Livestock oestrus intelligent prediction and breeding management system based on multi-modal sensing

Abstract

The invention relates to the technical field of livestock breeding intelligent management, and particularly discloses a livestock oestrus intelligent prediction and breeding management system based on multi-mode sensing. The system comprises a multi-mode sensing network, an edge intelligent computing unit and a cloud collaborative management platform. The sensing network collects livestock physiological behavior data; the edge unit performs local data preprocessing, feature extraction and real-time prediction, and generates early warning; the cloud platform is responsible for model training optimization and global data management. Through the edge and cloud collaborative distributed architecture, local real-time processing and decision making of the data are realized, dependence on the network is reduced, and the real-time performance and accuracy of prediction and the robustness of the system are improved.

Description

Technical Field

[0001] This invention belongs to the field of intelligent livestock breeding management technology, specifically involving an intelligent prediction and breeding management system for livestock estrus based on multimodal sensing. Background Technology

[0002] In modern large-scale animal husbandry, improving reproductive efficiency and animal welfare through technological means is a core development goal. Among these, accurately identifying and predicting the estrus cycle of livestock is a key link in achieving efficient breeding management. Traditional methods mainly rely on manual observation of changes in livestock behavior or physical characteristics, which not only consumes a lot of manpower but also suffers from problems such as strong subjectivity, delayed identification, and easy omissions, making it difficult to meet the real-time management needs of large-scale, intensive farms.

[0003] Intelligent livestock management systems based on the Internet of Things (IoT) and artificial intelligence (AI) have become an important technological direction. These systems aim to collect multi-dimensional physiological and behavioral data of livestock by deploying various sensors, and to use data analysis models to automatically identify and predict estrus states, thereby assisting managers in making scientific decisions.

[0004] Existing technologies typically employ a centralized cloud platform architecture, uploading multimodal data collected by front-end sensors to a cloud server for processing and analysis. This architecture faces significant challenges in practical applications on livestock farms: First, many farms are located in remote areas with poor network infrastructure, resulting in limited data transmission stability and bandwidth. Real-time uploading of large amounts of raw data can easily cause network congestion and high latency. Second, centralizing all computing tasks in the cloud leads to an excessively long link between data collection and the generation of early warning results. In critical estrus prediction scenarios, this delay can cause missed opportunities to determine the optimal mating time, directly impacting conception rates and production efficiency. Furthermore, complete reliance on the cloud also renders the system incapable of functioning properly during network outages, resulting in insufficient robustness.

[0005] Therefore, how to build a localized edge computing solution that can adapt to harsh network environments and achieve low-latency intelligent prediction has become an urgent technical challenge. Summary of the Invention

[0006] This invention provides a livestock estrus intelligent prediction and breeding management system based on multimodal sensing, in order to solve the problems of strong network dependence, high data processing latency and insufficient system robustness caused by reliance on centralized cloud processing in the prior art.

[0007] The technical solution of this invention is to construct a distributed system integrating an edge intelligent computing unit, a multimodal sensor network, and a cloud-based collaborative management platform. This system achieves localized real-time data processing and preliminary decision-making by moving the core livestock estrus prediction algorithm model to an edge intelligent computing unit deployed locally on the farm. Simultaneously, it uses a cloud platform for model updates, long-term data analysis, and global management, thereby significantly reducing reliance on network bandwidth and stability while ensuring the real-time performance and accuracy of predictions.

[0008] The system includes a multimodal sensor network, an edge intelligent computing unit, and a cloud-based collaborative management platform.

[0009] A multimodal sensor network is deployed within the farm to continuously collect multidimensional physiological and behavioral data from individual livestock. This network consists of multiple heterogeneous sensor nodes, each fixed to the livestock's body surface in a non-invasive or minimally invasive manner. Sensor node types include, but are not limited to: miniature thermal sensors for collecting core body temperature data; a triaxial accelerometer and gyroscope combination sensor for collecting activity levels, posture angles, and movement trajectories; infrared proximity sensors for collecting feeding and drinking behavior data; and miniature microphones for collecting vocal characteristics data. All sensor nodes transmit the raw sensor data to the corresponding data aggregation gateway via a low-power wireless communication protocol at a preset sampling frequency.

[0010] The edge intelligent computing unit is a core computing device deployed locally on the farm, physically close to the multimodal sensor network. This edge intelligent computing unit includes a data preprocessing module, a feature extraction engine, a local prediction model library, and a local decision-making and early warning module.

[0011] The data preprocessing module receives the raw data stream from the multimodal sensor network. First, it performs timestamp alignment and packet reassembly on the received multi-source heterogeneous data to ensure that data from different sensors at the same time can be correctly correlated. Further, the module performs data cleaning operations, specifically including filtering out abnormal body temperature readings based on preset threshold values ​​within a reasonable range of physiological parameters; and identifying and removing motion artifact noise caused by temporary sensor detachment or violent collisions based on the magnitude variation patterns of the triaxial accelerometer data.

[0012] The feature extraction engine performs multi-level feature calculations on the preprocessed data. The engine operates in two parallel threads. The first thread focuses on time-domain and frequency-domain feature extraction. For acceleration data, it calculates the integral of activity per unit time, the proportion of stillness time, and the variance of exercise intensity. For body temperature data, it calculates the mean, standard deviation, and diurnal temperature range within a 24-hour sliding window. For sound data, it extracts the fundamental frequency, harmonic energy distribution, and the energy proportion of specific frequency bands. The second thread focuses on behavioral pattern feature extraction. This thread incorporates a behavior state recognizer based on a Hidden Markov Model. By analyzing the sequence of acceleration and posture data, it segments and identifies the continuous behavior of livestock into discrete states such as "lying down," "standing," "walking," "running," and "mounting," and calculates the duration and transition frequency of each state within a preset time window.

[0013] A local prediction model library stores and runs a lightweight machine learning model for predicting livestock estrus. This local prediction model takes the multi-dimensional feature vector output by the aforementioned feature extraction engine as input. It employs a gradient boosting decision tree algorithm architecture, containing 128 decision trees of depth 5. Model training and updates are handled by a cloud-based collaborative management platform and periodically synchronized to edge computing units via encrypted differential updates. During the inference phase, the model calculates the estrus probability value between 0 and 1 from the input feature vector, and also outputs a score representing the prediction confidence level.

[0014] The local decision-making and early warning module receives the estrus probability value and confidence score output by the local prediction model. This module has a preset dynamic decision threshold, initially set to 0.7. The module compares the real-time calculated estrus probability value with the dynamic decision threshold. When the estrus probability value is greater than or equal to the dynamic decision threshold, and the confidence score is greater than the preset confidence threshold of 0.8, the module immediately generates a high-priority estrus warning command. This high-priority estrus warning command directly activates the audible and visual alarm devices deployed in the corresponding livestock pen via the farm's local communication network, and simultaneously sends a warning message containing the livestock number, predicted probability, confidence score, and timestamp to a preset management personnel mobile terminal via a wireless communication module. If network conditions permit, the warning message is also asynchronously uploaded to the cloud platform for recording. Furthermore, the local decision-making and early warning module also integrates a feedback learning mechanism, which records the trigger time of each early warning and compares it with the actual estrus time subsequently confirmed by manual verification or feedback of conception results. Based on the statistical results of the prediction time error, the dynamic decision threshold is adaptively fine-tuned on a weekly basis with a fine-tuning step size of 0.05.

[0015] The cloud-based collaborative management platform is deployed in a remote data center and interacts with edge intelligent computing units distributed across various locations via intermittent network connections. This platform includes a model training and optimization center, a global data warehouse, and a farm management interface.

[0016] The Model Training and Optimization Center is responsible for building and optimizing livestock estrus prediction models. This center gathers anonymized feature data samples and corresponding real estrus labels uploaded from all networked edge computing units. This data constitutes a continuously growing distributed training dataset. The center uses a federated learning framework for model training. The specific process is as follows: the current global model parameters are distributed to each edge node participating in the training; each edge node uses locally stored historical data to calculate the update gradient of the model parameters; each node uploads the encrypted gradient to the cloud; the cloud platform securely aggregates the encrypted gradient and updates the global model parameters. After multiple iterations, a new version of the model with improved performance is generated and then distributed to each edge computing unit using differential update technology.

[0017] A global data warehouse is used to store long-term, aggregated livestock data. It not only stores raw early warning records but also integrates livestock pedigree information, mating records, calving history, and environmental temperature and humidity data from the farm management system. Based on this multi-dimensional data, the global data warehouse supports complex retrospective analysis and statistical report generation, such as analyzing the estrus cycle patterns of different breeds and parities, or evaluating the performance of predictive models in different seasons.

[0018] The farm management interface provides managers with visual system monitoring and management functions. This interface displays the status of all networked farms in map form, showing the real-time online status, battery level, and storage space utilization of each farm's edge computing units. Managers can remotely query the recent activity curves, body temperature trends, and prediction history of any livestock through this interface, and can manually confirm or correct system alerts. This human feedback information will be sent back as important tags to the model training and optimization center.

[0019] In one embodiment of the present invention, the data aggregation gateway in the multimodal sensor network adopts a star topology. Each gateway manages up to 32 sensor nodes within a radius of 50 meters. Communication between the gateway and the sensor nodes is achieved using the 2.4GHz low-power Bluetooth 5.0 protocol, with the communication interval dynamically adjusted between 1 and 5 minutes based on the livestock's activity status. The gateway itself connects to the edge intelligent computing unit via wired Ethernet or a 4G / 5G cellular network.

[0020] In one embodiment of the present invention, the edge intelligent computing unit is implemented using an industrial-grade embedded system. Its core processor is a quad-core ARM architecture chip with a neural network processing unit and a clock speed of no less than 1.8GHz. This edge intelligent computing unit is equipped with no less than 4GB of RAM and 64GB of embedded solid-state storage for caching at least 30 days of local feature data and event logs. Its operating system is a trimmed real-time Linux kernel, ensuring deterministic latency in data processing task scheduling.

[0021] As one embodiment of the present invention, the local prediction model employs a gradient boosting decision tree algorithm, whose training objective function is the sum of the logistic regression loss function and L1 and L2 regularization terms. The regularization coefficients are determined through cross-validation, with the L1 regularization coefficient being 0.01 and the L2 regularization coefficient being 0.1, thereby controlling model complexity, preventing overfitting, and ensuring that the model maintains high generalization ability even with limited computing resources on edge devices.

[0022] In one embodiment of the present invention, the system operates within a hierarchical decision-making framework. This framework clearly defines a real-time layer, an analysis layer, and a planning layer. The real-time layer is hosted by edge intelligent computing units, processing data from seconds to minutes, with the core objective of achieving an early warning delay of less than 5 minutes. The analysis layer is hosted by a cloud-based collaborative management platform, processing data from hours to days, and is responsible for trend analysis and model iteration. The planning layer, also located in the cloud, processes data from months to years, supporting long-term decisions such as breeding planning and production capacity forecasting.

[0023] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention achieves localized real-time processing and decision-making of sensor data by moving the core livestock estrus prediction algorithm model to the edge intelligent computing unit on the farm. Data does not need to be entirely uploaded to the cloud, greatly reducing reliance on network bandwidth and stability, making it particularly suitable for remote farms with poor network conditions. The entire link latency from data acquisition to early warning generation is compressed to less than 5 minutes, ensuring timely detection of estrus signs and effectively improving the accuracy of determining the optimal mating time, thereby directly contributing to increased conception rate and reproductive efficiency.

[0024] 2. The distributed system architecture adopted in this invention significantly enhances the robustness and availability of the overall system. The edge intelligent computing unit possesses independent computing, storage, and decision-making capabilities, enabling continuous local data collection, processing, and early warning even during temporary network outages, ensuring the continuous operation of the farm's core management functions. The collaborative design of cloud and edge computing leverages both the low latency advantage of edge computing and the powerful computing capabilities of the cloud for model optimization and big data analysis, forming a complementary and elastic computing system.

[0025] 3. This invention constructs a more comprehensive and refined livestock status perception system through multimodal sensor network fusion and an advanced feature extraction engine. The system not only analyzes basic body temperature and activity levels but also delves into behavioral pattern sequences and vocal characteristics, thereby capturing earlier and more subtle estrus-related biological signals. Combining a lightweight local prediction model based on gradient boosting decision trees with a federated learning-driven cloud-based model continuous optimization mechanism, the system can continuously learn from massive amounts of field data, adapting to estrus patterns under different breeds, individuals, and environmental conditions, continuously improving the accuracy and generalization ability of the prediction model, and providing a solid technical foundation for achieving precision livestock management. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention; Figure 2 This is a schematic diagram of the core principle framework of local real-time prediction and decision-making based on edge intelligent computing units in this invention; Figure 3 This is a flowchart illustrating the logical process of multimodal sensing data preprocessing and multi-level feature extraction in this invention. Figure 4 This is a schematic diagram of the multi-level interaction relationship and data flow between the edge intelligent computing unit and the cloud collaborative management platform in this invention; Figure 5 This is a schematic diagram of the principle framework of the cloud-based model training and optimization center based on federated learning in this invention. Detailed Implementation

[0027] Example 1: Please refer to the appendix Figure 1 To be continued Figure 5 The overall technical architecture of the intelligent prediction and management system for livestock estrus based on multimodal sensing proposed in this invention is shown in the attached figure. Figure 1 As shown, the system comprises three main parts: a multimodal sensor network, an edge intelligent computing unit, and a cloud-based collaborative management platform. This forms a layered, distributed, closed-loop intelligent system with both local real-time decision-making capabilities and cloud-based global optimization capabilities. By deploying the core estrus prediction algorithm at the edge, close to the data source, the system achieves low-latency perception, analysis, and response to livestock physiological and behavioral states. Simultaneously, it relies on the cloud for continuous model evolution and systematic data insights. This significantly reduces dependence on network infrastructure while ensuring high-accuracy predictions and improves the system's robustness and availability in complex farming environments.

[0028] A multimodal sensor network is deployed within the farm, covering all target livestock individuals, to continuously and non-invasively collect multi-dimensional raw data reflecting their physiological state and behavioral patterns. This multimodal sensor network consists of several heterogeneous sensor nodes, each securely attached to the animal's body surface in the form of a flexible patch, ear tag, or collar, ensuring long-term stable wear without affecting normal activity. Sensor node types include miniature thermal sensors, a combination sensor of a three-axis accelerometer and gyroscope, an infrared proximity sensor, and a miniature microphone. The miniature thermal sensor is embedded inside the ear tag, close to the skin at the base of the ear, for continuous monitoring of the animal's core body temperature, sampling at a frequency of once per minute, with a measurement range of 35.0℃ to 42.0℃ and an accuracy of ±0.1℃. A triaxial accelerometer and gyroscope combination sensor is integrated into the collar to simultaneously collect linear acceleration and angular velocity data of livestock in three-dimensional space. The sampling frequency is 10 Hz, and the measurement ranges are ±8g and ±2000 degrees / second, respectively, to accurately reconstruct its posture changes and movement trajectory. Infrared proximity sensors are installed at the feed trough and waterer inlets. By detecting the infrared reflection signal when the livestock's head approaches, they record the start and end times and duration of each feeding and drinking event, with a time resolution of 1 second. A miniature microphone is integrated into the side of the collar to collect sound signals such as vocalizations and panting emitted by the livestock. The sampling rate is 8000 Hz, and the dynamic range is no less than 60 dB, for subsequent sound feature analysis.

[0029] All sensor nodes establish connections with their corresponding data aggregation gateways via low-power wireless communication protocols. (Combined with the attached...) Figure 1 As shown, the data aggregation gateway adopts a star topology for networking, with each gateway managing up to 32 sensor nodes within a 50-meter radius. Communication between the gateway and sensor nodes uses the 2.4GHz low-power Bluetooth 5.0 protocol, and the communication interval is dynamically adjusted based on the livestock's current activity status: when the system detects that the livestock is resting or in a low-activity state, the communication interval is extended to 5 minutes to save energy; when a significant increase in activity is detected (such as running or mounting), the communication interval is automatically shortened to 1 minute to ensure the timely capture of critical behavioral data. The data aggregation gateway itself establishes a stable connection with the edge intelligent computing unit via wired Ethernet or 4G / 5G cellular networks, forwarding the received raw sensor data streams to the edge for processing after timestamping.

[0030] The edge intelligent computing unit, serving as the core local processing hub of this system, is physically deployed in the farm's server room or control room. Its hardware platform is implemented using an industrial-grade embedded system. The core processor of this edge intelligent computing unit is a quad-core ARM architecture chip with a dedicated neural network processing unit, a clock speed of no less than 1.8GHz, and is equipped with no less than 4GB of RAM and 64GB of embedded solid-state storage, sufficient to cache at least 30 days of local feature data, event logs, and model versions. The operating system uses a deeply customized and real-time optimized Linux kernel to ensure deterministic scheduling latency for critical data processing tasks, meeting the system's stringent real-time requirements. (See attached image) Figure 2 As shown, the edge intelligent computing unit contains four major functional components: a data preprocessing module, a feature extraction engine, a local prediction model library, and a local decision-making and early warning module. The components interact with each other through a high-speed internal bus to form a complete local inference and decision-making pipeline.

[0031] The data preprocessing module first receives the raw, multi-source, heterogeneous data stream from the data aggregation gateway. The first step in this preprocessing module is timestamp alignment and data packet reassembly. Because different sensor nodes have different sampling frequencies and communication delays, multimodal data at the same moment may be scattered across multiple data packets. To address this, the module maintains a sliding time window (10 seconds wide), sorts all data entering the window by timestamp, and performs linear interpolation or forward padding for missing sensor channels to ensure that the output data frames are strictly aligned in the time dimension. Subsequently, the module performs data cleaning. For body temperature data transmitted from the miniature thermal sensor, the system filters out outliers based on a preset physiologically reasonable threshold range (37.5℃ to 40.5℃). Readings outside this range are considered to be caused by sensor malfunction or environmental interference and are discarded and marked. For triaxial accelerometer data, the module calculates its instantaneous magnitude. If this instantaneous magnitude experiences a drastic change within a short period (such as a step change exceeding 3g) and the duration is less than 0.5 seconds, it is identified as motion artifact noise caused by sensor collision or detachment, and is smoothed or removed. The cleaned data is packaged into standardized time-series data blocks for subsequent feature extraction.

[0032] The feature extraction engine performs multi-level, multi-dimensional feature calculations on the preprocessed data. Its operation mechanism consists of two parallel threads, as shown in the appendix. Figure 3As shown. The first thread focuses on extracting statistical features in the time and frequency domains. For acceleration data, the thread calculates the integral of activity per unit time (default 1 hour), that is, integrating the acceleration modulus over time to obtain a scalar reflecting the overall motion intensity; it also calculates the percentage of stationary time, defined as the proportion of cumulative time with an acceleration modulus below 0.2g to the total duration; in addition, it calculates the variance of motion intensity to characterize the volatility of motion patterns. For body temperature data, the thread calculates the mean, standard deviation, and diurnal temperature range (defined as the difference between the highest and lowest body temperatures of the day) within a 24-hour sliding window. These indicators can effectively reflect changes in the physiological rhythms of livestock. For sound data, the thread extracts the fundamental frequency, harmonic energy distribution (i.e., the energy proportion of the first 5 harmonics), and energy proportion of specific frequency bands (such as 500Hz to 2000Hz) through short-time Fourier transform. These acoustic features are closely related to the calling behavior during estrus.

[0033] The second thread focuses on extracting behavioral pattern features. This thread incorporates a behavior state recognizer based on a Hidden Markov Model (HMM), whose state space is defined by five discrete behavior categories: "reclining," "standing," "walking," "running," and "mounting." The model's observation sequence is a combined feature vector of acceleration and gyroscope data, including acceleration magnitude, pitch angle, yaw angle, and their first-order differences. The Viterbi algorithm decodes the observation sequence, segmenting the livestock's continuous behavior into sequences of these discrete states. Based on this, the feature extraction engine statistically analyzes the duration, frequency, and state transition frequency (e.g., the number of transitions from "reclining" to "mounting") of each behavioral state within a preset time window (e.g., 6 hours). Specifically, the frequency and duration of "mounting" behavior are given high weight, as it is one of the most typical overt behaviors during estrus. The feature vectors output by the two threads are finally concatenated into a high-dimensional feature vector with approximately 85 dimensions, which serves as input to the local prediction model.

[0034] This local prediction model library stores and runs a lightweight machine learning model for predicting livestock estrus. The local prediction model employs a gradient boosting decision tree algorithm architecture, specifically containing 128 decision trees of depth 5. The model's training objective function is the sum of the logistic regression loss function and L1 and L2 regularization terms, expressed mathematically as follows: ; in, For the sample size, The labels are real (1 indicates estrus, 0 indicates no estrus). For the model to the first The predicted score for each sample. For model parameters, For the total number of parameters, and These are the L1 and L2 regularization coefficients, respectively. Determined through cross-validation, The model is designed to strike a good balance between complexity and generalization ability, ensuring efficient operation and avoiding overfitting even with limited computing resources on edge devices. During the inference phase, the model receives an 85-dimensional feature vector from the feature extraction engine, traverses 128 decision trees layer by layer, and ultimately outputs an estrus probability value between 0 and 1, along with a score representing the prediction confidence. This score is calculated based on the consistency of prediction results across all trees within the model.

[0035] The local decision-making and early warning module receives the estrus probability value and confidence score mentioned above and executes the final decision logic. This module has a preset dynamic decision threshold, initially set to 0.7. When the real-time calculated estrus probability value is greater than or equal to this preset dynamic decision threshold, and the confidence score is greater than the preset confidence threshold of 0.8, the module immediately triggers a high-priority estrus warning. The warning command directly drives the audible and visual alarm device deployed in the corresponding pen via the farm's internal communication network (such as an RS485 bus or LAN), emitting a flashing red light and a specific frequency beep to alert the on-site handlers. Simultaneously, the module sends a structured warning message to a preset management personnel mobile terminal via its built-in wireless communication module (supporting 4G / 5G or Wi-Fi). The message includes the livestock number, predicted estrus probability (e.g., 0.85), confidence score (e.g., 0.92), warning generation timestamp (accurate to the second), and suggested handling measures. If the current network connection is normal, the warning message is also asynchronously uploaded to the cloud-based collaborative management platform for archiving.

[0036] Furthermore, the local decision-making and early warning module integrates a feedback learning mechanism to continuously optimize decision thresholds. This mechanism records the trigger time of each early warning and compares it with the actual estrus time confirmed later by manual verification (e.g., veterinary examination) or pregnancy results. The system calculates the absolute error between the predicted and actual times and averages the error weekly. If the average error exceeds 12 hours for two consecutive weeks, the system automatically fine-tunes the dynamic decision threshold: if predictions are generally too early, the threshold is increased by 0.05; if predictions are generally too late, the threshold is decreased by 0.05. This adaptive mechanism ensures that the system maintains optimal early warning sensitivity and specificity under different seasons and physiological conditions.

[0037] The cloud-based collaborative management platform is deployed in a remote data center and interacts with edge intelligent computing units in various locations via intermittent network connections. The overall interaction relationship is shown in the attached figure. Figure 4 As shown in the attached diagram, the platform comprises three main subsystems: a model training and optimization center, a global data warehouse, and a farm management interface. The model training and optimization center employs a federated learning framework for model iteration, the principle of which is illustrated in the appendix. Figure 5As shown, the central system first encrypts and distributes the current global model parameters to all online edge nodes. Each edge node uses locally stored historical feature data and real estrus tags (derived from manual confirmation or breeding records) to calculate the update gradient of the model parameters locally, and then uploads the gradient to the cloud after homomorphic encryption. The cloud platform securely aggregates all encrypted gradients (e.g., using a weighted average), decrypts them, and updates the global model parameters. After multiple iterations (usually 5 to 10 rounds), a new version of the model with improved performance is generated, and then distributed to each edge unit through differential update technology (transmitting only the changes in model parameters), significantly reducing the amount of data transmitted and protecting data privacy.

[0038] A global data warehouse stores long-term, structured livestock data. Its data sources include early warning records, feature summaries, and equipment status logs uploaded from edge units, as well as livestock pedigree information, mating records, calving history, and vaccination records synchronized from the farm management system. The warehouse also integrates environmental parameters such as temperature, humidity, and ammonia concentration provided by the environmental monitoring system. Based on this multi-dimensional data, the system supports complex retrospective analyses, such as: calculating the average estrous cycle length by breed (e.g., Holstein, Simmental); analyzing the impact of parity (1st, 2nd, 3rd, and above) on estrus intensity; and evaluating the accuracy differences of prediction models under summer high temperatures and winter low temperatures. These analytical results are output in the form of statistical reports or visual charts, providing data support for adjustments to breeding and feeding strategies.

[0039] The farm management interface provides managers with a unified, visual operating portal. The interface displays the geographical location and operational status of all networked farms in the form of an electronic map, showing in real-time key indicators such as the online status of each farm's edge computing units, remaining battery power (if using backup power), and storage space utilization (e.g., 45GB / 64GB used). Managers can click on any farm area to drill down to individual livestock, viewing their activity curves, body temperature trends, sound energy distribution maps, and historical prediction probability curves for the past 7 days. The interface also supports manual confirmation or correction of system alerts: if the system reports a false positive, managers can mark it as a "false positive"; if the system misses an alert, they can manually add a "true estrus event." This manual feedback is automatically timestamped and includes the operator's ID, and is sent back to the model training and optimization center as a high-quality tag for the next round of model training, forming a closed-loop learning mechanism.

[0040] The entire system operates within a clearly defined hierarchical decision-making framework. The real-time layer, hosted by edge intelligent computing units, processes data streams ranging from seconds to minutes, with the core objective of achieving an end-to-end latency of less than 5 minutes from data acquisition to alert generation. The analysis layer, hosted by a cloud-based collaborative management platform, processes aggregated data from hours to days, responsible for trend identification, anomaly detection, and model iteration. The planning layer, also located in the cloud, processes long-term data from months to years, supporting high-level decisions such as breeding planning, production capacity forecasting, and resource allocation optimization. The three layers collaborate through clearly defined data interfaces and interaction protocols, ensuring efficient information flow across different time scales.

[0041] In summary, this embodiment constructs a highly robust, low-latency, and high-precision intelligent prediction and management system for livestock estrus through the fine deployment of multimodal sensor networks, the localized real-time processing capabilities of edge intelligent computing units, and the global optimization mechanism of a cloud-based collaborative management platform. This system not only solves the problem of high network dependence in traditional solutions but also achieves early and accurate detection of estrus signs through deep analysis of behavioral patterns and federated learning-driven model evolution, providing solid technical support for modern precision animal husbandry.

[0042] Example 2: Building upon the previous examples, this example further introduces a dynamic binding mechanism between environmental perception and individual identity to enhance the system's adaptability and accuracy in large-scale mixed-group farming scenarios. Specifically, the data aggregation gateway in the multimodal sensor network is no longer merely a passive data relay device, but is upgraded to an intelligent edge node with local environmental perception and individual identification capabilities.

[0043] Each data aggregation gateway, in addition to maintaining its original wireless communication functions, integrates an ultra-wideband (UWB) positioning base station and an environmental multi-parameter sensor. The UWB positioning base station uses bidirectional ranging with UWB tags integrated into the livestock collars to achieve centimeter-level real-time positioning of individual livestock within the barn, updating its location every 2 seconds. The environmental multi-parameter sensor continuously monitors temperature, humidity, carbon dioxide concentration, and light intensity within the barn, sampling once per minute. The gateway uses the positioning data to dynamically construct a mapping relationship between individual livestock and their microenvironment. For example, when the system detects that a livestock has been staying in a poorly ventilated area with high ammonia concentration for an extended period, its body temperature reading may rise due to environmental heat stress. In this case, the gateway will add environmental parameters as contextual information to the livestock's body temperature data packet, allowing the edge intelligent computing unit to perform environmental deviation correction during the data preprocessing stage.

[0044] Furthermore, to address the issue of easily confused individual livestock identities in mixed-group farming, the system introduces a dynamic identity binding mechanism based on behavioral fingerprints. While extracting behavioral pattern features, the feature extraction engine calculates a "behavioral fingerprint vector" for each individual livestock. This vector consists of stable behavioral features such as typical resting time distribution, average walking speed, and the timing of mounting behaviors. When the system detects a sensor node signal loss (e.g., collar detachment) followed by reappearance in a nearby area, the edge intelligent computing unit performs a similarity match (using cosine similarity) between the newly appearing node's behavioral fingerprint and fingerprints in the historical database. If the match exceeds 0.9, it automatically binds the animal to its original identity, preventing identity confusion caused by brief signal interruptions. This dynamic identity binding mechanism significantly improves the data continuity and individual tracking reliability of the system in high-density, dynamically grouped farming environments.

[0045] Regarding the local prediction model, this embodiment expands the input feature vector by adding three new feature dimensions: environmentally corrected body temperature deviation, the percentage of time an individual spends in their comfort zone, and the behavioral fingerprint stability index, bringing the total feature dimensions to 88. During model training, these new features have proven to significantly contribute to improving the specificity of estrus prediction, especially during the hot summer season, effectively reducing the false positive rate caused by heat stress.

[0046] The cloud-based collaborative management platform has also been upgraded accordingly, adding an "Environment-Behavior-Eros" correlation analysis module. This module utilizes environmental and behavioral data accumulated in the global data warehouse to construct a multivariate regression model, quantifying the influence weights of different environmental factors on estrus behavior. For example, the analysis results show that when the ambient temperature exceeds 28℃, the frequency of typical estrus behaviors in livestock (such as mounting) decreases by an average of 35%. At this time, the system will automatically send environmental compensation coefficients to the edge units of the corresponding farms to dynamically adjust the decision thresholds of the local prediction model, ensuring that the sensitivity of the early warning is not affected by environmental factors.

[0047] Through the above enhancements, this embodiment maintains the advantages of low latency and high robustness of the original system while further improving the individual identification accuracy and prediction accuracy in complex and dynamic breeding environments, providing a feasible technical path for the refined management of large-scale ranches with tens of thousands of head of livestock.

Claims

1. A livestock estrus intelligent prediction and breeding management system based on multimodal sensing, characterized in that, include: A multimodal sensor network, deployed inside the farm, is used to continuously collect multi-dimensional physiological and behavioral data of individual target livestock. The multimodal sensor network includes multiple heterogeneous sensor nodes that are fixed to the surface of the livestock in a non-invasive or minimally invasive manner. The heterogeneous sensing nodes include a miniature thermal sensor for collecting core body temperature data of livestock, a triaxial accelerometer and gyroscope combination sensor for collecting data on livestock activity, posture angle and motion trajectory, an infrared proximity sensor for collecting data on livestock eating and drinking behavior, and a miniature microphone for collecting data on livestock sound characteristics. All sensing nodes send the collected raw sensing data to the corresponding data aggregation gateway through a low-power wireless communication protocol. An edge intelligent computing unit, physically located near the multimodal sensor network and deployed locally in the farm, is used to receive and process raw sensor data from the multimodal sensor network. The edge intelligent computing unit includes a data preprocessing module, a feature extraction engine, a local prediction model library, and a local decision-making and early warning module. The data preprocessing module is used to perform timestamp alignment and data packet reassembly on the received multi-source heterogeneous raw sensor data, and to perform data cleaning operations. The data cleaning operations include filtering out abnormal body temperature readings based on preset physiological parameter reasonable range thresholds, and identifying and removing motion artifact noise based on the modulus change law of triaxial accelerometer data. The cloud-based collaborative management platform is deployed in a remote data center and interacts with the edge intelligent computing units distributed in various locations through intermittent network connections. The cloud-based collaborative management platform includes a model training and optimization center, a global data warehouse, and a farm management interactive interface.

2. The intelligent prediction and management system for livestock estrus based on multimodal sensing according to claim 1, characterized in that, The feature extraction engine is used to perform multi-level feature calculations on the preprocessed data. The operation of the feature extraction engine is divided into two parallel threads. The first thread focuses on time domain and frequency domain feature extraction. For acceleration data, it calculates the integral of activity, the proportion of stillness time, and the variance of exercise intensity per unit time. For body temperature data, it calculates the mean, standard deviation, and diurnal temperature difference within a 24-hour sliding window. For sound data, it extracts the fundamental frequency, harmonic energy distribution, and energy proportion of specific frequency bands. The second thread focuses on behavioral pattern feature extraction. The thread has a built-in behavior state recognizer based on a hidden Markov model. By analyzing the sequence of acceleration and posture data, it segments and identifies the continuous behavior of livestock into discrete states including lying down, standing, walking, running, and mounting, and counts the duration and transition frequency of each state within a preset time window. The local prediction model library stores and runs a lightweight livestock estrus prediction machine learning model. The model takes the multi-dimensional feature vector output by the feature extraction engine as input. The model adopts a gradient boosting decision tree algorithm architecture, which contains 128 decision trees with a depth of 5. The training and updating of the model are handled by the cloud collaborative management platform and are periodically synchronized to the edge intelligent computing unit through encrypted differential updates. During the inference phase, the model calculates the input feature vector and outputs an estrus probability value between 0 and 1 and a score representing the prediction confidence. The local decision-making and early warning module receives the estrus probability value and confidence score output by the local prediction model. The module has a preset dynamic decision threshold. It compares the real-time calculated estrus probability value with the dynamic decision threshold. When the estrus probability value is greater than or equal to the dynamic decision threshold and the confidence score is greater than the preset confidence threshold of 0.8, the module immediately generates a high-priority estrus early warning command. This command activates the audible and visual alarm devices deployed in the corresponding livestock pens via the farm's local communication network. Simultaneously, it sends an early warning message containing the livestock number, predicted probability, confidence score, and timestamp to a preset management personnel mobile terminal via a wireless communication module. The module also integrates a feedback learning mechanism, recording the trigger time of each early warning and comparing it with the actual estrus time subsequently confirmed manually or based on pregnancy results. Based on the statistical results of the prediction time error, the dynamic decision threshold is adaptively fine-tuned weekly.

3. The intelligent prediction and management system for livestock estrus based on multimodal sensing according to claim 1, characterized in that, The model training and optimization center is responsible for building and optimizing the livestock estrus prediction model. The center collects anonymized feature data samples and corresponding real estrus labels uploaded from all networked edge intelligent computing units to form a distributed training dataset, and uses a federated learning framework for model training. The specific process of the federated learning framework is as follows: Distribute the current global model parameters to each edge node participating in training; Each edge node uses locally stored historical data to calculate the update gradient of model parameters; Each node uploads the encrypted gradient to the cloud; The cloud platform securely aggregates the encrypted gradients and updates the global model parameters. After multiple iterations, a new version of the model is generated and distributed to each edge intelligent computing unit through differential update technology. The global data warehouse is used to store long-term, aggregated breeding data, including early warning records, livestock pedigree information, mating records, calving history, and environmental temperature and humidity data, and supports retrospective analysis and statistical report generation based on this data. The farm management interface provides managers with visual system monitoring and management functions. The interface displays the status of all networked farms in map form, and displays the online status, battery level, and storage space utilization of each farm's edge computing unit in real time. Managers can remotely query the recent activity curves, body temperature trends, and prediction history of any livestock through the interface, and can manually confirm or correct the warnings issued by the system. The manual feedback information will be sent back to the model training and optimization center as tags.

4. The intelligent prediction and management system for livestock estrus based on multimodal sensing according to claim 1, characterized in that, The data aggregation gateway in the multimodal sensor network adopts a star topology network. Each gateway is responsible for managing up to 32 sensor nodes within a radius of 50 meters. The gateway and the sensor nodes communicate using the 2.4GHz low-power Bluetooth 5.0 protocol. The communication interval is dynamically adjusted between 1 minute and 5 minutes according to the livestock activity status. The gateway itself is connected to the edge intelligent computing unit through wired Ethernet or 4G / 5G cellular network. The edge intelligent computing unit is implemented using an industrial-grade embedded system. Its core processor is a quad-core ARM architecture chip with a neural network processing unit and a main frequency of not less than 1.8GHz. The edge intelligent computing unit is equipped with no less than 4GB of running memory and 64GB of embedded solid-state storage to cache at least 30 days of local feature data and event logs. Its operating system is a trimmed real-time Linux kernel.

5. The intelligent prediction and management system for livestock estrus based on multimodal sensing according to claim 1, characterized in that, The local prediction model uses a gradient boosting decision tree algorithm, whose training objective function is the sum of the logistic regression loss function and L1 and L2 regularization terms. The regularization coefficients are determined through cross-validation, where the L1 regularization coefficient is 0.01 and the L2 regularization coefficient is 0.

1. The initial value of the dynamic decision threshold is 0.7, and the fine-tuning step size of the feedback learning mechanism for adaptively fine-tuning the dynamic decision threshold is 0.

05.

6. The intelligent prediction and management system for livestock estrus based on multimodal sensing according to claim 1, characterized in that, The system operates within a hierarchical decision-making framework, which includes a real-time layer, an analysis layer, and a planning layer. The real-time layer is carried by the edge intelligent computing unit, which processes data from the second to the minute level, with the core objective of achieving an early warning delay of less than 5 minutes. The analysis layer is hosted by the cloud-based collaborative management platform, which processes data from hourly to daily levels and is responsible for trend analysis and model iteration. The planning layer is located in the cloud and processes data from the monthly to the grade level, supporting breeding planning and production capacity forecasting.

7. The intelligent prediction and management system for livestock estrus based on multimodal sensing according to claim 1, characterized in that, The data preprocessing module performs motion artifact noise identification on the triaxial accelerometer data as follows: Calculate the instantaneous magnitude of the acceleration data. If the instantaneous magnitude undergoes a step change of more than 3g within a short period of time and the duration is less than 0.5 seconds, it is identified as motion artifact noise and is removed.

8. The intelligent prediction and management system for livestock estrus based on multimodal sensing according to claim 1, characterized in that, The feature extraction engine uses a hidden Markov model-based behavior state recognizer, whose state space is defined as five discrete behavior categories: lying down, standing, walking, running, and climbing. The model's observation sequence is a feature vector combining acceleration and gyroscope data. The Viterbi algorithm is used to decode the observation sequence, dividing the livestock's continuous behavior into a sequence of discrete states.

9. The intelligent prediction and management system for livestock estrus based on multimodal sensing according to claim 1, characterized in that, In the federated learning framework, each edge node calculates the updated gradient of the model parameters locally, then homomorphically encrypts the gradient before uploading it to the cloud.

10. The intelligent prediction and management system for livestock estrus based on multimodal sensing according to claim 1, characterized in that, The farm management interface also supports manual marking of system warnings. If the system reports a false alarm, it is marked as a false positive. If the system misses a warning, a real estrus event is manually added. The marking information is sent back to the model training and optimization center as a high-quality label for model training.

Images