Health assessment system and method based on cattle rumination voiceprint recognition

Abstract

The invention belongs to the technical field of intelligent animal husbandry and biosensing, and discloses a health assessment system and method based on cattle rumination voiceprint recognition, and the system comprises an intelligent necklace terminal, an edge calculation unit and a cloud health assessment engine. The intelligent necklace integrates voiceprint collection, jaw movement perception and environment compensation units, and obtains ruminant voiceprint, jaw movement and environment data of cattle in real time; the edge calculation unit optimizes voiceprint signals through a dynamic noise suppression module in combination with environmental parameters, and synchronously analyzes voiceprint and motion features by using a feature joint extraction module; and the cloud end adopts a lightweight hybrid model to perform state classification on ruminant voiceprints, fuses motion features and environmental data to generate health indexes, and triggers early warning or links a feed library to generate an accurate feeding scheme. The cattle digestion health real-time monitoring system realizes cattle digestion health real-time monitoring through non-intrusive voiceprint sensing and multi-modal data analysis, and is especially suitable for pasture accurate management, disease early warning and feeding strategy optimization.

Description

Technical Field

[0001] The present invention relates to the field of smart animal husbandry and biosensor technology, and in particular to a health assessment system and method based on cattle rumination voiceprint recognition. Background Art

[0002] Traditional cattle health monitoring technologies have the following limitations:

[0003] 1. Highly manual: Relying on experience to judge the frequency and duration of rumination, it is impossible to quantify the correlation between voiceprint characteristics and health status;

[0004] 2. Disadvantages of invasive testing: Rumen sensors require surgical implantation, which carries infection risks and is costly;

[0005] 3. Existing equipment lacks accuracy: A single sensor (such as an accelerometer collar) has difficulty distinguishing between rumination, chewing, and environmental noise, with a false alarm rate exceeding 30%. Summary of the Invention

[0006] The purpose of the present invention is to provide a health assessment system and method based on cattle rumination voiceprint recognition, which can realize real-time monitoring of cattle digestive health through non-invasive voiceprint sensing and multimodal data analysis. It is particularly suitable for precise pasture management, early disease warning and feeding strategy optimization.

[0007] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0008] In a first aspect, the present invention provides a health assessment system based on cattle rumination voiceprint recognition, comprising a smart collar terminal, an edge computing unit, and a cloud-based health assessment engine; wherein:

[0009] The smart collar terminal includes:

[0010] Voiceprint collection unit, used to collect original voiceprints;

[0011] Motion sensing unit, used to collect jaw movement characteristics;

[0012] Environmental compensation unit, used to collect environmental data;

[0013] The edge computing unit is connected to the smart collar terminal signal, and the edge computing unit includes:

[0014] a dynamic noise suppression module, which dynamically filters and adjusts the original voiceprint and jaw movement characteristics in response to the input environmental data;

[0015] a feature joint extraction module connected to the dynamic noise suppression module signal, which responds to the input adjusted original voiceprint and jaw movement features and extracts voiceprint features and movement features therefrom respectively;

[0016] The cloud health assessment engine is signal-connected to the edge computing unit to obtain the voiceprint features, motion features, and environmental data. The cloud health assessment engine is configured with a lightweight hybrid model, which is used to:

[0017] Performing preliminary classification based on the voiceprint features to obtain preliminary classification results, wherein the preliminary classification results include rumination and non-rumination voiceprint features;

[0018] Based on the preliminary classification results, the motion characteristics and environmental data corresponding to the rumination voiceprint characteristics are extracted, and the health index is calculated based on the motion characteristics and environmental data;

[0019] Based on the health index, an early warning signal is triggered or an adjustment plan is generated by automatically matching the feed library.

[0020] Furthermore, the environmental compensation unit includes a temperature and humidity sensor and a barometer, and the environmental data includes ambient temperature, ambient humidity and atmospheric pressure.

[0021] Furthermore, the feature joint extraction module extracts voiceprint features in response to the input adjusted original voiceprint based on Mel-frequency cepstral coefficients and wavelet packet energy entropy.

[0022] Furthermore, the feature joint extraction module calculates the jaw motion period variance and the three-dimensional trajectory complexity of the jaw motion feature in response to the input adjusted jaw motion feature as motion features.

[0023] Furthermore, the lightweight hybrid model includes a pre-trained MobileNetV2 network, an XGBoost model and a decision output module; wherein, the pre-trained MobileNetV2 network is used to perform preliminary classification according to the voiceprint features to obtain preliminary classification results, the XGBoost model is used to extract the motion features and environmental data corresponding to the rumination voiceprint features based on the preliminary classification results, and fuse the motion features and environmental parameters to calculate the health index, and the decision output module is used to trigger an early warning signal based on the health index and automatically match the feed library to generate an adjustment plan.

[0024] Furthermore, the calculation formula of the health index is:

[0025]

[0026] Where HI is the health index, ω1 is the first weight, ω2 is the second weight, ω1+ω2=1, f real is the measured chewing frequency, f base is the variety benchmark chewing frequency, E normalFor effective rumination energy, E total is the total ruminant energy.

[0027] Furthermore, the decision output module is used to:

[0028] Based on the set first and second thresholds, when the health index is less than the first threshold, while triggering the early warning signal, it is determined whether the health index is greater than the second threshold. If the health index is greater than the second threshold, it is recommended to increase the neutral detergent fiber by 5%-10%.

[0029] In a second aspect, the present invention provides a health assessment method based on cattle rumination voiceprint recognition, the method comprising:

[0030] Obtain the cattle's original voiceprint, collect jaw movement characteristics and environmental data;

[0031] Dynamically filtering and adjusting the original voiceprint and jaw movement characteristics according to the environmental data;

[0032] According to the adjusted original voiceprint and jaw movement features, voiceprint features and movement features are extracted respectively;

[0033] Performing preliminary classification based on the voiceprint features to obtain preliminary classification results, wherein the preliminary classification results include rumination and non-rumination voiceprint features;

[0034] Based on the preliminary classification results, the motion characteristics and environmental data corresponding to the rumination voiceprint characteristics are extracted, and the health index is calculated based on the motion characteristics and environmental data;

[0035] Based on the health index, an early warning signal is triggered or an adjustment plan is generated by automatically matching the feed library.

[0036] Furthermore, the calculation formula of the health index is:

[0037]

[0038] Where HI is the health index, ω1 is the first weight, ω2 is the second weight, ω1+ω2=1, f real is the measured chewing frequency, f base is the variety benchmark chewing frequency, E normal For effective rumination energy, E total is the total ruminant energy.

[0039] Furthermore, triggering an early warning signal based on the health index or automatically matching the feed library to generate an adjustment plan includes:

[0040] Based on the set first and second thresholds, when the health index is less than the first threshold, while triggering the early warning signal, it is determined whether the health index is greater than the second threshold. If the health index is greater than the second threshold, it is recommended to increase the neutral detergent fiber by 5%-10%.

[0041] The beneficial effects of the present invention are:

[0042] 1. The present invention fuses voiceprint features with multimodal data of jaw movement, combined with an environmental compensation algorithm, to effectively distinguish between rumination and non-rumination states (with an accuracy increase of approximately 30%), and can accurately identify early health risks such as abnormal rumination cycles and abnormal chewing frequency.

[0043] 2. The present invention adopts an edge computing unit for localized signal processing, dynamically adjusts the filtering parameters through environmental data, significantly improves the data signal-to-noise ratio in complex outdoor environments, and ensures the effective extraction of voiceprint features.

[0044] 3. This invention proposes a cloud-based hybrid model to achieve a closed loop of "monitoring-assessment-intervention", reducing the false alarm rate of health index-triggered early warnings to below 5%. It can also generate customized nutrition plans based on the status of individual cattle, significantly improving feed utilization.

[0045] 4. The collaborative architecture of edge computing and cloud computing reduces the amount of raw data transmission by more than 80%. The lightweight model enables a single server to support concurrent analysis of 100,000 cows, reducing cloud computing resource consumption by 60% compared to traditional solutions.

[0046] 5. The present invention integrates multi-dimensional environmental compensation of temperature / humidity / air pressure, and can maintain a monitoring stability of more than 94% in extreme environments of -20°C to 45°C, effectively solving the technical implementation difficulties of outdoor scenes in animal husbandry.

[0047] 6. The present invention adopts a non-invasive monitoring method to avoid the stress response of traditional gastric tube detection. Combined with health warning, it can increase the early detection rate of digestive system diseases by 3 times and reduce the use of antibiotics by about 25%, which meets the requirements of modern ranch welfare breeding. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 The figure shows a structural diagram of a health assessment system based on cattle rumination voiceprint recognition according to an embodiment of the present invention.

[0049] Figure 2 The figure shows the structural layout of the environment compensation unit in a health assessment system based on cattle rumination voiceprint recognition according to an embodiment of the present invention.

[0050] Figure 3A flowchart of a health assessment system based on cattle rumination voiceprint recognition according to an embodiment of the present invention is shown to implement dynamic filtering and adjustment of the original voiceprint and jaw movement characteristics.

[0051] Figure 4 A flowchart of extracting motion features using a health assessment system based on cattle rumination voiceprint recognition according to an embodiment of the present invention is shown.

[0052] Figure 5 A flowchart of a health assessment method based on cattle rumination voiceprint recognition according to an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0053] The following describes the embodiments of the present invention through specific examples. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments. The details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the following embodiments and features in the embodiments can be combined with each other unless they conflict.

[0054] In the description of the present invention, unless otherwise specified, "plurality" means two or more; terms such as "upper," "lower," "left," "right," "inner," "outer," "front end," "rear end," "head," and "tail" indicate positions or relationships based on those shown in the accompanying drawings. These terms are intended solely to facilitate the description of the present invention and simplify the description. They do not indicate or imply that the devices or components referred to must have a specific orientation, be constructed, or operate in a specific orientation. Therefore, they should not be construed as limiting the present invention. Furthermore, terms such as "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0055] In the description of the present invention, it should be noted that, unless otherwise expressly specified or limited, the terms "connected" and "connection" should be understood in a broad sense. For example, they can refer to fixed connection, detachable connection, or integral connection; mechanical connection, electrical connection; direct connection, or indirect connection through an intermediary. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on the specific circumstances.

[0056] The specific implementation of the present invention is further described in detail below with reference to the accompanying drawings and examples.

[0057] Example 1:

[0058] The embodiment of the present invention provides a health assessment system based on cattle rumination voiceprint recognition, such as Figure 1As shown, the health assessment system based on cattle rumination voiceprint recognition includes a smart collar terminal 100, an edge computing unit 200 and a cloud health assessment engine 300; wherein the smart collar terminal 100 includes:

[0059] Voiceprint collection unit 110, used to collect original voiceprints;

[0060] A motion sensing unit 120 is used to collect jaw movement characteristics;

[0061] The environment compensation unit 130 is used to collect environment data.

[0062] The edge computing unit 200 is connected to the smart collar terminal 100 by signal, and the edge computing unit 200 includes:

[0063] A dynamic noise suppression module 210, which dynamically filters and adjusts the original voiceprint and jaw movement characteristics in response to the input environmental data;

[0064] a feature joint extraction module 220 signal-connected to the dynamic noise suppression module 210, which responds to the input adjusted original voiceprint and jaw movement feature and extracts voiceprint features and movement features therefrom respectively;

[0065] The cloud-based health assessment engine 300 is signal-connected to the edge computing unit 200 for acquiring the voiceprint features, motion features, and environmental data. The cloud-based health assessment engine 300 is configured with a lightweight hybrid model 310, which is used to: perform preliminary classification based on the voiceprint features to obtain preliminary classification results, where the preliminary classification results include rumination and non-rumination voiceprint features; extract motion features and environmental data corresponding to the rumination voiceprint features based on the preliminary classification results, and calculate a health index based on the motion features and environmental data; and trigger an early warning signal based on the health index or automatically match the feed library to generate an adjustment plan.

[0066] In some embodiments, the voiceprint collection unit 110 may be a directional microphone array, whose parameters may be, for example, a frequency response of 50 Hz-12 kHz and a dynamic range ≥ 80 dB. The directional microphone array is placed close to the throat of the cow to collect the original voiceprint.

[0067] In some embodiments, motion sensing unit 120 can be a 9-axis IMU sensor (accelerometer + gyroscope + magnetometer) with a sampling rate of 200 Hz to capture jaw motion characteristics. The 9-axis IMU sensor is fixed to the cow's mandible or zygomatic bone to ensure that the signal truly reflects jaw motion.

[0068] In some embodiments, as Figure 2As shown, the environmental compensation unit 130 includes a temperature and humidity sensor 131 and a barometer 132. The environmental data includes ambient temperature, humidity, and atmospheric pressure. The temperature and humidity sensor 131 is used to collect ambient temperature and humidity, while the barometer 132 is used to collect atmospheric pressure. The environmental compensation unit 130 can be placed close to the cattle, for example, by attaching the environmental compensation unit 130 to a collar placed around the cattle's neck.

[0069] In some embodiments, the dynamic noise suppression module 210 is provided to dynamically filter and adjust the original voice print and jaw movement characteristics in response to the inputted environmental data.

[0070] Specifically, ambient temperature, ambient humidity and atmospheric pressure are used as environmental data, such as Figure 3 As shown, the dynamic noise suppression module 210 is configured to implement dynamic filtering and adjustment of the original voiceprint and jaw movement characteristics in response to the inputted environmental data through the following steps S301-S306:

[0071] S301: Get original voiceprint x k , jaw movement characteristics k And the environmental data vector u k ,u k =[T k ,H k ,P k ] T , T k ,H k ,P k are ambient temperature, ambient humidity, and atmospheric pressure, respectively. The superscript T represents the matrix transpose, and the subscript k represents the discrete time step (sampling point k = 1, 2, ..., N).

[0072] S302: Normalize the environmental data vector and convert the original voiceprint x k , jaw movement characteristics k It is concatenated with the normalized environment data vector to obtain the input vector.

[0073] In this embodiment, the environmental data vector is normalized by mapping the temperature, humidity, and atmospheric pressure to the interval [-1, 1]: Among them, u′ k is the normalized environmental data vector, u u and u u are the mean and standard deviation of the environmental data respectively. The input vector is represented as

[0074] S303: Set the initial weight w0 and initialize the inverse autocorrelation matrix R0 = δI. The initial weight w0 is usually a zero vector or a random small value, and δ is a small positive number, such as 10 -6 , I is the identity matrix.

[0075] S304: For each time step k, calculate the gain vector, filter output and error in sequence, and update the weight matrix and the inverse autocorrelation matrix.

[0076] In this embodiment, the calculation formula of the gain vector is: Among them, K k is the gain vector at time step k, R k-1 is the inverse matrix of the autocorrelation matrix of time step k-1, λ is the forgetting factor, 0<λ≤1, which is used to weaken the influence of old data.

[0077] The filter output is calculated as: in, is the output after filtering.

[0078] The error is calculated as: Among them, e k is the filtering error, d k is a supervisory value (such as a historical stable eigenvalue or an environment-independent reference feature). If a supervisory value is missing, the current eigenvalue is used as the approximate value of d k , suppressing environmental noise through error.

[0079] The update formula of the weight matrix is: k =w k-1 +K k e k .

[0080] The update formula of the inverse matrix of the autocorrelation matrix is:

[0081] S305: When the filtering error reaches the set value or the number of iterations reaches the maximum set number, the weight vector w after filtering is obtained. k The optimal weight for separating voiceprint features and jaw movement features:

[0082]

[0083] in, is the optimal weight of the voiceprint feature, is the optimal weight of the jaw motion feature, dian(x) is the dimension of the voiceprint feature vector, and dim(y) is the dimension of the jaw motion feature vector.

[0084] S306: Obtaining adjusted original voiceprint and jaw movement features based on the optimal weights of the voiceprint and jaw movement features:

[0085]

[0086] Where, and They are the adjusted original voiceprint and jaw movement characteristics respectively.

[0087] In some embodiments, the joint feature extraction module extracts voiceprint features based on Mel-frequency cepstral coefficients and wavelet packet energy entropy in response to the input adjusted original voiceprint.

[0088] Specifically, this embodiment can extract voiceprint features in the following ways:

[0089] Based on the original voiceprint after rectification, the cepstral coefficients are obtained through pre-emphasis, frame windowing, FFT transform, Mel filter bank weighting, and DCT transform. The first 12-13 coefficients are taken as static features, and the first and second order differences are added to form dynamic features, which are the MFCC features. Subsequently, the wavelet basis function (such as db4) and the number of decomposition layers (for example, 3 layers, resulting in 8 sub-bands) are selected; the signal is decomposed into wavelet packets to obtain the coefficients of each sub-band; the energy of each sub-band is calculated, and the entropy value is calculated after normalization to represent the complexity of the energy distribution; the calculated entropy value is used as the wavelet packet energy entropy. The MFCC feature wavelet packet energy entropy is combined into a joint feature vector as the voiceprint feature.

[0090] In some embodiments, the joint feature extraction module calculates the jaw motion period variance and the three-dimensional trajectory complexity of the input adjusted jaw motion feature as motion features.

[0091] In this embodiment, Figure 4 As shown, motion features can be obtained through the following steps:

[0092] Step S401: Periodic detection.

[0093] The adjusted time series data of jaw movement is obtained; wherein, the adjusted time series data of jaw movement includes acceleration data and displacement data, and the acceleration data is low-pass filtered (the cut-off frequency is set according to the motion characteristics, and is set to 5Hz in this embodiment), high-frequency noise is removed, and the displacement data is smoothed. Then, cycle segmentation is performed, including peak detection and zero-crossing detection. During peak detection, local maximum points are detected in the main motion direction, such as the vertical acceleration, and the time interval between adjacent maximum points is one cycle. Zero-crossing detection is to find the zero-crossing point from positive to negative or negative to positive in the acceleration signal to define the cycle boundary. Record the start time of each cycle and end time

[0094] S402: Calculate the variance of the jaw movement cycle.

[0095] Calculate the duration of each cycle, that is, the cycle length

[0096] Extract the maximum acceleration or displacement amplitude A in each cycle k .

[0097] Calculate the average rate of change of displacement during the period.

[0098] Cycle variance calculation: Select key features, such as cycle length D k , and calculate its variance as the jaw movement cycle variance. The multi-dimensional variance (such as the joint variance matrix of cycle duration, amplitude, and average rate) can also be calculated simultaneously as the jaw movement cycle variance.

[0099] S403: Calculate the three-dimensional trajectory complexity.

[0100] If the input is acceleration data, the displacement trajectory is obtained by double integration. If the input is displacement data, the displacement trajectory is obtained directly based on the displacement data, where the displacement trajectory r(t) = [x(t), y(t), z(t)]; x(t), y(t), and z(t) are the displacements in the three directions at time t, respectively.

[0101] Discretize the displacement trajectory r(t) = [x(t), y(t), z(t)] into N points, select multiple scales ∈1, ∈2, …, ∈ M ; For each scale ∈ i , calculate the number of three-dimensional cubes N(∈ i ), the slope of logN(∈) and logN(1 / ∈) is fitted, which is the fractal dimension, and the fractal dimension is used to characterize the complexity of the three-dimensional trajectory.

[0102] S404: Splicing jaw motion period variance and 3D trajectory complexity as motion features.

[0103] In this embodiment, the movement characteristics include jaw movement period variance and three-dimensional trajectory complexity. The jaw movement period variance reflects whether the cattle's movement is regular. An increase in variance indicates irregular movement, which can manifest as chewing weakness or pain. The three-dimensional trajectory complexity indicates the cattle's movement coordination. An increase in complexity (increase in the fractal dimension) indicates a decrease in movement coordination, such as jaw stiffness.

[0104] In some embodiments, the lightweight hybrid model includes a pre-trained MobileNetV2 network, an XGBoost model, and a decision output module; wherein the pre-trained MobileNetV2 network is used to perform preliminary classification according to the voiceprint features to obtain preliminary classification results, the XGBoost model is used to extract the motion features and environmental data corresponding to the rumination voiceprint features based on the preliminary classification results, and fuse the motion features and environmental parameters to calculate the health index, and the decision output module is used to trigger an early warning signal based on the health index and automatically match the feed library to generate an adjustment plan.

[0105] In this embodiment, the voiceprint features are voiceprint feature vectors extracted using MFCC and wavelet packet energy entropy (e.g., 12-dimensional MFCC + 1-dimensional WPEE). Motion features include jaw movement period variance and three-dimensional trajectory complexity. Environmental parameters include real-time environmental data such as temperature, humidity, and atmospheric pressure. The voiceprint features are converted into a voiceprint spectrogram, which serves as the input image for MobileNetV2. The motion features and environmental parameters are combined into a structured data table for XGBoost processing.

[0106] The pre-trained MobileNetV2 network (ImageNet weights) is fine-tuned through transfer learning, and the input is a voiceprint spectrogram. The last fully connected layer of the MobileNetV2 network is modified to output a binary classification result: rumination (1) or non-rumination (0). The voiceprint spectrogram of rumination has periodic energy peaks in the low frequency band (1-2kHz), and MobileNetV2 captures this spatial pattern through the convolutional layer. The output probability P belongs to [0,1], and a threshold is set to determine the valid rumination state. In this embodiment, the threshold is set to P>0.8.

[0107] The XGBoost model achieves high-precision cattle health assessment and decision support by integrating domain knowledge (health index formula) and data-driven modeling (XGBoost nonlinear learning). Specifically, the input features of the XGBoost model include rumination voiceprint features, motion features, and environmental parameters; the rumination voiceprint features are the MFCC and WPEE of the voiceprint classified as "rumination" by MobileNetV2. The motion features are normalized (Z-score) and the environmental parameters are normalized to the range [0,1]. Construct a joint feature vector, MFCC1 is the first dimension of the voiceprint MFCC feature, WPEE is the energy entropy value, is the variance of the jaw movement cycle, FD is the three-dimensional trajectory complexity, and T, H, and P are temperature, humidity, and atmospheric pressure. XGBoost, based on a gradient boosted tree, learns the mapping between features and health status (a regression task). Output is the health index.

[0108] In some embodiments, the health index is calculated as follows:

[0109]

[0110] Where HI is the health index, ω1 is the first weight, ω2 is the second weight, ω1+ω2=1, f real is the measured chewing frequency, f base is the variety benchmark chewing frequency (pre-stored in the database, directly retrieved), E normal For effective rumination energy, E total is the total ruminant energy.

[0111] Based on the above health index calculation method, XGBoost receives multimodal features (voiceprint, movement, environment) as input, splits and combines the features through the tree structure, generates intermediate prediction values, and finally the output layer linearly combines these intermediate values ​​according to the weights to obtain the health index HI. Among them, the intermediate prediction value is f real 、E normal and E total ;f real By calculating the variance of the jaw movement cycle, E normal is the low entropy frequency band energy extracted by the voiceprint wavelet packet energy entropy, E total is the total energy of the voiceprint signal.

[0112] The XGBoost model can be pre-trained to determine the first weight and the second weight. Specifically, a training data set is constructed for training the XGBoost model, where each sample in the training data set includes rumination voiceprint features, movement features, and environmental parameters. Based on the above training data set, the XGBoost model is trained to determine the first weight and the second weight in the following manner:

[0113] The input features of the XGBoost model are rumination voiceprint features, movement features and environmental parameters. The XGBoost model is equipped with a pre-data processing module, which calculates the measured chewing frequency f through the variance of the jaw movement cycle. real , retrieve the variety benchmark chewing frequency from the database, and use the low entropy frequency band energy extracted from the voiceprint wavelet packet energy entropy as the effective rumination energy E normal , the total energy of the voiceprint signal is used as the total rumination energy; after being processed by the pre-data processing module, the normalized chewing frequency ratio and effective ruminant energy ratio As key features, motion features and environmental parameters are input into the XGBoost model as auxiliary features to train the XGBoost model.

[0114] XGBoost is trained through supervised learning, taking the health index HIHI as the regression target, and optimizing the following loss function:

[0115]

[0116] Where L is the loss value, Ω is the model complexity, and f real,i is the measured chewing frequency of the i-th sample, f base,i is the benchmark chewing frequency of the i-th sample (pre-stored in the database, directly retrieved), E normal,i is the effective rumination energy of the i-th sample, E total,i is the total rumination energy of the i-th sample, i is the sample number, and N is the number of samples.

[0117] Based on the gradient descent algorithm, the XGBoost model is trained until the set number of iterations is reached or the loss value converges (the loss value reaches the set loss threshold range), and the first weight and the second weight are output.

[0118] The inference process of this XGBoost model is as follows:

[0119] Each decision tree splits nodes according to the eigenvalues, where the eigenvalues ​​include standardized chewing frequency ratio, effective rumination energy ratio and temperature. Taking the standardized chewing frequency ratio as an example, if And the temperature is greater than 30℃, then it is assigned to the left subtree (higher health risk). Each leaf node outputs weight ω j , reflecting the contribution of the path to the health index. The final health index is the weighted sum of all tree outputs.

[0120] In this embodiment, the auxiliary features input to the XGBoost model are used to dynamically adjust the first weight and the second weight, and to assist the feed library in generating the adjustment plan. Specifically, for temperature, high temperature may cause the cow to chew less frequently (heat stress response), and directly use It will be misjudged as a health problem. After introducing the temperature parameter, the model can dynamically adjust the weight of the frequency ratio. For example, when the temperature is >30℃, the model reduces ω1 (frequency weight) to avoid misjudging the physiological frequency drop caused by high temperature as a disease. As for humidity, high humidity may affect the signal quality of the voiceprint sensor (such as increased low-frequency noise), resulting in The humidity parameter helps the model filter out invalid energy bands.

[0121] The 3D trajectory complexity in the motion features is used to capture independent health signals of motion behavior. The jaw motion trajectory of healthy cattle is usually regular, that is, the complexity is low, while disease or pain may cause movement disorder, which will lead to increased complexity. Therefore, if If the trajectory complexity is significantly increased (e.g. fractal dimension > 1.8), the model will still trigger an early warning.

[0122] In some embodiments, the decision output module is used to: based on the set first threshold and second threshold, when the health index is less than the first threshold, while triggering the early warning signal, determine whether the health index is greater than the second threshold; if the health index is greater than the second threshold, it is recommended to increase the neutral detergent fiber by 5%-10%.

[0123] In this embodiment, a graded early warning mechanism based on the health index is combined with the adjustment of neutral detergent fiber (NDF) intake to achieve preventive health management. The first threshold (early warning critical value) is set as the health risk warning line, indicating that the physiological indicators begin to deviate from the normal range, but have not yet reached a dangerous level. Triggering an early warning prompt requires intervention. The second threshold (low-risk boundary value) is between the normal range and the early warning threshold, which is used to distinguish between "mild risk" and "moderate risk". When the health index is in this range, it indicates that the risk is controllable, but intervention is still required.

[0124] When the health index falls between the first and second thresholds, NDF supplementation is recommended. The scientific basis for this recommendation includes the following: NDF improves intestinal motility by increasing dietary fiber, alleviating decreased digestive function due to metabolic stress (e.g., rumen health in ruminants). The slow fermentation of fiber stabilizes blood sugar and energy supply, reducing the risk of metabolic disorders (e.g., preventing diabetes). Fiber breakdown products (short-chain fatty acids) have anti-inflammatory effects and can downregulate markers of chronic inflammation (e.g., C-reactive protein).

[0125] The adjustment range of 5%-10% is based on the principle of minimum effective dose, avoiding nutrient absorption inhibition caused by excessive fiber, and only supplementing NDF within the warning range rather than full-area intervention, reflecting personalized management logic.

[0126] The embodiment of the present invention also provides a health assessment method based on cattle rumination voiceprint recognition, such as Figure 5 As shown, the method includes the following steps:

[0127] S501, obtaining the original voiceprint of the cattle, collecting jaw movement characteristics and environmental data;

[0128] S502: Dynamically filter and adjust the original voiceprint and jaw movement characteristics according to the environmental data;

[0129] S503, extracting voiceprint features and movement features from the adjusted original voiceprint and jaw movement features respectively;

[0130] S504: Perform preliminary classification based on the voiceprint features to obtain preliminary classification results, wherein the preliminary classification results include rumination and non-rumination voiceprint features;

[0131] S505: Based on the preliminary classification results, extract the motion features and environmental data corresponding to the rumination voiceprint features, and calculate the health index based on the motion features and environmental data;

[0132] S506: triggering an early warning signal based on the health index or automatically matching the feed library to generate an adjustment plan.

[0133] In some embodiments, the health index is calculated as follows:

[0134]

[0135] Where HI is the health index, ω1 is the first weight, ω2 is the second weight, ω1+ω2=1, f real is the measured chewing frequency, f base is the variety benchmark chewing frequency, E normal For effective rumination energy, E total is the total ruminant energy.

[0136] In some embodiments, triggering an early warning signal based on the health index or automatically matching the feed library to generate an adjustment plan includes:

[0137] Based on the set first and second thresholds, when the health index is less than the first threshold, while triggering the early warning signal, it is determined whether the health index is greater than the second threshold. If the health index is greater than the second threshold, it is recommended to increase the neutral detergent fiber by 5%-10%.

[0138] It should be noted that the health assessment method based on cattle rumination voiceprint recognition provided by the present invention belongs to the same technical concept as the previous assessment system, and can achieve the same technical effect, so it will not be repeated here.

[0139] The above embodiments are only used to illustrate the present invention, and are not intended to limit the present invention. Ordinary technicians in the relevant technical field may make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, all equivalent technical solutions also fall within the scope of the present invention. The scope of patent protection of the present invention should be defined by the claims.

Claims

1. A health assessment system based on cattle rumination voiceprint recognition, characterized in that: It includes a smart collar terminal, an edge computing unit, and a cloud-based health assessment engine; among which: The smart collar terminal includes: Voiceprint collection unit, used to collect original voiceprints; Motion sensing unit, used to collect jaw movement characteristics; Environmental compensation unit, used to collect environmental data; The edge computing unit is connected to the smart collar terminal signal, and the edge computing unit includes: a dynamic noise suppression module, which dynamically filters and adjusts the original voiceprint and jaw movement characteristics in response to the input environmental data; a feature joint extraction module connected to the dynamic noise suppression module signal, which responds to the input adjusted original voiceprint and jaw movement features and extracts voiceprint features and movement features therefrom respectively; The cloud health assessment engine is signal-connected to the edge computing unit to obtain the voiceprint features, motion features, and environmental data. The cloud health assessment engine is configured with a lightweight hybrid model, which is used to: Performing preliminary classification based on the voiceprint features to obtain preliminary classification results, wherein the preliminary classification results include rumination and non-rumination voiceprint features; Based on the preliminary classification results, the motion characteristics and environmental data corresponding to the rumination voiceprint characteristics are extracted, and the health index is calculated based on the motion characteristics and environmental data; Based on the health index, an early warning signal is triggered or an adjustment plan is generated by automatically matching the feed library.

2. The health assessment system based on cattle rumination voiceprint recognition according to claim 1, characterized in that: The environmental compensation unit includes a temperature and humidity sensor and a barometer, and the environmental data includes environmental temperature, environmental humidity and atmospheric pressure.

3. The health assessment system based on cattle rumination voiceprint recognition according to claim 1, characterized in that: The feature joint extraction module extracts voiceprint features in response to the input adjusted original voiceprint based on Mel-frequency cepstral coefficients and wavelet packet energy entropy.

4. The health assessment system based on cattle rumination voiceprint recognition according to claim 1, characterized in that: The feature joint extraction module calculates the jaw motion period variance and the three-dimensional trajectory complexity as motion features in response to the input adjusted jaw motion features.

5. The health assessment system based on cattle rumination voiceprint recognition according to claim 1, characterized in that: The lightweight hybrid model includes a pre-trained MobileNetV2 network, an XGBoost model and a decision output module; wherein the pre-trained MobileNetV2 network is used to perform preliminary classification according to the voiceprint features to obtain preliminary classification results, the XGBoost model is used to extract the motion features and environmental data corresponding to the rumination voiceprint features based on the preliminary classification results, and to fuse the motion features and environmental parameters to calculate the health index, and the decision output module is used to trigger an early warning signal based on the health index and automatically match the feed library to generate an adjustment plan.

6. The health assessment system based on cattle rumination voiceprint recognition according to claim 1 or 5, characterized in that: The calculation formula of the health index is: Where HI is the health index, ω1 is the first weight, ω2 is the second weight, ω1+ω2=1, f real is the measured chewing frequency, f base is the variety benchmark chewing frequency, E normal For effective rumination energy, E total is the total ruminant energy.

7. The health assessment system based on cattle rumination voiceprint recognition according to claim 5, characterized in that: The decision output module is used to: Based on the set first and second thresholds, when the health index is less than the first threshold, while triggering the early warning signal, it is determined whether the health index is greater than the second threshold. If the health index is greater than the second threshold, it is recommended to increase the neutral detergent fiber by 5%-10%.

8. A health assessment method based on cattle rumination voiceprint recognition, characterized in that: The method comprises: Obtain the original voiceprint of cattle, collect jaw movement characteristics and environmental data; Dynamically filtering and adjusting the original voiceprint and jaw movement characteristics according to the environmental data; According to the adjusted original voiceprint and jaw movement features, voiceprint features and movement features are extracted respectively; Performing preliminary classification based on the voiceprint features to obtain preliminary classification results, wherein the preliminary classification results include rumination and non-rumination voiceprint features; Based on the preliminary classification results, the motion characteristics and environmental data corresponding to the rumination voiceprint characteristics are extracted, and the health index is calculated based on the motion characteristics and environmental data; Based on the health index, an early warning signal is triggered or an adjustment plan is generated by automatically matching the feed library.

9. The health assessment method based on cattle rumination voiceprint recognition according to claim 8, characterized in that: The calculation formula of the health index is: Where HI is the health index, ω1 is the first weight, ω2 is the second weight, ω1+ω2=1, f real is the measured chewing frequency, f base is the variety benchmark chewing frequency, E normal For effective rumination energy, E total is the total ruminant energy.

10. The health assessment system based on cattle rumination voiceprint recognition according to claim 8, characterized in that: Based on the health index, an early warning signal is triggered or an adjustment plan is automatically matched with the feed library, including: Based on the set first and second thresholds, when the health index is less than the first threshold, while triggering the early warning signal, it is determined whether the health index is greater than the second threshold. If the health index is greater than the second threshold, it is recommended to increase the neutral detergent fiber by 5%-10%.

Images