module dynamically adjusts the feed ratio, pen space allocation, and environmental control parameters for the corresponding duck flock based on the expected slaughter time window, generating a breeding resource allocation strategy. For duck flocks with an expected slaughter time window between day 38 and day 40, the feed ratio is adjusted to 65% energy feed, 30% protein feed, and 5% mineral feed, and the pen space allocation is adjusted to 0.15m² per duck. 2 The environmental control parameters were adjusted to a temperature of 22°C and a humidity of 60%.

[0053] In some embodiments, the resource allocation module aggregates the estimated slaughter time windows of all individuals and, combined with cold chain logistics capacity and slaughter line scheduling, generates a slaughter logistics scheduling plan. The farm's cold chain logistics capacity is 500 animals transported daily, and the slaughter line schedule is 600 animals processed daily. After aggregation, it is estimated that 300 animals will be slaughtered on day 38, 400 on day 39, and 300 on day 40. The generated logistics scheduling plan involves booking cold chain vehicles on day 37 and dispatching vehicles daily from day 38 to day 40.

[0054] Optionally, the formula is used to calculate the adjustment factor for the expected slaughter time window, and the formula is:

[0055] in, To adjust the coefficient, This represents the actual cumulative change in the dimension of shape features throughout the entire growth trajectory. This refers to the cumulative change corresponding to the preset standard exit shape threshold. This represents the actual intensity trend of basic behavioral patterns related to foraging and drinking. This represents the standard intensity trend of basic behavioral patterns related to feeding and drinking. Adjustment coefficient. When the value is greater than 1, the expected slaughter time window is earlier; when the value is less than 1, the expected slaughter time window is later.

[0056] It is understandable that the resource allocation module achieves precise allocation of aquaculture resources through the above operations, avoiding resource waste or insufficient supply. For example, when the adjustment coefficient... At that time, the expected slaughter time window is 2 days earlier than the standard time, and the feed ratio will be reduced by 2% for energy feed, and the pen space allocation will be adjusted to 0.14m² per duck. 2 .

[0057] It is understandable that the generation of the slaughter logistics scheduling plan needs to take into account the fluctuation of cold chain logistics capacity. When the cold chain logistics capacity decreases by 10%, the slaughter logistics scheduling plan will be adjusted to slaughter 270 animals on the 38th day, 430 animals on the 39th day, and 300 animals on the 40th day, to ensure that it matches the slaughter production line schedule.

[0058] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A monitoring and management system for the entire growth trajectory of meat ducks, characterized in that, The system includes: The multi-source sensing module performs real-time sensing of the duck flock based on a multi-source sensor network, forming an original sensing stream. The feature decoupling module parses the original sensory stream, extracts independent sensory segments related to individual ducks, performs deep feature decoupling on each independent sensory segment, and separates static body features from dynamic behavior features. The state modeling module performs cross-timestamp association and fusion on the separated static body features to generate a continuous body evolution sequence, performs pattern recognition and quantification on the separated dynamic behavior features to generate a quantized behavior pattern spectrum, and spatiotemporally aligns and couples the continuous body evolution sequence with the quantized behavior pattern spectrum to construct an individual growth state diagram. The trajectory generation module calculates an internal coordination index reflecting growth consistency based on the individual growth state map, performs feedback iterative optimization on the individual growth state map based on the internal coordination index, forms a final individual growth state map, integrates and smooths the final individual growth state map in the time dimension, and outputs the full-cycle growth trajectory of the meat duck individual. The resource allocation module is used to predict the slaughter time of individual meat ducks based on the full-cycle growth trajectory, and dynamically adjust the allocation strategy of breeding resources and the slaughter logistics plan according to the prediction results.

2. The meat duck full-cycle growth trajectory monitoring and management system as described in claim 1, characterized in that, The method of performing real-time perception of the duck flock based on a multi-source sensor network to form an original perception stream includes: Deploy a heterogeneous sensor network that includes a non-contact weight sensor array, a distributed acoustic sensor network, and a millimeter-wave radar scanning device; The non-contact weight sensor array is used to acquire time-series data of the pressure distribution generated by the duck flock in the ground activity area; The distributed acoustic sensor network synchronously collects sound spectrum information within the aquaculture environment to form acoustic fingerprint time-series data; The millimeter-wave radar scanning device periodically transmits and receives modulated signals to extract point cloud time-series data containing the duck flock's location, speed, and direction of movement. The pressure distribution time-series data, the voiceprint time-series data, and the point cloud time-series data are spliced ​​and fused together using a unified time reference to generate the original perception stream.

3. The meat duck full-cycle growth trajectory monitoring and management system as described in claim 1, characterized in that, The original sensory stream was analyzed, and independent sensory fragments related to individual meat ducks were extracted from it, including: The original sensory stream is subjected to joint time-frequency domain analysis to identify signal segments that match preset individual activity primitives; For each marked signal segment, the pressure centroid position and pressure distribution profile are calculated from the pressure distribution time-series data channel of the original sensing stream; By combining the pressure center of gravity position with the pressure distribution contour, motion trajectory segments belonging to the same individual are separated from the point cloud time series data channel of the original sensing stream. From the acoustic feature data channel of the original sensing stream, extract the acoustic feature segment that is time-synchronized with the motion trajectory segment; The pressure distribution profile, the motion trajectory segment, and the acoustic feature segment are packaged together to form an independent sensing segment.

4. The meat duck full-cycle growth trajectory monitoring and management system as described in claim 1, characterized in that, Deep feature decoupling is performed on each of the independent perceptual segments to separate static shape features from dynamic behavior features, including: The pressure distribution contour in the independent sensing segment is fitted with a contour curve, and the contour curvature change sequence is extracted as a preliminary shape feature. Kinematic differential analysis is performed on the motion trajectory segments in the independent sensing segments to calculate the temporal values ​​of velocity, acceleration, and changes in motion direction, forming a motion feature sequence; Mel frequency cepstral coefficients are extracted from the acoustic feature segments in the independent sensing segments to obtain acoustic feature vectors; The contour curvature change sequence, the motion feature sequence, and the acoustic feature vector are input into a pre-trained feature decoupling network; The feature decoupling network outputs two mutually orthogonal latent variables. One latent variable is decoded as the static body features representing the individual's body shape and posture at the current moment, and the other latent variable is decoded as the dynamic behavioral features representing the individual's actions and gait at the current moment.

5. The whole-cycle growth trajectory monitoring and management system for meat ducks as described in claim 1, characterized in that, The step of performing cross-timestamp association and fusion on the separated static shape features to generate a continuous body evolution sequence includes: Map the static shape features from different timestamps to a high-dimensional feature space, and calculate the feature distance between any two static shape features; Using time sequence as nodes and the feature distance as edge weights, a feature spatiotemporal correlation graph is constructed; On the aforementioned spatiotemporal correlation graph, a path search algorithm based on minimum spanning tree is used to find an optimal correlation path that runs through all time nodes; According to the node order of the optimal association path, nonlinear smooth interpolation is performed on all the static body features on the path to generate the continuous body evolution sequence that is continuous and smooth in time.

6. The meat duck full-cycle growth trajectory monitoring and management system as described in claim 1, characterized in that, The step of performing pattern recognition and quantization on the separated dynamic behavioral features to generate a quantized behavioral pattern spectrum includes: Cluster analysis is performed on the dynamic behavioral features within multiple time windows to identify basic behavioral patterns; Define a unique pattern code for each of the identified basic behavioral patterns; Statistics were compiled on the frequency, average duration, and intensity distribution of each basic behavioral pattern within a preset evaluation period. The pattern encoding, frequency of occurrence, average duration, and intensity distribution are integrated to construct a multi-dimensional behavior quantification matrix; Singular value decomposition is performed on the behavior quantization matrix to extract principal components, forming a low-dimensional spectrum of the quantized behavior patterns.

7. The meat duck full-cycle growth trajectory monitoring and management system as described in claim 1, characterized in that, The continuous body evolution sequence is spatiotemporally aligned and coupled with the quantified behavioral pattern spectrum to construct an individual growth state map, including: Establish a unified spatiotemporal coordinate system with time as the horizontal axis, the feature dimension in the continuous body evolution sequence as the first vertical axis, and the pattern intensity in the quantified behavior pattern spectrum as the second vertical axis; In the unified spatiotemporal coordinate system, the continuous volume evolution sequence is represented as a multidimensional feature evolution trajectory line; In the unified spatiotemporal coordinate system, the quantized behavior pattern spectrum is represented as a set of time-varying multidimensional pattern intensity surfaces; Calculate the correlation strength between the feature evolution trajectory and the multidimensional mode intensity surface at each time point to form a correlation strength time series; Using the points on the feature evolution trajectory as master nodes, the corresponding points on the multidimensional pattern intensity surface as attribute nodes, and the association strength as edge weights, a dynamic heterogeneous graph network is constructed, which is the individual growth state graph.

8. The whole-cycle growth trajectory monitoring and management system for meat ducks as described in claim 1, characterized in that, Based on the individual growth status map, an intrinsic coordination index reflecting growth consistency is calculated, including: In the individual growth state diagram, the mutual information value between the set of master nodes and the set of attribute nodes is calculated; In the individual growth state diagram, traverse all connection paths between master nodes and attribute nodes, and calculate the average path length and clustering coefficient of all paths; Extract the topological structure of the individual growth state map on multiple consecutive time slices, calculate the similarity of the topological structure between adjacent time slices, and form a topological evolution consistency sequence. The mutual information value, the average path length, the clustering coefficient, and the topological evolution consistency sequence are weighted and fused to calculate the intrinsic coordination index, which characterizes the degree of synergy between physical changes and behavioral patterns during individual growth and development.

9. The whole-cycle growth trajectory monitoring and management system for meat ducks as described in claim 1, characterized in that, Based on the aforementioned intrinsic coordination index, the individual growth state map is subjected to feedback iterative optimization to form the final individual growth state map, including: Set an optimization target threshold for the intrinsic coordination index; Compare the intrinsic coordination index corresponding to the current individual growth state diagram with the optimization target threshold; If the intrinsic coordination index does not reach the optimization target threshold, then the edges with association strength lower than the preset threshold in the individual growth state graph are pruned, and the main node and attribute node connected by the edge with the highest association strength are enhanced. Based on the graph structure after pruning and feature enhancement, the intrinsic coordination index is recalculated. Repeat the steps of comparing indicators, pruning and enhancing the graph structure, and recalculating the indicators until the intrinsic coordination indicator reaches or exceeds the optimization target threshold. The resulting graph structure is the final individual growth state graph. The final individual growth status map is integrated and smoothed over time to output the full-cycle growth trajectory of the meat duck, including: In the final individual growth state diagram, piecewise linear integration is performed on the change curves of all feature dimensions of the master node set along the time axis to calculate the cumulative change of each feature dimension within the observation period. The pattern intensity change curve of each attribute node is processed by time window moving average to generate a smoothed behavior pattern evolution curve. The cumulative change and the smoothed behavior pattern evolution curve are aligned in time and combined into a multidimensional, time-continuous vector sequence. Principal component analysis is performed on the vector sequence to extract its first few principal components, forming a low-dimensional trajectory curve that represents the overall growth and development process of a meat duck individual from the initial stage to the end of the observation. The trajectory curve is the full-cycle growth trajectory.

10. The whole-cycle growth trajectory monitoring and management system for meat ducks as described in claim 1, characterized in that, The resource allocation module is used to perform the following operations: Extract the cumulative changes in each shape feature dimension in the full-cycle growth trajectory, and the intensity trends of basic feeding and drinking behavior patterns in the smoothed behavior pattern evolution curve; The cumulative change is matched with a preset standard slaughter size threshold, and combined with the intensity trend of the behavior pattern, the expected slaughter time window of a single meat duck is predicted. Based on the expected slaughter time window, the feed ratio, pen space allocation and environmental control parameters of the corresponding duck flock are dynamically adjusted to generate a breeding resource allocation strategy. The estimated slaughter time window for all individuals is aggregated, and a slaughter logistics scheduling plan is generated by combining cold chain logistics capacity and slaughter production line schedules.

Images