Intelligent monitoring system for beef cattle feeding behavior based on machine learning

Abstract

The application relates to the technical field of intelligent livestock breeding, in particular to a beef cattle feeding behavior intelligent monitoring system based on machine learning. The system comprises a posture decoupling module, which is used for decomposing original acceleration data into reference axis vibration intensity and reference plane swing intensity based on a quasi-static posture reference vector of a target cow; an effectiveness auditing module, which is used for determining feeding confidence based on the difference between the reference axis vibration intensity and a pre-generated resting vibration benchmark; a group difference module, which is used for constructing a group average sweeping ratio; and an excessive sorting deviation determination module, which is used for determining an excessive sorting deviation based on the group average sweeping ratio, a sweeping and chewing ratio of the target cow and the feeding confidence. The application is used for the quantitative monitoring of sorting behavior in the beef cattle feeding process.

Description

Technical Field

[0001] This invention relates to the field of intelligent livestock farming technology, specifically to an intelligent monitoring system for beef cattle feeding behavior based on machine learning. Background Technology

[0002] In large-scale beef cattle farming, total mixed rations (TMRs) aim to provide beef cattle with a nutritionally balanced diet by physically mixing roughage and concentrate. However, in actual feeding, beef cattle, based on their instinct to maximize energy intake, tend to use specific neck-swinging behaviors to push aside the upper layer of roughage and preferentially consume the concentrate that has sunk to the bottom. This behavior, known as sorting, causes the actual roughage-to-concentrate ratio ingested by beef cattle to deviate from the formula, leading to metabolic diseases such as rumen acidosis and feed waste. Therefore, quantitative monitoring of individual sorting behavior is a key link in achieving precision farming.

[0003] Current technologies primarily rely on accelerometers worn around the necks of cattle to monitor their activity. By analyzing the acceleration data collected by the sensors, they identify behaviors such as feeding and rumination, and attempt to quantify sorting behavior. However, in practical engineering applications, existing technologies face several technical obstacles: First, due to loose collars, the sensors may rotate randomly around the neck, causing a misalignment between the sensor coordinate system and the cattle's anatomical coordinate system. Directly using raw axial data cannot distinguish between effective vertical chewing vibrations and lateral sweeping movements. Second, lateral swaying behavior in beef cattle is not solely caused by sorting; non-feeding behaviors such as swatting flies and social interactions also produce significant lateral swaying, leading to signal confusion. Furthermore, during the initial herd feeding phase, all cattle exhibit high levels of sweeping behavior, a common behavior caused by environmental factors rather than individual pathological characteristics. These factors combined result in low monitoring accuracy in existing technologies, making it difficult to meet the practical needs of large-scale farming for precise individual health management. Summary of the Invention

[0004] To address the technical problem that existing technologies have low monitoring accuracy, making it difficult to meet the practical needs of large-scale farming for precise individual health management, the present invention aims to provide an intelligent monitoring system for beef cattle feeding behavior based on machine learning. The specific technical solution adopted is as follows:

[0005] In a first aspect, the present invention provides an intelligent monitoring system for beef cattle feeding behavior based on machine learning. The system includes: a posture decoupling module, used to decompose raw acceleration data into reference axial vibration intensity and reference planar oscillation intensity based on a quasi-static posture reference vector of the target cattle; the quasi-static posture reference vector is extracted from raw acceleration data collected by sensors worn on the neck of the target cattle; the reference axial vibration intensity is used to characterize the cattle's chewing behavior; the reference planar oscillation intensity is used to characterize the cattle's sweeping behavior; the target cattle is any cattle in the pen; an effectiveness auditing module, used to determine the feeding confidence level based on the difference between the reference axial vibration intensity and a pre-generated resting vibration benchmark; the feeding confidence level is used to characterize the probability that the target cattle's current action includes chewing behavior; a group differentiation module, used to construct a group average sweeping ratio based on the reference axial vibration intensity and reference planar oscillation intensity of each cattle in the same pen as the target cattle; and an excess sorting deviation determination module, used to determine the excess sorting deviation characterizing abnormal sorting behavior of the target cattle relative to the group based on the group average sweeping ratio, the target cattle's sweeping-chewing ratio, and the feeding confidence level.

[0006] In conjunction with the first aspect mentioned above, in one possible implementation, the attitude decoupling module includes: a reference axis extraction submodule, used to filter the original acceleration data and determine the low-frequency component obtained after filtering as the quasi-static attitude reference vector; an orthogonal decomposition submodule, used to decompose the original acceleration data into instantaneous axial vibration acceleration and instantaneous planar oscillation acceleration based on the quasi-static attitude reference vector; the instantaneous axial vibration acceleration is the projection component of the original acceleration data in the direction of the quasi-static attitude reference vector; the instantaneous planar oscillation acceleration is the projection component of the original acceleration data in a plane perpendicular to the quasi-static attitude reference vector; and a feature integration submodule, used to accumulate and sum the instantaneous axial vibration acceleration and instantaneous planar oscillation acceleration within a preset micro-sampling period, respectively, and determine the accumulated sum as the reference axial vibration intensity and the reference planar oscillation intensity, respectively.

[0007] In conjunction with the first aspect mentioned above, in one possible implementation, the effectiveness audit module includes: a benchmark maintenance submodule, used to collect a reference axial vibration intensity sequence of the target cattle within a preset resting period and determine the variance of the reference axial vibration intensity sequence; the benchmark maintenance submodule is used to determine a resting vibration benchmark based on the lower quartile of the reference axial vibration intensity sequence when the variance is less than or equal to a preset resting fluctuation threshold; and a confidence calculation submodule, used to determine the feeding confidence based on the difference between the current reference axial vibration intensity and the resting vibration benchmark, and a preset mapping function; the feeding confidence is positively correlated with the degree of difference.

[0008] In conjunction with the first aspect mentioned above, in one possible implementation, the population differential module includes: an active screening submodule, used to mark cattle with a feeding confidence level greater than or equal to a preset validity threshold as active individuals; a benchmark construction submodule, used to calculate the sweeping-chewing ratio of each active individual based on the reference axial vibration intensity and reference planar oscillation intensity of all active individuals at the current moment; and a benchmark construction submodule, used to determine the population average sweeping ratio based on the sweeping-chewing ratio of all active individuals.

[0009] In conjunction with the first aspect mentioned above, in one possible implementation, the excess sorting deviation determination module is specifically used to: determine the sweeping and chewing ratio of the target cow based on the reference axial vibration intensity and the reference planar oscillation intensity of the target cow; and determine the excess sorting deviation of the target cow based on the difference between the sweeping and chewing ratio of the target cow and the average sweeping ratio of the herd, as well as the feeding confidence level.

[0010] In conjunction with the first aspect above, in one possible implementation, the baseline construction submodule is further used to: determine whether the number of active individuals is less than a preset minimum sample size threshold; if the number of active individuals is less than the preset minimum sample size threshold, determine the most recent valid historical population average scan ratio or the preset default population baseline as the population average scan ratio.

[0011] In conjunction with the first aspect above, in one possible implementation, the benchmark construction submodule is further configured to: determine the initial ratio of each active individual based on the reference plane oscillation intensity of each active individual, the reference axial vibration intensity of each active individual, and a preset minimum positive threshold; the initial ratio is used to characterize the relative magnitude of the original sweeping chewing intensity of the active individual relative to its chewing intensity; the initial ratio of each active individual is subjected to amplitude limiting processing, and the value obtained after amplitude limiting processing is determined as the sweeping chewing ratio of each active individual; the amplitude limiting processing is used to make the initial ratio less than or equal to the preset ratio upper limit.

[0012] In conjunction with the first aspect mentioned above, in one possible implementation, the system further includes: a risk decision module for acquiring the excess sorting deviation of the target cattle; and a risk decision module for generating a risk index characterizing the health risk of the target cattle's sorting behavior based on time-series data of the excess sorting deviation of the target cattle changing over time.

[0013] In conjunction with the first aspect mentioned above, in one possible implementation, the risk decision-making module includes: a session encapsulation submodule, used to encapsulate a data sequence that meets preset feeding behavior conditions into a feeding session based on the time-series data of the target cattle's excess sorting deviation; a trend calculation submodule, used to determine the sorting persistence trend of the target cattle in the feeding session based on the statistical difference between the excess sorting deviation of the first half and the second half of the time series within the feeding session; and a risk synthesis submodule, used to synthesize the risk index of the target cattle in the feeding session based on the cumulative amount of excess sorting deviation and the sorting persistence trend within the feeding session.

[0014] In conjunction with the first aspect mentioned above, in one possible implementation, the risk decision module is also used to: generate health warning information for the target cow when the risk index of the target cow exceeds a preset individual risk threshold; and generate adjustment instructions for the feed processing technology when the risk index distribution of multiple cows in the same pen as the target cow meets preset group risk conditions within a preset period after feeding.

[0015] The present invention has the following beneficial effects:

[0016] This invention forms a complete dual-differential processing closed loop by constructing an attitude decoupling module, an effectiveness auditing module, a population differential module, and an excess sorting deviation determination module. The attitude decoupling module decomposes the raw acceleration data based on a quasi-static attitude reference vector, fundamentally eliminating the impact of sensor wearing attitude uncertainty on monitoring accuracy. This allows the same algorithm to adapt to devices with different tightness and wearing angles, significantly reducing the difficulty of on-site installation and maintenance. The effectiveness auditing module introduces feeding confidence to achieve longitudinal gating of non-feeding noise, effectively suppressing invalid oscillation signals generated by non-feeding behaviors such as swatting flies and social interactions. The population differential module constructs a population average sweeping ratio to eliminate lateral background interference from common environmental behaviors such as population feeding. The excess sorting deviation determination module integrates the above processing results to finally obtain the excess sorting deviation, which characterizes the abnormal sorting behavior of an individual relative to the population. This indicator is stripped of sensor error, non-feeding noise, and environmental background interference, specifically quantifying the individual's true pathological sorting behavior and providing a reliable data foundation for subsequent health risk assessment. This solves the technical problem that existing technologies have low monitoring accuracy and cannot meet the actual needs of large-scale farming for precise individual health management. Attached Figure Description

[0017] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of the system architecture of a machine learning-based intelligent monitoring system for beef cattle feeding behavior provided in an embodiment of the present invention.

[0019] Figure 2 This invention provides an intelligent monitoring device for beef cattle feeding behavior based on machine learning, as one embodiment of the present invention. Detailed Implementation

[0020] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of the intelligent monitoring system for beef cattle feeding behavior based on machine learning proposed in this invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0021] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0022] The specific solution of the intelligent monitoring system for beef cattle feeding behavior based on machine learning provided by the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Please see Figure 1 This diagram illustrates the system architecture of a machine learning-based intelligent monitoring system 100 for beef cattle feeding behavior according to an embodiment of the present invention. The system includes:

[0024] The attitude decoupling module 101 is used to decompose the original acceleration data into reference axial vibration intensity and reference planar oscillation intensity based on the quasi-static attitude reference vector of the target cow.

[0025] In one possible implementation, the attitude decoupling module 101 is deployed on a smart collar terminal worn around the neck of the target cow. This module extracts behavioral features with clear biomechanical meaning from the raw acceleration signal when the sensor's wearing posture changes randomly. Because the collar rotates circumferentially due to skin laxity or movement inertia during free movement, the measurement coordinate system of the sensor's built-in triaxial accelerometer continuously changes relative to the target cow's anatomical coordinate system. Directly using the raw axial data would fail to distinguish between vertical chewing movements and lateral sweeping movements. To address this issue, a dynamic decoupling method based on a quasi-static attitude reference vector is employed. This method extracts components from the signal that characterize slow neck posture changes to construct an absolute reference system independent of the physical orientation of the sensor housing.

[0026] For example, the triaxial accelerometer built into the sensor continuously acquires the raw acceleration vector sequence of the target cow's neck at a preset sampling frequency. The attitude decoupling module 101 processes the raw acceleration signal and extracts the quasi-static component that changes slowly with the target cow's body posture as a quasi-static attitude reference vector, which is engineeringly approximated as the current gravity direction. After obtaining a stable quasi-static attitude reference vector, the attitude decoupling module 101 decomposes the raw acceleration data into two components with clear physical meanings based on the biomechanical characteristics of the target cow's feeding behavior: reference axial vibration intensity and reference planar oscillation intensity. The reference axial vibration intensity is used to characterize the mechanical vibration along the gravity line direction caused by chewing and swallowing; the reference planar oscillation intensity is used to characterize the lateral oscillation caused by the left and right swaying of the neck and the forward arching action. For feeding behaviors characterized by chewing and swallowing, the resulting mechanical vibrations are mainly transmitted along the direction of the quasi-static attitude reference vector, and are therefore accurately extracted as the reference axial vibration intensity. For behaviors characterized by neck swinging, such as sorting, arching, and driving away mosquitoes and flies, the energy is mainly concentrated in the plane perpendicular to the quasi-static attitude reference vector, and is therefore accurately quantified as the swing intensity in the reference plane direction.

[0027] The validity audit module 102 is used to determine the feeding confidence level based on the difference between the reference axial vibration intensity and the pre-generated resting vibration baseline.

[0028] In one possible implementation, the validity audit module 102 is deployed at the cloud server entry point to receive and process asynchronous sample streams uploaded from the edge of the multi-cattle system. By maintaining a reliable resting vibration benchmark and combining it with biological laws, the module performs preliminary cleaning and weighting of the original behavioral characteristics, thereby solving the problems of data temporal disorder and non-feeding noise interference.

[0029] In one possible implementation, due to network transmission latency, data packets uploaded by different servers may arrive out of order. The server maintains a global sliding window for data synchronization. The window duration is set to 5 seconds, and the step size is also 5 seconds. For each synchronization cycle, the system uses the current system time as a reference and extracts all samples whose timestamps fall within the current window.

[0030] For example, to accurately assess feeding behavior, the effectiveness audit module 102 collects a reference axial vibration intensity sequence of the target cow within a preset time period and determines the resting vibration benchmark for the day based on this sequence. After completing the temporal alignment and aggregation processing of multi-source samples, the effectiveness audit module 102 calculates the feeding confidence level for the target cow using the difference between its aggregated reference axial vibration intensity and the resting vibration benchmark for the day. The feeding confidence level is used to characterize the probability that the target cow's current action includes chewing behavior, and its value is positively correlated with the degree of difference between the reference axial vibration intensity and the resting vibration benchmark. For a target cow that is shaking its head to drive away flies, although its lateral sweeping intensity is high, because it is not accompanied by continuous chewing, the reference axial vibration intensity is close to the resting vibration benchmark, resulting in a low feeding confidence level. This invalid sway will be suppressed in subsequent calculations. For a target cow that is feeding, its chewing action produces significant axial vibration, and the reference axial vibration intensity is significantly higher than the resting vibration benchmark, resulting in a high feeding confidence level. In this case, the detected lateral sway will be regarded as an effective signal related to feeding. Through this longitudinal gating mechanism based on feeding confidence, the effectiveness audit module 102 effectively filters out non-feeding oscillations.

[0031] The group differential module 103 is used to construct the group average sweep ratio based on the reference axial vibration intensity and reference planar oscillation intensity of each cow in the same pen as the target cow.

[0032] In one possible implementation, the population difference module 103 is deployed on a cloud server to eliminate the common environmental background caused by the physical properties of the feed and the competitive feeding atmosphere in the early stage of feeding. By constructing a dynamic population benchmark, it lays the foundation for achieving accurate quantification without environmentalization.

[0033] For example, the group differential module 103 first selects cattle in an effective feeding state from the sample set output by the effectiveness audit module 102 and marks them as active individuals, ensuring that only cattle truly in a feeding state are included in the group baseline calculation. Based on this, the group differential module 103 determines the sweeping-chewing ratio of each active individual according to the reference axial vibration intensity and reference plane oscillation intensity of all active individuals at the current moment. The sweeping-chewing ratio physically characterizes how much sweeping effort an individual expends to obtain a unit amount of chewed food. After obtaining the sweeping-chewing ratio of all active individuals, the group differential module 103 performs statistical processing on the sweeping-chewing ratio of all active individuals, determining the resulting statistical value as the group average sweeping ratio. In the initial feeding stage, due to the freshness and fluffiness of the feed, the sweeping ratio of the entire herd is generally high, and the group average sweeping ratio also rises accordingly. As the feed is consumed for a period of time, the concentrate sinks and the roughage floats, increasing the difficulty of sweeping, and the group average sweeping ratio decreases accordingly. Through this real-time dynamic calculation of the population baseline, the system can accurately capture the common environmental background caused by changes in the physical properties of feed and the competitive feeding atmosphere among the population.

[0034] The excess sorting deviation determination module 104 is used to determine the excess sorting deviation of the target cattle based on the group average sweeping ratio, the sweeping and chewing ratio of the target cattle, and the feeding confidence level.

[0035] In one possible implementation, the excess sorting deviation determination module 104 is deployed on a cloud server to integrate the results of vertical gating and horizontal comparison, and finally remove environmental background and noise interference to output core indicators that characterize the actual pathological sorting behavior of individuals.

[0036] For example, the excess sorting deviation determination module 104 first determines the target cow's sweeping-chewing ratio based on the reference axial vibration intensity and the reference planar oscillation intensity. Based on this, the excess sorting deviation determination module 104 calculates the difference between the target cow's sweeping-chewing ratio and the group's average sweeping ratio. This difference, in the horizontal dimension, characterizes the degree of deviation of the target cow's sweeping intensity from the common level of the current environment. Subsequently, this difference is fused with the target cow's feeding confidence to obtain the target cow's excess sorting deviation. This process incorporates the core logic of dual difference: in the horizontal dimension, the difference between the individual sweeping-chewing ratio and the group's average sweeping ratio offsets the common environmental influences caused by feed physical properties and group competition for food; in the vertical dimension, the difference is weighted by the feeding confidence to further suppress invalid oscillation noise that is not accompanied by chewing rhythm. In the initial feeding phase, during the initial herd competition for food, even when the target cattle were feeding normally as part of the herd, their sweeping-and-chewing ratio was close to the group average, with minimal difference. Combined with a high confidence level in feeding, the excess sorting deviation was small, and the system successfully filtered out false positives caused by environmental factors. In the scenario where the target cattle were shaking their heads to avoid flies, although their sweeping-and-chewing ratio might be higher due to the large head-shaking motion, the difference was small after integrating the difference with the confidence level due to the low confidence level in feeding. The system successfully filtered out non-feeding noise. In scenarios where the target cattle exhibited pathological sorting, their sweeping-and-chewing ratio was significantly higher than the group average, and they also had a high confidence level in feeding due to continuous chewing. In this case, the excess sorting deviation showed a significant positive value, directly reflecting the individual's specific sorting intention.

[0037] The technical solution provided in the above embodiments can bring at least the following beneficial effects: This embodiment forms a complete dual-differential processing closed loop by constructing an attitude decoupling module, an effectiveness auditing module, a group differential module, and an excess sorting deviation determination module. The attitude decoupling module decomposes the original acceleration data based on the quasi-static attitude reference vector, fundamentally eliminating the impact of sensor wearing attitude uncertainty on monitoring accuracy, enabling the same algorithm to adapt to devices with different tightness and wearing angles, and significantly reducing the difficulty of on-site installation and maintenance. The effectiveness auditing module, by introducing feeding confidence, achieves longitudinal gating of non-feeding noise, which can effectively suppress invalid oscillation signals generated by non-feeding behaviors such as chasing away mosquitoes and flies and social interactions. The group differential module, by constructing the group average sweeping ratio, achieves lateral background elimination of common environmental behaviors such as group food grabbing. The excess sorting deviation determination module integrates the above processing results and finally obtains the excess sorting deviation, which represents the abnormal sorting behavior of an individual relative to the group. This indicator is stripped of sensor error, non-feeding noise, and environmental background interference, specifically quantifying the individual's true pathological sorting behavior, and providing a reliable data foundation for subsequent health risk assessment. This solves the technical problem that existing technologies have low monitoring accuracy and cannot meet the actual needs of large-scale farming for precise individual health management.

[0038] In one possible implementation, the attitude decoupling module 101 includes a reference axis extraction submodule, which is used to filter the original acceleration data and determine the low-frequency component obtained after filtering as the quasi-static attitude reference vector.

[0039] In one possible implementation, a reference axis extraction submodule is deployed within a smart collar worn around the neck of the target cow. This submodule extracts quasi-static components from the raw acceleration data that change slowly with the cow's body posture, serving as a reference for subsequent orthogonal decomposition. Because the collar rotates circumferentially due to skin laxity or inertia during free movement, the measurement coordinate system of the sensor's built-in triaxial accelerometer continuously changes relative to the cow's anatomical coordinate system. Directly using the raw axial data would fail to distinguish between vertical chewing and lateral sweeping movements. To address this issue, the reference axis extraction submodule employs low-pass filtering to separate low-frequency components approximating the direction of gravity from the raw acceleration signal, constructing an absolute reference system independent of the physical orientation of the sensor housing.

[0040] For example, the reference axis extraction submodule uses a first-order low-pass filter to track the changing trend of the original acceleration signal in real time, and extracts the quasi-static attitude reference vector using the following formula 1:

[0041] Formula 1

[0042] in, express The quasi-static attitude reference vector at time t is a three-dimensional vector with the same dimensions as the acceleration, used to characterize... The direction of the slow posture of the target cow's neck is, in engineering terms, approximated by the direction of gravity; Represents the quasi-static attitude reference vector at the previous moment; express The original acceleration vector at time t; Represents the filter coefficients, which are dimensionless constants with a range of values. It is determined by both the cutoff frequency and the sampling frequency.

[0043] When t=0, set the initial conditions. = That is, during the system cold start, the original acceleration vector at the initial moment is used as the initial value of the quasi-static attitude reference vector to ensure that the recursive algorithm can be executed.

[0044] The orthogonal decomposition submodule is used to decompose the original acceleration data into instantaneous axial vibration acceleration and instantaneous planar oscillation acceleration based on the quasi-static attitude reference vector. The instantaneous axial vibration acceleration is the projection component of the original acceleration data in the direction of the quasi-static attitude reference vector; the instantaneous planar oscillation acceleration is the projection component of the original acceleration data in the plane perpendicular to the quasi-static attitude reference vector.

[0045] In one possible implementation, the orthogonal decomposition submodule and the reference axis extraction submodule are deployed together in the smart collar terminal. Based on the extracted quasi-static attitude reference vector, the original acceleration data is decomposed into two orthogonal components with clear biomechanical meaning, which respectively represent chewing behavior and sweeping behavior.

[0046] For example, after the reference axis extraction submodule outputs the quasi-static attitude reference vector at the current moment, the orthogonal decomposition submodule first calculates the difference between the original acceleration data and the quasi-static attitude reference vector to obtain the pure dynamic acceleration vector after removing the static component. Based on the biomechanical characteristics of the target cattle's feeding behavior, effective feeding is mainly driven by the periodic opening and closing of the mandible. This mechanical vibration is transmitted along the cervical spine and manifests as high-frequency vibration along the direction of the quasi-static attitude reference vector. Sorting, rooting, and swatting flies are mainly driven by neck muscles and manifest as large-amplitude swaying in a plane perpendicular to the quasi-static attitude reference vector. Accordingly, the orthogonal decomposition submodule orthogonally decomposes the pure dynamic acceleration vector into two components: instantaneous axial vibration acceleration and instantaneous planar swaying acceleration. The instantaneous axial vibration acceleration is the projection modulus of the pure dynamic acceleration vector in the direction of the quasi-static attitude reference vector. This component specifically responds to the mechanical vibration along the gravity line caused by chewing and swallowing. The instantaneous planar oscillation acceleration is the projection modulus of the pure dynamic acceleration vector in a plane perpendicular to the quasi-static attitude reference vector. This component responds to the left and right swaying and forward arching movements of the target bull's head, eliminating the interference of vertical vibration.

[0047] The feature integration submodule is used to accumulate and sum the instantaneous axial vibration acceleration and instantaneous planar oscillation acceleration within the micro-sampling period based on a preset micro-sampling period, and to determine the accumulated sum as the reference axial vibration intensity and the reference planar oscillation intensity, respectively.

[0048] In one possible implementation, the feature integration submodule, the reference axis extraction submodule, and the orthogonal decomposition submodule are jointly deployed in the smart collar terminal to convert high-frequency instantaneous signals into low-frequency statistical features, thereby reducing wireless transmission power consumption and effectively capturing the instantaneous intensity of behavior.

[0049] For example, to reduce wireless transmission power consumption and extract statistical features, the feature integration submodule integrates the instantaneous components output by the orthogonal decomposition submodule according to a preset micro-sampling period. The micro-sampling period is set to a duration sufficient to cover multiple complete chewing or sweeping cycles, effectively capturing the instantaneous intensity of the behavior. For any given micro-sampling period, the feature integration submodule calculates the cumulative sum of all instantaneous axial vibration acceleration sampling points within that period, as the reference axial vibration intensity; simultaneously, it calculates the cumulative sum of all instantaneous planar oscillation acceleration sampling points within that period, as the reference planar oscillation intensity. At the end of the period, the feature integration submodule generates a standardized neck motion sample and sends it to the cloud server. This sample includes the target cow's identification, the absolute timestamp of the end of the sampling period, the calculated reference axial vibration intensity, and the calculated reference planar oscillation intensity.

[0050] The technical solution provided in the above embodiments can bring at least the following beneficial effects: This embodiment specifically defines the attitude decoupling module. Through the synergistic effect of the reference axis extraction subunit, the orthogonal decomposition subunit, and the feature integration subunit, it achieves a precise conversion from raw acceleration data to reference axial vibration intensity and reference planar oscillation intensity with clear biomechanical meaning. The reference axis extraction subunit uses filtering to extract low-frequency components as quasi-static attitude reference vectors, which are approximated as the gravity direction in engineering, constructing an absolute reference system independent of the physical orientation of the sensor housing. Based on this reference system, the orthogonal decomposition subunit decomposes dynamic acceleration into axial vibration representing chewing and planar oscillation representing sweeping, achieving precise separation of signals in terms of physical meaning. The feature integration subunit, through cumulative summation within a micro-sampling period, reduces wireless transmission power consumption and effectively captures the instantaneous intensity of the behavior. The processing of the entire unit ensures that the input signals relied upon for subsequent analysis have consistent physical meaning, laying a solid physical foundation for the accuracy of the entire system.

[0051] In one possible implementation, the validity audit module 102 includes: a benchmark maintenance submodule, used to collect a reference axial vibration intensity sequence of the target cow within a preset resting period and determine the variance of the reference axial vibration intensity sequence.

[0052] In one possible implementation, the baseline maintenance submodule is deployed at the cloud server entry point to establish and dynamically maintain the background vibration level of the target cattle in a non-feeding state, providing a reliable reference zero point for subsequent calculation of feeding confidence. Due to the influence of factors such as sensor aging, changes in wearing tightness, and differences in environmental micro-vibrations, the resting vibration baseline needs to be dynamically updated daily to ensure its accuracy.

[0053] For example, the benchmark maintenance submodule presets a daily non-feeding period (e.g., 2:00 AM to 4:00 AM) as the benchmark acquisition window. During this period, it continuously collects reference axial vibration intensity data of the target cattle, forming a reference axial vibration intensity sequence. To assess whether the target cattle are truly at rest during this period, the benchmark maintenance submodule calculates the variance of this sequence, which characterizes the degree of fluctuation in the sequence. If the target cattle are resting or standing during this period, their reference axial vibration intensity should remain relatively stable with a small variance; if the target cattle are active due to illness, fighting, or environmental stress, their reference axial vibration intensity will fluctuate significantly with a large variance. By calculating the variance of the reference axial vibration intensity sequence, the benchmark maintenance submodule can effectively determine whether the data for this period is representative of a resting state, providing a quantitative basis for subsequent benchmark update decisions.

[0054] The benchmark maintenance submodule is used to determine the resting vibration benchmark based on the lower quartile of the reference axial vibration intensity sequence when the variance is less than or equal to a preset resting fluctuation threshold.

[0055] In one possible implementation, after completing the variance calculation, the benchmark maintenance submodule compares the variance with a preset resting fluctuation threshold and decides whether to update the resting vibration benchmark based on the comparison result, so as to ensure the reliability and accuracy of the benchmark value.

[0056] For example, the benchmark maintenance submodule compares the calculated variance with a preset resting volatility threshold. If the variance is greater than the resting volatility threshold (e.g., 0.05 g²·s²), it indicates that the target cattle are active during that period, and the data is not representative of resting conditions. In this case, the benchmark maintenance submodule abandons the update for the day and uses the previous day's resting vibration benchmark or the preset default resting benchmark as the current resting vibration benchmark. If the variance is less than or equal to the resting volatility threshold, it indicates that the target cattle are relatively still, and the data is valid. In this case, the benchmark maintenance submodule performs statistical processing on the reference axial vibration intensity sequence and extracts the lower quartile of the sequence as the resting vibration benchmark for the day. The reason for choosing the lower quartile instead of the mean is that the lower quartile can further eliminate the transient vibration interference caused by occasional head-raising or rumination movements, ensuring that the benchmark value closely matches the actual sensor noise floor. Through this mechanism combining volatility auditing and statistical updates, the benchmark maintenance submodule achieves robust maintenance of the resting vibration benchmark, laying a solid foundation for the accurate calculation of feeding confidence.

[0057] The confidence calculation submodule is used to determine the feeding confidence based on the difference between the current reference axial vibration intensity and the resting vibration benchmark, as well as a preset mapping function.

[0058] In one possible implementation, the confidence calculation submodule and the benchmark maintenance submodule are deployed together at the cloud server entry point to convert the current chewing intensity of the target cow into a feeding confidence that represents the probability that its current action includes chewing behavior, thereby achieving longitudinal gating of non-feeding noise.

[0059] For example, after completing the temporal alignment and aggregation of multi-source samples, the confidence calculation submodule obtains the reference axial vibration intensity of the target cow within the current synchronization window and calls the resting vibration baseline maintained by the baseline maintenance submodule. The confidence calculation submodule calculates the difference between the current reference axial vibration intensity and the resting vibration baseline, which reflects the degree of deviation of the target cow's current chewing intensity from the resting background level. Subsequently, the confidence calculation submodule uses a preset mapping function to map this difference into a feeding confidence level ranging from 0 to 1. This mapping function is configured to make the feeding confidence level positively correlated with the degree of difference. When the reference axial vibration intensity is close to the resting vibration baseline, the difference is small, and the feeding confidence level approaches 0, indicating that any oscillation detected at this time is likely unrelated to feeding; when the reference axial vibration intensity is significantly higher than the resting vibration baseline, the difference is large, and the feeding confidence level approaches 1, indicating that the oscillation detected at this time is highly likely related to feeding behavior. For a target cow that is head-shaking to shoo away flies, although its lateral sweeping intensity is high, the reference axial vibration intensity is close to the resting vibration baseline because it is not accompanied by continuous chewing, resulting in a feeding confidence level close to 0. This invalid oscillation will be strongly suppressed in subsequent calculations. For a target cow that is actually eating, its chewing action produces significant axial vibration, with a reference axial vibration intensity much higher than the resting vibration baseline, resulting in a feeding confidence level close to 1. In this case, any detected lateral oscillation will be considered a valid signal related to feeding. Through this longitudinal gating mechanism based on feeding confidence, the confidence calculation submodule effectively filters out non-feeding oscillations such as those caused by shooing away flies.

[0060] For example, the confidence level of feed intake for the kth cow. The following formula 2 is satisfied:

[0061] Formula 2

[0062] in, The reference axial vibration intensity of the kth cow after aggregation through the synchronization window has the same dimension as acceleration × time and is provided by the time-aligned aggregated sample. Let be the resting vibration reference for the kth cow, with the same dimensions as acceleration × time, representing the background vibration level of the cow when it is not eating; A bias constant (dimensions of acceleration × time) with the same dimensions as the reference axial vibration intensity is used to control the center position of the activation function. The specific value can be determined by the maximum fluctuation amplitude of the reference axial vibration intensity under historical resting conditions. For example, the 95th percentile or maximum value of the reference axial vibration intensity sequence of the target cattle during the resting period can be statistically analyzed and used as the reference value. The reference value ensures that minor fluctuations under normal resting conditions do not lead to false activation of feeding confidence; λ is a coefficient with the reciprocal of the reference axial vibration intensity (dimension: × ), used to control the steepness of the activation function; It is an exponential function with the natural constant e as its base.

[0063] Formula 2 uses a sigmoid function to map the difference in chewing intensity into a probability value. First, calculate... This represents the degree to which the current chewing intensity exceeds the resting baseline and activation bias; then, this difference is multiplied by the sensitivity coefficient λ and negatively taken as the exponent of the exponential function, mapping any real number to the interval (0, 1). The calculation results are consistent with... There is a positive correlation; the larger the difference, the closer the confidence level of feeding is to 1; the smaller the difference (especially when it is negative), the closer the confidence level of feeding is to 0.

[0064] The technical solution provided in the above embodiments can bring at least the following beneficial effects: This embodiment specifically defines the validity audit module. Through the cooperation of the benchmark maintenance subunit and the confidence calculation subunit, reliable calculation of feeding confidence is achieved. The benchmark maintenance subunit collects the reference axial vibration intensity sequence and determines its variance within a preset resting period. The resting vibration benchmark is updated by comparing the variance with the resting fluctuation threshold. This volatility audit mechanism can effectively prevent the benchmark from being contaminated by abnormally active data, ensuring that the resting benchmark always closely matches the actual sensor noise floor. The confidence calculation subunit calculates the feeding confidence based on the difference between the reference axial vibration intensity at the current moment and the resting vibration benchmark through a mapping function. This confidence is positively correlated with the degree of difference, and can smoothly convert chewing intensity into probability weights. Through this vertical gating mechanism, the system can accurately identify the biological law of not eating without chewing, filter out invalid oscillations that are not accompanied by chewing rhythms in the time domain, and effectively solve the signal confusion problem between non-feeding noise such as fly repellency and real sorting behavior.

[0065] In one possible implementation, the population differential module 103 includes an active screening submodule, used to mark cattle with a feeding confidence level greater than or equal to a preset validity threshold as active individuals.

[0066] In one possible implementation, the active screening submodule is deployed on a cloud server to screen individuals that are truly in a feeding state from the sample set output by the validity audit module 102, providing a clean data foundation for the subsequent construction of the population benchmark.

[0067] For example, after the validity audit module 102 calculates the feeding confidence level for each cow at the current moment, the active screening submodule obtains the feeding confidence level of all cows in the same pen and compares the feeding confidence level of each cow with a preset validity threshold. This validity threshold is a pre-set threshold (e.g., 0.1; when the confidence level reaches or exceeds 0.1, it indicates that the target cow has begun to exhibit certain chewing characteristics and has a relatively high probability of engaging in feeding behavior), used to determine whether a cow is currently in a valid feeding state. When a cow's feeding confidence level is greater than or equal to the validity threshold, it indicates that the cow has a high probability of currently performing chewing feeding behavior, and the active screening submodule marks it as an active individual; when the feeding confidence level is less than the validity threshold, it indicates that the cow may currently be in a non-feeding state such as resting, ruminating, or simply swatting flies, and the active screening submodule marks it as an inactive individual, not including it in the subsequent construction of the group baseline.

[0068] The benchmark construction submodule is used to calculate the sweeping and chewing ratio of each active individual based on the reference axial vibration intensity and reference planar oscillation intensity of all active individuals at the current moment.

[0069] In one possible implementation, the baseline construction submodule and the active screening submodule are deployed together on a cloud server to calculate the chewing ratio of each active individual based on the behavioral characteristic data of active individuals, providing basic data for subsequent group baseline statistics.

[0070] For example, after the active screening submodule completes the marking of active individuals, the benchmark construction submodule obtains the reference axial vibration intensity and reference planar oscillation intensity of all active individuals at the current moment. For each active individual, the benchmark construction submodule calculates the sweeping-chewing ratio based on its reference planar oscillation intensity and reference axial vibration intensity. The sweeping-chewing ratio is calculated by dividing the reference planar oscillation intensity by the reference axial vibration intensity. This ratio physically represents how much sweeping effort the individual expends to obtain a unit amount of chewed food. Simultaneously, to avoid individual outliers skewing subsequent population statistical characteristics, the benchmark construction submodule also limits the initial calculated ratio to ensure it does not exceed a preset upper limit. Through these processes, the benchmark construction submodule determines a stable and reliable sweeping-chewing ratio for each active individual, laying the foundation for the subsequent statistical calculation of the population average sweeping ratio.

[0071] For example, this represents the sweeping and chewing ratio of the p-th cow. The following formula 3 is satisfied:

[0072] Formula 3

[0073] in, Let be the oscillation intensity of the p-th cow in the reference plane within the current synchronization window, with the same dimensions as acceleration × time; Let be the reference axial vibration intensity of the p-th cow within the current synchronization window, with the same dimensions as acceleration × time; It is a preset threshold for extremely small positive numbers (e.g., 0.01), with the same dimensions as acceleration × time, used to prevent the denominator from being zero; The upper limit of the preset ratio (e.g., 10.0, the value is determined based on the ratio of sweeping intensity to chewing intensity in normal feeding behavior, which can effectively prevent individual abnormal data from skewing the statistical characteristics of the group) is a dimensionless constant used to limit the calculation results.

[0074] Formula 3 involves two key processing steps: the first step is to protect the denominator by taking... and The larger value in the first step is used as the denominator to ensure that even if the chewing intensity is close to zero, it will not lead to a division-by-zero error or the calculation result tending to infinity; the second step is to perform numerator division, dividing the oscillation intensity by the protected denominator to obtain the initial ratio; the third step is to perform amplitude limiting processing, taking the initial ratio and... The smaller value in the result is taken as the final result. The calculation result is compared with... There is a positive correlation with There is a negative correlation.

[0075] The baseline construction submodule is used to determine the population average sweep ratio based on the sweep-chew ratio of all active individuals.

[0076] In one possible implementation, after the baseline construction submodule completes the calculation of the sweeping and chewing ratio of all active individuals, it further performs statistical processing on these ratios to obtain the group average sweeping ratio that can represent the common behavior of the current environment.

[0077] For example, the baseline construction submodule obtains the sweeping and chewing ratios of all active individuals at the current moment, forming a set of ratios. To eliminate the influence of individual extreme values ​​on the group baseline, the baseline construction submodule performs statistical processing on this set, calculates the median of the sweeping and chewing ratios of all active individuals, and determines this median as the current group average sweeping ratio. The reason for choosing the median instead of the arithmetic mean is that, in pathological monitoring scenarios, a few sick individuals may exhibit extreme sweeping behavior. If these outliers are calculated using the average, it will raise the group baseline, causing the baseline itself to deviate from the normal majority; while the median has stronger statistical robustness and can more realistically represent the common behavioral baseline of normal individuals. In the early stages of feeding, due to the freshness and fluffiness of the feed, the sweeping ratio of all cattle in the herd is generally high, and the group average sweeping ratio also rises accordingly; after the feed has been consumed for a period of time, the concentrate sinks and the roughage floats, increasing the difficulty of sweeping, and the group average sweeping ratio decreases accordingly. Through this real-time dynamic calculation of the population baseline, the baseline construction submodule can accurately capture the common environmental background caused by changes in feed physical properties and the competitive feeding atmosphere of the group, providing a reliable horizontal comparison basis for subsequent calculation of the excess sorting deviation of individuals relative to the group.

[0078] The technical solution provided by the above embodiments can bring at least the following beneficial effects: This embodiment specifically defines the group differential module, and through the cooperation of the active screening subunit and the benchmark construction subunit, the accurate construction of the group average feed sweeping ratio is achieved. The active screening subunit marks active individuals based on the comparison results of feeding confidence and validity threshold, ensuring that only cattle that are truly in a feeding state are included in the calculation of the group benchmark, avoiding the data of inactive individuals from polluting the group characteristics. The benchmark construction subunit calculates the feed sweeping and chewing ratio of each active individual based on the reference axial vibration intensity and reference planar oscillation intensity of all active individuals, and determines the group average feed sweeping ratio based on the feed sweeping and chewing ratio of all active individuals. This process enables the group benchmark to reflect the current physical properties of the feed (such as fluffiness and moisture) and the common environmental background caused by the group's competitive feeding atmosphere at the beginning of feeding in real time. At the same time, since the group benchmark is determined by statistical methods, the influence of extreme outliers can be automatically eliminated, ensuring that the group benchmark truly represents the common behavior of the normal majority.

[0079] In one possible implementation, the excess sorting deviation determination module is specifically used to determine the excess sorting deviation of the target cow. This process can be implemented through the following S101-S102, which will be described in detail below.

[0080] S101, the excess sorting deviation determination module determines the sweeping and chewing ratio of the target cow based on the reference axial vibration intensity and the reference planar oscillation intensity of the target cow.

[0081] In one possible implementation, the excess sorting deviation determination module is deployed on a cloud server to calculate the sweeping-chewing ratio required for the target cattle to be compared with the herd benchmark. This ratio is the basic input parameter for subsequent calculation of excess sorting deviation.

[0082] For example, the excess sorting deviation determination module obtains the reference axial vibration intensity and reference planar oscillation intensity of the target cow at the current moment. Similar to the process of the benchmark construction submodule calculating the sweeping-chewing ratio for active individuals, the excess sorting deviation determination module also employs a numerical protection mechanism to ensure the stability of the calculation. Specifically, to prevent calculation anomalies caused by the reference axial vibration intensity approaching zero, the excess sorting deviation determination module introduces a preset minimum positive threshold (e.g., 0.01), takes the larger value between the reference axial vibration intensity and this threshold as the denominator, and then divides the reference planar oscillation intensity by this denominator to obtain the initial ratio of the target cow. To further eliminate the influence of occasional outliers, the excess sorting deviation determination module compares this initial ratio with a preset upper limit of the ratio (e.g., 10.0), and takes the smaller value as the final sweeping-chewing ratio of the target cow. The sweeping-chewing ratio physically represents the sweeping cost borne by the target cow at the current moment in order to obtain a unit amount of chewing. Its value directly reflects the intensity of sweeping behavior relative to chewing behavior, and has a completely consistent definition and dimension with the sweeping-chewing ratio used in the population benchmark, ensuring the fairness and accuracy of subsequent comparisons.

[0083] S102, the excess sorting deviation determination module determines the excess sorting deviation of the target cattle based on the difference between the target cattle's sweeping and chewing ratio and the group's average sweeping ratio, as well as the feeding confidence level.

[0084] In one possible implementation, the excess sorting deviation determination module is deployed on a cloud server. After obtaining the sweeping and chewing ratio of the target cattle, it integrates horizontal group comparison and vertical confidence weighting, and introduces a non-negative constraint mechanism to finally output the core indicators that are stripped of environmental background and noise interference and focus only on the excess part.

[0085] For example, the excess sorting deviation determination module first obtains the current-moment average sweeping ratio of the herd constructed by the herd differentiation module. This ratio represents the typical sweeping level exhibited by a normally consuming individual under the current feed conditions and environmental atmosphere. Simultaneously, the excess sorting deviation determination module obtains the current-moment feeding confidence score of the target cow calculated by the validity audit module. This confidence score, in the vertical dimension, characterizes the probability that the target cow's current action includes chewing behavior. Based on this, the excess sorting deviation determination module performs the following calculation steps: First, it calculates the difference between the target cow's sweeping-chewing ratio and the herd average sweeping ratio. This difference, in the horizontal dimension, characterizes the degree of deviation of the target cow's sweeping intensity from the common level of the current environment. Second, it takes the larger value between this difference and 0, i.e., only positive deviations are retained; if the difference is negative, it is set to zero. The introduction of this non-negative constraint has a clear physical meaning: this scheme focuses on the portion of individual sweeping intensity exceeding the herd benchmark; sweeping behavior below the herd benchmark is not considered pathological and should not be included in subsequent risk accumulation. The third step is to multiply the difference after non-negation by the feeding confidence level of the target cattle to obtain the excess sorting bias of the target cattle.

[0086] For example, excess sorting deviation The following formula 4 is satisfied:

[0087] Formula 4

[0088] in, Let be the feed chewing ratio of the kth cow, which is a dimensionless ratio. The group average sweep ratio is a dimensionless ratio. Let be the confidence level of the kth cow's feeding, and be a dimensionless probability value with a range of (0, 1). Pick For values ​​larger than 0, a non-negativity constraint is introduced.

[0089] The technical solution provided in the above embodiments can bring at least the following beneficial effects: This embodiment defines the specific implementation method of the excess sorting deviation determination module, and clarifies the calculation logic of determining excess sorting deviation based on the difference between the target cow's sweeping and chewing ratio and the group's average sweeping ratio, as well as the feeding confidence level. This calculation process fully embodies the core inventive concept of dual difference: In the horizontal dimension, the difference between the individual sweeping and chewing ratio and the group's average sweeping ratio offsets the common environmental influences caused by feed physical properties and the group's competitive feeding atmosphere, enabling the system to specifically quantify the individual's sorting intention; in the vertical dimension, the difference is weighted by the feeding confidence level, further suppressing invalid oscillation noise that is not accompanied by chewing rhythm. The synergistic effect of the two makes the final excess sorting deviation a net value after removing environmental background and noise interference. Only when the individual shows a sweeping intensity significantly higher than the group and the action is confirmed as a feeding process will it show a significant positive value, realizing the accurate capture of pathological sorting behavior.

[0090] In one possible implementation, the baseline construction submodule is also used to determine whether the number of active individuals is less than a preset minimum sample size threshold. This process can be specifically implemented through the following S201-S202, which will be explained in detail below.

[0091] S201. Determine whether the number of active individuals is less than the preset minimum sample size threshold.

[0092] In one possible implementation, the baseline construction submodule is also used to determine whether the number of active individuals is less than a preset minimum sample size threshold. If the number of active individuals is less than the preset minimum sample size threshold, the most recent valid historical population average scan ratio or the preset default population baseline is determined as the population average scan ratio.

[0093] For example, after the active screening submodule marks the active individuals at the current moment, the baseline construction submodule obtains the number of all active individuals at the current moment. The baseline construction submodule compares the number of all active individuals with a preset minimum sample size threshold to determine whether the current number of active individuals meets the minimum requirements for statistical calculation. The minimum sample size threshold is a preset constant, for example, it can be set to 5. This value is determined based on the minimum sample size required for median calculation in statistics, ensuring that the population baseline has sufficient representativeness. If the number of all active individuals is greater than or equal to the preset minimum sample size threshold, it indicates that the current number of active individuals is sufficient, and the baseline construction submodule continues to execute the normal process, calculating the population average sweeping ratio based on the sweeping and chewing ratio of the active individuals at the current moment. If the number of all active individuals is less than the preset minimum sample size threshold, it indicates that the current number of active individuals is insufficient, and reliable population characteristics cannot be extracted from the current data. At this time, the baseline construction submodule triggers a cold start or insufficient sample handling process.

[0094] S202. If the number of active individuals is less than the preset minimum sample size threshold, the most recent valid historical average population scanning ratio or the preset default population baseline is determined as the population average scanning ratio.

[0095] In one possible implementation, when the baseline construction submodule determines that the number of active individuals is insufficient, an alternative baseline determination mechanism is initiated to ensure that the system can output a stable population average sweep ratio under any operating condition, thus guaranteeing the continuity and robustness of the algorithm.

[0096] For example, when the number of all active individuals is less than a preset minimum sample size threshold, the baseline construction submodule first attempts to call the most recently valid historical average feed sweeping ratio. This historical baseline is the average feed sweeping ratio calculated and stored by the system at a previous point in time (e.g., the last feeding peak), reflecting the common behavioral characteristics of the enclosure under normal feeding conditions. If the system is running for the first time and no historical data is available, or the historical data has expired due to the passage of time, the baseline construction submodule calls the preset default population baseline. This default baseline is the default value set during system initialization, for example, it can be set to 0.5. This value is determined based on the normal feed sweeping ratio calibrated in previous experiments, representing the feed sweeping-chewing ratio level under normal circumstances. The baseline construction submodule uses the historical baseline or default baseline obtained from the call as the current average feed sweeping ratio for use by the subsequent excess sorting deviation determination module 104.

[0097] The technical solution provided by the above embodiments can bring at least the following beneficial effects: This embodiment limits the handling method of the baseline construction subunit when there are insufficient samples. By introducing a cold start strategy, it ensures the stable operation of the system under various working conditions. When the number of active individuals is less than the preset minimum sample size threshold (commonly seen during non-feeding periods or when the system has just started), the system cannot extract reliable population characteristics from the current data. At this time, the most recent valid historical population average sweep ratio or the preset default population baseline is called as the current population average sweep ratio. This mechanism avoids the failure or deviation of population baseline calculation due to insufficient statistical samples, ensures that the algorithm will not be interrupted or crashed under any working condition, and at the same time ensures that it can quickly switch to the real-time calculated population baseline during peak feeding periods, taking into account both the stability and accuracy of the system.

[0098] In one possible implementation, the baseline construction submodule is also used to determine the sweeping-chewing ratio of each active individual. This process can be specifically implemented through the following S301-S302, which will be described in detail below.

[0099] S301. Determine the initial ratio of each active individual based on the reference plane oscillation intensity, the reference axial vibration intensity, and the preset minimum positive threshold of each active individual.

[0100] In one possible implementation, after obtaining the reference axial vibration intensity and reference planar oscillation intensity of all active individuals at the current moment, the benchmark construction submodule calculates the initial ratio for each active individual. This initial ratio is used to characterize the relative magnitude of the individual's original sweeping intensity with respect to its chewing intensity at the current moment.

[0101] For example, for a given active individual, the baseline construction submodule obtains its current reference planar oscillation intensity and reference axial vibration intensity. To prevent calculation anomalies caused by the reference axial vibration intensity approaching zero, the baseline construction submodule introduces a preset, extremely small positive threshold ε (e.g., 0.01), which is much smaller than the reference axial vibration intensity value under normal feeding conditions. The baseline construction submodule first compares the reference axial vibration intensity with ε, takes the larger value as the denominator, and then divides the reference planar oscillation intensity by the denominator to obtain the initial ratio for the active individual. The initial ratio physically represents the original scraping cost borne by the individual at the current moment to obtain a unit amount of chewed food; its magnitude directly reflects the intensity of scraping behavior relative to chewing behavior.

[0102] S302. The initial ratio of each active individual is subjected to amplitude limiting processing, and the value obtained after amplitude limiting processing is determined as the chewing ratio of each active individual.

[0103] In one possible implementation, after obtaining the initial ratio of each active individual, the baseline construction submodule further performs amplitude limiting processing on it to eliminate the influence of individual extreme outliers on subsequent population statistical characteristics, thus obtaining the final sweeping and chewing ratio used for population baseline calculation.

[0104] For example, for a given active individual, the baseline construction submodule compares its initial ratio with a preset upper limit (e.g., 10.0). If the initial ratio is less than or equal to the preset upper limit, it indicates that the initial ratio is within the normal range, and the baseline construction submodule directly determines the initial ratio as the active individual's sweeping-chewing ratio. If the initial ratio is greater than the preset upper limit, it indicates that the initial ratio has deviated significantly from the normal range due to sensor malfunction, occasional violent movements, or other abnormal factors, and the baseline construction submodule uses the preset upper limit as the active individual's sweeping-chewing ratio, i.e., truncating abnormally high values. This limiting mechanism effectively prevents the risk of individual extreme outliers skewing the subsequent population median statistical results, ensuring the robustness of the population average sweeping ratio. The upper limit of the ratio is determined based on statistical analysis of the ratio between sweeping intensity and chewing intensity in normal feeding behavior, which can effectively remove outliers while retaining the true characteristics of high-intensity sorting behavior.

[0105] The technical solution provided by the above embodiments can bring at least the following beneficial effects: This embodiment limits the numerical protection mechanism for the benchmark construction subunit when calculating the chewing ratio. By introducing a minimum positive threshold and amplitude limiting processing, it effectively prevents calculation anomalies and outlier interference. Specifically, the reference plane oscillation intensity is divided by the larger of the reference axial vibration intensity and the preset minimum positive threshold, avoiding the problem of division by zero error or the calculation result tending to infinity that may occur when the reference axial vibration intensity is close to zero; the initial ratio is amplitude limited so that it does not exceed the preset upper limit of the ratio, preventing individual abnormal data (such as pure lateral vibration caused by sensor failure) from skewing the subsequent group statistical characteristics. This numerical protection mechanism ensures the robustness of the group average chewing ratio calculation, so that the group benchmark can truly reflect the chewing characteristics of normal feeding behavior, and provides a guarantee for the accurate calculation of subsequent excess sorting deviation.

[0106] In one possible implementation, the machine learning-based intelligent monitoring system for beef cattle feeding behavior further includes a risk decision module, used to acquire the excess sorting deviation of the target cattle. Based on time-series data of the excess sorting deviation of the target cattle changing over time, a risk index is generated to characterize the health risk of the target cattle's sorting behavior.

[0107] In one possible implementation, the risk decision-making module is deployed on a cloud server to receive the core indicators output by the excess sorting deviation determination module 104 and use them as the basic input data for subsequent health risk assessment.

[0108] For example, the excess sorting deviation determination module 104 calculates the excess sorting deviation for each cow in real time and transmits the calculation results to the risk decision module. The risk decision module maintains a time-series data buffer for each cow in the pen, storing its historical excess sorting deviation values ​​in chronological order. For the target cow, the risk decision module continuously acquires its excess sorting deviation at each moment and appends the new deviation value to the cow's time-series data buffer. This buffer maintains historical data within a certain time window according to the first-in, first-out principle, ensuring the amount of data required for trend analysis while avoiding unlimited storage pressure. Through this mechanism, the risk decision module establishes a complete time-series dataset of the target cow's excess sorting deviation, laying the data foundation for subsequent behavioral pattern analysis and risk assessment.

[0109] The risk decision-making module is used to generate a risk index that characterizes the health risks of the target cattle's sorting behavior based on time-series data of the target cattle's excess sorting deviation over time.

[0110] In one possible implementation, after accumulating sufficient time-series data, the risk decision-making module conducts an in-depth analysis of the target cattle's excess sorting deviation, comprehensively assesses its health risk from two dimensions: the intensity and evolution trend of sorting behavior, and finally outputs a quantitative risk index.

[0111] For example, the risk decision-making module first performs session segmentation on the time-series data of the target cattle. Data sequences in which the excess sorting deviation continuously exceeds a preset noise threshold over multiple consecutive time periods are encapsulated into independent feeding sessions, and invalid sessions with too short a duration are filtered out to ensure that each feeding session participating in the analysis has sufficient statistical significance. For each complete feeding session, the risk decision-making module extracts features from two dimensions: first, the cumulative sorting amount, which is the sum of all excess sorting deviations within the session, representing the total amount of extra sorting costs borne by the individual during this feeding process; second, the sorting persistence trend, which quantifies the individual's behavioral response to increased sorting difficulty in the later stages of feeding by comparing the mean difference between the excess sorting deviation in the second half and the first half of the session—if the mean in the second half is lower than that in the first half, it indicates that the individual gradually abandons sorting, which is consistent with health logic; if the mean in the second half is the same as or higher than that in the first half, it indicates that the individual exhibits a persistent sorting tendency, strongly pointing to pathological characteristics.

[0112] The technical solution provided in the above embodiments can bring at least the following beneficial effects: This embodiment introduces a risk decision-making unit, transforming the core indicator of excess sorting deviation into a health risk index for sorting behavior with business value. The risk decision-making unit obtains the excess sorting deviation of the target cattle and generates a risk index based on its time-series data changing over time, realizing a leap from discrete instantaneous deviation to continuous health risk assessment. This processing enables the system not only to monitor the intensity of individual sorting behavior in real time, but also to assess its cumulative effect and evolution pattern over a period of time, providing a comprehensive quantitative indicator for subsequent early warning decisions. By introducing the risk index, the system can transform the underlying sensor signals into health information that is easy for livestock managers to understand and use, enhancing the practical value of the system.

[0113] In one possible implementation, the risk decision module includes a session encapsulation submodule, which encapsulates data sequences that meet preset feeding behavior conditions into feeding sessions based on time-series data of the target cattle’s excess sorting deviation.

[0114] In one possible implementation, the session encapsulation submodule is deployed within the risk decision module to segment the continuous excess sorting deviation data stream into biologically significant independent feeding events, providing complete data units for subsequent trend analysis.

[0115] For example, the risk decision module continuously stores newly calculated excess sorting deviation values ​​in the time-series data buffer maintained for the target cattle. The session encapsulation submodule scans this time-series data sequence in real time, identifying data segments where the excess sorting deviation is greater than a preset noise threshold (for example, the preset noise threshold is 0.05) for multiple consecutive time periods. When a continuous data segment meeting the above conditions is detected, the session encapsulation submodule encapsulates all excess sorting deviation values ​​from the beginning to the end of the data segment into a feeding session, which corresponds to one complete feeding process of the target cattle. During the encapsulation process, the session encapsulation submodule simultaneously records the start time, end time, and all deviation values ​​included in the session. After encapsulation is completed, the session encapsulation submodule performs quality screening on the encapsulated sessions, filtering out feeding sessions whose duration is less than the preset minimum session length. For example, only sessions lasting more than 50 seconds are retained for subsequent analysis to avoid fragmented data caused by brief interruptions such as head raising or drinking interfering with the accuracy of trend analysis. Through the above processing, the session encapsulation submodule transforms discrete time-series data into feeding sessions with clear boundaries, providing structured input data for subsequent calculation of sorting persistence trends.

[0116] The trend calculation submodule is used to determine the sorting persistence trend of the target cattle during the feeding session based on the statistical difference between the excess sorting deviation in the first half and the second half of the time series within the feeding session.

[0117] In one possible implementation, the trend calculation submodule is deployed within the risk decision module to analyze the evolution of the target cattle's sorting behavior during a single feeding process and quantify the degree to which it maintains sorting in the later stages of feed stratification.

[0118] For example, for each feeding session output by the session encapsulation submodule, the trend calculation submodule first divides the time series within the session into two equal segments in chronological order: the first half corresponds to the initial feeding period, and the second half corresponds to the later feeding period. The trend calculation submodule calculates the mean of all excess sorting deviations in the first half of the time series and the mean of all excess sorting deviations in the second half of the time series. Subsequently, the trend calculation submodule subtracts the mean of the second half from the mean of the first half to obtain the sorting persistence trend of the feeding session. The sign and magnitude of this trend value have a clear biological interpretation: if the trend value is negative, it indicates that as feeding progresses, the sinking of concentrate and the rising of roughage increase the difficulty of sorting, and the target cattle gradually abandon sorting behavior, which conforms to the health logic of minimizing energy; if the trend value is positive or close to zero, it indicates that even in the later feeding period, when the difficulty of sorting increases, the target cattle still maintain or even intensify their sorting intensity. This "stubborn" behavior that violates the logic of energy consumption strongly points to a pathological craving for concentrate and is an important indicator of the risk of rumen acidosis. Through this analysis, the trend calculation submodule transforms the original excess sorting deviation sequence into a sorting persistence trend with clear behavioral semantics, providing key dimensional features for risk synthesis.

[0119] The risk synthesis submodule is used to synthesize the risk index of the target cattle in the feeding session based on the cumulative amount of excess sorting deviation and the sorting persistence trend during the feeding session.

[0120] In one possible implementation, the risk synthesis submodule is deployed within the risk decision module to integrate the intensity and trend dimensions of sorting behavior into a comprehensive quantitative indicator of health risk.

[0121] For example, for each feeding session, the risk synthesis submodule first calculates the cumulative sum of all excess sorting deviations within that session to obtain the cumulative sorting amount, which represents the total amount of additional sorting costs borne by the target cow during that feeding session. Simultaneously, the risk synthesis submodule obtains the sorting persistence trend of that session output by the trend calculation submodule. Based on this, the risk synthesis submodule fuses the cumulative sorting amount with the sorting persistence trend. Specifically, using the cumulative sorting amount as the base value, if the sorting persistence trend is positive, the base value is amplified according to a preset penalty coefficient (e.g., 1.5); if the sorting persistence trend is negative or zero, the base value remains unchanged. Through this multiplicative fusion method, the risk index can reflect the amplification effect of persistent trends on the base risk: when an individual only exhibits sorting behavior in the early stages of feeding but abandons it later, the risk index only reflects its total sorting amount; when an individual exhibits persistent sorting that continues into the later stages of feeding, the risk index is significantly amplified, thereby accurately identifying individuals with pathological characteristics. The resulting risk index is a dimensionless comprehensive indicator, and its value range is related to the sorting accumulation and penalty coefficient.

[0122] For example, quantitative indicators of health risks The following formula 5 is satisfied:

[0123] Formula 5

[0124] in, It is the cumulative sum (cumulative amount) of excess sorting deviation within a single feeding session, which is a dimensionless net value. Its physical meaning is the total amount of extra sorting cost borne by the individual in this feeding session. The sorting persistence trend in a single feeding session is a dimensionless net value. The preset penalty coefficient is a dimensionless positive constant used to control the amplification factor of persistent trends on risk. This indicates that positive value operations are performed, and a penalty is introduced only when the trend is positive.

[0125] Formula 5 uses a multiplicative structure to integrate sorting accumulation and persistence trends. First, it calculates... If the trend is negative or zero, the factor is 1; if the trend is positive, the factor is greater than 1 and positively correlated with the trend value. The cumulative sorting volume is then multiplied by this factor to obtain the risk index. The calculation results are... There is a positive correlation with The positive values ​​of are positively correlated; This is a quantitative value for health risk that takes into account both the total sorting volume and persistent trends.

[0126] The technical solution provided by the above embodiments can bring at least the following beneficial effects: This embodiment specifically defines the risk decision-making unit, and through the synergistic effect of the session encapsulation subunit, trend calculation subunit, and risk synthesis subunit, it achieves a refined evaluation of sorting behavior. The session encapsulation subunit encapsulates the data sequence that meets the preset feeding behavior conditions into a feeding session, and aggregates fragmented time slices into biologically meaningful complete feeding events, providing a complete data unit for evaluating a single feeding process. The trend calculation subunit determines the sorting persistence trend based on the statistical difference between the excess sorting deviation of the first half and the second half of the time series within the feeding session, which can quantify the behavioral evolution of individuals in the feed stratification process—if the sorting persistence trend is negative, it indicates that the individual gradually abandons sorting as the sorting difficulty increases, which is consistent with health logic; if the sorting persistence trend is positive, it indicates that the individual maintains or even intensifies sorting in the later stages of feeding, and this "stubbornness" that violates the energy consumption logic strongly points to a pathological craving for concentrate. The risk synthesis subunit combines the cumulative sorting volume and the sorting persistence trend to synthesize a risk index, so that the basic risk is determined by the total sorting volume, but if an individual shows a persistent trend, the risk value is significantly amplified, thus achieving accurate identification of pathological sorting behavior.

[0127] In one possible implementation, the risk decision module is also used to generate adjustment instructions for the feed processing technology. This process can be specifically implemented through the following S401-S402, which are described in detail below:

[0128] S401. When the risk index of the target cow exceeds the preset individual risk threshold, generate health warning information for the target cow.

[0129] In one possible implementation, the risk decision-making module is deployed on a cloud server to monitor the risk index of each cow in real time and issue timely warnings to the breeding managers when abnormal risks are detected, thereby achieving individual-level health management.

[0130] For example, the risk decision-making module maintains a risk index for the target cattle across multiple consecutive feeding sessions. This risk index can be filtered using methods such as sliding windows or exponential smoothing to obtain a stable risk assessment value. The risk decision-making module compares this assessment value with a preset individual risk threshold, which can be calibrated based on the farm's historical disease data, veterinary experience, or breed characteristics; for example, it can be set to 100. When the target cattle's risk index exceeds this threshold multiple times consecutively, or when a single risk index significantly exceeds the threshold, the risk decision-making module determines that the cattle have a high risk of rumen acidosis and immediately generates a health warning message for that target cattle. This warning message can include key data such as the target cattle's identification, risk index value, time of abnormal occurrence, and duration, and can also incorporate trend analysis results to label behavioral characteristics such as persistent sorting. The risk decision-making module pushes this warning message to the terminal devices of veterinarians or farm managers, such as computer management systems or on-site displays, prompting managers to promptly conduct clinical examinations, isolate and observe the cattle, or adjust the feeding plan.

[0131] S402. When the risk index distribution of multiple cattle in the same pen as the target cattle meets the preset group risk conditions within a preset time period after feeding, an adjustment instruction for the feed processing technology is generated.

[0132] In one possible implementation, the risk decision module is also used to analyze the distribution characteristics of the risk index at the group level. When a high risk is detected in the entire group, it infers that there may be problems with the feed processing technology and generates corresponding process adjustment instructions to achieve closed-loop feedback from individual monitoring to production management.

[0133] For example, the risk decision-making module uses the first peak feeding period after each feeding as the analysis window, such as 1-2 hours after feeding, to obtain the risk index of all cattle in the pen during that period. The risk decision-making module performs statistical analysis on these risk indices to determine whether their distribution meets the preset group risk conditions. These group risk conditions may include multiple dimensions, such as: the proportion of cattle with risk indices exceeding individual risk thresholds exceeding a certain percentage (e.g., above 70%). When the risk index distribution is detected to meet the group risk conditions, the risk decision-making module determines that the current problem does not originate from the pathological characteristics of a few individuals, but is caused by common environmental factors. Specifically, it is highly likely that there are defects in the feed processing technology, such as uneven mixing of total mixed rations, insufficient mixing time leading to inadequate mixing of concentrates and roughage, or excessively low feed moisture making it easy to sort. Based on this judgment, the risk decision-making module generates adjustment instructions for the feed processing technology. These instructions may include specific process parameter suggestions, such as increasing the mixing time by 5 minutes, adding molasses to adjust the moisture, or checking the wear of the mixer blades. The risk decision-making module sends the adjustment instruction to the central control system of the feed processing workshop or the terminal equipment of the relevant responsible personnel, guiding them to optimize the feed processing technology in a timely manner.

[0134] The technical solution provided in the above embodiments can bring at least the following beneficial effects: This embodiment limits the decision output function of the risk decision-making unit, and realizes a complete closed loop from monitoring to management through a two-level feedback mechanism of individual-level health early warning and group-level process instructions. At the individual level, when the risk index of a target cow exceeds a preset individual risk threshold, a health early warning message for that target cow is generated and pushed to the veterinary terminal, enabling farm managers to pay attention to and intervene in individuals at risk in a timely manner. At the group level, when the risk index distribution of multiple cows in the same pen within a preset time period after feeding meets the preset group risk conditions, an adjustment instruction for the feed processing technology is generated and sent to the feed processing workshop to guide the adjustment of mixing time or humidity. This two-level feedback mechanism not only focuses on individual health, but also uses the statistical characteristics of group behavior to reverse evaluate the feed processing technology, realizing the closed-loop value from individual monitoring to production management, and providing complete technical support for precision farming.

[0135] Please see Figure 2This illustration shows a schematic diagram of a machine learning-based intelligent monitoring device 200 for beef cattle feeding behavior according to an embodiment of the present invention. The device includes: a processing unit 201, used to decompose raw acceleration data into reference axial vibration intensity and reference planar oscillation intensity based on a quasi-static posture reference vector of the target cattle; the quasi-static posture reference vector is extracted from raw acceleration data collected by sensors worn on the neck of the target cattle; the reference axial vibration intensity is used to characterize the cattle's chewing behavior; the reference planar oscillation intensity is used to characterize the cattle's sweeping behavior; the target cattle is any cattle in the pen; based on the difference between the reference axial vibration intensity and a pre-generated resting vibration baseline, a feeding confidence level is determined; the feeding confidence level is used to characterize the probability that the target cattle's current action includes chewing behavior; based on the reference axial vibration intensity and reference planar oscillation intensity of each cattle in the same pen as the target cattle, a group average sweeping ratio is constructed; based on the group average sweeping ratio, the target cattle's sweeping-chewing ratio, and the feeding confidence level, an excess sorting deviation is determined to characterize the target cattle exhibiting abnormal sorting behavior relative to the group.

[0136] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0137] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

Claims

1. A machine learning-based intelligent monitoring system for beef cattle feeding behavior, characterized in that, include: The attitude decoupling module is used to decompose the raw acceleration data into reference axial vibration intensity and reference planar oscillation intensity based on the quasi-static attitude reference vector of the target cow. The quasi-static attitude reference vector is extracted from the raw acceleration data collected by a sensor worn on the neck of the target cow; the reference axial vibration intensity is used to characterize the cow's chewing behavior. The reference plane oscillation intensity is used to characterize the cattle's sweeping behavior; The target cow is any cow in the pen; The validity audit module is used to determine the feeding confidence level based on the difference between the reference axial vibration intensity and the pre-generated resting vibration baseline; the feeding confidence level is used to characterize the probability that the target cow's current action includes chewing behavior; The group differential module is used to construct the group average sweeping ratio based on the reference axial vibration intensity and reference planar oscillation intensity of each cow in the same pen as the target cow; The excess sorting deviation determination module is used to determine the excess sorting deviation, which characterizes the abnormal sorting behavior of the target cattle relative to the group, based on the average sweeping ratio of the herd, the sweeping and chewing ratio of the target cattle, and the feeding confidence level. The excess sorting deviation determination module is specifically used to: determine the sweeping and chewing ratio of the target cow based on the reference axial vibration intensity and the reference planar oscillation intensity of the target cow; and determine the excess sorting deviation of the target cow based on the difference between the sweeping and chewing ratio of the target cow and the average sweeping ratio of the herd, as well as the feeding confidence level.

2. The intelligent monitoring system for beef cattle feeding behavior based on machine learning according to claim 1, characterized in that, The attitude decoupling module includes: The reference axis extraction submodule is used to filter the original acceleration data and determine the low-frequency component obtained after filtering as the quasi-static attitude reference vector. The orthogonal decomposition submodule is used to decompose the original acceleration data into instantaneous axial vibration acceleration and instantaneous planar oscillation acceleration based on the quasi-static attitude reference vector; the instantaneous axial vibration acceleration is the projection component of the original acceleration data in the direction of the quasi-static attitude reference vector; the instantaneous planar oscillation acceleration is the projection component of the original acceleration data in a plane perpendicular to the quasi-static attitude reference vector. The feature integration submodule is used to accumulate and sum the instantaneous axial vibration acceleration and the instantaneous planar oscillation acceleration within the micro-sampling period based on a preset micro-sampling period, and to determine the accumulated sum as the reference axial vibration intensity and the reference planar oscillation intensity, respectively.

3. The intelligent monitoring system for beef cattle feeding behavior based on machine learning according to claim 1, characterized in that, The validity audit module includes: The benchmark maintenance submodule is used to collect the reference axial vibration intensity sequence of the target cow during a preset resting period and determine the variance of the reference axial vibration intensity sequence. The benchmark maintenance submodule is used to determine the resting vibration benchmark based on the lower quartile of the reference axial vibration intensity sequence when the variance is less than or equal to a preset resting fluctuation threshold. The confidence calculation submodule is used to determine the feeding confidence based on the difference between the reference axial vibration intensity at the current moment and the resting vibration benchmark, and a preset mapping function; the feeding confidence is positively correlated with the degree of difference.

4. The intelligent monitoring system for beef cattle feeding behavior based on machine learning according to claim 1, characterized in that, The population difference module includes: The active screening submodule is used to mark cattle with a feeding confidence level greater than or equal to a preset validity threshold as active individuals; The benchmark construction submodule is used to calculate the sweeping and chewing ratio of each active individual based on the reference axial vibration intensity and reference planar oscillation intensity of all active individuals at the current moment. The baseline construction submodule is used to determine the population average sweep ratio based on the sweep-chewing ratio of all active individuals.

5. The intelligent monitoring system for beef cattle feeding behavior based on machine learning according to claim 4, characterized in that, The benchmark construction submodule is also used for: Determine whether the number of active individuals is less than a preset minimum sample size threshold; If the number of active individuals is less than the preset minimum sample size threshold, the most recent valid historical average population scan ratio or the preset default population benchmark will be determined as the average population scan ratio.

6. The intelligent monitoring system for beef cattle feeding behavior based on machine learning according to claim 4, characterized in that, The benchmark construction submodule is also used for: The initial ratio of each active individual is determined based on the reference plane oscillation intensity, the reference axial vibration intensity, and a preset minimum positive threshold. The initial ratio is used to characterize the relative magnitude of the initial sweeping chewing intensity of an active individual with respect to its chewing intensity; The initial ratio of each active individual is subjected to a limiting process, and the value obtained after the limiting process is determined as the chewing ratio of each active individual; the limiting process is used to make the initial ratio less than or equal to a preset upper limit of the ratio.

7. The intelligent monitoring system for beef cattle feeding behavior based on machine learning according to claim 1, characterized in that, The system also includes: The risk decision-making module is used to obtain the excess sorting deviation of the target cow; The risk decision module is used to generate a risk index that characterizes the health risk of the target cow's sorting behavior based on time-series data of the excess sorting deviation of the target cow over time.

8. The intelligent monitoring system for beef cattle feeding behavior based on machine learning according to claim 7, characterized in that, The risk decision-making module includes: The session encapsulation submodule is used to encapsulate data sequences that meet preset feeding behavior conditions into feeding sessions based on the time-series data of the excess sorting deviation of the target cattle. The trend calculation submodule is used to determine the sorting persistence trend of the target cattle in the feeding session based on the statistical difference between the excess sorting deviation in the first half of the time series and the second half of the time series within the feeding session. The risk synthesis submodule is used to synthesize the risk index of the target cattle in the feeding session based on the cumulative amount of excess sorting deviation and the sorting persistence trend during the feeding session.

9. The intelligent monitoring system for beef cattle feeding behavior based on machine learning according to claim 7, characterized in that, The risk decision-making module is also used for: When the risk index of the target cow exceeds a preset individual risk threshold, a health warning message is generated for the target cow. When the risk index distribution of multiple cattle in the same pen as the target cattle meets the preset group risk conditions within a preset time period after feeding, an adjustment instruction for the feed processing technology is generated.

Images