Animal individual health monitoring system based on real-time data

Abstract

The invention relates to the technical field of animal health monitoring, and discloses an animal individual health monitoring system based on real-time data. According to the system physiological data acquisition assembly, vital sign original data flow of an animal individual is synchronously obtained through an implantable biosensor and a non-contact optical sensor, and the data flow comprises a cortex temperature change curve, a microcirculation blood flow frequency spectrum and a three-dimensional motion acceleration sequence; the multi-modal fusion analysis component is used for carrying out space-time alignment and feature level fusion processing on the original data stream to generate a comprehensive physiological state map; the health state evaluation component is used for performing multi-scale pattern recognition on the atlas by adopting a deep neural network model and outputting a health risk probability value and a potential pathological region mark; the dynamic self-adaptive compensation component is used for carrying out individualized calibration on the risk probability value according to the animal variety physiological constant database; and the real-time feedback execution component is used for generating a health intervention instruction set based on the pathological region marks.

Description

Animal individual health monitoring system based on real-time data Technical Field

[0001] This invention relates to the field of animal health monitoring technology, specifically to an individual animal health monitoring system based on real-time data. Background Technology

[0002] In intensive farming, companion animal health management, and wildlife conservation, continuous and accurate monitoring of individual animal health is crucial. Traditional animal health monitoring primarily relies on methods such as periodic manual observation, behavioral assessment, or invasive blood tests. These methods have significant limitations: manual observation is highly subjective, difficult to maintain consistently, and prone to missing early abnormalities; invasive tests can cause stress in animals, disrupting their normal physiological state, and cannot achieve high-frequency, long-term continuous monitoring. For animals, especially those unable to express discomfort verbally, disease development is usually a gradual process, often manifesting in subtle changes in physiological parameters and behavioral patterns in the early stages, such as small fluctuations in core body temperature, changes in local microcirculatory blood flow velocity, or abnormal increases or decreases in activity levels. These early signals are weak and easily masked by environmental noise, making them extremely difficult to capture and identify using traditional methods.

[0003] With the development of sensor technology, implantable biosensors can monitor deep physiological parameters within the body, while non-contact optical sensors can acquire physiological information from the body surface without disturbance. However, data generated by different types of sensors differ in spatiotemporal resolution and physical meaning, forming multimodal data silos. Existing technologies often analyze a single type of data in isolation, lacking the ability to effectively fuse multi-source data to construct a unified and in-depth understanding of an individual's overall physiological state. For example, a decrease in exercise volume may be due to fatigue, disease, or environmental factors, but combining it with abnormal microcirculation data and changes in skin temperature in specific areas can more accurately point to specific health problems such as muscle damage or local inflammation. Furthermore, the normal physiological constant ranges differ between different species, and even between different individuals of the same species; using a uniform standard for health assessment may lead to misjudgments. Currently, a comprehensive intelligent health monitoring system capable of automatically fusing multimodal data, calibrating based on individual characteristics, and automatically generating personalized intervention plans based on assessment results is still lacking. Summary of the Invention

[0004] The purpose of this invention is to provide an animal individual health monitoring system based on real-time data to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides an animal individual health monitoring system based on real-time data, the system comprising:

[0006] The physiological data acquisition component synchronously acquires raw data streams of vital signs of an individual animal through implanted biosensors and non-contact optical sensors. The raw data streams of vital signs include cortical temperature change curves, microcirculatory blood flow spectra, and three-dimensional motion acceleration sequences.

[0007] The multimodal fusion analysis component performs spatiotemporal alignment and feature-level fusion processing on the raw data stream of vital signs to generate a comprehensive physiological state map of an individual animal. The comprehensive physiological state map includes a metabolic equivalent distribution map, a neural excitability index, and a fatigue accumulation index.

[0008] The health status assessment component uses a deep neural network model to perform multi-scale pattern recognition on the comprehensive physiological status map, and outputs health risk probability values ​​and potential pathological area markers.

[0009] The dynamic adaptive compensation component performs individualized calibration of the health risk probability value based on a database of physiological constants of animal species, and generates a calibrated health risk level.

[0010] The real-time feedback execution component generates a multi-channel health intervention instruction set based on the potential pathological area markers. The health intervention instruction set includes nutritional supplementation strategies, environmental parameter adjustment schemes, and physical rehabilitation training plans.

[0011] Preferably, the specific steps for feature-level fusion processing performed by the multimodal fusion analysis component include:

[0012] The cortical temperature change curve is divided into multiple temperature segments according to the physiological cycle, and each temperature segment corresponds to a complete cardiopulmonary circulation cycle.

[0013] For each of the cardiopulmonary circulation cycles, the following operations are performed: a hemodynamic model of the capillary network is constructed using the microcirculation blood flow spectrum, the model including the blood flow velocity field and vessel wall elastic parameters;

[0014] The three-dimensional motion acceleration sequence is projected onto the animal body coordinate system, and the kinematic feature vectors of each joint are calculated.

[0015] By fusing the hemodynamic model and kinematic feature vectors through an attention mechanism, a metabolic equivalent distribution map, neural excitability index, and fatigue accumulation index for the current cycle are generated.

[0016] Preferably, the attention mechanism fusion process includes:

[0017] A nonlinear mapping relationship between microcirculatory blood flow velocity and cortical temperature changes was established, and spatiotemporal features were extracted through a gated circulation unit;

[0018] Based on the power spectral density distribution of three-dimensional motion acceleration, motion mode characteristics are identified and energy consumption estimates are calculated.

[0019] A feature cross-network is used to fuse spatiotemporal features and energy consumption estimates to generate pixel-level data of metabolic equivalent distribution map;

[0020] Numerical representations of neural excitability indices and fatigue accumulation indices are generated using a multilayer perceptron decoder.

[0021] Preferably, the method for multi-scale pattern recognition by the health status assessment component includes:

[0022] The metabolic equivalent distribution map is input into a convolutional neural network architecture to extract local tissue metabolic abnormalities.

[0023] Long short-term memory networks were used to analyze time-series patterns of neural excitability indicators to detect the functional status of the central nervous system.

[0024] By modeling the propagation path of the fatigue accumulation index using graph neural networks, the load distribution of the musculoskeletal system can be identified.

[0025] By integrating the abnormal metabolic features of the local tissues, the functional state of the central nervous system, and the load distribution of the musculoskeletal system, a health risk probability value and a potential pathological area marker are output.

[0026] Preferably, the process for individualized calibration of the dynamic adaptive compensation component includes:

[0027] Query the standard physiological constants database for animal breeds to obtain the basal metabolic rate range and neural excitability threshold;

[0028] Calculate the deviation between the current metabolic equivalent distribution map and the basal metabolic rate range, and generate a metabolic compensation coefficient;

[0029] By analyzing the fluctuation characteristics of neural excitability indices relative to thresholds, neural function regulatory factors are generated.

[0030] The probability value of health risk is calibrated by weighted fusion of the metabolic compensation coefficient and the neural function regulation factor.

[0031] Preferably, the generation of the metabolic compensation coefficient further includes:

[0032] The sliding window algorithm was used to analyze the spatiotemporal variation patterns of metabolic equivalent distribution maps and identify abnormal metabolic hotspots.

[0033] Based on the geometric characteristics and intensity distribution of hotspot areas, an index of metabolic imbalance degree is calculated;

[0034] By combining environmental temperature and humidity data, the indicators of metabolic imbalance are normalized to generate standardized metabolic compensation coefficients.

[0035] The metabolic compensation coefficient is smoothed using a Kalman filter to eliminate the influence of measurement noise.

[0036] Preferably, the process by which the real-time feedback execution component generates a health intervention instruction set includes:

[0037] Analyze the spatial distribution characteristics of potential pathological area markers to delineate emergency intervention areas and routine monitoring areas;

[0038] Design precise nutrition supplementation strategies for emergency intervention areas, including specific nutrient ratios and feeding time planning;

[0039] Adjust environmental parameter regulation schemes according to animal behavior patterns, and control light cycle and temperature and humidity gradients;

[0040] A physical rehabilitation training plan is customized based on the results of the exercise ability assessment, and parameters for training intensity and duration are set.

[0041] Preferably, the specific method for dividing the cortical temperature change curve into multiple temperature segments according to the physiological cycle includes:

[0042] The extreme points in the cortical temperature change curve are identified by the peak detection algorithm, and the interval between adjacent extreme points is defined as a basic temperature range.

[0043] Wavelet transform is performed on each basic temperature range to extract the frequency domain features of the temperature waveform;

[0044] The basic temperature range is merged or divided based on the dominant frequency component in the frequency domain characteristics to form a standardized temperature range that precisely matches the cardiopulmonary circulation cycle.

[0045] Preferably, the process of fusing spatiotemporal features and energy consumption estimates using the feature cross-network further includes:

[0046] Construct a feature cross network with a dual-tower structure, where the first tower processes spatiotemporal features and the second tower processes energy consumption estimates.

[0047] Residual connections and batch normalization layers are used in each tower to ensure training stability, and the correlation weights between the output features of the two towers are calculated through an attention matching mechanism.

[0048] Preferably, the generation of the precision nutrition supplementation strategy further includes:

[0049] Establish a database of nutrient-physiological state mapping relationships, including the required micronutrient ratios under different pathological states;

[0050] Reinforcement learning algorithms are used to optimize feeding time planning, taking into account the animal's digestion cycle and activity patterns;

[0051] A feedback adjustment mechanism is set up to dynamically adjust the nutritional supplementation plan based on changes in physiological indicators after intervention, generating a multi-objective optimized supplementation strategy that takes into account nutritional balance, palatability, and absorption efficiency.

[0052] Compared with the prior art, the beneficial effects of the present invention are:

[0053] This invention's system achieves refined and intelligent management of individual animal health by constructing a complete technology chain from multi-source data acquisition and fusion analysis to intelligent assessment and intervention execution. The physiological data acquisition component, combined with implantable and non-contact sensors, can simultaneously acquire various vital sign data from within the body to the surface, and from chemical to physical parameters, providing rich raw information for comprehensive assessment. The multimodal fusion analysis component performs spatiotemporal alignment and feature-level fusion of heterogeneous data, generating a comprehensive physiological state map that reflects the body's metabolic level, nervous system state, and physical exertion. This map integrates fragmented information from different sources into an organic whole, deepening the understanding of physiological states.

[0054] The health status assessment component utilizes deep neural networks to process comprehensive physiological state maps. Its powerful pattern recognition capabilities can extract deep, non-linear patterns related to health risks from complex, high-dimensional data, outputting quantified risk probabilities and specific suspected pathological areas, thus shifting health assessment from qualitative judgment to quantitative analysis. The dynamic adaptive compensation component incorporates a species physiological constant database for individualized calibration, considering the impact of individual differences on normal physiological baselines, making risk assessment results more targeted and accurate. The real-time feedback execution component transforms assessment results into an actionable set of health intervention instructions, covering multiple aspects such as nutrition, environment, and rehabilitation, forming a closed loop of "monitoring-assessment-intervention." This system shifts animal health management from post-treatment diagnosis and passive response to pre-emptive warning and proactive intervention, helping to identify potential health problems early, implement personalized care plans, and improve animal welfare and management efficiency. Attached Figure Description

[0055] Figure 1 shows the overall system monitoring view;

[0056] Figure 2 is a flowchart of the feature-level fusion processing of the multimodal fusion analysis component;

[0057] Figure 3 is a flowchart of the multi-scale pattern recognition method for the health status assessment component;

[0058] Figure 4 shows the dynamic adaptive compensation analysis diagram. Detailed Implementation

[0059] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0060] Please refer to Figure 1. This invention provides an animal individual health monitoring system based on real-time data. The system includes: a physiological data acquisition component, a multimodal fusion analysis component, a health status assessment component, a dynamic adaptive compensation component, and a real-time feedback execution component. Specific implementation methods are as follows:

[0061] The physiological data acquisition component simultaneously acquires raw vital sign data streams from individual animals via implanted biosensors and non-contact optical sensors. These raw data streams include cortical temperature change curves, microcirculatory blood flow spectra, and three-dimensional motion acceleration sequences. The multimodal fusion analysis component performs spatiotemporal alignment and feature-level fusion processing on the raw vital sign data streams to generate a comprehensive physiological state map of the individual animal. This map includes a metabolic equivalent distribution map, neural excitability indicators, and a fatigue accumulation index. The health status assessment component uses a deep neural network model to perform multi-scale pattern recognition on the comprehensive physiological state map, outputting health risk probability values ​​and potential pathological region markers. The dynamic adaptive compensation component performs individualized calibration of the health risk probability values ​​based on a physiological constant database for the animal species, generating a calibrated health risk level. The real-time feedback execution component generates a multi-channel health intervention instruction set based on potential pathological region markers. This instruction set includes nutritional supplementation strategies, environmental parameter adjustment schemes, and physical rehabilitation training plans.

[0062] Example 1: Referring to Figure 2, the peak detection algorithm identifies extreme points in the cortical temperature change curve, defining the interval between adjacent extreme points as a base temperature segment. The peak detection algorithm uses a first-order difference combined with threshold comparison to locate local maxima and minima. The difference operation calculates the rate of change of values ​​between adjacent sampling points, and a dynamic threshold is set to filter out minor fluctuations caused by sensor noise. The interval between adjacent extreme points constitutes a base temperature segment, representing a complete fluctuation unit of cortical temperature from trough to peak and back to trough or from peak to trough and back to peak. Wavelet transform is performed on each base temperature segment to extract the frequency domain features of the temperature waveform. The wavelet transform uses the Morlet wavelet basis function to decompose the time domain signal of the base temperature segment into frequency components of different scales. The scale parameter is selected to cover the typical frequency range relevant to animal physiology. The frequency domain feature extraction process calculates the amplitude and phase information of each frequency component to generate a spectrum of the temperature waveform. Based on the dominant frequency component in the frequency domain features, the base temperature segments are merged or segmented to form standardized temperature segments that precisely match the cardiopulmonary circulation cycle. The dominant frequency component is determined by analyzing the power distribution of the spectrum, and the frequency component with the highest power is selected as the dominant frequency. The cardiopulmonary circulation cycle corresponds to a characteristic frequency band in the frequency domain. Dominant frequencies within the characteristic frequency band of a baseline temperature segment are retained as standard segments; dominant frequencies deviating from the characteristic frequency band trigger merging or segmentation operations. The merging operation combines multiple consecutive baseline temperature segments with similar dominant frequencies into a longer standardized temperature segment. The segmentation operation divides a baseline temperature segment with mixed dominant frequencies into multiple clean segments in the time domain. Standardized temperature segments maintain consistency and comparability in time length, and each standardized temperature segment precisely corresponds to a complete cardiopulmonary circulation cycle.

[0063] For each cardiopulmonary circulation cycle, the following operations were performed: A hemodynamic model of the capillary network was constructed using the microcirculation blood flow spectrum. The model included the blood flow velocity field and vessel wall elastic parameters. The microcirculation blood flow spectrum data came from a non-contact optical sensor, and the spectral information reflected the laser Doppler frequency shift caused by red blood cell movement. The hemodynamic model was established based on a simplified form of the Navier-Stokes equations, treating the capillary network as a porous medium. The blood flow velocity field described the three-dimensional flow characteristics of blood within the microvessels. The vessel wall elastic parameters were derived by analyzing the harmonic components of the blood flow spectrum; the pulsatile expansion and contraction of the vessels generated characteristic sidebands on the spectrum. The model solution process used the finite element method to discretize the computational domain and iteratively calculate the blood flow velocity and vessel wall stress distribution at each grid point. The three-dimensional motion acceleration sequence was projected onto the animal body coordinate system, and the kinematic eigenvectors of each joint were calculated. The three-dimensional motion acceleration sequence came from the inertial measurement unit of the implanted biosensor, and the raw data was based on the sensor coordinate system. The animal body coordinate system was established as a standard anatomical coordinate system with the animal spine as the vertical axis. The projection transformation used a direction cosine matrix to achieve coordinate system rotation. The positions of each joint are determined based on an animal skeletal model, and the kinematic feature vectors include the joint angles, angular velocities, and the magnitude and direction of linear acceleration. The calculation of feature vectors involves the principles of rigid body kinematics, simplifying limb movements into multi-link mechanism movements.

[0064] By fusing hemodynamic models and kinematic feature vectors through an attention mechanism, a metabolic equivalent distribution map, neural excitability index, and fatigue accumulation index for the current cycle are generated. The attention mechanism fusion process involves establishing a nonlinear mapping relationship between microcirculatory blood flow velocity and cortical temperature changes, and extracting spatiotemporal features through a gated circulation unit. Microcirculatory blood flow velocity field data is spatiotemporally registered with synchronously acquired cortical temperature change curves. The nonlinear mapping relationship is modeled using a multilayer perceptron, with blood flow velocity as input and predicted temperature changes as output. The gated circulation unit processes spatiotemporal sequence data, resetting the degree of forgetting historical information controlled by the gate and updating the proportion of new information incorporated into the gate management. Spatiotemporal feature extraction focuses on the dynamic evolution of the blood flow-temperature coupling pattern, capturing the long-term dependence of physiological rhythms. Based on the power spectral density distribution of three-dimensional motion acceleration, motion pattern features are identified and energy consumption estimates are calculated. The power spectral density distribution is calculated using the Welch method, segmenting the acceleration sequence and applying a Hanning window to reduce spectral leakage. Motion pattern feature identification is based on the energy distribution of characteristic frequency bands; different activity states such as static, walking, and running have unique signatures in the frequency domain. Energy consumption estimates are derived from the area of ​​the integral acceleration power spectrum within a specific frequency band, with the conversion formula based on a biomechanical model. A feature cross-network is employed to fuse spatiotemporal features and energy consumption estimates, generating pixel-level data of the metabolic equivalent distribution map. The feature cross-network constructs a dual-tower structure, where the first tower processes spatiotemporal features and the second tower processes energy consumption estimates. The spatiotemporal feature input to the first tower is processed by a graph convolutional network, with the graph structure constructed based on the capillary network topology, and node features including blood flow velocity and temperature information. The energy consumption estimate input to the second tower is encoded using a fully connected network to upscale the scalar estimate into a high-dimensional feature vector. Residual connections and batch normalization layers are used in each tower. Residual connections propagate gradients across layers to mitigate network degradation, while batch normalization layers adjust the activation value distribution to accelerate training convergence. An attention matching mechanism is used to calculate the correlation weights between the output features of the two towers, and the attention matching mechanism calculates the dot product similarity of the feature vectors of the two towers, which is then normalized to attention weights using the Softmax function. A weighted summation of dual-tower features generates a fused feature vector. This fused feature vector is then input to a deconvolutional network for sampling, generating pixel-level data for a metabolic equivalent distribution map. The metabolic equivalent distribution map presents the intensity of local tissue metabolic activity as a two-dimensional heatmap, with pixel values ​​representing the oxygen consumption equivalent per unit tissue. A multilayer perceptron decoder generates numerical representations of neural excitability indices and fatigue accumulation indices. The multilayer perceptron decoder receives the fused feature vector output from a feature crossover network, and the hidden layers use the ReLU activation function for nonlinear transformation. The neural excitability index output layer uses a Sigmoid activation function to constrain the values ​​to the range of 0 to 1, reflecting the balance between the sympathetic and parasympathetic nervous systems. The fatigue accumulation index output layer uses a linear activation function to generate unbounded continuous values, with the index representing the accumulation level of metabolic waste products such as lactic acid in muscles.The decoder training uses a mean squared error loss function, and the optimization algorithm employs adaptive moment estimation. The entire fusion process progresses sequentially along the cardiopulmonary circulation cycle, generating a serialized comprehensive physiological state map.

[0065] The process of fusing spatiotemporal features and energy consumption estimates using a feature cross-network further includes parameter configuration of a dual-tower structure. The first tower's graph convolutional network has three convolutional layers, each followed by a batch normalization layer and a LeakyReLU activation function. Mean pooling is used for neighborhood aggregation in the graph convolutions, and a gating mechanism is used for the node update function. The second tower's fully connected network is designed as a four-layer structure, with gradually increasing inter-layer dimensions to capture complex mapping relationships. Residual connections are added between every two fully connected layers, and skip connections achieve identity mapping. A temperature coefficient is introduced into the similarity calculation of the attention matching mechanism to adjust the sharpness of the distribution; this temperature coefficient gradually decays during training. A gating mechanism is used to control the information flow in the feature weighting fusion stage; the gating signal is jointly generated by the features from both towers. A deconvolutional network is used to upsample and fuse the feature vectors to the target spatial resolution. The deconvolutional layers use transposed convolution operations, and the kernel size and stride match the size requirements of the metabolic equivalent distribution map. The number of hidden units in the multilayer perceptron decoder is set according to the task complexity. The neural excitability index decoding branch uses a bottleneck structure to compress features, while the fatigue accumulation index decoding branch uses a wide structure to preserve details. Mini-batch gradient descent is used during training, and the learning rate scheduler is dynamically adjusted based on the validation set loss. Model regularization techniques include weight decay and Dropout to prevent overfitting and improve generalization ability. During the inference phase, multimodal data from one cardiopulmonary cycle are input, and the output is a comprehensive physiological state map including a metabolic equivalent distribution map, neural excitability index, and fatigue accumulation index. This comprehensive physiological state map is used as a standardized feature representation input to the downstream health status assessment component.

[0066] Example 2: Referring to Figure 3, the metabolic equivalent distribution map represents the energy metabolism intensity of different regions on the animal's body surface in the form of a two-dimensional matrix. The value of each pixel in the matrix represents the metabolic equivalent value of the corresponding anatomical location. The convolutional neural network architecture adopts a multi-branch design. The first branch uses a small-scale convolutional kernel to capture subtle texture changes in the metabolic equivalent distribution map. The size of the small-scale convolutional kernel is set to 3x3 pixels, and the stride is set to 1 pixel. After the convolution operation, a nonlinear transformation is introduced through the ReLU activation function. The second branch uses a large-scale convolutional kernel to identify regional metabolic abnormality patterns. The size of the large-scale convolutional kernel is set to 7x7 pixels. Dilated convolution expands the receptive field without increasing the number of parameters. The convolutional outputs of the two branches are concatenated in the channel dimension to form a multi-scale feature map. Local tissue metabolic abnormality features are extracted through spatial pyramid pooling layers to extract feature representations at different resolutions. The pooling window sizes are 4x4, 8x8, and 16x16 pixels, respectively, converting the variable-size input into a fixed-length feature vector. The feature vector is mapped to a high-dimensional space through a fully connected layer, and Dropout technology is used to randomly shield some neurons to prevent overfitting. Long Short-Term Memory (LSTM) networks are used to analyze time-series patterns of neural excitability indices to detect the functional state of the central nervous system (CNS). The neural excitability indices are input into the LSM network as continuous numerical sequences, with a fixed sequence length of 64 time steps, each corresponding to a standardized temperature range period. The input gates of the LSM network control the inflow of information at the current moment, the forget gate manages the retention of historical memory units, and the output gate determines the output content of the hidden state. The hidden state vector captures the short-term fluctuations and long-term trends of the neural excitability indices, and time-series pattern analysis focuses on the periodic changes, amplitude fluctuations, and baseline drift characteristics of the index values. CNS functional state detection is achieved by analyzing the statistical properties of the hidden state vector, calculating the moving average, sliding variance, and autocorrelation coefficient. Abnormal state identification is based on preset thresholds; when index fluctuations exceed the normal range, they are marked as CNS functional disorders. The hidden state vector at the last moment of the LSM network serves as a comprehensive representation of the CNS functional state and is input into the fully connected layer for classification decisions.

[0067] This study uses a graph neural network to model the propagation path of the fatigue accumulation index and identify the load distribution in the musculoskeletal system. The musculoskeletal system is modeled as a graph structure, where nodes represent skeletal joints and edges represent muscle connections and mechanical transmission relationships. Node features include the fatigue accumulation index, joint angle, and range of motion, while edge features include the rate of change of muscle length and tension estimates. The graph neural network employs a message-passing mechanism, where each node aggregates feature information from its neighbors, and the update function uses a gated graph neural network layer. The propagation path of the fatigue accumulation index is simulated through multiple rounds of message passing, with node feature updates after each round reflecting the diffusion effect of fatigue levels. The load distribution in the musculoskeletal system is identified through spatial pattern recognition of node features; high-load areas are characterized by clusters of nodes with abnormally high fatigue accumulation indices and persistently high muscle tension. The output of the graph neural network is a load score for each node, which is mapped back to a 3D animal anatomy model to generate a load distribution heatmap.

[0068] This method integrates local tissue metabolic abnormality features, central nervous system functional state, and musculoskeletal system load distribution to output health risk probability values ​​and potential pathological region markers. The fusion process employs a feature concatenation approach, combining the local tissue metabolic abnormality feature vector extracted by a convolutional neural network, the central nervous system functional state vector generated by a long short-term memory network, and the musculoskeletal system load distribution vector output by a graph neural network into a unified multimodal feature vector. This multimodal feature vector undergoes dimensionality compression, and principal component analysis retains 95% of the variance information, reducing feature redundancy. The health risk probability value is calculated using a multilayer perceptron (MLP), which contains three hidden layers with decreasing numbers of neurons in each layer. The output layer uses a sigmoid activation function to constrain the result between 0 and 1. Potential pathological region markers are generated by deconvolution of feature maps based on a convolutional neural network, mapping high-level features back to the original image space. The local tissue metabolic abnormality feature map is spatially aligned with the musculoskeletal system load distribution map, and element-wise multiplication is used to highlight overlapping abnormal regions. Potential pathological regions were marked using binarization, and a dynamic threshold was set to distinguish between normal and abnormal tissues. The marking results were superimposed on the metabolic equivalent distribution map in the form of outline boxes.

[0069] The convolutional neural network architecture is trained using a cross-entropy loss function, with the Adam algorithm chosen as the optimizer and an exponential decay strategy for the learning rate. Training data augmentation techniques include random rotation, mirror flipping, and brightness adjustment to improve the model's robustness to pose changes. The long short-term memory network is trained using a mean squared error loss function, with gradients calculated via backpropagation and gradient pruning to prevent gradient explosion. The graph neural network is trained based on a node classification task, with the loss function incorporating cross-entropy and graph structure regularization to maintain topological continuity. The multimodal feature fusion network is trained end-to-end, with a weighted summation of the loss functions for three branches: local tissue metabolic abnormalities, central nervous system functional state, and musculoskeletal system load distribution. The weight coefficients are adjusted based on validation set performance. Health risk probability values ​​are calibrated using the Platt scaling method, converting the neural network output into probability estimates. The calibration set data comes from historical health records. Post-processing for potential pathological region marking includes morphological closure operations to fill empty regions and non-maximum suppression to merge overlapping bounding boxes. The output interface quantifies health risk probability values ​​into five risk levels, and the potential pathological region markings are associated with confidence scores and spatial coordinate information. The inference process of the health status assessment component is computed in real time. The multi-scale pattern recognition module adopts a parallel computing architecture, with convolutional neural networks, long short-term memory networks, and graph neural networks simultaneously processing data from their respective modalities. The system can dynamically load and update model parameters during runtime to adapt to the physiological characteristics of different animal species.

[0070] Example 3: The animal breed standard physiological constants database stores baseline values ​​of physiological parameters for different species using a tree-like index structure. Each animal breed corresponds to a leaf node, and the node data includes a breed identifier, an age segmentation matrix, a sex classification vector, and upper and lower bound vectors of basal metabolic rate (BMR), as well as upper and lower bound scalar values ​​of neural excitability threshold. Query operations match database records based on the individual animal identification information output by the health status assessment component. The BMR range is expressed in kilocalories per square meter per hour, and the neural excitability threshold is expressed as a dimensionless standardized fraction. Database record updates employ a version control mechanism; each data revision retains historical versions, and new data undergoes distribution consistency checks and outlier detection before being added to the database.

[0071] The deviation between the current metabolic equivalent distribution map and the basal metabolic rate range is calculated to generate a metabolic compensation coefficient. The metabolic equivalent distribution map is derived from the comprehensive physiological state map output by the multimodal fusion analysis component. The distribution map represents the metabolic intensity of each region of the animal's body surface as a floating-point matrix. The basal metabolic rate range is obtained from a database query. The deviation calculation uses a pixel-by-pixel comparison method, matching the metabolic equivalent value of each pixel with the basal metabolic rate range. A spatial weighting factor is introduced into the deviation calculation. The metabolic activity of different anatomical regions has different physiological significance, and the weighting factor for the core trunk region is set higher than that for the extremities. The formula for generating the metabolic compensation coefficient is expressed as:

[0072]

[0073] in: This represents the metabolic compensation coefficient. This represents the total number of pixels in the metabolic equivalent distribution map. Represents the spatial weight factor of the i-th pixel. This represents the metabolic deviation of the i-th pixel. This represents the reference metabolic deviation. This represents the normalization factor for the body surface area corresponding to the i-th pixel. The hyperbolic tangent function tanh ensures that the ratio is mapped to the [-1, 1] interval, representing the metabolic deviation. The calculation uses a piecewise function approach, taking zero values ​​when pixel values ​​are within the basal metabolic rate range, and calculating based on deviation distance when values ​​exceed the range. Reference metabolic deviation. Take one-tenth of the range of basal metabolic rate, and the body surface area normalization factor. Determined based on the anatomical location corresponding to the pixel.

[0074] The fluctuation characteristics of neural excitability indices relative to thresholds were analyzed to generate neural function regulation factors. Neural excitability indices were derived from time-series data output by the health status assessment component, while threshold information was obtained from a database of standard physiological constants for animal breeds. Fluctuation characteristic analysis employed a multi-scale approach: short-term fluctuations were calculated using the coefficient of variation within a sliding window, and long-term trends were analyzed using linear regression slope analysis. The generation of neural function regulation factors was based on a comprehensive assessment of the duration and magnitude of deviations, with sustained deviations from the threshold receiving higher weight than instantaneous peak values. A time decay function was incorporated into the regulation factor calculation, giving greater weight to recent deviation events than historical ones. The neural function regulation factor values ​​were set between 0.5 and 2.0; values ​​greater than 1 indicated a hyperactive state requiring downward regulation, while values ​​less than 1 indicated an inhibited state requiring upward regulation. Health risk probability values ​​were calibrated by weighted fusion of metabolic compensation coefficients and neural function regulation factors. The weighted fusion employed an adaptive weight allocation mechanism, with weight coefficients dynamically adjusted based on the confidence levels of the metabolic compensation coefficients and neural function regulation factors. Confidence assessment was based on data quality indicators, assigning higher weights to modalities with higher signal-to-noise ratios. The health risk probability value calibration formula employs a multiplicative model, multiplying the original health risk probability value by a fused compensation factor. A saturation function is introduced in the compensation factor calculation to prevent overcorrection, and upper and lower limit protection intervals are set to avoid the influence of extreme values. The calibrated health risk probability value is requantized to a standard probability range, maintaining the standardization and comparability of the numerical range.

[0075] The generation of metabolic compensation coefficients further includes analyzing the spatiotemporal variation patterns of the metabolic equivalent distribution map using a sliding window algorithm to identify abnormal metabolic hotspot regions. The sliding window algorithm sets a fixed-length observation window in the time dimension, corresponding to 24 complete cardiopulmonary circulation cycles. Spatiotemporal variation pattern analysis calculates the inter-frame difference matrix of the metabolic equivalent distribution map, and extracts the main variation patterns through principal component analysis for dimensionality reduction. Abnormal metabolic hotspot region identification is based on a local anomaly factor algorithm, calculating the density deviation of each pixel from its spatiotemporal neighborhood. Hotspot regions are defined based on two conditions: the local anomaly factor exceeds a threshold and the continuous occurrence time reaches a minimum duration requirement. Geometric feature extraction of hotspot regions includes calculating the region area, perimeter, circularity, and spatial distribution moments. The region area is obtained by multiplying the pixel count by the actual area represented by a single pixel; the perimeter is calculated using a boundary tracing algorithm; and the circularity is calculated by multiplying 4π by the area and dividing by the square of the perimeter. Intensity distribution analysis calculates the statistical characteristics of pixel values ​​within the hotspot region, including mean, variance, skewness, and kurtosis. The intensity distribution uses a kernel density estimation method to fit a continuous probability distribution and compares the distribution differences with the background region.

[0076] The metabolic imbalance index is calculated by combining geometric features and intensity distribution to generate a multivariate scoring vector. The calculation of the metabolic imbalance index incorporates Mahalanobis distance to measure the deviation of hotspot region features from a health baseline, which is derived from typical metabolic pattern templates in a database of standard physiological constants for animal breeds. The multivariate scoring vector includes three dimensions: spatial anomaly score, temporal persistence score, and intensity anomaly score, each normalized to the range of 0-100. The spatial anomaly score is based on the spatial relationship between the hotspot region and known anatomical structures, with regions closer to vital organs receiving higher scores. The temporal persistence score calculates the frequency and duration of hotspot regions within a sliding window. The intensity anomaly score is weighted by combining pixel value deviation and region area. Environmental temperature and humidity data are used to normalize the metabolic imbalance index, generating standardized metabolic compensation coefficients. Environmental temperature and humidity data are acquired in real-time from sensors deployed in the animal's living area, with the data sampling frequency synchronized with the physiological data acquisition components. Normalization processing establishes a calibration model of environmental factors and metabolic activity, and the model parameters are calibrated experimentally using controlled variables. Temperature and humidity compensation employs a multiple linear regression method to correct the original metabolic imbalance index to its equivalent value under standard environmental conditions. Standard environmental conditions are defined as a baseline state of 25 degrees Celsius and 60% relative humidity. The calculation of the standardized metabolic compensation coefficient incorporates an environmental fitness factor, whose value reflects the degree of difference between the current environment and the standard environment.

[0077] The metabolic compensation coefficient is smoothed using a Kalman filter to eliminate the influence of measurement noise. The Kalman filter is designed with the true value of the metabolic compensation coefficient as the state variable and the estimated value of the metabolic compensation coefficient obtained from the preceding steps as the observed variable. The state transition model assumes that the metabolic compensation coefficient follows a linear change within a short time window, and the process noise covariance is determined statistically based on historical data. The observation model establishes a linear relationship between the state variable and the observed variable, and the observation noise covariance is calculated using sensor accuracy indicators. The Kalman filter performs two phases: prediction and update. In the prediction phase, the prior state at the current time is estimated based on the state at the previous time step. In the update phase, the prior state and the current observation value are fused to generate a posterior state estimate. The smoothed metabolic compensation coefficient is output to the weighted fusion module and used together with the neural function modulation factor for health risk probability value calibration. The Kalman filter parameters are periodically recalibrated, and the statistical characteristics of process noise and observation noise are re-estimated using the latest acquired data.

[0078] The weighted fusion of metabolic compensation coefficient and neural function regulation factor adopts dynamic weight allocation, with weight coefficients inversely proportional to the coefficient of variation of each factor. The coefficient of variation is calculated based on historical numerical sequences within a time window, with factors exhibiting lower fluctuations receiving higher weights. A nonlinear transformation function is introduced for health risk probability value calibration, ensuring the calibrated probability values ​​maintain monotonicity and do not alter the original probability order. An outlier detection mechanism is implemented during calibration; when the metabolic compensation coefficient or neural function regulation factor exceeds a reasonable range, a backup calibration strategy is activated. The backup calibration strategy uses historical averages instead of real-time calculated values ​​to ensure system robustness in the event of sensor data anomalies. The output interface of the dynamic adaptive compensation component generates the calibrated health risk level, which is divided into five standard levels, each corresponding to specific management recommendations and intervention priorities.

[0079] Referring to Figure 4, the working mechanism and effect of the dynamic adaptive compensation component are illustrated. The metabolic compensation coefficient in the figure reflects the deviation between the current metabolic state and the standard physiological constant of the species; its fluctuations reflect the system's real-time compensatory regulation of metabolic abnormalities. The neural function regulatory factor shows the dynamic changes in the functional state of the central nervous system; a value greater than 1 indicates that neural function is hyperactive and needs downward regulation, while a value less than 1 indicates that function is inhibited and needs upward regulation. The metabolic hotspot intensity curve (represented using the maximum metabolic equivalent) reveals the regional distribution and intensity changes of local tissue metabolic abnormalities. The shaded areas in the figure mark high-risk periods, which are usually accompanied by significant fluctuations in the metabolic compensation coefficient and abnormal changes in the neural function regulatory factor, indicating that the system can accurately identify health risks and promptly activate the compensation mechanism.

[0080] Example 4: The process of generating a health intervention instruction set by the real-time feedback execution component includes analyzing the spatial distribution characteristics of potential pathological region markers and dividing the area into emergency intervention areas and routine monitoring areas. Potential pathological region markers are derived from the output of the health status assessment component, and the marker information includes region contour coordinates, abnormality intensity scores, and pathological type classifications. Spatial distribution feature analysis employs morphological image processing methods to calculate the centroid location, area proportion, and spatial clustering degree of the marked regions. Emergency intervention region division is based on a multi-factor decision model, with model input parameters including an abnormality intensity score threshold, region expansion rate, and distance to nearby vital organs. Routine monitoring areas are defined as regions with abnormality intensities below the threshold but exhibiting morphological abnormalities; the monitoring frequency is set to half that of the emergency intervention areas. The region division results are stored in the form of spatial masks, with each pixel labeled with an intervention priority level. Precision nutritional supplementation strategies are designed for emergency intervention areas, including specific nutrient ratios and feeding time planning. The generation of precision nutritional supplementation strategies further includes establishing a nutrient-physiological state mapping database, containing the required micronutrient ratios for different pathological states. The nutrient-physiological state mapping database adopts a relational database structure. The main table stores pathological state codes, and the secondary tables associate nutrient types and recommended dosages. Database queries match records based on the pathological type classification of potential pathological areas and return a basic nutrient formula. Specific nutrient ratios are personalized based on the basic formula, with adjustment factors including animal weight, age, and current metabolic level. Feeding time planning is designed based on the animal's digestive physiological rhythm, taking into account gastric emptying time and nutrient absorption rate. Precision nutrition supplementation strategies are presented in the form of daily menus, specifying the precise weight and feeding time for each feed component.

[0081] A reinforcement learning algorithm was employed to optimize feeding time planning, taking into account the animal's digestive cycle and activity patterns. The reinforcement learning algorithm's state space included real-time physiological indicators, environmental time variables, and historical feeding records. The action space was defined as the feeding time decision-making space, and the reward function was based on the improvement in health indicators after intervention. Animal digestive cycle data were obtained from a physiological constant database, with gastric emptying time and intestinal peristalsis frequency for different species serving as state transition constraints. Activity pattern analysis was based on pattern recognition of three-dimensional motion acceleration sequences, dividing daily activity into feeding, activity, and resting periods. The goal of feeding time optimization was to find a balance between maximizing digestive and absorptive efficiency and minimizing animal activity disturbances. The reinforcement learning agent was trained using a proximal policy optimization algorithm, incorporating noise perturbations to avoid local optima. A feedback adjustment mechanism was implemented to dynamically adjust the nutritional supplementation plan based on changes in physiological indicators after intervention, generating a multi-objective optimized supplementation strategy that considered nutritional balance, palatability, and absorption efficiency. The feedback adjustment mechanism was designed as a closed-loop control system, with the input being the raw vital sign data stream collected after intervention and the output being the parameter adjustments for the nutritional supplementation strategy. Monitoring of physiological indicators focuses on the normalized difference of the metabolic equivalent distribution map, the trend slope of the neural excitability index, and the decay rate of the fatigue accumulation index. Dynamic adjustment employs an incremental PID control algorithm, with the proportional term responding to immediate deviations, the integral term eliminating steady-state errors, and the derivative term predicting trends. Multi-objective optimization constructs a Pareto front: the nutritional balance objective aims to ensure the proportions of various nutrients meet standard requirements; the palatability objective maximizes the animal's voluntary feed intake; and the absorption efficiency objective improves nutrient utilization. The optimization algorithm uses a non-dominated sorting genetic algorithm, where individuals in the population encode nutrient ratio parameters, and the fitness function calculates the weighted sum of the three objectives.

[0082] Environmental parameter regulation schemes were adjusted based on animal behavior patterns, controlling the light cycle and temperature and humidity gradients. Animal behavior pattern analysis was based on long-term observed three-dimensional motion acceleration sequences, and typical behavior pattern sequences were identified through clustering algorithms. The environmental parameter regulation scheme was output in the form of control commands. The light cycle regulation adopted a gradual brightening and dimming method to simulate natural sunrise and sunset, and the light intensity curve used a sine function for smooth transition. A three-dimensional spatial model was established for temperature and humidity gradient control, and gradient parameters were set according to the thermodynamic characteristics of different altitude layers. The status of environmental regulation equipment was monitored in real time, and the deviation between the current environmental parameters and the set values ​​was fed back. An alarm signal was triggered when the deviation exceeded the threshold. A physical rehabilitation training plan was customized based on the results of the exercise capacity assessment, and the training intensity and duration parameters were set. The exercise capacity assessment results came from the musculoskeletal system load distribution analysis of the health status assessment component, and the assessment indicators included joint range of motion, muscle endurance, and balance ability. The physical rehabilitation training plan was designed following the principle of gradual progression, with the initial intensity set at 30% of the animal's maximum capacity, increasing by 5% each week. The training content included passive joint movement, active weight-bearing training, and balance coordination exercises, and the duration of each exercise was dynamically adjusted according to the fatigue accumulation index. The training plan generates detailed instructions for the movement sequences, including the starting position, movement trajectory, number of repetitions, and rest time between sets for each movement. See Table 1 for the detailed structure of the nutrient-physiological state mapping database.

[0083] Table 1: Mapping Relationship between Nutrients and Pathological States

[0084]

[0085] The feedback regulation mechanism employs an adaptive control strategy, with the control system automatically adjusting gain parameters based on the intervention effect. A time-series model of physiological index changes after intervention is established, with unit root tests used for stationarity and sliding window regression for trend analysis. Dynamic adjustments to the nutritional supplementation strategy are constrained by safety boundaries, with each adjustment not exceeding 20% ​​of the current value to prevent drastic changes that could cause digestive discomfort. Weight coefficients in multi-objective optimization are adaptively adjusted based on individual animal preferences, obtained by monitoring animal behavior in selecting different feeds. Palatability assessment uses a two-basin test, recording the feeding time and proportion of different feed formulations. The environmental parameter regulation scheme employs a multi-channel LED system for lighting control, with independent controllable spectral components in different bands. Ultraviolet intensity is adjusted based on vitamin D synthesis requirements, while infrared intensity considers the need to maintain body surface temperature. Temperature and humidity gradient control uses a tiered ventilation strategy: a higher temperature is maintained at the ground level to promote rest, a suitable temperature is maintained at the activity level to support movement, and a slightly lower temperature at the top level to create natural convection. A network of temperature and humidity sensors is deployed at different heights within the animal activity space, with data sampling every minute. The control algorithm fine-tunes the air conditioning system output based on real-time data. The effectiveness of the physical rehabilitation training program is monitored in real time using wearable sensors, with motion trajectory tracking employing an inertial measurement unit combined with a geomagnetic sensor. An automatic training intensity adjustment mechanism dynamically adjusts based on real-time physiological feedback, automatically reducing intensity when the fatigue accumulation index exceeds a safe threshold. An adaptive scheduling algorithm manages the training program's progress, appropriately increasing the proportion of programs with good training results and analyzing the reasons for poor results while adjusting training methods. Long-term tracking of training data generates rehabilitation progress reports, including trends in various indicators and interim assessment conclusions. The output interface of the real-time feedback execution component is standardized as a JSON data packet, containing detailed parameters for three modules: nutritional supplementation strategy, environmental parameter adjustment scheme, and physical rehabilitation training program. Execution instructions are transmitted to each execution device via a wireless communication network, with a device status feedback confirmation mechanism established, awaiting device completion feedback after instruction delivery. The system operation log records the detailed content and execution results of each intervention instruction; log data is used for subsequent strategy optimization and effect evaluation. The real-time feedback execution component and the dynamic adaptive compensation component form a closed-loop control loop, continuously optimizing animal health management strategies.

[0086] Example 5: A specific method for dividing the cortical temperature change curve into multiple temperature segments according to the physiological cycle includes identifying extreme points in the cortical temperature change curve using a peak detection algorithm, and defining the interval between adjacent extreme points as a base temperature segment. The peak detection algorithm adopts a hybrid method based on first-order difference and dynamic threshold. The first-order difference calculates the temperature change rate of adjacent sampling points, and the dynamic threshold is adaptively adjusted according to the local statistical characteristics of the temperature curve. The extreme point determination condition requires that the sign of the temperature change rate changes and the amplitude exceeds the threshold to avoid false detections caused by noise. The interval between adjacent extreme points constitutes the base temperature segment, which represents a complete monotonic change unit of temperature from trough to peak or from peak to trough. Taking canines as an example, the cortical temperature sampling frequency is 10Hz, and the typical base temperature segment duration is in the range of 2-5 seconds, corresponding to the respiratory cycle in a normal resting state. Wavelet transform is performed on each base temperature segment to extract the frequency domain features of the temperature waveform. The wavelet transform uses the Morlet wavelet basis function to decompose the time domain signal of the base temperature segment into frequency components of different scales. The scale parameters were selected to cover the typical frequency range relevant to animal physiology; the main physiological frequencies of canines are concentrated in the 0.1-2Hz range. The frequency domain feature extraction process calculates the modulus maxima and zero-crossing information of the wavelet coefficients at each scale, generating a time-frequency distribution map of the temperature waveform. Feature extraction focuses on characteristic frequency bands related to cardiopulmonary activity; the scale coefficient amplitude corresponding to heart rate reflects the functional state of the circulatory system, while the scale coefficient phase corresponding to respiratory rate contains ventilation efficiency information.

[0087] The baseline temperature segments are merged or segmented based on the dominant frequency components in the frequency domain features to form standardized temperature segments that precisely match the cardiopulmonary circulation cycle. The dominant frequency components are determined by analyzing the power distribution of the wavelet scale spectrum, and the frequency component with the highest power is selected as the dominant frequency. In canines, the cardiopulmonary circulation cycle is characterized by a coupling pattern of heart rate and respiratory rate. The merging operation of standardized temperature segments combines multiple consecutive baseline temperature segments with similar dominant frequencies into longer segments, while the segmentation operation precisely divides the baseline temperature segments with mixed dominant frequencies in the time domain. The duration adjustment of the standardized temperature segments is achieved through resampling technology, unifying all temperature segments to the same number of sampling points, facilitating subsequent feature alignment and comparison. The process of fusing spatiotemporal features and energy consumption estimates through a feature cross-network further includes constructing a dual-tower feature cross-network, where the first tower processes spatiotemporal features and the second tower processes energy consumption estimates. The dual-tower feature cross-network adopts a parallel architecture design; the input spatiotemporal features of the first tower include frequency domain features extracted by wavelet transform and time-domain statistics of the temperature curve. The spatiotemporal feature vector includes statistical features such as mean, variance, skewness, and kurtosis, as well as wavelet energy distribution features, with a fixed vector dimension of 128. The energy consumption estimate for the second tower is input from the integral calculation results of three-dimensional motion acceleration data. The estimate is based on a biomechanical model that converts acceleration power into metabolic equivalents. The energy consumption estimate is mapped to a high-dimensional feature space through a fully connected layer, maintaining consistency with the feature dimension of the first tower.

[0088] Residual connections and batch normalization layers are used in each tower. Residual connections propagate gradients across layers, alleviating the vanishing gradient problem during deep network training. Batch normalization layers standardize the activation values ​​of each layer, accelerating network training convergence. The residual block of the first tower contains two fully connected layers, with a non-linear transformation introduced between layers using the ReLU activation function. The residual block of the second tower is similarly designed but its parameters are learned independently; the parameters of the two towers are not shared, maintaining the independence of feature extraction. The momentum parameter of the batch normalization layer is set to 0.9, and the mean and variance statistics are dynamically updated during training. A correlation weight between the output features of the two towers is calculated through an attention matching mechanism, which also calculates the cosine similarity of the feature vectors of the two towers. The similarity calculation uses dot product operation combined with L2 norm normalization to generate an attention weight matrix. The weight matrix is ​​normalized using the Softmax function so that the sum of all weights is 1. The attention weights reflect the correlation strength between spatiotemporal features and energy consumption estimates; high weights indicate a significant correlation between the two modality features. In the weighted fusion stage, the attention weights are outer productd with the features of the two towers to generate a cross-feature matrix. The cross-feature matrix contains information on intermodal interactions, capturing the coupling relationship between temperature changes and motion energy consumption. The feature cross-network is trained using a contrastive loss function; positive samples are derived from synchronously acquired multimodal data, while negative samples are generated through time offsets. The optimization algorithm employs adaptive moment estimation, with an initial learning rate of 0.001 and an exponential decay strategy. Training utilizes mini-batch gradient descent with a batch size of 256 samples. Network regularization techniques include weight decay and Dropout to prevent overfitting and improve model generalization ability. After training, the feature cross-network effectively fuses heterogeneous sensor data to generate robust feature representations.

[0089] In a practical application example, taking the endurance training monitoring of working dogs as an example, a cortical temperature sensor is implanted in the subcutaneous tissue of the scapula, continuously recording temperature changes at a sampling frequency of 10Hz. A motion acceleration sensor is fixed to the chest harness, and three-dimensional acceleration data is transmitted to the processing unit in real time. During the data acquisition phase, the working dogs perform standardized exercise tasks, including various movement states such as resting, walking, and running. Temperature curve analysis identifies the time period corresponding to each respiratory cycle, and wavelet transform extracts frequency domain features related to cardiopulmonary function. Acceleration data integration calculates exercise energy consumption, and a feature cross-network fuses temperature features and energy consumption estimates to output a comprehensive physiological state index. During the network inference phase, streaming data is processed in real time, with a sliding window length of 30 seconds and a window overlap rate of 50%. Temperature data within each window is processed in segments, and acceleration data integration calculates the average energy consumption. The forward propagation time of the feature cross-network is completed within 10 milliseconds, meeting real-time requirements. The output features are input into a downstream health status assessment model to detect early signs of fatigue and the risk of overexertion. The system demonstrates good stability in actual deployment and can adapt to individual differences among different dog breeds and environmental changes. The feature crossover network employs a modular design, allowing network parameters to be fine-tuned for different animal species. An online learning mechanism enables the network to continuously optimize based on newly acquired data, adapting to long-term physiological changes. Model interpretability tools provide feature importance analysis, helping to understand the correlation between temperature changes and energy expenditure during exercise. System integration tests demonstrate that the multimodal features fused by the feature crossover network achieve higher health status prediction accuracy than single-modal features. The network architecture's scalability supports the integration of new sensor modalities simply by adding corresponding tower layers to the dual-tower structure and adjusting the attention matching mechanism. This design flexibility allows the system to adapt to evolving sensing technologies, maintaining long-term technological competitiveness. The core algorithm of the feature crossover network has been hardware-accelerated, running efficiently on embedded processors and meeting the low-power requirements of mobile monitoring devices. Real-world application data validates the network's robustness in handling individual differences and environmental interference, providing a reliable technological foundation for animal health monitoring.

[0090] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art 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 appended claims and their equivalents.

Claims

1. An animal individual health monitoring system based on real-time data, characterized in that, The system includes: a physiological data acquisition component, which synchronously acquires raw vital sign data streams of an individual animal through implanted biosensors and non-contact optical sensors, the raw vital sign data streams including cortical temperature change curves, microcirculatory blood flow spectra, and three-dimensional motion acceleration sequences; a multimodal fusion analysis component, which performs spatiotemporal alignment and feature-level fusion processing on the raw vital sign data streams to generate a comprehensive physiological state map of the individual animal, the comprehensive physiological state map including a metabolic equivalent distribution map, neural excitability index, and fatigue accumulation index; a health status assessment component, which uses a deep neural network model to perform multi-scale pattern recognition on the comprehensive physiological state map, outputting health risk probability values ​​and potential pathological region markers; a dynamic adaptive compensation component, which performs individualized calibration of the health risk probability values ​​based on a physiological constant database of the animal species, generating calibrated health risk levels; and a real-time feedback execution component, which generates a multi-channel health intervention instruction set based on the potential pathological region markers, the health intervention instruction set including nutritional supplementation strategies, environmental parameter adjustment schemes, and physical rehabilitation training plans.

2. The animal individual health monitoring system based on real-time data as described in claim 1, characterized in that, The specific steps of the feature-level fusion processing of the multimodal fusion analysis component include: dividing the cortical temperature change curve into multiple temperature segments according to the physiological cycle, with each temperature segment corresponding to a complete cardiopulmonary circulation cycle; performing the following operations for each cardiopulmonary circulation cycle: constructing a hemodynamic model of the capillary network using the microcirculation blood flow spectrum, the model including the blood flow velocity field and vascular wall elastic parameters; projecting the three-dimensional motion acceleration sequence onto the animal body coordinate system and calculating the kinematic feature vectors of each joint; and fusing the hemodynamic model and kinematic feature vectors through an attention mechanism to generate the metabolic equivalent distribution map, neural excitability index, and fatigue accumulation index for the current cycle.

3. The animal individual health monitoring system based on real-time data as described in claim 2, characterized in that, The process of the attention mechanism fusion includes: establishing a nonlinear mapping relationship between microcirculatory blood flow velocity and cortical temperature changes; extracting spatiotemporal features through a gated circulation unit; identifying motion pattern features and calculating energy consumption estimates based on the power spectral density distribution of three-dimensional motion acceleration; fusing spatiotemporal features and energy consumption estimates using a feature cross-network to generate pixel-level data of metabolic equivalent distribution maps; and generating numerical expressions of neural excitability indices and fatigue accumulation indices through a multilayer perceptron decoder.

4. The animal individual health monitoring system based on real-time data as described in claim 1, characterized in that, The method for multi-scale pattern recognition in the health status assessment component includes: inputting the metabolic equivalent distribution map into a convolutional neural network architecture to extract local tissue metabolic abnormality features; using a long short-term memory network to analyze the time series pattern of neural excitability indicators to detect the functional state of the central nervous system; modeling the propagation path of the fatigue accumulation index through a graph neural network to identify the load distribution of the musculoskeletal system; and fusing the local tissue metabolic abnormality features, the functional state of the central nervous system, and the load distribution of the musculoskeletal system to output a health risk probability value and potential pathological area markers.

5. The animal individual health monitoring system based on real-time data as described in claim 4, characterized in that, The process of individualized calibration of the dynamic adaptive compensation component includes: querying the standard physiological constants database of animal breeds to obtain the basal metabolic rate range and neural excitability threshold; calculating the deviation between the current metabolic equivalent distribution map and the basal metabolic rate range to generate a metabolic compensation coefficient; analyzing the fluctuation characteristics of neural excitability indicators relative to the threshold to generate a neural function regulation factor; and calibrating the health risk probability value by weighted fusion of the metabolic compensation coefficient and the neural function regulation factor.

6. The animal individual health monitoring system based on real-time data as described in claim 5, characterized in that, The generation of the metabolic compensation coefficient further includes: using a sliding window algorithm to analyze the spatiotemporal variation pattern of the metabolic equivalent distribution map and identify abnormal metabolic hotspots; calculating the degree of metabolic imbalance index based on the geometric characteristics and intensity distribution of the hotspots; normalizing the degree of metabolic imbalance index in conjunction with environmental temperature and humidity data to generate standardized metabolic compensation coefficients; and smoothing the metabolic compensation coefficients using a Kalman filter to eliminate the influence of measurement noise.

7. The animal individual health monitoring system based on real-time data as described in claim 1, characterized in that, The process by which the real-time feedback execution component generates a health intervention instruction set includes: analyzing the spatial distribution characteristics of potential pathological area markers and dividing the area into emergency intervention areas and routine monitoring areas; designing precise nutritional supplementation strategies for emergency intervention areas, including specific nutrient ratios and feeding time planning; adjusting environmental parameter regulation schemes based on animal behavior patterns and controlling light cycles and temperature and humidity gradients; and customizing physical rehabilitation training plans based on the results of motor ability assessments, setting training intensity and duration parameters.

8. The animal individual health monitoring system based on real-time data as described in claim 2, characterized in that, The specific method for dividing the cortical temperature change curve into multiple temperature segments according to the physiological cycle includes: identifying the extreme points in the cortical temperature change curve through a peak detection algorithm, defining the interval between adjacent extreme points as a basic temperature segment; performing wavelet transform on each basic temperature segment to extract the frequency domain features of the temperature waveform; and merging or segmenting the basic temperature segments according to the dominant frequency components in the frequency domain features to form standardized temperature segments that precisely match the cardiopulmonary circulation cycle.

9. The animal individual health monitoring system based on real-time data as described in claim 3, characterized in that, The process of fusing spatiotemporal features and energy consumption estimates in the feature cross-network further includes: constructing a dual-tower feature cross-network, wherein the first tower processes spatiotemporal features and the second tower processes energy consumption estimates; using residual connections and batch normalization layers in each tower to ensure training stability; and calculating the correlation weights between the output features of the two towers through an attention matching mechanism.

10. The animal individual health monitoring system based on real-time data as described in claim 7, characterized in that, The generation of the precision nutrition supplementation strategy further includes: establishing a nutrient-physiological state mapping database, which includes the required micronutrient ratios under different pathological states; using reinforcement learning algorithms to optimize feeding time planning, taking into account the animal's digestive cycle and activity patterns; setting up a feedback regulation mechanism to dynamically adjust the nutrition supplementation plan according to changes in physiological indicators after intervention, generating a multi-objective optimized supplementation strategy that simultaneously considers nutritional balance, palatability, and absorption efficiency.

Images