Live pig breeding management system and method based on Internet of Things

Abstract

The invention relates to a pig breeding management system and method based on the Internet of Things, and relates to the technical field of Internet of Things breeding management, and the system comprises a healthy sample collection module which is used for collecting Internet of Things sample monitoring information of healthy pig ingestion; the rhythm feature construction module is used for extracting double-level features and constructing rhythm features; the health behavior baseline establishing module is used for establishing a dynamic health behavior baseline of the target pig; the real-time feature acquisition module is used for acquiring real-time monitoring information of the Internet of Things and acquiring real-time double-layer features and real-time rhythm features; and the abnormity identification response module is used for identifying ingestion abnormity and executing breeding management response according to a result. The problems that in traditional live pig breeding management, the ingestion state of pigs depends on manual observation, early health abnormalities are difficult to capture in time, a dynamic health benchmark is lacked, abnormal early warning lags, management efficiency is low, and breeding benefits are affected are solved.

Description

Technical Field

[0001] This application relates to the field of Internet of Things (IoT) in livestock farming management, and in particular to an IoT-based pig farming management system and method. Background Technology

[0002] As pig farming develops towards large-scale and refined operations, early monitoring and efficient management of pig health status have become key technological requirements for improving farming efficiency.

[0003] Currently, traditional pig farming management relies on manual observation of pigs' feeding status, without fully integrating IoT technology to achieve accurate collection and analysis of multi-dimensional data. This makes it impossible to capture subtle health abnormalities in feeding behavior in a timely manner, and it also lacks dynamic health behavior benchmarks as a basis for judgment. This not only easily leads to delayed identification of health problems, but also reduces the efficiency and pertinence of farming management. Summary of the Invention

[0004] This application provides an Internet of Things-based pig farming management system and method, which improves the timeliness of capturing health abnormalities in pigs' feeding status through manual observation in traditional pig farming, as well as the speed of abnormal early warning response and farming management efficiency, and reduces the impact of delayed early warning and inefficient management on farming benefits.

[0005] The embodiments of this application disclose the following technical solutions: In a first aspect, embodiments of this application provide an Internet of Things-based pig farming management system, the system comprising: The health sample collection module is used to synchronously collect IoT sample monitoring information of healthy pigs during the feeding process through an IoT sensor network set up in the pig feeding area. The IoT sample monitoring information includes at least mechanical vibration signals, audio signals and video images. The rhythm feature construction module is used to process the mechanical vibration signal, the audio signal and the video image, extract two-level features, and construct rhythm features based on the two-level features, wherein the two-level features include physical signal features and kinematic signal features; The health behavior baseline establishment module is used to establish a dynamic health behavior baseline for target pigs based on the dual-level features and the rhythm features, combined with statistical analysis methods. The dynamic health behavior baseline includes a first type of baseline and a second type of baseline. The real-time feature acquisition module is used to synchronously acquire real-time IoT monitoring information during the feeding process of the target pigs through the IoT sensor network, and correspondingly obtain real-time dual-level features and real-time rhythm features. The anomaly identification and response module is used to identify feeding anomalies based on the real-time dual-level features, the real-time rhythm features, and the dynamic health behavior baseline, combined with a pre-built feeding anomaly identification component, and to execute aquaculture management responses based on the feeding anomaly identification results.

[0006] Secondly, embodiments of this application provide a method for pig farming management based on the Internet of Things, the method comprising: The IoT sensor network installed in the pig feeding area is used to synchronously collect IoT sample monitoring information of healthy pigs during the feeding process. The IoT sample monitoring information includes at least mechanical vibration signals, audio signals and video images. The mechanical vibration signal, the audio signal, and the video image are processed to extract dual-level features, and rhythm features are constructed based on the dual-level features, wherein the dual-level features include physical signal features and kinematic signal features; Based on the aforementioned dual-level features and rhythmic features, and combined with statistical analysis methods, a dynamic health behavior baseline for the target pigs is established, wherein the dynamic health behavior baseline includes a first type of baseline and a second type of baseline. Through the IoT sensor network, real-time IoT monitoring information during the feeding process of the target pigs is collected synchronously in real time, and real-time dual-level features and real-time rhythm features are obtained accordingly. Based on the real-time dual-level features, the real-time rhythm features, and the dynamic health behavior baseline, feeding anomalies are identified using a pre-built feeding anomaly identification component, and aquaculture management responses are executed based on the feeding anomaly identification results.

[0007] One or more technical solutions provided in this application have at least the following technical effects or advantages: This application proposes an Internet of Things (IoT)-based pig farming management system and method. By deploying an IoT sensor network in the pig feeding area, mechanical vibration signals, audio signals, and video images of healthy pigs during feeding are simultaneously collected. Signal processing extracts a two-level feature structure consisting of physical and kinematic signal features. This structure is then combined with feeding behavior event detection and statistical analysis to construct rhythmic features. Based on these features and statistical analysis methods, a dynamic health behavior baseline containing a first-type baseline and a second-type baseline is established. Subsequently, the IoT sensor network collects real-time feeding monitoring information and corresponding real-time features of target pigs. A pre-constructed feeding anomaly identification component is used to compare the real-time features with the dynamic health behavior baseline, identifying feeding anomalies and executing tiered farming management responses. This achieves real-time monitoring and early warning of pig feeding health.

[0008] The technical solution proposed in this application solves the problems of delayed detection of abnormal feeding, subjective judgment of health status, and insufficient targeted management response caused by reliance on manual observation in traditional pig farming. It realizes intelligent and personalized health monitoring of pigs, and provides technical support for timely detection of sub-health or disease states in pigs, reducing breeding losses, and improving the efficiency and precision of pig farming management. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying 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.

[0010] Figure 1 A schematic diagram of the structure of an Internet of Things-based pig farming management system provided in this application embodiment; Figure 2 This is a flowchart illustrating an Internet of Things-based pig farming management method provided in an embodiment of this application.

[0011] The components represented by each number in the attached diagram are explained below: Health sample collection module 01, rhythm feature construction module 02, health behavior baseline establishment module 03, real-time feature acquisition module 04, anomaly identification and response module 05. Detailed Implementation

[0012] This application provides an Internet of Things-based pig farming management system and method to solve the technical problems in the existing technology of traditional pig farming that rely on manual observation of pigs' feeding status, making it difficult to capture early health abnormalities in a timely manner, and lacking dynamic health benchmarks, resulting in delayed abnormal warnings, low management efficiency, and thus affecting farming benefits.

[0013] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0014] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0015] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0016] Example 1, as shown in the appendix Figure 1 As shown, this application provides an Internet of Things-based pig farming management system, the method of which includes the following steps: The health sample collection module 01 is used to synchronously collect IoT sample monitoring information of healthy pigs during the feeding process through an IoT sensor network set in the pig feeding area. The IoT sample monitoring information includes at least mechanical vibration signals, audio signals and video images. In this embodiment of the application, in the scenario of early monitoring and management of the health status of pigs during the pig farming process, in order to obtain basic data that can accurately reflect the feeding behavior of healthy pigs, it is necessary to collect multiple types of monitoring information in different scenarios through the Internet of Things sensor network and perform data preprocessing to ensure the accuracy of subsequent construction of dynamic health behavior baseline and anomaly identification.

[0017] Specifically, firstly, IoT sensor units are deployed on the surface of the pigs' bodies. These IoT sensor units collect mechanical vibration signals and audio signals generated when the pigs are eating in real time, in order to reflect the detailed characteristics of the pigs' eating actions such as chewing and swallowing.

[0018] Meanwhile, IoT imaging units are set up around the pig feeding area to capture video images of the pigs feeding process, so as to obtain intuitive behavioral information such as the pigs' head movements.

[0019] Furthermore, from all the collected signal and image data, mechanical vibration signals, audio signals, and video images corresponding to healthy pigs were screened out to eliminate the interference of data from unhealthy pigs on the sample baseline.

[0020] Furthermore, based on a unified time benchmark, the three types of data after screening are processed for time synchronization to ensure that mechanical vibration, audio and video information at the same point in time can be accurately matched, forming a complete and synchronized IoT sample monitoring dataset.

[0021] This module provides sample data support for subsequent extraction of dual-level features and construction of rhythm features through categorized data collection and time-synchronized processing. This enables the health behavior baseline established based on the sample data to truly reflect the feeding patterns of healthy pigs, laying a reliable foundation for anomaly identification in real-time monitoring.

[0022] In the system provided in this application embodiment, the health sample collection module 01 includes: The mechanical vibration signal and the audio signal are collected by an Internet of Things (IoT) sensor unit installed on the surface of the pig's body; The video images are captured by IoT imaging units installed around the pig feeding area; The mechanical vibration signal, audio signal, and video image corresponding to healthy pigs are selected, and the mechanical vibration signal, audio signal, and video image are synchronized in time based on a unified time reference.

[0023] In this embodiment of the application, in order to obtain multi-dimensional basic data that can accurately reflect the feeding behavior of healthy pigs and avoid data deviation or invalid calculations in the subsequent feature extraction, rhythm analysis and health baseline construction process, it is necessary to form a complete and synchronous healthy pig feeding sample dataset through scene-specific collection and preprocessing operations, so as to ensure the accuracy of the subsequent dynamic health behavior baseline and the reliability of anomaly identification.

[0024] Specifically, mechanical vibration signals and audio signals are first collected by an Internet of Things (IoT) sensor unit installed on the surface of the pig's body.

[0025] When deploying IoT sensing units, it is necessary to select appropriate installation locations based on the characteristics of pigs' feeding behavior. For example, fixing the sensing unit in the area near the pig's jawbone can more directly capture jawbone vibrations generated during the pig's chewing and swallowing process, ensuring the accuracy of mechanical vibration signal acquisition.

[0026] Meanwhile, the IoT sensing unit needs to have a high-sensitivity audio acquisition function, which can accurately record characteristic audio such as chewing sounds and swallowing sounds when pigs are eating.

[0027] After the deployment location is determined, the acquisition parameters of the sensing unit also need to be adapted and set. For example, the acquisition frequency of the mechanical vibration signal is set to 500-1000Hz, which can fully cover the main frequency range of the pig's chewing vibration; the sampling rate of the audio signal is set to 44kHz to ensure that the audio signal can clearly reproduce the sound details during the feeding process.

[0028] Simultaneously, video images are captured by IoT imaging units installed around the pig feeding area. The installation of IoT imaging units needs to cover the entire feeding area to avoid blind spots. For example, high-definition cameras are deployed directly above and on both sides of the feeding trough to ensure that the movement trajectory of the pig's head during the feeding process can be fully captured.

[0029] At the same time, the shooting parameters of the IoT imaging unit need to be adjusted accordingly, setting the resolution to 1080P or higher and the frame rate to 25fps, so as to ensure that the video images can clearly show the subtle movements of the pig's head.

[0030] In addition, it is necessary to configure the IoT image unit with automatic supplemental lighting function according to the lighting conditions of the feeding area to avoid image blurring due to excessively dark or bright light, which would affect subsequent coordinate tracking and feature extraction.

[0031] Furthermore, after collecting mechanical vibration signals, audio signals, and video images, it is necessary to screen the three types of data corresponding to healthy pigs. During the screening process, the healthy pig population must first be identified through manual observation combined with routine health testing indicators. Then, based on the individual pig identifiers, the collected three types of data are precisely matched with the healthy pigs, eliminating data from unhealthy pigs and pigs without clear identifiers to avoid interference from unhealthy data on the sample benchmark.

[0032] After the filtering is completed, the three types of data are synchronized based on a unified time benchmark. First, a timestamp is added to each type of data. The timestamp accuracy needs to be at the millisecond level. For example, the clock of the Internet of Things sensor network is used as the benchmark to label the corresponding time information for each mechanical vibration signal, audio signal and video frame collected.

[0033] Furthermore, using the timestamp of the video image as a reference, the timestamps of the mechanical vibration signal and the audio signal are calibrated to ensure that the mechanical vibration, audio and video information at the same point in time can correspond accurately.

[0034] During the time synchronization process, the duration of the three types of data also needs to be adjusted for consistency. For example, if the duration of a certain mechanical vibration signal and audio signal is 10 minutes, the duration of the corresponding video image also needs to be adjusted to 10 minutes, and the start time and end time must be completely consistent. In the end, a healthy pig feeding sample dataset with unified time dimension and complete data is formed, which provides synchronous data support for subsequent extraction of dual-level features and construction of rhythm features.

[0035] The rhythm feature construction module 02 is used to process the mechanical vibration signal, the audio signal and the video image, extract dual-level features, and construct rhythm features based on the dual-level features, wherein the dual-level features include physical signal features and kinematic signal features; In this embodiment of the application, in order to extract features that reflect the essence of pigs’ feeding behavior from the collected raw signals and images, and then construct regular indicators that can be used to judge health status, it is necessary to first process the raw data to obtain two-level features, and then analyze feeding behavior events based on the two-level features and construct rhythmic features to provide accurate feature support for establishing a dynamic health behavior baseline and identifying feeding abnormalities.

[0036] Specifically, the mechanical vibration signal and audio signal are first processed to extract physical signal features. During the processing, filtering is performed on both types of signals to remove irrelevant interference signals from the environment, ensuring the accuracy of subsequent feature extraction.

[0037] Furthermore, based on the filtered signal, the obtained time-domain features, frequency-domain features, and nonlinear features are analyzed and integrated to output physical signal features.

[0038] Meanwhile, kinematic signal features are extracted from video images, and coordinate tracking is performed using the IoT sensor unit on the pig's head as the key point to generate and output a coordinate time history curve that reflects the head's movement trajectory.

[0039] Furthermore, after acquiring the dual-level features consisting of physical signal features and kinematic signal features, feeding behavior events are detected based on the dual-level features to identify events such as chewing and swallowing.

[0040] Furthermore, statistical analysis was performed on the occurrence sequence of feeding behavior events to construct rhythmic features, including a first rhythmic feature based on the swallowing-chewing rhythm ratio and rhythm variance, and a second rhythmic feature based on the mean and distribution skewness of a single effective feeding session, thereby characterizing the periodic regularity of feeding behavior.

[0041] This module extracts dual-level features and constructs rhythmic features, transforming raw IoT monitoring data into quantifiable feature indicators, providing a basis for establishing a dynamic health behavior baseline for target pigs.

[0042] In the system provided in this application embodiment, the rhythm feature construction module 02 includes: The mechanical vibration signal and the audio signal are filtered, and the time domain features, frequency domain features and nonlinear features are obtained based on the filtering results, and the output is the physical signal features. Using the IoT sensing unit on the pig's head as the key point for coordinate tracking, the coordinate time history curve is extracted from the video image and output as the kinematic signal feature.

[0043] Based on the physical signal features and the kinematic signal features, feeding behavior events are detected, and feeding behavior event detection results are obtained. The feeding behavior events include at least swallowing and chewing. Statistical analysis is performed on the occurrence sequence of the feeding behavior events to construct rhythmic features characterizing the periodic regularity of feeding behavior, wherein the rhythmic features include at least: First rhythmic features based on swallowing-chewing rhythm ratio and swallowing-chewing rhythm variance; The second rhythmic feature is constructed based on the mean duration of a single effective feeding session and the distribution skewness.

[0044] In this embodiment of the application, in order to transform the raw signal and image data collected during the pig feeding process into feature indicators that can be used to judge the health status of pigs, it is necessary to process the raw data step by step, extract features, detect behavioral events and construct rhythmic features to form a core basis that can accurately reflect the feeding behavior pattern of pigs.

[0045] Specifically, filtering is first performed on the mechanical vibration signal and the audio signal. The type of filtering needs to be determined based on the frequency characteristics of the signals related to pig feeding. For example, for mechanical vibration signals, a low-pass filter is selected to filter out high-frequency environmental interference higher than the dominant frequency of pig chewing vibration (usually 50-200Hz); for audio signals, a high-pass filter is selected to remove low-frequency environmental noise lower than the dominant frequency of pig swallowing sound (usually 300-800Hz), to ensure that the filtered signal can accurately retain key information related to feeding behavior.

[0046] After filtering is completed, the time-domain features, frequency-domain features, and nonlinear features are obtained by analyzing the filtering results.

[0047] Among them, the analysis of time-domain features requires the calculation of root mean square and amplitude. The root mean square is obtained by averaging the squares of the vibration or audio signal and then taking the square root. It can intuitively reflect the overall vibration energy of the signal. When healthy pigs are eating, this value will remain in a relatively stable range. Amplitude is the peak value of the signal, which can directly correspond to the force of a pig's single chewing. Abnormal changes in force may indicate health problems.

[0048] In addition, the analysis of frequency domain features requires converting the time domain signal into a frequency domain signal through Fourier transform in existing technologies, and then calculating the spectral centroid and frequency standard deviation. The spectral centroid can reflect the distribution of the main frequency of the sound. The main frequency will be significantly different when pigs swallow and when they chew. The frequency standard deviation reflects the degree of dispersion of the spectrum. This value fluctuates less when healthy pigs eat.

[0049] Furthermore, the analysis of nonlinear features mainly calculates approximate entropy. By quantifying the complexity and regularity of vibration signals, when pigs are in a sub-healthy or diseased state, their behavioral patterns become rigid or weak, leading to a decrease in the approximate entropy value. Integrating these analyzed features yields the physical signal characteristics.

[0050] Simultaneously, kinematic signal features are extracted from the acquired video images. During extraction, the IoT sensor unit on the pig's head must first be located using existing image recognition algorithms. This unit is then used as the key point for coordinate tracking, with the coordinates established using a fixed reference point in the feeding area as the origin to ensure the stability and consistency of the coordinates.

[0051] Furthermore, keypoint coordinates are extracted from the video images frame by frame, and the X-axis and Y-axis coordinates of keypoints in each frame are recorded. The coordinates of consecutive frames are then arranged in chronological order to generate a coordinate time-history curve. This curve can clearly show the movement trajectory of the pig's head during feeding. For example, when healthy pigs are feeding, their heads will move steadily around the feeding trough, while sick pigs may exhibit abnormal trajectories such as frequent head shaking and short dwell times. Outputting this coordinate time-history curve is the kinematic signal feature.

[0052] Furthermore, after acquiring physical signal features and kinematic signal features, feeding behavior events are detected based on these two types of features.

[0053] Specifically, when detecting swallowing and chewing events, it is necessary to combine the complementarity of the two types of features for comprehensive judgment: for chewing events, the amplitude peak of the vibration signal in the physical signal features (the vibration amplitude will increase significantly during chewing) and the approximate entropy valley value (the chewing action is regular and the signal complexity is reduced) are used as the basis for judgment. At the same time, the small amplitude and high frequency movement of the pig's head in the kinematic signal features (the head will move up and down slightly during chewing) are referred to for verification.

[0054] In addition, for swallowing events, specific frequency domain feature bursts of audio signals in physical signal characteristics (a brief and concentrated sound signal is generated during swallowing) are used as the basis for judgment, combined with the brief stillness of the pig's head in kinematic signal characteristics (the head stops moving during swallowing) to assist in confirmation, so as to ensure accurate differentiation and recording of swallowing and chewing events, forming a complete detection result of feeding behavior events.

[0055] After obtaining the detection results of feeding behavior events, statistical analysis is performed based on the event occurrence sequence to construct rhythmic features.

[0056] Specifically, for the first rhythm characteristics based on the swallowing-chewing rhythm ratio and the swallowing-chewing rhythm variance, a continuous feeding period needs to be determined first. This period is divided from the start of the pig entering the feeding area to stop eating to the end of the period leaving the feeding area. It is usually set to 15-30 minutes to ensure that it can cover a complete feeding process.

[0057] During this period, the total number of swallowing events and the total number of chewing events are counted, and the ratio between the two is calculated, which is the swallowing-chewing rhythm ratio. Simultaneously, swallowing-chewing rhythm ratio data from the same feeding period over the past 7-14 days for this pig are retrieved, and the variance of these ratios is calculated, which is the swallowing-chewing rhythm variance. In healthy pigs, the swallowing-chewing rhythm ratio remains within a fixed range with a small variance. However, when pigs have digestive tract diseases, reduced swallowing leads to a lower ratio, and rhythm disturbances lead to an increased variance.

[0058] In addition, for the second rhythm feature constructed based on the mean and distribution skewness of the single effective feeding time, it is necessary to first define the single effective feeding time. The definition criteria combine physical signal features and kinematic signal features: when the root mean square value of the vibration signal is higher than the preset threshold (i.e. the pig is chewing) and the pig's head coordinates are within the feeding trough range in the kinematic signal features, it is determined to be the start of "head-down feeding". When neither of the two conditions is met, it is determined to be the end. Thus, multiple discrete single effective feeding times are obtained.

[0059] Furthermore, the mean and skewness of all single effective feeding sessions within a unit of time (e.g., one feeding period) are calculated. The mean reflects the pig's sustained feeding capacity in a single session, while the skewness reflects the degree of symmetry in the duration distribution. Healthy pigs have a larger mean duration of single effective feeding sessions and a skewness close to 0 (symmetrical distribution), while sick pigs will have a smaller mean due to decreased appetite and insufficient energy, and will exhibit a large number of extremely short feeding attempts, causing the duration distribution to show a right skewness (positive skewness value).

[0060] Furthermore, in constructing rhythmic characteristics, parameters need to be adapted in conjunction with basic attributes such as breed and age of the pigs. For example, for pigs in the growing and fattening period, the average threshold of the duration of a single effective feeding session needs to be higher than that for pigs in the nursery period; for large breed pigs, the standard range of the swallowing-chewing rhythm ratio needs to be distinguished from that of small breed pigs to avoid deviations in the baseline of rhythmic characteristics due to differences in breed and age.

[0061] At the same time, outliers in the statistical analysis process need to be removed, such as extremely short (less than 2 seconds) or extremely long (more than 5 minutes) single effective feeding time caused by signal interference, in order to ensure the accuracy of the statistical results.

[0062] Ultimately, the rhythmic features constructed through the above steps can accurately analyze the feeding status of pigs from the perspective of behavioral cycle patterns, providing a quantitative basis for establishing a dynamic health behavior baseline.

[0063] The health behavior baseline establishment module 03 is used to establish a dynamic health behavior baseline for the target pigs based on the dual-level features and the rhythm features, combined with statistical analysis methods. The dynamic health behavior baseline includes a first type of baseline and a second type of baseline. In this embodiment of the application, in order to establish a clear reference standard that can be used as a basis for health judgment and avoid misjudgment or omission of health abnormalities due to the lack of a benchmark during subsequent real-time monitoring, it is necessary to combine statistical analysis methods to extract patterns from feature data and construct a dynamic health behavior baseline to ensure that deviations from the health status can be accurately identified when comparing real-time features with the baseline.

[0064] Specifically, based on dual-level features and rhythmic features, a set of typical feature lines for the target pigs is obtained by combining mathematical modeling methods. This set of typical feature lines can reflect the standard level of each feature under healthy conditions.

[0065] Furthermore, based on the obtained set of typical feature lines, the volatility indicators and confidence intervals of the two-level features are calculated, and they are associated with the corresponding typical feature lines to obtain the first type of baseline, so as to accurately define the normal fluctuation range of the two-level features in a healthy state.

[0066] Simultaneously, the volatility indicators and confidence intervals of rhythmic characteristics are calculated and correlated with the corresponding typical characteristic lines to obtain the second type of baseline, so as to clearly delineate the stable interval of rhythmic characteristics in a healthy state.

[0067] Finally, two types of baselines are output as dynamic health behavior baselines, which are updated using a sliding window method according to a preset period to adapt to changes in the pigs' condition.

[0068] This module establishes a dynamic health behavior baseline, providing a precise reference for identifying health abnormalities in subsequent real-time monitoring. This effectively avoids judgment bias caused by the lack of fixed or lagging benchmarks and improves the accuracy of health monitoring.

[0069] In the system provided in this application embodiment, the health behavior baseline establishment module 03 includes: Based on the dual-level features and the rhythmic features, a set of typical feature lines for the target pigs is obtained by combining mathematical modeling methods. Based on the set of typical feature lines, calculate the volatility indicators and confidence intervals of the two-level features, and associate them with the corresponding typical feature lines to obtain the first type of baseline; Based on the determined set of typical feature lines, the volatility indicators and confidence intervals of the rhythm features are calculated, and the results are correlated with the corresponding typical feature lines to obtain the second type of baseline. The first type of baseline and the second type of baseline are output as the dynamic health behavior baseline, wherein the dynamic health behavior baseline is updated using a sliding window method according to a preset period.

[0070] In this embodiment of the application, in order to provide an accurate and comparable reference standard for judging the health status of target pigs, it is necessary to extract typical patterns in the health status from the acquired dual-level features and rhythmic features, and construct a dynamic baseline that can adapt to changes in the pig's status, so as to ensure that the abnormal health signals in the pig's feeding behavior can be accurately captured by comparing the real-time features with the baseline.

[0071] First, based on the obtained dual-level features and constructed rhythmic features, a set of typical feature markings for the target pigs is obtained by combining mathematical modeling methods.

[0072] Specifically, mathematical modeling methods need to select the appropriate type based on the attributes of the feature data. For example, for numerical features such as physical signal features, kinematic signal features, and rhythm features in dual-level features, the K-means clustering model is used to model and analyze multiple sets of feature data of the target pigs under the historical health status of 7-14 days.

[0073] During the modeling process, it is necessary to select the feature data clusters with the highest proportion in the clustering results that conform to the common sense of feeding behavior of healthy pigs. The mean, median and trend fitting curve of each feature in the feature data cluster are calculated to form the standard value or change law line corresponding to each feature. By integrating these standard values ​​and law lines, a typical feature line set is obtained to intuitively reflect the baseline level of each feature of the target pig in a healthy state.

[0074] Furthermore, based on the typical feature criterion set, the volatility indicators and confidence intervals of the two-level features are calculated, thereby obtaining the first type of baseline.

[0075] Among them, the calculation of volatility indicators needs to be based on the dispersion of the historical health characteristic data of the target pigs. For example, for the root mean square in the two-level characteristics, the standard deviation and coefficient of variation of its historical health data are calculated. The standard deviation reflects the absolute fluctuation range of the data, and the coefficient of variation reflects the relative fluctuation degree. The two together constitute the volatility indicator, which is used to quantify the normal fluctuation level of the characteristic under the health state.

[0076] In addition, the confidence interval is based on statistical principles, with the mean of the features in the typical feature line as the center, and the calculated standard deviation is used to set a 95% confidence interval to ensure that the two-level feature data in a healthy state has a 95% probability of falling within this interval.

[0077] After the calculation is completed, the volatility indicators and confidence intervals of each type of dual-level feature are associated with the corresponding typical feature benchmarks. For example, the typical mean, standard deviation, coefficient of variation and 95% confidence interval of the root mean square are bound and output to form the first type of baseline, which is used to measure whether the real-time dual-level features deviate from the healthy range.

[0078] Simultaneously, based on the determined set of typical feature lines, the volatility indicators and confidence intervals of rhythmic features are calculated, thereby obtaining the second type of baseline.

[0079] Among them, the calculation of the fluctuation indicators of rhythm characteristics needs to be carried out on its core indicators. For example, for the swallowing-chewing rhythm ratio, the variance and range of its historical health data should be calculated. The variance reflects the stability of the data fluctuation around the mean, and the range reflects the maximum fluctuation amplitude of the data. For the distribution skewness of the duration of a single effective feeding, the standard deviation of its historical health data should be calculated to quantify the normal fluctuation range of the skewness value.

[0080] In addition, the confidence intervals are also based on the typical curves of the rhythm characteristics and are set in combination with the corresponding volatility indicators. For example, for the typical mean of the swallowing-chewing rhythm ratio, the upper and lower limits are set at 1.96 times the standard deviation to form a 95% confidence interval.

[0081] Furthermore, the fluctuation indicators and confidence intervals of rhythmic characteristics are associated with the corresponding typical characteristic benchmarks. For example, the typical mean, variance, range and confidence interval of the swallowing-chewing rhythm ratio are bound together, and the typical values ​​of the mean of a single effective feeding time and the distribution skewness are bound together with the corresponding fluctuation indicators and confidence intervals to jointly form a second type of baseline, which is used to determine whether there are abnormalities in real-time rhythmic characteristics.

[0082] Finally, the constructed first-class baseline and second-class baseline are output together as dynamic health behavior baselines.

[0083] Furthermore, in order to avoid the baseline being fixed for a long time and unable to adapt to the characteristic differences caused by the growth and development of pigs or environmental changes, it is necessary to update the baseline using the sliding window method according to a preset cycle.

[0084] The preset cycle needs to be adjusted according to the growth stage of the pigs. For example, pigs in the nursery period grow quickly and their feeding characteristics change frequently, so the cycle is set to 5-7 days; pigs in the fattening period have relatively stable feeding characteristics, so the cycle can be extended to 10-14 days.

[0085] Specifically, when updating using the sliding window method, the window length is set to a preset period. Each time the update is performed, the earliest historical health feature data in the window is removed, and the latest health feature data is added. The volatility indicators and confidence intervals of the typical feature line set, the two-level features and the rhythm features are recalculated, thereby generating new first-type baselines and second-type baselines.

[0086] For example, when the window length is 7 days, when updating the baseline on the 8th day, the data from the 1st day is deleted and the data from the 7th day is added to ensure that the dynamic health behavior baseline is always consistent with the current health status of the target pigs.

[0087] The real-time feature acquisition module 04 is used to synchronously acquire real-time IoT monitoring information during the feeding process of the target pigs through the IoT sensor network, and correspondingly acquire real-time dual-level features and real-time rhythm features. In this embodiment of the application, after establishing a dynamic health behavior baseline for the target pigs, in order to capture changes in the pigs' behavioral characteristics during the feeding process in real time, it is necessary to continuously collect multiple types of real-time data during the feeding process through the Internet of Things sensor network and simultaneously convert them into real-time features that can be compared with the baseline, so as to ensure that abnormal signals deviating from the health baseline can be quickly identified in the future.

[0088] Specifically, real-time synchronous data collection is first conducted through an Internet of Things (IoT) sensor network. Relying on IoT sensor units on the pig's body surface and imaging units around the feeding area, the system continues to operate according to the parameters used in the previous health sample collection: the body surface sensor units collect mechanical vibration signals and audio signals at a fixed frequency, and the imaging units capture video at a stable resolution and frame rate to ensure that the dimensions of the collected data are consistent with those in the previous stage.

[0089] Similarly, during data collection, the clock of the IoT sensor network is used as a reference, and millisecond-level timestamps are marked on various types of data to prevent time misalignment due to transmission delays and ensure that multiple types of data corresponding to the same feeding behavior can be accurately matched.

[0090] Furthermore, after acquiring real-time monitoring information, corresponding real-time dual-level features are extracted. That is, the same filtering operation as before is performed on the vibration and audio signals, and then the time domain features, frequency domain features, and nonlinear features are analyzed to obtain real-time physical signal features; taking the head sensor unit as the key point, the real-time coordinate time history curve is extracted from the video as the real-time kinematic signal feature, and the two types of features are integrated to form real-time dual-level features.

[0091] Furthermore, based on the obtained real-time dual-level features, swallowing and chewing events are identified and statistically analyzed using the same construction method as the previous rhythm features, thereby obtaining real-time rhythm features.

[0092] Ultimately, the real-time dual-level features and real-time rhythm features obtained through the above-mentioned real-time monitoring information collection and feature extraction process can accurately reflect the actual state of the target pig's current feeding behavior. Moreover, the data dimensions are consistent with the features of previous healthy samples and the dynamic health behavior baseline, providing real-time data with a unified format for subsequent comparison of real-time features with the health baseline and identification of abnormal feeding behavior.

[0093] The anomaly identification and response module 05 is used to identify feeding anomalies based on the real-time dual-level features, the real-time rhythm features, and the dynamic health behavior baseline, combined with a pre-built feeding anomaly identification component, and to execute aquaculture management responses based on the feeding anomaly identification results.

[0094] In this embodiment of the application, in order to accurately determine whether the current feeding status of pigs deviates from the healthy range and avoid the deterioration of health problems due to the inability to identify abnormalities in time, it is necessary to complete the abnormality determination by steps such as calculating feature residuals and inputting feeding abnormality identification components, and take targeted management measures based on the determination results to ensure the health of pigs and reduce breeding losses.

[0095] First, the residuals between the real-time bi-level features, the real-time rhythm features, and the dynamic health behavior baseline are calculated separately to obtain multi-dimensional residual features.

[0096] Furthermore, the obtained multidimensional residual features are vectorized to transform them into a vector format that can be recognized by the feeding anomaly detection component.

[0097] Furthermore, the vectorized multidimensional residual features are input into a pre-built feeding anomaly identification component. This component analyzes and calculates the input feature vector based on a trained regression model, and determines whether the current feeding status of the target pig is abnormal based on the model output. If the output exceeds the threshold for normal feeding status, it is determined to be feeding anomaly; otherwise, it is determined to be feeding normal. Finally, the feeding anomaly identification result is obtained.

[0098] Furthermore, corresponding aquaculture management responses are executed based on the results of abnormal feeding identification to ensure the efficiency and targeted nature of aquaculture management, avoiding over-management due to misjudgment and preventing delays in handling abnormalities due to missed judgment.

[0099] This module transforms the comparison between real-time features and the health baseline into clear judgment results and management actions for abnormal feeding through residual calculation, feature recognition and response execution. It effectively solves the problem that it is difficult to accurately and timely identify abnormal feeding by relying on manual observation in traditional farming, and provides intelligent technical support for pig health monitoring.

[0100] In the system provided in this application embodiment, the anomaly identification and response module 05 includes: Calculate the residuals between the real-time dual-level features, the real-time rhythm features, and the dynamic health behavior baseline to obtain multidimensional residual features; The multidimensional residual features are vectorized and input into a pre-constructed feeding anomaly detection component to obtain the feeding anomaly detection result. The multidimensional residual features include multiple volatility dimensions, multiple binary confidence bias dimensions, and multiple relative confidence bias dimensions.

[0101] In this embodiment of the application, in order to avoid missing or misjudging the health problems of pigs due to a lack of precise feeding anomaly identification, it is necessary to first quantify the difference between real-time features and the health baseline through residual calculation, and then use a pre-trained feeding anomaly identification component to complete the anomaly determination, so as to carry out scientific and efficient feeding anomaly identification and provide a reliable basis for timely implementation of breeding management response.

[0102] Specifically, the residuals between real-time dual-level features, real-time rhythm features, and dynamic health behavior baseline are calculated first to obtain multidimensional residual features.

[0103] Specifically, the residual calculation process requires precise matching between different types of features and their corresponding baselines: for real-time dual-level features, the physical signal features and kinematic signal features need to be compared with the feature values ​​corresponding to the first type of baseline in the dynamic health behavior baseline.

[0104] For example, the residual of the physical signal feature is obtained by subtracting the typical datum value of the feature in the first type of baseline from the root mean square of the real-time mechanical vibration signal.

[0105] In addition, for real-time rhythmic features, the swallowing-chewing rhythm ratio, swallowing-chewing rhythm variance, mean of single effective feeding duration, and distribution skewness need to be calculated with the corresponding typical datum values ​​of rhythmic features in the second type of baseline to obtain the rhythmic feature residuals.

[0106] Furthermore, the physical signal feature residuals and rhythm feature residuals obtained by calculation together constitute a multidimensional residual feature, which covers multiple volatility dimensions, multiple binary confidence bias dimensions and multiple relative confidence bias dimensions.

[0107] Specifically, the volatility dimension reflects the magnitude of residual change over time, such as the fluctuation of the swallowing-chewing rhythm ratio residual over a certain period of time; the binary confidence bias dimension marks whether the residual exceeds the baseline confidence interval with 0 or 1, with 1 indicating that the residual exceeds the baseline confidence interval and 0 indicating that it does not; the relative confidence bias dimension calculates the proportion of the residual that exceeds the baseline confidence interval, for example, if the residual of a certain real-time feature exceeds the upper limit of the confidence interval by 20%, the proportion is 20%. Through multi-dimensional residuals, the degree of deviation between real-time features and the healthy baseline is intuitively reflected.

[0108] Furthermore, after obtaining the multidimensional residual features, they need to be vectorized to transform the multidimensional and dispersed residual information into a unified vector format that can be recognized by the pre-built feeding anomaly identification component.

[0109] Specifically, during vectorization, multiple volatility dimension residuals, binary confidence bias dimension residuals, and relative confidence bias dimension residuals are arranged sequentially according to a preset feature order, and then transformed into a one-dimensional numerical vector.

[0110] For example, the fluctuation residuals of all physical signal features are arranged first, then the fluctuation residuals of kinematic signal features are arranged, and then the binary confidence bias residuals and relative confidence bias residuals of various features are arranged to ensure that the vector can completely retain all the information of the multidimensional residuals, and the format is consistent with the input requirements of the pre-built feeding anomaly identification component, so as to avoid the component being unable to parse features normally due to format incompatibility.

[0111] Subsequently, the vectorized multidimensional residual features are input into a pre-built feeding anomaly identification component, so as to analyze the deviation between the real-time feeding characteristics reflected by the residual features and the healthy baseline by means of a regression model trained on samples in the component.

[0112] The system provided in this application embodiment includes the following construction steps for the "abnormal feeding identification component": Constrained by the breed characteristics of the target pigs, pig feed samples are collected, wherein the pig feed samples include abnormal feed samples and normal feed samples; Feature extraction is performed on the abnormal feeding samples to obtain the first sample's two-level features, and the first sample's rhythm features are constructed based on the first sample's two-level features. Feature extraction is performed on the normal feeding samples to obtain the second sample's dual-level features, and the second sample's rhythm features are constructed based on the second sample's dual-level features. Based on the normal feeding samples, a sample health behavior baseline is constructed, and the first sample residual between the sample health behavior baseline and the first sample dual-level features and the first sample rhythm features is calculated. The second sample residual between the sample health behavior baseline and the second sample dual-level features and the second sample rhythm features is also calculated. Using the first sample residual and the second sample residual as training input data, and feeding status as supervision, a regression-based feeding anomaly identification component is constructed and trained, wherein feeding status includes feeding anomaly and feeding normality.

[0113] In this embodiment of the application, in order to ensure that the abnormal feeding identification component can accurately adapt to the feeding behavior characteristics of different breeds of target pigs, it is necessary to carry out sample collection and model training with breed characteristics as constraints. Through standardized feature extraction, residual calculation and regression training, a feeding abnormality identification component that can accurately distinguish between normal and abnormal feeding states is constructed.

[0114] Specifically, the first step is to collect feed samples from the target pigs, using the breed characteristics as a constraint.

[0115] Among them, the breed characteristic constraint needs to be set according to the differences in feeding behavior of different pig breeds. For example, for lean pigs with fast growth rate and large feed intake, the collection cycle is set to 3 feeding periods per day, with each collection period lasting no less than 45 minutes, to ensure that the entire feeding process is covered; for fat pigs with high feeding frequency and small single feed intake, the collection cycle is set to 4 feeding periods per day, with each collection period lasting no less than 30 minutes.

[0116] In addition, the collected pig feed samples must include abnormal feed samples and normal feed samples: abnormal feed samples must cover different types of abnormal feed intake caused by different health problems, such as slow feed intake caused by digestive tract diseases (single feed intake time shortened by more than 50% compared with the normal level), and weak chewing caused by respiratory diseases (root mean square vibration signal lower than the normal level by more than 30%). The number of samples collected for each type of abnormality shall not be less than 50 groups.

[0117] Meanwhile, normal feed intake samples should be selected from healthy pigs, and their feed intake data at different ages should be collected. Each sample group should simultaneously record mechanical vibration signals, audio signals and video images, and the number of samples should not be less than 100 groups to ensure that the samples cover the typical feeding scenarios of the breed of pigs.

[0118] Furthermore, after obtaining the pig feed samples, feature extraction is performed based on the abnormal feed samples to obtain the first sample's two-level features and construct the first sample's rhythm features.

[0119] Specifically, feature extraction must follow the same parameters and procedures as real-time feature acquisition. That is, mechanical vibration signals in abnormal feeding samples are filtered with a 200Hz low-pass filter (to remove high-frequency environmental noise) and audio signals are filtered with a 300Hz high-pass filter (to remove low-frequency interference). Then, time-domain features, frequency-domain features and nonlinear features are analyzed and integrated to form the physical signal features of the first sample.

[0120] Meanwhile, taking the head sensor unit of the pig as the key point, the coordinate time history curve is extracted from the video image as the kinematic signal feature of the first sample. The two types of features together constitute the dual-level feature of the first sample.

[0121] Furthermore, when constructing the rhythm features of the first sample based on the dual-level features of the first sample, the detection parameters for feeding behavior events need to be adapted to the characteristics of abnormal feeding. For example, the peak threshold of vibration amplitude for chewing event recognition is set to 1 / 2 of the threshold for normal samples, and the duration of the audio frequency domain feature burst for swallowing event recognition is set to more than 1.5 times that of normal samples. During statistical analysis, the calculation period for the swallowing-chewing rhythm ratio is set to 10 minutes / segment, and the root mean square threshold of the vibration signal defined by the duration of a single effective feeding is set to 1 / 3 of that of normal samples, thus forming the rhythm features of the first sample.

[0122] Simultaneously, feature extraction was performed based on normal feeding samples to obtain the second sample's dual-level features, and the second sample's rhythmic features were constructed.

[0123] Similarly, the feature extraction parameters must be consistent with those for anomaly sample extraction to ensure uniformity of feature dimensions: the filtering type and frequency range of mechanical vibration signals and audio signals remain unchanged, and the calculation window, threshold, frequency band, and other parameters of time-domain features, frequency-domain features, and nonlinear features are the same as those for anomaly sample extraction.

[0124] In addition, the coordinate sampling frequency and key point tracking rules of the kinematic signal features are kept consistent, forming a second sample of dual-level features.

[0125] Furthermore, when constructing the rhythm features of the second sample, the parameters for detecting feeding behavior events need to be set based on the statistics of normal samples. For example, the peak threshold of the vibration amplitude of chewing events is set to 1 / 2 of the mean of normal samples (to ensure accurate identification of slight chewing movements), and the burst frequency band of the audio frequency domain features of swallowing events is set to 300-800Hz. The calculation time period of the swallowing-chewing rhythm ratio and the threshold for defining the duration of a single effective feeding are the same as those for abnormal samples, thereby ensuring the comparability of the rhythm features of the first sample and the rhythm features of the second sample.

[0126] Furthermore, after obtaining the dual-level and rhythmic characteristics of the two types of samples, a baseline of healthy behavior for the samples was constructed based on the normal feeding samples.

[0127] Similarly, the construction process also requires the use of statistical analysis methods. For the physical signal characteristics and kinematic signal characteristics in the second sample's two-level characteristics, the mean, standard deviation, and 95% confidence interval are calculated respectively to form the sample's physical signal baseline and kinematic signal baseline. For the swallowing-chewing rhythm ratio, the mean and distribution skewness of the single effective feeding time in the second sample's rhythm characteristics, the mean, standard deviation, and 95% confidence interval are also calculated to form the sample's rhythm characteristic baseline. The three types of baselines are integrated into the sample's health behavior baseline.

[0128] Furthermore, the first-sample residuals of the sample health behavior baseline and the first-sample bi-level features and the first-sample rhythm features are calculated.

[0129] Specifically, the residual calculation adopts the absolute difference method. For example, the root mean square residual in the physical signal feature of the first sample = |root mean square of the first sample - root mean square of the sample physical signal baseline|, and the swallowing-chewing rhythm ratio residual in the rhythm feature of the first sample = |rhythm ratio of the first sample - rhythm ratio of the sample rhythm feature baseline|. All residuals are integrated into the residual of the first sample.

[0130] Similarly, the second sample residuals between the sample health behavior baseline and the second sample bi-level features and second sample rhythm features are calculated using the same method to ensure that the obtained residuals quantify the degree of deviation between the sample features and the health baseline.

[0131] Finally, using the residuals of the first and second samples as training input data and feeding status as supervision, a regression-based feeding anomaly identification component was constructed and trained.

[0132] Specifically, the regression model uses the logistic regression algorithm, and the model parameters are set as follows: regularization strength is set to 0.8 (to balance model complexity and generalization ability), number of iterations is set to 1000 (to ensure model convergence), and classification threshold is set to 0.5 (output probability ≥0.5 is judged as abnormal feeding, <0.5 is judged as normal feeding).

[0133] During training, the residuals of the first sample (labeled as "abnormal feeding") and the residuals of the second sample (labeled as "normal feeding") were divided into training and testing sets in a 7:3 ratio. Five-fold cross-validation was used to optimize the model parameters to ensure that the model's accuracy on the testing set was no less than 92%, recall no less than 90%, and precision no less than 88%.

[0134] After training, the model needs to be tested for stability by inputting 10 new samples that were not used in training (5 normal samples and 5 abnormal samples). If the model's recognition results are completely consistent with the actual state, the abnormal feeding recognition component is considered to have been successfully constructed. If there is a deviation, the corresponding type of samples need to be added and the model needs to be retrained until it meets the performance requirements (i.e., test set accuracy ≥ 92%, recall ≥ 90%, precision ≥ 88%).

[0135] In addition, the collection equipment needs to be calibrated regularly during the sample collection process. For example, the vibration acquisition accuracy of the body surface sensing unit should be calibrated every 7 days, and the coordinate tracking accuracy of the imaging unit should be calibrated every 10 days to ensure the accuracy of the sample data.

[0136] Furthermore, the vectorized multidimensional residual features are input into the pre-constructed feeding anomaly identification component. After receiving the vector, the feeding anomaly identification component performs calculations on the vector based on the trained logistic regression model to obtain the feeding anomaly probability value. If the feeding anomaly probability value is ≥0.5, the identification result of "feeding anomaly" is output, and the abnormal dimension with the highest contribution is marked. For example, the relative confidence bias dimension of the swallowing-chewing rhythm ratio residual is 0.7 (far exceeding the normal threshold of 0.3). This dimension contributes 60% to the anomaly judgment, and the feeding rhythm disorder reflected by this dimension needs to be given special attention.

[0137] Conversely, if the probability value of abnormal feeding is less than 0.5, the recognition result of "normal feeding" is output, and the difference between the current vector's dimension values ​​and the model's judgment threshold is recorded. For example, the volatility dimension value of the root mean square residual of mechanical vibration signal is 0.15, which is 0.05 away from the model's judgment upper limit of 0.2 for normal volatility. The binary confidence bias dimension value of the centroid residual of audio signal spectrum is 0, which is 1 away from the threshold of 1. By recording these differences, data references are provided for subsequent tracking of feature change trends and timely detection of potential anomalies.

[0138] Furthermore, after the abnormal feeding identification component outputs the identification results, the results need to be verified a second time. For example, when "abnormal feeding" is output, the historical identification results of the pig in the past 24 hours are retrieved. If an abnormality occurs only once and the relative confidence deviation of the abnormal dimension is small (e.g., below 0.3), it is judged as "suspected abnormality" and subsequent feeding characteristics need to be continuously monitored. If an abnormality occurs three or more times consecutively or the relative confidence deviation of a single abnormal dimension is large (e.g., above 0.6), it is judged as "confirmed abnormality" and subsequent breeding management response is immediately triggered. The secondary verification avoids misjudgment caused by accidental factors and improves the reliability of the abnormal feeding identification results.

[0139] Furthermore, for the "confirmed abnormal" feeding abnormality identification results, targeted breeding management response should be implemented immediately, that is, send an early warning notification to the staff containing the pen of the abnormal pig and the type of abnormality, and guide them to check the health status of the pigs on site as soon as possible; if the farm has automatic isolation equipment, the equipment instruction can be triggered simultaneously to transfer the abnormal pigs to a separate observation pen to avoid the possible risk of disease transmission.

[0140] In addition, for the "suspected abnormal" feeding abnormality identification results, there is no need for emergency warning. It is only necessary to adjust the monitoring frequency of the pig, for example, from collecting feeding data once every 6 hours to once every 2 hours, continuously track the changes in characteristics, and if the abnormal characteristics disappear, resume regular monitoring. If the abnormality worsens, upgrade to "confirmed abnormality" and execute the corresponding response.

[0141] In addition, if the abnormal feeding identification result is "normal feeding", the monitoring data and identification result will be entered into the pig's health record and the historical characteristic database will be updated to provide data support for the subsequent update of the dynamic health behavior baseline. At the same time, the routine breeding management process will be maintained without additional intervention.

[0142] The embodiments of this application, through the specific implementation methods described above, achieve the following technical effects: This application proposes an IoT-based pig farming management system. First, an IoT sensor network is deployed in the pig feeding area using a health sample collection module. Mechanical vibration and audio signals are collected using sensors on the pigs' bodies, and video images are captured by imaging units around the feeding area. Healthy pig data is filtered and synchronized in time to form a complete health sample dataset. Next, a rhythm feature construction module processes the sample data, filtering the vibration and audio signals to extract time-domain, frequency-domain, and nonlinear features as physical signal features. Kinematic signal features are generated by tracking head coordinates in the video. Combining these two types of features, swallowing and chewing events are identified, and statistical analysis is performed. The process involves analyzing and constructing rhythmic features. Subsequently, the health behavior baseline establishment module, based on dual-level features and rhythmic features, obtains a set of typical feature lines through mathematical modeling, calculates feature volatility indicators and confidence intervals, and constructs and dynamically updates a health behavior baseline containing first and second-type baselines. Then, the real-time feature acquisition module collects monitoring information of target pigs in real-time, according to parameters collected from healthy samples, and extracts real-time dual-level features and real-time rhythmic features. Finally, the anomaly identification and response module calculates the residual between the real-time features and the health baseline to obtain multi-dimensional residual features, which are then vectorized and input into a pre-constructed feed intake anomaly identification component to obtain the identification results. Based on the results, a tiered breeding management response is executed.

[0143] The system provided in this application, through the technical solution of "health sample collection - feature and rhythm construction - dynamic baseline establishment - real-time feature collection - anomaly identification and response", solves the problems of delayed detection of abnormal feeding, subjective health judgment, and insufficient targeted management response caused by reliance on manual observation in traditional pig farming. It realizes intelligent and precise health monitoring of pigs, and provides reliable technical support for timely detection of sub-health or disease states in pigs, reducing breeding losses, and improving the efficiency and refinement of pig farming management.

[0144] Example 2, as shown in the appendix Figure 2 As shown, based on the inventive concept of an IoT-based pig farming management system provided in Embodiment 1, this application also provides an IoT-based pig farming management method, specifically including: An IoT sensor network deployed in the pig feeding area synchronously collects IoT sample monitoring information of healthy pigs during their feeding process. This IoT sample monitoring information includes at least mechanical vibration signals, audio signals, and video images. The mechanical vibration signals, audio signals, and video images are processed to extract dual-level features, and rhythmic features are constructed based on these dual-level features. These dual-level features include physical signal features and kinematic signal features. Based on these dual-level features and the rhythmic features, combined with statistical analysis methods, a dynamic health behavior baseline for the target pigs is established. This dynamic health behavior baseline includes a first-type baseline and a second-type baseline. Real-time IoT monitoring information of the target pigs during their feeding process is synchronously collected through the IoT sensor network, and corresponding real-time dual-level features and real-time rhythmic features are obtained. Based on these real-time dual-level features, real-time rhythmic features, and the dynamic health behavior baseline, a pre-constructed feeding anomaly identification component is used to identify feeding anomalies, and a breeding management response is executed based on the feeding anomaly identification results.

[0145] In one embodiment, an Internet of Things (IoT) sensor network installed in the pig feeding area synchronously collects IoT sample monitoring information during the feeding process of healthy pigs. The IoT sample monitoring information includes at least mechanical vibration signals, audio signals, and video images, and further includes: The mechanical vibration signal and the audio signal are collected by an IoT sensor unit installed on the surface of the pig's body; the video image is collected by an IoT imaging unit installed around the pig's feeding area; the mechanical vibration signal, the audio signal and the video image corresponding to healthy pigs are selected, and the mechanical vibration signal, the audio signal and the video image are synchronized in time based on a unified time reference.

[0146] In one embodiment, the mechanical vibration signal, the audio signal, and the video image are processed to extract dual-level features, and rhythmic features are constructed based on the dual-level features. The dual-level features include physical signal features and kinematic signal features, and further include: The mechanical vibration signal and the audio signal are filtered, and the time-domain features, frequency-domain features, and nonlinear features are obtained based on the filtering results, and the output is the physical signal features. Coordinate tracking is performed using the IoT sensing unit on the pig's head as a key point, and the coordinate time-history curve is extracted from the video image and output as the kinematic signal features. Based on the physical signal features and the kinematic signal features, feeding behavior events are detected, and the detection results are obtained. The feeding behavior events include at least swallowing and chewing. Statistical analysis is performed on the occurrence sequence of the feeding behavior event detection results to construct the rhythmic features characterizing the periodic regularity of feeding behavior. The rhythmic features include at least: a first rhythmic feature based on the swallowing-chewing rhythm ratio and the swallowing-chewing rhythm variance; and a second rhythmic feature constructed based on the mean and distribution skewness of a single effective feeding duration.

[0147] In one embodiment, based on the dual-level features and the rhythmic features, combined with statistical analysis methods, a dynamic health behavior baseline for the target pigs is established. This dynamic health behavior baseline includes a first-type baseline and a second-type baseline, and further includes: Based on the dual-level features and the rhythmic features, a set of typical feature lines for the target pigs is obtained using mathematical modeling methods. According to the set of typical feature lines, the volatility indicators and confidence intervals of the dual-level features are calculated and correlated with the corresponding typical feature lines to obtain the first type of baseline. According to the determined set of typical feature lines, the volatility indicators and confidence intervals of the rhythmic features are calculated and correlated with the corresponding typical feature lines to obtain the second type of baseline. The first type of baseline and the second type of baseline are output as the dynamic health behavior baseline, wherein the dynamic health behavior baseline is updated using a sliding window method according to a preset period.

[0148] In one embodiment, based on the real-time dual-level features, the real-time rhythm features, and the dynamic health behavior baseline, and combined with a pre-constructed feeding anomaly identification component, feeding anomaly identification is performed, and aquaculture management responses are executed based on the feeding anomaly identification results. The method further includes: The residuals between the real-time dual-level features, the real-time rhythm features, and the dynamic health behavior baseline are calculated respectively to obtain multidimensional residual features; the multidimensional residual features are vectorized and input into a pre-constructed feeding anomaly identification component to obtain the feeding anomaly identification result, wherein the multidimensional residual features include multiple volatility dimensions, multiple binary confidence bias dimensions, and multiple relative confidence bias dimensions.

[0149] Furthermore, based on the real-time dual-level features, the real-time rhythm features, and the dynamic health behavior baseline, and combined with a pre-constructed feeding anomaly identification component, feeding anomalies are identified, and aquaculture management responses are executed based on the feeding anomaly identification results. This is also used for: Constrained by the breed characteristics of the target pigs, feed intake samples are collected, including abnormal feed intake samples and normal feed intake samples. Feature extraction is performed on the abnormal feed intake samples to obtain a first sample dual-level feature, and a first sample rhythm feature is constructed based on the first sample dual-level feature. Feature extraction is performed on the normal feed intake samples to obtain a second sample dual-level feature, and a second sample rhythm feature is constructed based on the second sample dual-level feature. A sample health behavior baseline is constructed according to the normal feed intake samples, and a first sample residual is calculated between the sample health behavior baseline and the first sample dual-level feature and the first sample rhythm feature. A second sample residual is also calculated between the sample health behavior baseline and the second sample dual-level feature and the second sample rhythm feature. Using the first sample residual and the second sample residual as training input data, and with feed intake status as supervision, a regression-based abnormal feed intake identification component is constructed and trained, where feed intake status includes abnormal feed intake and normal feed intake.

[0150] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0151] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0152] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and variations fall within the scope of this application and its equivalents, this application intends to include such modifications and variations.

Claims

1. A pig farming management system based on the Internet of Things, characterized in that, include: The health sample collection module is used to synchronously collect IoT sample monitoring information of healthy pigs during the feeding process through an IoT sensor network set up in the pig feeding area. The IoT sample monitoring information includes at least mechanical vibration signals, audio signals and video images. The rhythm feature construction module is used to process the mechanical vibration signal, the audio signal and the video image, extract two-level features, and construct rhythm features based on the two-level features, wherein the two-level features include physical signal features and kinematic signal features; The health behavior baseline establishment module is used to establish a dynamic health behavior baseline for target pigs based on the dual-level features and the rhythm features, combined with statistical analysis methods. The dynamic health behavior baseline includes a first type of baseline and a second type of baseline. The real-time feature acquisition module is used to synchronously acquire real-time IoT monitoring information during the feeding process of the target pigs through the IoT sensor network, and correspondingly obtain real-time dual-level features and real-time rhythm features. The anomaly identification and response module is used to identify feeding anomalies based on the real-time dual-level features, the real-time rhythm features, and the dynamic health behavior baseline, combined with a pre-built feeding anomaly identification component, and to execute aquaculture management responses based on the feeding anomaly identification results.

2. The IoT-based pig farming management system as described in claim 1, characterized in that, The execution steps of the health sample collection module involve synchronously collecting IoT sample monitoring information of healthy pigs during their feeding process through an IoT sensor network installed in the pig feeding area. This IoT sample monitoring information includes at least mechanical vibration signals, audio signals, and video images, including: The mechanical vibration signal and the audio signal are collected by an Internet of Things (IoT) sensor unit installed on the surface of the pig's body; The video images are captured by IoT imaging units installed around the pig feeding area; The mechanical vibration signal, audio signal, and video image corresponding to healthy pigs are selected, and the mechanical vibration signal, audio signal, and video image are synchronized in time based on a unified time reference.

3. The IoT-based pig farming management system as described in claim 2, characterized in that, The execution steps of the rhythm feature construction module include processing the mechanical vibration signal, the audio signal, and the video image to extract two-level features, including: The mechanical vibration signal and the audio signal are filtered, and the time domain features, frequency domain features and nonlinear features are obtained based on the filtering results, and the output is the physical signal features. Using the IoT sensing unit on the pig's head as the key point for coordinate tracking, the coordinate time history curve is extracted from the video image and output as the kinematic signal feature.

4. The IoT-based pig farming management system as described in claim 3, characterized in that, The execution steps of the rhythm feature construction module, which constructs rhythm features based on the dual-level features, include: Based on the physical signal features and the kinematic signal features, feeding behavior events are detected, and feeding behavior event detection results are obtained. The feeding behavior events include at least swallowing and chewing. Statistical analysis is performed on the occurrence sequence of the feeding behavior events to construct rhythmic features characterizing the periodic regularity of feeding behavior, wherein the rhythmic features include at least: First rhythmic features based on swallowing-chewing rhythm ratio and swallowing-chewing rhythm variance; The second rhythmic feature is constructed based on the mean duration of a single effective feeding session and the distribution skewness.

5. The IoT-based pig farming management system as described in claim 4, characterized in that, The execution steps of the health behavior baseline establishment module are as follows: based on the dual-level features and the rhythm features, and combined with statistical analysis methods, a dynamic health behavior baseline for the target pigs is established. This dynamic health behavior baseline includes a first type of baseline and a second type of baseline, comprising: Based on the dual-level features and the rhythmic features, a set of typical feature lines for the target pigs is obtained by combining mathematical modeling methods. Based on the set of typical feature lines, calculate the volatility indicators and confidence intervals of the two-level features, and associate them with the corresponding typical feature lines to obtain the first type of baseline; Based on the determined set of typical feature lines, the volatility indicators and confidence intervals of the rhythm features are calculated, and the results are correlated with the corresponding typical feature lines to obtain the second type of baseline. The first type of baseline and the second type of baseline are output as the dynamic health behavior baseline, wherein the dynamic health behavior baseline is updated using a sliding window method according to a preset period.

6. The IoT-based pig farming management system as described in claim 5, characterized in that, The execution steps of the anomaly identification and response module, based on the real-time dual-level features, the real-time rhythm features, and the dynamic health behavior baseline, combined with a pre-constructed feeding anomaly identification component, include: Calculate the residuals between the real-time dual-level features, the real-time rhythm features, and the dynamic health behavior baseline to obtain multidimensional residual features; The multidimensional residual features are vectorized and input into a pre-constructed feeding anomaly identification component to obtain the feeding anomaly identification result; The multidimensional residual features include multiple volatility dimensions, multiple binary confidence bias dimensions, and multiple relative confidence bias dimensions.

7. A pig farming management system based on the Internet of Things as described in claim 6, characterized in that, The construction steps of the feeding abnormality detection component include: Constrained by the breed characteristics of the target pigs, pig feed samples are collected, wherein the pig feed samples include abnormal feed samples and normal feed samples; Feature extraction is performed on the abnormal feeding samples to obtain the first sample's two-level features, and the first sample's rhythm features are constructed based on the first sample's two-level features. Feature extraction is performed on the normal feeding samples to obtain the second sample's dual-level features, and the second sample's rhythm features are constructed based on the second sample's dual-level features. Based on the normal feeding samples, a sample health behavior baseline is constructed, and the first sample residual between the sample health behavior baseline and the first sample dual-level features and the first sample rhythm features is calculated. The second sample residual between the sample health behavior baseline and the second sample dual-level features and the second sample rhythm features is also calculated. Using the first sample residual and the second sample residual as training input data, and feeding status as supervision, a regression-based feeding anomaly identification component is constructed and trained, wherein feeding status includes feeding anomaly and feeding normality.

8. A method for pig farming management based on the Internet of Things, characterized in that, The method is applied to the IoT-based pig farming management system according to any one of claims 1-7, the method comprising: The IoT sensor network installed in the pig feeding area is used to synchronously collect IoT sample monitoring information of healthy pigs during the feeding process. The IoT sample monitoring information includes at least mechanical vibration signals, audio signals and video images. The mechanical vibration signal, the audio signal, and the video image are processed to extract dual-level features, and rhythm features are constructed based on the dual-level features, wherein the dual-level features include physical signal features and kinematic signal features; Based on the aforementioned dual-level features and rhythmic features, and combined with statistical analysis methods, a dynamic health behavior baseline for the target pigs is established, wherein the dynamic health behavior baseline includes a first type of baseline and a second type of baseline. Through the IoT sensor network, real-time IoT monitoring information during the feeding process of the target pigs is collected synchronously in real time, and real-time dual-level features and real-time rhythm features are obtained accordingly. Based on the real-time dual-level features, the real-time rhythm features, and the dynamic health behavior baseline, feeding anomalies are identified using a pre-built feeding anomaly identification component, and aquaculture management responses are executed based on the feeding anomaly identification results.

Images