Captive panda oestrus starting prediction method and system based on machine learning

Abstract

The invention provides a captive panda oestrus starting prediction method and system based on machine learning, and relates to the technical field of panda breeding, and the method comprises the steps: obtaining original data; isomerous data stream synchronization is carried out according to the original data, and a synchronized spatio-temporal data matrix is obtained; performing nonlinear interaction relation modeling according to the synchronized spatio-temporal data matrix to construct a spatio-temporal dependency relation graph; performing environment constraint feature dimension reduction according to the space-time dependency graph to obtain a low-dimensional state representation vector; performing oestrus state transition probability calculation according to the low-dimensional state representation vector to generate a probability evolution trajectory; and according to the probability evolution trajectory, turning point identification is carried out, and an estrus starting moment prediction value is output. Through collaborative inflection point detection of progesterone gradient characteristics and an ingestion mutation mode in the probability evolution trajectory, an estrus starting moment prediction value is accurately output, and finally the biological interpretability of estrus critical state evolution of captive pandas is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of giant panda breeding, in particular to a method and system for predicting the estrus initiation of captive giant pandas based on machine learning. BACKGROUND

[0002] As an endangered species, the reproductive success rate of giant pandas is highly dependent on the accurate determination of the estrus initiation window. This species faces unique challenges such as endocrine rhythm disorders and frequent stereotyped behaviors in captive environments. Traditional prediction techniques mainly develop along two paths: one is the biochemical analysis method based on hormone detection, which establishes a fixed threshold model through a single physiological marker such as progesterone, but ignores the dynamic interference of behavior mode mutation on hormone metabolism caused by space compression; the other is the application of conventional time series algorithms to process feeding and drinking behavior data, which is difficult to capture the coordinated features of hormone gradient transition and sudden change in feeding mode within 48 hours before estrus due to the lack of cross-scale coupling mechanism between physiological indicators and behavior events. Especially in data processing, the existing methods have essential defects: the conflict between low-frequency hormone sampling (6-12 hours / time) and second-level behavior monitoring in time-frequency scale has not been effectively resolved, and the distortion effect of the cage physical boundary on biological signal propagation has never been quantitatively modeled, leading to the probability evolution trajectory being easily contaminated by exogenous noise such as feeding intervention in the turning point identification stage, ultimately restricting the prediction accuracy.

[0003] Based on the above-mentioned shortcomings of the prior art, there is an urgent need for a method and system for predicting the estrus initiation of captive giant pandas based on machine learning. SUMMARY

[0004] The purpose of the present application is to provide a method for predicting the estrus initiation of captive giant pandas based on machine learning to improve the above-mentioned problems. In order to achieve the above-mentioned purpose, the technical solutions adopted by the present application are as follows: In a first aspect, the present application provides a method for predicting the estrus initiation of captive giant pandas based on machine learning, comprising: obtaining raw data, the raw data including the concentration sequence of fecal progesterone metabolites, the change curve of unit food intake, and the instantaneous drinking frequency sequence of captive giant pandas; synchronizing the heterogeneous data streams according to the raw data, aligning the physiological and behavioral indicators of different sampling frequencies based on the dynamic time warping algorithm, and obtaining a synchronized spatiotemporal data matrix; modeling the nonlinear interaction relationship according to the synchronized spatiotemporal data matrix, learning the spatial dependence of behavior and physiological indicators within a preset time slice through a graph convolution network, and constructing a spatiotemporal dependence graph; performing environmental constraint feature dimension reduction according to the spatiotemporal dependence graph, and obtaining a low-dimensional state representation vector; According to the low-dimensional state representation vector, estrus state transition probability calculation is performed, transition probability from the current state to the estrus starting state is derived, and a probability evolution trajectory is generated; According to the probability evolution trajectory, a turning point is identified, an extreme point of a first-order derivative of probability is detected by using a kernel density estimation function, and an estrus starting time prediction value is output.

[0005] In a second aspect, the application further provides a captive giant panda estrus starting prediction system based on machine learning, comprising: An acquisition module is configured to acquire original data, wherein the original data comprises a fecal progesterone metabolite concentration sequence, a unit food intake change curve, and an instantaneous drinking frequency sequence of a captive giant panda; A synchronization module is configured to synchronize heterogeneous data streams according to the original data, align physiological and behavioral indicators of different sampling frequencies based on a dynamic time warping algorithm, and obtain a synchronized spatiotemporal data matrix; A modeling module is configured to model nonlinear interaction relationships according to the synchronized spatiotemporal data matrix, learn spatial dependencies of behavioral and physiological indicators within a preset time slice by using a graph convolution network, and construct a spatiotemporal dependency graph. A dimension reduction module is configured to perform environment constraint feature dimension reduction according to the spatiotemporal dependency graph, and obtain a low-dimensional state representation vector; A calculation module is configured to perform estrus state transition probability calculation according to the low-dimensional state representation vector, derive transition probability from a current state to an estrus starting state, and generate a probability evolution trajectory; A prediction module is configured to identify a turning point according to the probability evolution trajectory, detect an extreme point of a first-order derivative of probability by using a kernel density estimation function, and output an estrus starting time prediction value.

[0006] The application has the following beneficial effects: The application synchronizes cross-scale data streams of the fecal progesterone metabolite concentration sequence, the unit food intake change curve, and the instantaneous drinking frequency sequence, realizes dynamic coupling of hormone rhythm and behavioral events, quantifies environment constraints based on a spatiotemporal dependency graph, significantly enhances signal robustness under artificial noise interference, detects a cooperative inflection point of progesterone gradient features and feeding mutation patterns in a probability evolution trajectory, accurately outputs an estrus starting time prediction value, and finally improves biological explainability of evolution of a captive giant panda estrus critical state. BRIEF DESCRIPTION OF DRAWINGS

[0007] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments. It should be understood that the following drawings only show some of the embodiments of the present application, and therefore should not be considered as limiting the scope. For those skilled in the art, other related drawings can also be obtained without creative labor.

[0008] Figure 1 A flow chart of a machine learning-based captive giant panda estrus initiation prediction method described in the embodiments of the present application; Figure 2 A structure diagram of a machine learning-based captive giant panda estrus initiation prediction system described in the embodiments of the present application; Figure 3 A structure diagram of a machine learning-based captive giant panda estrus initiation prediction device described in the embodiments of the present application.

[0009] In the figure, 800 is a machine learning-based captive giant panda estrus initiation prediction device; 801 is a processor; 802 is a memory; 803 is a multimedia component; 804 is an I / O interface; 805 is a communication component; 901 is an acquisition module; 902 is a synchronization module; 903 is a modeling module; 904 is a dimension reduction module; 905 is a calculation module; and 906 is a prediction module. DETAILED DESCRIPTION

[0010] In order to make the objectives, technical solutions and advantages of the embodiments of the present application clearer, the following will combine the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are some of the embodiments of the present application, but not all the embodiments. The components of the embodiments of the present application described and shown in the drawings can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the claimed present application, but only represents selected embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0011] It should be noted that: similar labels and letters represent similar items in the following drawings, therefore, once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings. Meanwhile, in the description of the present application, the terms "first", "second", etc. are only used to distinguish the description, and cannot be understood as indicating or implying relative importance.

[0012] Embodiment 1

[0013] This embodiment provides a machine learning-based method for predicting the onset of estrus in captivity in giant pandas.

[0014] See Figure 1 The figure shows that the method includes steps S100 to S600.

[0015] Step S100: Obtain raw data, which includes fecal progesterone metabolite concentration sequence, unit food intake change curve, and instantaneous water drinking frequency sequence of captive giant pandas. In real-world scenarios, data collection strictly adheres to the biological characteristics and breeding operation standards of captive giant pandas. The fecal progesterone metabolite concentration sequence is obtained through a standardized sample collection process: keepers collect fresh fecal samples regularly, which are then cryopreserved and analyzed in the laboratory using enzyme-linked immunosorbent assay (ELISA) to generate a time-stamped concentration sequence. The unit feed intake variation curve is automatically generated by the intelligent feeding system, using an electronic scale to measure the weight difference before and after each feeding in real time. The cumulative feed intake is recorded every 5 minutes, and the variation curve is output with minute-level precision through differential calculation. The instantaneous water drinking frequency sequence is collected using a micro-pressure sensor array in the cage's water trough. By monitoring changes in the liquid level and identifying licking events indicating a drop in liquid level, the number of effective drinking actions per minute is counted to generate a frequency sequence.

[0016] Step S200: Synchronize heterogeneous data streams based on the original data, and align physiological and behavioral indicators with different sampling frequencies based on the dynamic time warping algorithm to obtain a synchronized spatiotemporal data matrix. Understandably, to address the time-frequency mismatch between low-frequency hormone sampling (6 hours / time) and high-frequency behavioral monitoring (second-level), this step employs a dynamic time warping algorithm to construct a nonlinear timescale mapping: the hormone sequence is non-uniformly stretched based on the sudden point of drinking events, while simultaneously preserving the physiological rising edge characteristics of progesterone pulses; the second derivative of the feed intake curve is used as a dynamic constraint (quantifying the acceleration change of feeding behavior) to ensure that the time anchoring of key biological events does not drift due to noise in the captive environment.

[0017] Step S300: Model nonlinear interaction relationships based on the synchronized spatiotemporal data matrix, and learn the spatial dependence of behavior and physiological indicators within a preset time slice through graph convolutional network to construct a spatiotemporal dependency relationship map. It should be noted that this step defines the adjacency matrix weights of the graph convolutional network as the spatial propagation function of the rate of change in food intake, so that the differences in behavioral patterns among individuals in the same cage can be transformed into learnable topological dependencies. By constructing the tensor product of the fecal progesterone gradient vector and the frequency mutation of drinking water, the asynchronous association between hormones and behaviors is encoded within a preset time slice, fundamentally solving the problem of the failure of traditional models to model the time-delay coupling effect.

[0018] Step S400: Perform dimensionality reduction of environmental constraint features based on the spatiotemporal dependency graph to obtain a low-dimensional state representation vector; Understandably, this step models the physical boundaries of the cages (such as the spacing between cages) as Dirichlet boundary conditions in the hormone diffusion equation, constructing a Riemannian manifold with spatial compression effect; and uses variational geometric constraints to inject information entropy into the curvature adjustment of the manifold, so that the dimensionality reduction process actively resists the signal distortion caused by the captive environment and retains the weak state transition characteristics before estrus.

[0019] Step S500: Calculate the estrus state transition probability based on the low-dimensional state representation vector, and generate a probability evolution trajectory by deriving the transition probability from the current state to the estrus initiation state. Preferably, this step uses the set of pre-estrus states decomposed by continuous homology theory as discrete nodes, and designs a time modulation factor for the state transition kernel function to automatically enhance the critical transition probability during the breeding season. The Falk-Planck equation is used to simulate the directional diffusion of the state vector in the progesterone concentration potential field, generating a probabilistic evolution trajectory consistent with reproductive endocrine dynamics.

[0020] Step S600: Identify turning points based on the probability evolution trajectory, use the kernel density estimation function to detect the extreme points of the first derivative of the probability, and output the predicted value of the estrus initiation time.

[0021] Understandably, this step constructs the Hamilton-Jacobi equation boundary conditions based on the progesterone metabolism rate threshold, and uses the Lie derivative tensor of the first derivative of the probability density to capture trajectory curvature abrupt changes in the joint time-frequency domain, making the inflection point identification process essentially immune to exogenous disturbances such as nutritional intervention, and ultimately achieving physical interpretability anchoring of the estrus initiation moment.

[0022] Further, step S200 includes steps S210 to S230.

[0023] Step S210: Perform dynamic stretching of the time axis based on the original data. By establishing a nonlinear mapping function between the fecal progesterone metabolite concentration sequence and the instantaneous drinking water frequency sequence, a dual-channel time-aligned basis vector is generated. Understandably, to address the time-domain mismatch between low-frequency fecal progesterone sequences and high-frequency sequences from instantaneous water drinking, this step establishes an exponential nonlinear mapping function based on biodynamic metabolic characteristics. This function dynamically stretches the time axis of the progesterone sequence through sudden increases in water drinking frequency. The key consideration is compensating for the physiological lag in hormone detection. Sudden changes in drinking behavior, as an early marker of neural excitation, use their time anchor point to drive non-uniform scaling of the progesterone data window, automatically aligning the peak of the behavioral event with the rising edge of the hormone, while avoiding metabolic rate distortion caused by linear interpolation. The dynamic time-axis stretching function is as follows: ; in, t0 represents the new time coordinate after stretching; ρ represents the original time coordinate; g(•) represents the response function for sudden changes in drinking water volume; Θ(•) represents the progesterone gradient constraint function. ξ represents the instantaneous change in drinking frequency; ξ represents the bandwidth control parameter of the drinking response function; P(t0) represents the fecal progesterone metabolite concentration at time t0.

[0024] Step S220: Construct the feature distance matrix based on the dual-channel time-aligned basis vectors, and calculate the minimum path loss values ​​of physiological and behavioral indicators by using the second derivative of the unit food intake change curve as a dynamic regularization constraint. It should be noted that the rate of change in the acceleration of food intake (i.e., the rate of food intake decay) in captive giant pandas is the most stable stereotyped behavioral marker during proestrus. This derivative serves as a weight regulator for path loss calculation; that is, when an acceleration in food intake decay is detected, the constraint strength is increased, forcing the synchronization path to bypass periods of external interference such as artificial feeding; during periods of behavioral stability, the constraint is reduced to preserve the intrinsic association between hormones and water intake. This step injects biological prior knowledge into cross-modal alignment through the dynamic characteristics of feeding behavior.

[0025] Step S230: Perform topology optimization based on minimizing path loss value, and output a synchronized spatiotemporal data matrix with physical timescale consistency by removing pseudo-synchronization points caused by interference from the captive environment.

[0026] To address pseudo-synchronization points caused by environmental disturbances specific to captive rearing (such as handler movement and cage washing), this step employs algebraic topology theory for signal purification. The focus is on analyzing the lifecycle of the homology structure in the synchronization point cloud: only low-dimensional homology classes lasting longer than two hours (corresponding to genuine biological rhythms) are retained, while short-lifecycle clusters (typically less than 30 minutes) generated by rearing operations are filtered out as high-dimensional noise. This method transforms physical timescale consistency into a problem of the continuous stability of biological signals in topological space, ensuring that the output matrix simultaneously satisfies clock timing accuracy and intrinsic rhythmic synchronization.

[0027] Further, step S300 includes steps S310 to S330.

[0028] Step S310: Perform cross-modal feature fusion processing based on the synchronized spatiotemporal data matrix. Generate a spatiotemporal fusion feature matrix by constructing an interaction tensor between the fecal progesterone concentration gradient vector and the frequency mutation of drinking water. Preferably, this step calculates the differential gradient between adjacent sampling points based on the fecal progesterone concentration sequence obtained by enzyme-linked immunosorbent assay (ELISA) to capture changes in hormone metabolic rate; simultaneously, it extracts mutations exceeding three standard deviations from the baseline in the instantaneous water intake frequency sequence. A high-order tensor product is then performed on both to establish a non-Euclidean spatial correlation between the sudden onset of the water intake event and the physiological response of the progesterone rise. By preserving the asynchronous coupling strength of hormone-behavioral signals through tensor structures (such as the 48-hour delay effect of a sudden increase in water intake on progesterone secretion), the problem of time-delayed correlation breaks caused by traditional feature splicing is resolved.

[0029] Step S320: Perform dynamic neighborhood topology construction processing based on the spatiotemporal fusion feature matrix, and establish an iterative update mechanism for node attributes and edge weights with the rate of change of food intake as the spatial dependency weight to generate a spatiotemporal heterogeneous graph. Specifically, this step utilizes the instantaneous derivative of the unit feed intake change curve to quantify the acceleration of feeding behavior. When an acceleration in the feeding decay rate is detected (second derivative > 0.1), the topological connection weights between individuals in the same cage are automatically strengthened. Preferably, an edge weight iterative update mechanism is designed: every 15 minutes (the preset time slice in the claim), the graph adjacency matrix is ​​reconstructed based on the latest feeding rate, so that nodes with abrupt changes in drinking frequency dynamically aggregate towards nodes with feeding collapse.

[0030] Step S330: Perform spectral domain convolution processing on the spatiotemporal heterogeneous graph, learn the cross-scale dependency patterns in the graph Laplacian matrix through deep spatiotemporal operators, and output a spatiotemporal dependency graph with spectral clustering optimization characteristics.

[0031] Understandably, this step embeds a spatiotemporal migration kernel into standard spectral domain convolution, approximating the signal attenuation function caused by the physical barriers of the cages using Chebyshev polynomials. Specifically, it achieves cross-scale dependency pattern separation—low-order polynomials capture short-range hormone synchronization between adjacent individuals within the cage (e.g., progesterone fluctuation correlation within 7 meters), while higher-order terms learn long-range resonances in drinking behavior across cage areas (e.g., pacing throughout the cage during specific periods of estrus). The output spatiotemporal dependency graph exhibits spectral clustering optimization characteristics; specifically, when the graph cut modularity exceeds a set threshold, it automatically merges node clusters with similar progesterone gradients, accurately isolating spurious dependency clusters caused by artificial feeding.

[0032] Further, step S400 includes steps S410 to S430.

[0033] Step S410: Perform captive physical constraint modeling based on the spatiotemporal dependency graph, and generate environmental embedding basis functions by constructing a coupled field between the cage space topology and the hormone diffusion equation. Understandably, this step establishes a three-dimensional spatial feature vector field by analyzing the geometric layout parameters of the stainless steel cage (preferably such as grid spacing and cage corner curvature); and couples this field with the diffusion dynamics equation of fecal progesterone metabolism in the intestine, forming an environmental constraint embedding function. The biological basis lies in the fact that the path folding effect of the cage's physical boundaries on the stereotypical pacing behavior of giant pandas distorts the propagation pattern of hormone fluctuation signals. This coupled field mathematically encodes the signal attenuation law caused by spatial compression, providing a physical constraint basis for noise separation. The coupled field model of the cage spatial topology and the hormone diffusion equation is as follows: ; in, Represents the environment embedding basis functions; Indicates the spatial coordinates of the cages; Represents spatial topological constraint functions; Represents the spatial hormone concentration field; The Laplace operator represents hormone concentration; λ represents the metabolic decay coefficient. Step S420: Perform mode decoupling processing based on the environmental embedding basis function, and separate the intrinsic mode components of estrus-related hormone fluctuations and environmental noise through the spectral embedding algorithm to obtain the interference-free feature manifold; It should be noted that this step uses the Fourier-Bessel spectral embedding algorithm to decompose the intrinsic mode components. Specifically, a cylindrical coordinate system is constructed based on the environmental embedding basis function (the axis corresponds to the feeding point in the cage), and the eigenvalues ​​of the Bessel function are solved to achieve radial decomposition of the signal. A dual-band stop filter is used to block the high-frequency noise band of 0.8–1.2 Hz (corresponding to the sound of the zookeeper's footsteps and fluctuations in the washing water pressure) and amplify the low-frequency biological rhythm band of 0.05–0.2 Hz (the dominant frequency of estrus-related progesterone fluctuations).

[0034] Step S430: Perform variational geometric constraint processing based on the interference-removing feature manifold, and achieve adaptive curvature compression of the state space in the Riemannian manifold space using information entropy as a metric, outputting a low-dimensional state representation vector with biological interpretability.

[0035] An information entropy gradient field is constructed on the interference-free feature manifold obtained in step S420. When the slope of the entropy change exceeds the threshold, the local curvature of the manifold is automatically increased to preserve the abrupt state change features; while in the steady state, the flatness metric is maintained to compress redundant oscillations. Through geometric invariance, the feeding disturbance is transformed into curvature parameters, so that the output low-dimensional vector carries key biological features such as the progesterone gradient direction (the estrus initiation indicator) while removing environmental noise such as cage door opening and closing and flushing water flow.

[0036] Further, step S500 includes steps S510 to S530.

[0037] Step S510: Discretize the hormone-behavioral state based on the low-dimensional state representation vector, and decompose the continuous state manifold into a set of biologically meaningful estrus precursor states using the continuous homology theory. Understandably, this step utilizes persistent homology theory to analyze the topologically invariant characteristics of low-dimensional manifolds. By calculating the Betty number and lifespan of the state vector, only stable connected components lasting longer than 24 hours (corresponding to typical stages of the estrous cycle) are retained. The biological mechanism lies in capturing the co-occurrence pattern of the slow rise in progesterone and the sudden drop in water intake, decomposing the continuous hormone-behavioral state space into a set of discrete state nodes with clear physiological significance. This set includes "hormonal preparatory states" and "critical transition states."

[0038] Step S520: Based on the set of pre-estrus states, construct the transition probability tensor and establish a dynamic adjustment mechanism for the state transition kernel function based on the time non-homogeneous characteristics of the reproductive endocrine rhythm of captive giant pandas. It should be noted that this step uses an endocrine rhythm function as a modulator to automatically adjust the state transition probability according to the reproductive cycle. The core lies in constructing a fourth-order transition probability tensor: the row and column dimensions correspond to the predecessor state nodes, the depth dimension encodes the spatial location of the cages, and the time dimension binds to the 24-hour light cycle (increasing the state transition probability by 22% at night). This design quantifies the physical constraints of the captive environment (cage spacing) and biological rhythms (season / day / night) into endogenous parameters of the probability tensor, avoiding the static transition defects of traditional Markov chains. Specifically, the fourth-order tensor formula is: ; in, Let be a time-nonhomogeneous transition probability tensor, representing the probability of transitioning from state i to state j given time t and spatial location k to l; i and j are the indices of the pre-estrus state, i representing the initial state and j representing the target state; k and l are the spatial location indices of the cage. It is an endocrine rhythm modulator; Let be the state-space distance function, representing Similarity between physiological states; β is the state transition scale parameter; Let be a spatial constraint function, representing the state transition barrier effect caused by the physical structure of the cage (such as a fence), where Let λ(t) represent the physical distance from position k to l, and let λ(t) represent the diurnal rhythm function.

[0039] Step S530: Perform multi-scale trajectory generation processing according to the dynamic adjustment mechanism. Simulate the probability diffusion process of the state vector in the hormone potential field by solving the Fock-Planck equation, and output the probability evolution trajectory with timestamp.

[0040] Based on the hormonal potential-driven drift mechanism and the diffusion effect of behavioral perturbations, the specific equations are as follows: ; in, It is represented by a probability density function, which represents the probability distribution of state S at time t, and describes the evolution of the system from the "hormone preparatory state" to the "estrus critical state". denoted as the drift coefficient vector; S is a low-dimensional state representation vector containing the projections of progesterone-related features and feeding behavior features into the reduced-dimensional space; The diffusion coefficient tensor is dominated by behavioral perturbations; t represents the time index. Here, is the divergence operator, representing the directional aggregation / dissipation rate of the probability density driven by the progesterone potential field.

[0041] Further, step S600 includes steps S610 to S630.

[0042] Step S610: Perform multi-scale curvature analysis based on the probability evolution trajectory, extract the Lie derivative feature tensor of the trajectory function in the time-frequency joint domain, and construct the differential geometric invariants of the probability manifold; Preferably, the probabilistic trajectory function is decomposed using wavelet multi-resolution decomposition. This step extracts curvature tensor components at the hourly (behavioral mutation scale) and minute (hormonal response scale) levels, respectively. The translation invariance of the trajectory tangent bundle on the manifold is analyzed using the Lie derivative operator, and noise-resistant differential geometric invariants are constructed. The biological essence of this design is that pseudo-turning points in captivity (such as probability perturbations caused by artificial feeding) exhibit high-frequency curvature fluctuation characteristics (>0.5Hz) in the time-frequency domain, while the true estrus signal exhibits a stable curvature mutation pattern across scales (0.05-0.2Hz main frequency band). Noise is essentially isolated through differential invariants.

[0043] Step S620: Perform biological mutation marker detection processing based on differential geometric invariants, and solve the Hamilton-Jacobi equation of probability density with the physiological critical value of progesterone metabolism rate as the boundary condition. Understandably, when the rate of change in progesterone exceeds a critical threshold, the driving equation forms a saddle point solution in phase space that characterizes the estrus transition. This step incorporates endocrine dynamics constraints into the partial differential equation solution process, ensuring that mutation detection is strictly controlled within the reproductive physiology of captive giant pandas.

[0044] Step S630: Perform phase transition time anchoring processing based on the solution of the Hamilton-Jacobi equation, and output the predicted value of the estrus initiation time by analyzing the time convergence point of the strange attractor in the characteristic velocity field.

[0045] It should be noted that this step is based on the dynamic convergence characteristics of the strange attractor in the characteristic velocity field: when the state trajectory enters the attraction domain defined by the progesterone concentration and drinking frequency thresholds, the time point of the first boundary crossing is locked by inverse time integration. This process solves the problem of instantaneous trajectory drift caused by artificial feeding, ensuring that the predicted value of the estrus initiation time simultaneously meets the biological critical condition and the dynamic convergence criterion.

[0046] Example 2

[0047] like Figure 2 As shown, this embodiment provides a machine learning-based estrus initiation prediction system for captive giant pandas. The system includes: The acquisition module is used to acquire raw data, which includes fecal progesterone metabolite concentration sequences, unit food intake change curves, and instantaneous water drinking frequency sequences of captive giant pandas. The synchronization module is used to synchronize heterogeneous data streams based on the original data. It aligns physiological and behavioral indicators with different sampling frequencies based on the dynamic time warping algorithm to obtain a synchronized spatiotemporal data matrix. The modeling module is used to model nonlinear interaction relationships based on the synchronized spatiotemporal data matrix. It learns the spatial dependence of behavior and physiological indicators within a preset time slice through graph convolutional networks and constructs a spatiotemporal dependency graph. The dimensionality reduction module is used to reduce the dimensionality of environmental constraint features based on the spatiotemporal dependency graph to obtain a low-dimensional state representation vector. The calculation module is used to calculate the estrus state transition probability based on the low-dimensional state representation vector, and generate the probability evolution trajectory by deriving the transition probability from the current state to the estrus initiation state. The prediction module is used to identify turning points based on the probability evolution trajectory, and uses the kernel density estimation function to detect the extreme points of the first derivative of the probability, and outputs the predicted value of the estrus initiation time.

[0048] In one specific embodiment disclosed in this application, the synchronization module 902 includes: The first synchronization unit is used to perform dynamic stretching of the time axis based on the original data. It generates a dual-channel time-aligned basis vector by establishing a nonlinear mapping function between the fecal progesterone metabolite concentration sequence and the instantaneous drinking water frequency sequence. The second synchronization unit is used to construct the feature distance matrix based on the dual-channel time-aligned basis vectors. By using the second derivative of the unit food intake change curve as a dynamic regularization constraint, the minimum path loss value of physiological and behavioral indicators is calculated. The third synchronization unit is used to perform topology optimization based on minimizing path loss. By removing pseudo-synchronization points caused by interference from the captive environment, it outputs a synchronized spatiotemporal data matrix with physical timescale consistency.

[0049] In one specific embodiment disclosed in this application, the modeling module 903 includes: The first modeling unit is used to perform cross-modal feature fusion processing based on the synchronized spatiotemporal data matrix. It generates a spatiotemporal fusion feature matrix by constructing an interaction tensor between the fecal progesterone concentration gradient vector and the frequency mutation of drinking water. The second modeling unit is used to perform dynamic neighborhood topology construction based on the spatiotemporal fusion feature matrix. It uses the rate of change in food intake as the spatial dependency weight to establish an iterative update mechanism for node attributes and edge weights, and generates a spatiotemporal heterogeneous graph. The third modeling unit is used to perform spectral domain convolution processing based on the spatiotemporal heterogeneous graph. It learns cross-scale dependency patterns in the graph Laplacian matrix through deep spatiotemporal operators and outputs a spatiotemporal dependency graph with spectral clustering optimization characteristics.

[0050] In one specific embodiment disclosed in this application, the dimensionality reduction module 904 includes: The first dimensionality reduction unit is used to model the physical constraints of captive rearing based on the spatiotemporal dependency map. By constructing a coupled field between the cage space topology and the hormone diffusion equation, environmental embedding basis functions are generated. The second dimensionality reduction unit is used to perform mode decoupling processing based on the environmental embedding basis function. The intrinsic mode components of estrus-related hormone fluctuations and environmental noise are separated by the spectral embedding algorithm to obtain the interference-free feature manifold. The third dimensionality reduction unit is used to perform variational geometric constraint processing based on the interference-free feature manifold. In the Riemannian manifold space, the curvature of the state space is adaptively compressed using information entropy as a metric, and the output is a low-dimensional state representation vector with biological interpretability.

[0051] In one specific embodiment disclosed in this application, the computing module 905 includes: The first computational unit is used to discretize hormone-behavioral states based on low-dimensional state representation vectors and decompose the continuous state manifold into a set of biologically meaningful estrus precursor states using the theory of continuous homology. The second computational unit is used to construct the transition probability tensor based on the set of pre-estrus states. Based on the time non-homogeneous characteristics of the reproductive endocrine rhythm of captive giant pandas, a dynamic adjustment mechanism for the state transition kernel function is established. The third computational unit is used to perform multi-scale trajectory generation processing based on the dynamic adjustment mechanism. It simulates the probability diffusion process of the state vector in the hormone potential field by solving the Falk-Planck equation and outputs the probability evolution trajectory with timestamp.

[0052] In one specific embodiment disclosed in this application, the prediction module 906 includes: The first prediction unit is used to perform multi-scale curvature analysis based on the probability evolution trajectory, extract the Lie derivative feature tensor of the trajectory function in the time-frequency joint domain, and construct the differential geometric invariants of the probability manifold. The second prediction unit is used to perform biological mutation marker detection processing based on differential geometric invariants, and solves the Hamilton-Jacobi equation of probability density by using the physiological critical value of progesterone metabolism rate as the boundary condition. The third prediction unit is used to perform phase transition time anchoring processing based on the solution of the Hamilton-Jacobi equation, and outputs the predicted value of the estrus initiation time by analyzing the time convergence point of the strange attractor in the characteristic velocity field.

[0053] Example 3

[0054] Corresponding to the above method embodiments, this embodiment also provides a machine learning-based estrus initiation prediction device for captive giant pandas. The machine learning-based estrus initiation prediction device for captive giant pandas described below can be referred to in conjunction with the machine learning-based estrus initiation prediction method for captive giant pandas described above.

[0055] Figure 3 This is a block diagram illustrating a machine learning-based estrus initiation prediction device 800 for captive giant pandas, according to an exemplary embodiment. Figure 3 As shown, the machine learning-based estrus initiation prediction device 800 for captive giant pandas may include: a processor 801 and a memory 802. The machine learning-based estrus initiation prediction device 800 may also include one or more of the following: a multimedia component 803, an I / O interface 804, and a communication component 805.

[0056] The processor 801 controls the overall operation of the machine learning-based captive giant panda estrus initiation prediction device 800 to complete all or part of the steps in the aforementioned machine learning-based captive giant panda estrus initiation prediction method. The memory 802 stores various types of data to support the operation of the machine learning-based captive giant panda estrus initiation prediction device 800. This data may include, for example, instructions for any application or method operating on the machine learning-based captive giant panda estrus initiation prediction device 800, as well as application-related data, such as contact data, sent and received messages, images, audio, video, etc. The memory 802 can be implemented using any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The multimedia component 803 may include a screen and an audio component. The screen may be, for example, a touchscreen, and the audio component is used to output and / or input audio signals. For example, the audio component may include a microphone for receiving external audio signals. The received audio signals may be further stored in the memory 802 or transmitted via the communication component 805. The audio component also includes at least one speaker for outputting audio signals. I / O interface 804 provides an interface between processor 801 and other interface modules, such as keyboards, mice, and buttons. These buttons can be virtual or physical. Communication component 805 is used for wired or wireless communication between the machine learning-based captive giant panda estrus initiation prediction device 800 and other devices. Wireless communication methods include Wi-Fi, Bluetooth, Near Field Communication (NFC), 2G, 3G, or 4G, or a combination thereof. Therefore, the corresponding communication component 805 may include a Wi-Fi module, a Bluetooth module, and an NFC module.

[0057] In an exemplary embodiment, a machine learning-based estrus initiation prediction device 800 for captive giant pandas can be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to execute the aforementioned machine learning-based estrus initiation prediction method for captive giant pandas.

[0058] In another exemplary embodiment, a computer-readable storage medium including program instructions is also provided. When executed by a processor, these program instructions implement the steps of the aforementioned machine learning-based method for predicting the estrus initiation of captive giant pandas. For example, the computer-readable storage medium may be the aforementioned memory 802 including program instructions, which may be executed by a processor 801 of a machine learning-based device 800 for predicting the estrus initiation of captive giant pandas to complete the aforementioned machine learning-based method for predicting the estrus initiation of captive giant pandas.

[0059] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A machine learning-based method for predicting estrus initiation in captive giant pandas, characterized in that, include: Obtain raw data, including fecal progesterone metabolite concentration sequences, unit food intake variation curves, and instantaneous water drinking frequency sequences of captive giant pandas; Based on the original data, heterogeneous data streams are synchronized, and physiological and behavioral indicators with different sampling frequencies are aligned using a dynamic time warping algorithm to obtain a synchronized spatiotemporal data matrix. Based on the synchronized spatiotemporal data matrix, a nonlinear interaction relationship model is performed. The spatial dependence of behavior and physiological indicators within a preset time slice is learned through a graph convolutional network, and a spatiotemporal dependency relationship map is constructed. Based on the spatiotemporal dependency graph, the environmental constraint features are reduced in dimensionality to obtain a low-dimensional state representation vector. The estrus state transition probability is calculated based on the low-dimensional state representation vector. By deriving the transition probability from the current state to the estrus initiation state, a probability evolution trajectory is generated. Turning points are identified based on the probability evolution trajectory, and the extreme points of the first derivative of the probability are detected using the kernel density estimation function. The predicted value of the estrus initiation time is then output.

2. The machine learning-based method for predicting the estrus initiation in captive giant pandas according to claim 1, characterized in that, Based on the original data, heterogeneous data streams are synchronized. Physiological and behavioral indicators with different sampling frequencies are aligned using a dynamic time warping algorithm to obtain a synchronized spatiotemporal data matrix, including: Based on the original data, dynamic stretching of the time axis is performed, and a nonlinear mapping function between the fecal progesterone metabolite concentration sequence and the instantaneous drinking water frequency sequence is established to generate a dual-channel time-aligned basis vector. Based on the dual-channel time-aligned basis vectors, a feature distance matrix is ​​constructed. By using the second derivative of the unit food intake change curve as a dynamic normalization constraint, the minimum path loss values ​​of physiological and behavioral indicators are calculated. Based on the minimized path loss value, topology optimization is performed, and pseudo-synchronization points caused by captive environment interference are removed, outputting a synchronized spatiotemporal data matrix with physical timescale consistency.

3. The method for predicting the estrus initiation of captive giant pandas based on machine learning according to claim 1, characterized in that, Based on the synchronized spatiotemporal data matrix, a nonlinear interaction relationship model is performed. A graph convolutional network is used to learn the spatial dependence of behavior and physiological indicators within a preset time slice, constructing a spatiotemporal dependency graph, including: Based on the synchronized spatiotemporal data matrix, cross-modal feature fusion processing is performed, and a spatiotemporal fusion feature matrix is ​​generated by constructing an interaction tensor between the fecal progesterone concentration gradient vector and the frequency mutation of drinking water. Based on the spatiotemporal fusion feature matrix, dynamic neighborhood topology construction is performed. Using the rate of change in food intake as the spatial dependency weight, an iterative update mechanism for node attributes and edge weights is established to generate a spatiotemporal heterogeneous graph. Based on the spatiotemporal heterogeneous graph, spectral domain convolution processing is performed, and cross-scale dependency patterns in the graph Laplacian matrix are learned through deep spatiotemporal operators, outputting a spatiotemporal dependency graph with spectral clustering optimization characteristics.

4. The machine learning-based method for predicting the onset of estrus in captivity in giant pandas according to claim 1, characterized in that, Based on the spatiotemporal dependency graph, environmental constraint features are reduced in dimensionality to obtain a low-dimensional state representation vector, including: Based on the spatiotemporal dependency map, the physical constraints of captive rearing are modeled, and an environmental embedding basis function is generated by constructing a coupled field between the cage space topology and the hormone diffusion equation. Based on the environmental embedding basis function, mode decoupling processing is performed, and the intrinsic mode components of estrus-related hormone fluctuations and environmental noise are separated by the spectral embedding algorithm to obtain the interference-free feature manifold; Variational geometric constraints are applied to the de-interference characteristic manifold to achieve adaptive curvature compression of the state space in the Riemannian manifold space using information entropy as a metric, outputting a low-dimensional state representation vector with biological interpretability.

5. The machine learning-based method for predicting the onset of estrus in captivity in giant pandas according to claim 1, characterized in that, The estrus state transition probability is calculated based on the low-dimensional state representation vector. By deriving the transition probability from the current state to the estrus initiation state, a probability evolution trajectory is generated, including: Based on the low-dimensional state representation vector, the hormone-behavioral state is discretized, and the continuous state manifold is decomposed into a set of biologically meaningful estrus precursor states using the continuous homology theory. Based on the set of pre-estrus states, a transition probability tensor is constructed. Based on the temporal non-homogeneous characteristics of the reproductive endocrine rhythm of captive giant pandas, a dynamic adjustment mechanism for the state transition kernel function is established. Multi-scale trajectory generation is performed based on the dynamic adjustment mechanism. The probability diffusion process of the state vector in the hormone potential field is simulated by solving the Falk-Planck equation, and the probabilistic evolution trajectory with timestamp is output.

6. A machine learning-based system for predicting the onset of estrus in captivity in giant pandas, characterized in that, include: The acquisition module is used to acquire raw data, which includes the fecal progesterone metabolite concentration sequence, unit food intake change curve, and instantaneous water drinking frequency sequence of captive giant pandas. The synchronization module is used to synchronize heterogeneous data streams based on the original data, and align physiological and behavioral indicators with different sampling frequencies based on the dynamic time warping algorithm to obtain a synchronized spatiotemporal data matrix. The modeling module is used to model nonlinear interaction relationships based on the synchronized spatiotemporal data matrix, and to learn the spatial dependence of behavior and physiological indicators within a preset time slice through a graph convolutional network to construct a spatiotemporal dependency graph. The dimensionality reduction module is used to reduce the dimensionality of environmental constraint features based on the spatiotemporal dependency graph to obtain a low-dimensional state representation vector. The calculation module is used to calculate the estrus state transition probability based on the low-dimensional state representation vector, and generate a probability evolution trajectory by deriving the transition probability from the current state to the estrus initiation state. The prediction module is used to identify turning points based on the probability evolution trajectory, detect the extreme points of the first derivative of the probability using a kernel density estimation function, and output the predicted value of the estrus initiation time.

7. The machine learning-based estrus initiation prediction system for captive giant pandas according to claim 6, characterized in that, The synchronization module includes: The first synchronization unit is used to perform dynamic time-axis stretching processing based on the original data and generate a dual-channel time-aligned basis vector by establishing a nonlinear mapping function between the fecal progesterone metabolite concentration sequence and the instantaneous drinking frequency sequence. The second synchronization unit is used to construct a feature distance matrix based on the dual-channel time alignment basis vector, and calculate the minimum path loss value of physiological and behavioral indicators by using the second derivative of the unit food intake change curve as a dynamic normalization constraint. The third synchronization unit is used to perform topology optimization processing based on the minimized path loss value, and outputs a synchronized spatiotemporal data matrix with physical timescale consistency by removing pseudo-synchronization points caused by interference from the captive environment.

8. The machine learning-based estrus initiation prediction system for captive giant pandas according to claim 6, characterized in that, The modeling module includes: The first modeling unit is used to perform cross-modal feature fusion processing based on the synchronized spatiotemporal data matrix, and to generate a spatiotemporal fusion feature matrix by constructing an interaction tensor between the fecal progesterone concentration gradient vector and the frequency mutation of drinking water. The second modeling unit is used to perform dynamic neighborhood topology construction processing based on the spatiotemporal fusion feature matrix, and to establish an iterative update mechanism for node attributes and edge weights by using the rate of change of food intake as the spatial dependency weight, thereby generating a spatiotemporal heterogeneous graph. The third modeling unit is used to perform spectral domain convolution processing based on the spatiotemporal heterogeneous graph, learn cross-scale dependency patterns in the graph Laplacian matrix through deep spatiotemporal operators, and output a spatiotemporal dependency graph with spectral clustering optimization characteristics.

9. The machine learning-based estrus initiation prediction system for captive giant pandas according to claim 6, characterized in that, The dimensionality reduction module includes: The first dimensionality reduction unit is used to perform captive physical constraint modeling based on the spatiotemporal dependency graph, and to generate environmental embedding basis functions by constructing a coupled field between the cage space topology and the hormone diffusion equation. The second dimensionality reduction unit is used to perform mode decoupling processing based on the environmental embedding basis function, and to separate the intrinsic mode components of estrus-related hormone fluctuations and environmental noise through the spectral embedding algorithm to obtain the interference-free feature manifold. The third dimensionality reduction unit is used to perform variational geometric constraint processing based on the interference-removing feature manifold, and to achieve adaptive curvature compression of the state space in the Riemannian manifold space using information entropy as a metric, outputting a low-dimensional state representation vector with biological interpretability.

10. The machine learning-based estrus initiation prediction system for captive giant pandas according to claim 6, characterized in that, The computing module includes: The first computing unit is used to perform hormone-behavioral state discretization processing based on the low-dimensional state representation vector, and to decompose the continuous state manifold into a set of biologically meaningful estrus precursor states using the continuous homology theory. The second calculation unit is used to construct a transition probability tensor based on the set of pre-estrus states and to establish a dynamic adjustment mechanism for the state transition kernel function based on the time non-homogeneous characteristics of the reproductive endocrine rhythm of captive giant pandas. The third computing unit is used to perform multi-scale trajectory generation processing according to the dynamic adjustment mechanism. It simulates the probability diffusion process of the state vector in the hormone potential field by solving the Falk-Planck equation and outputs the probability evolution trajectory with timestamp.

Images