Attitude angle and track fusion dairy cow oestrus recognition system based on low-power-consumption Bluetooth ear tag

Abstract

The invention discloses a dairy cow oestrus recognition system based on attitude angle and track fusion of a low-power-consumption Bluetooth ear tag. The dairy cow oestrus recognition system comprises an attitude angle acquisition terminal module, a track behavior reconstruction module, a high-frequency sampling trigger module, an edge recognition and preprocessing module, a fusion judgment and early warning output module and a cloud platform collaboration module. The system periodically collects three-axis attitude angle data through Bluetooth ear tag equipment worn on the ears of a dairy cow, reconstructs a virtual track point string in combination with wireless signal intensity, and identifies behavior trends and spatial path characteristics of individuals respectively; in a suspected state, a short-time high-frequency sampling window is automatically triggered to capture subtle disturbance behaviors, and recognition of complex behaviors such as silent oestrus, short-time oestrus and passive oestrus is achieved. The system adopts a multi-channel collaborative recognition and fusion voting mechanism to output an oestrus judgment result, has the advantages of low power consumption, high adaptation, deployability, expandability and the like, and is suitable for behavior monitoring and intelligent breeding management of a large-scale cow farm.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent monitoring of animal husbandry, in particular to a posture angle and trajectory fusion estrus recognition system based on a low-power Bluetooth ear tag, which is a kind of intelligent system device for realizing multi-channel recognition and early warning output of estrus behavior state of dairy cows by using a wearable posture perception terminal and a wireless virtual positioning means, and belongs to the technical field of cross application of animal behavior recognition, intelligent breeding and edge perception calculation. BACKGROUND

[0002] With the development of intensive animal husbandry and the deepening of the concept of precision breeding, the efficiency of dairy cow reproduction management has become one of the important factors affecting the economic benefits of ranches. As a key link of reproduction management, the accuracy and timeliness of estrus recognition directly determine the success rate of breeding and labor cost. However, due to the significant differences in estrus performance among individual dairy cows, and the concealment or transience of some estrus behaviors, traditional manual patrol and single sensor monitoring methods have been difficult to meet the real-time, stable and low-power recognition needs of large-scale ranches.

[0003] At present, some studies have used inertial sensors, video monitoring or trajectory positioning to identify dairy cow estrus behavior, but there are the following key bottlenecks: 1. Single recognition dimension, easy to be disturbed by behavior differences: such as only through acceleration change or visual motion capture, it is difficult to accurately identify silent or passive estrus individuals; 2. High power consumption, difficult to realize long-term deployment: high-frequency continuous sampling or video stream processing requires a large amount of energy consumption, which is not suitable for small wearable devices such as ear tags; 3. Low system integration, lack of collaborative judgment mechanism: various sensing data cannot be effectively fused, and a robust and fault-tolerant comprehensive early warning system cannot be formed.

[0004] In order to solve the above problems, an estrus recognition system integrating three-axis posture angle perception, virtual trajectory reconstruction and high-frequency behavior capture is proposed, which realizes the complete path of data acquisition-edge recognition-cloud fusion-classified early warning relying on low-power Bluetooth ear tag terminal. Based on the "multi-channel recognition method" proposed by the applicant before, the system further implements its landing design in engineering structure, module cooperation, power consumption control and deployment mechanism, etc., to adapt to the practical needs of large-scale deployment, remote management and long-term operation of ranches, and to build a systematic solution for intelligent perception of estrus behavior. SUMMARY

[0005] The main purpose of the present application is to provide a dairy cow estrus recognition system suitable for large-scale ranch environment, with low power consumption, supporting multi-channel collaborative recognition and hierarchical early warning mechanism, to solve the problems of low recognition rate, high deployment cost and poor system stability of traditional recognition methods.

[0006] Specifically, the present application aims to achieve the following technical objectives: 1. Construct a systematic structure of fusion multi-channel recognition mechanism: by integrating three-axis attitude angle recognition module, virtual trajectory reconstruction module and high-frequency micro-disturbance action recognition module, the diversity of behavior performance of different types of cows in estrus process is fully covered, and the overall recognition rate and adaptability of the system are improved; 2. Realize the balance strategy of controllable power consumption and endurance: through the dynamic sampling scheduling mechanism of low-frequency sampling as the main and high-frequency triggering when necessary, combined with edge recognition and batch data upload design, the ear tag terminal energy consumption is significantly reduced, meeting the long-term wear and large-scale deployment requirements; 3. Establish an end-edge-cloud collaborative processing mechanism: the system supports terminal local prediction, edge module preliminary identification and cloud multi-channel fusion judgment, forming a closed loop process from data acquisition to result output, improving real-time performance and fault tolerance; 4. Adapt to multiple varieties of cows and special environmental scenarios: the system supports dynamic adjustment of recognition parameters and channel weights according to breed behavior differences, seasonal activity characteristics and changes in shed environment, ensuring stable operation and accurate recognition of the system under different individuals and environmental conditions; 5. Block the recognition scheme of non-system path: the system only performs behavior recognition based on attitude angle and virtual trajectory point string, does not rely on six-axis, nine-axis fusion devices or external vision sensing means, and builds a clear technical boundary to improve the specificity and patent anti-circumvention ability of the system solution.

[0007] In summary, the present application provides an engineering solution with high recognition rate, low power consumption, high adaptability and strong deployment capability for intelligent recognition of cow estrus through system-level architecture design and modular function integration, which is suitable for intelligent breeding systems of various scales in pastures.

[0008] I. System design logic basis and multi-channel collaborative recognition structure description

[0009] The attitude angle and trajectory fusion cow estrus recognition system based on low-power Bluetooth ear tag proposed by the present application has a system architecture and module configuration based on the deep understanding of the high difference and volatility of cow estrus behavior in physiological performance, behavior characteristics and data sensing level. In long-term practical application scenarios, it is found that the estrus behavior of cows often has the following common basis and recognition challenges:

[0010] (I) The basis of behavior and perception assumptions for system construction

[0011] 1. The inevitability of behavior occurrence As a high correlation of reproductive behavior, the external performance of estrus on individual cows has certain certainty, that is, every cow entering the estrus cycle will show perceptible behavior changes within a certain period of time, although the specific performance varies due to individual, physiological stage, and environmental factors.

[0012] 2. Heterogeneity of behavior characteristics The estrus behavior of cows varies significantly among individuals, in terms of action form, duration, intensity, and whether it is active or not. This heterogeneity determines that the recognition system cannot rely on a single data source or algorithm model, but must build a multi-path parallel perception and fusion decision mechanism.

[0013] (B) System recognition channel division and typical behavior type mapping

[0014] To cope with the above behavior heterogeneity and recognition complexity, the system designs three recognition channels, each of which runs independently and has autonomous recognition ability, and forms a parallel collaborative relationship in the system structure. The three channels correspond to the following typical estrus behavior types:

[0015] 1. Posture angle trend channel (low frequency) It mainly captures the structural posture behavior changes exhibited by cows in estrus, such as circling, frequent lying and standing, and turning left and right, which usually form obvious trends or periodic changes in three-axis posture angle data (Yaw, Pitch, Roll), and are suitable for intermittent low-power sampling mechanism extraction.

[0016] 2. Trajectory behavior channel (virtual positioning) It builds the relative activity path of cows through the cooperation of Bluetooth ear tags and multiple gateways, and is used to identify spatial behavior changes such as fence wandering, return path, and activity focus. This channel does not rely on physical positioning, but only constructs a "virtual trajectory point string" based on the RSSI signal strength trend to form the basis for spatial behavior modeling.

[0017] 3. High-frequency behavior perturbation channel It is suitable for identifying individuals with strong behavior concealment, such as "silent estrus", "passive estrus", and "climbed crossers", which are usually difficult to identify in regular sampling. When the system's recognition score reaches the suspected state, it triggers a short-time high-frequency posture angle sampling window to extract micro-motion features within a short period, thereby realizing the excavation of abnormal perturbation behavior.

[0018] (Three) Parallel collaborative structure design goals

[0019] The above three types of recognition channels run in parallel in the system structure, with the following design goals: Cover multiple types of estrus behavior: Ensure that different types of individuals can be perceived by the system, and improve the recognition coverage rate; Enhance recognition accuracy and confidence: channel independent scoring, synergistic fusion, support voting mechanism and confidence weighting; Guarantee recognition timeliness and power consumption control: through low-frequency main sampling + high-frequency trigger window mechanism, achieve low power consumption and high precision system balance; Realize edge and cloud collaborative recognition: part of the model can complete basic prediction in edge terminal, complex fusion recognition can be sent to cloud for execution, reduce delay and bandwidth consumption.

[0020] (Four) System module division and path closed loop structure Each module of the system is configured around the above identification channel to form the following closed loop logic: The attitude angle acquisition module and the trajectory point string construction module form the data input end; The edge recognition module undertakes basic state judgment and high-frequency window triggering; The cloud fusion module is responsible for cross-channel behavior trend modeling and final decision; The warning module adopts single-channel triggering and multi-channel fusion judgment dual mechanism to output results according to the three-channel scoring results; The power supply and dispatching module guarantees the stable operation of the system under low power consumption conditions, and the longest continuous deployment period can reach more than 36 months.

[0021] II. System composition and implementation method

[0022] The system includes the following six core functional modules:

[0023] (I) Ear tag type attitude angle acquisition module

[0024] This module is the front-end data acquisition unit of the system, integrated in the intelligent ear tag terminal worn by the cow, and its main function is to collect the spatial attitude change information of the individual cow in real time, and output in the form of three-axis attitude angle, which is used as the basic data input for the subsequent identification process. This module has the following structure and functional characteristics:

[0025] 1. Sensing parameter definition The attitude angle data includes but is not limited to the following three dimensions: Yaw: reflects the left and right rotation amplitude of the cow's head, suitable for identifying behaviors such as turning around in place and repeatedly turning the head; Pitch: reflects the inclination state of the cow's head, used to identify behaviors such as lifting the head, lowering the head, and getting up; Roll: reflects the lateral inclination state of the cow, suitable for identifying behaviors such as lying on the side, lying on the side, or asymmetric posture.

[0026] Three-axis attitude angle data is obtained through the built-in low-power gyroscope component, and the sampling frequency and accuracy can be dynamically adjusted according to the algorithm requirements.

[0027] 2. Sampling scheduling mechanism To balance the identification accuracy and terminal endurance, this module supports two sampling scheduling strategies:

[0028] a. Low-frequency intermittent sampling mechanism In normal monitoring state, short-time sampling is triggered at fixed period (e.g. 30 seconds to 2 minutes); Each sampling duration is 3-10 seconds; Used to capture daily behavior trend changes and abnormal behavior frequency characteristics.

[0029] b. High-frequency sampling window mechanism After suspected behavior is identified in low-frequency channel or trajectory channel, it can be triggered by the system; The typical sampling frequency of high-frequency window is 20-50 Hz, and the duration is about 15-60 seconds; Used to reveal high-timing details changes of perturbation type, short-time type and passive type estrus behavior.

[0030] The sampling scheduling is uniformly managed by the edge processor of the system, supporting dynamic switching and power optimization strategy control.

[0031] 3. Data preprocessing and feature structure design After the attitude angle raw data is preliminarily processed by this module, the following feature structures are generated: Attitude angle time series: used for subsequent behavior trend modeling; Sliding window standard deviation / mean change rate: used for frequent head turning and restlessness behavior identification; Extreme point density: identify repeated lying and in-place disturbance behavior; Derivative change sign: identify Yaw angle positive and negative switching features, suitable for circle turning and in-place rotation detection.

[0032] This module outputs the above feature summary data under edge computing conditions, and some of them can also be uploaded to the cloud platform for fusion identification.

[0033] 4. Communication and identification mechanism This module integrates a low-power Bluetooth (BLE 4.2 or BLE 5.0) communication unit, supporting the following communication functions: Periodic broadcast: ear tag terminal can broadcast device ID, attitude angle summary data and state marker regularly; Passive response upload: supports interactive upload mode with positioning gateway or mobile terminal; Signal strength output (RSSI): as the input basis of virtual positioning trajectory module, used for trajectory point string reconstruction.

[0034] 5. Structure and power consumption characteristics The module adopts an integrated packaging structure and has the following characteristics: Small size, light weight, suitable for long-term wearing on the ear of a dairy cow; Excellent power consumption control, a single battery can support operation for more than 180 days; Waterproof, dustproof and anti-impact ability, suitable for high humidity and high intensity environment in the pasture.

[0035] In summary, as the core perception component of the system, the module constitutes the basic support path of the attitude angle recognition channel, not only has high reliability data acquisition capability, but also realizes low power consumption and light deployment terminal monitoring function, providing key raw input and edge recognition data for subsequent recognition mechanism.

[0036] (II) Virtual positioning trajectory module

[0037] This module is used to build the motion trajectory point string of the individual dairy cow in space, and assist in identifying the common path feature changes in estrus behavior, such as fence wandering, repeated return, and activity area contraction. This module receives the broadcast signal of the low-power Bluetooth ear tag, and combines with the multi-point gateway cooperative positioning mechanism, to realize the trajectory trend recognition without GPS or high-precision positioning equipment, with the advantages of low deployment cost and high robustness.

[0038] 1. Trajectory construction basis and logical structure

[0039] The trajectory reconstruction logic of this module is based on the following path:

[0040] a. Bluetooth broadcast signal input (RSSI) The ear tag terminal periodically broadcasts the device ID and status; Multiple positioning gateways are arranged in key areas of the pasture to synchronously receive the broadcast signal; Each gateway records the received signal strength (RSSI), reception timestamp, and device ID.

[0041] b. Signal strength inversion positioning mechanism According to the RSSI signal strength received by different gateways, the approximate relative position of the dairy cow is inversed by using the trilateration or area estimation method; Construct a "virtual trajectory point string" on a two-dimensional plane; Each trajectory point is attached with a timestamp, signal reliability and other indicators.

[0042] c. Trajectory sliding modeling strategy The system aggregates the trajectory points according to a fixed time window (such as 30-60 minutes); Use fitting smoothing, outlier rejection and other algorithms to construct a continuous trajectory line; Support moving average, Kalman filter and other methods to improve trajectory stability and trend readability.

[0043] 2、Trajectory behavior feature recognition and output indicators

[0044] This module can identify the following typical estrus-related path features based on trajectory point strings:

[0045] a. Fence boundary wandering behavior Features: Trajectory points are concentrated on the boundary of the cage, and the back-and-forth turns are obvious. Criteria: Trajectory offset barycenter is close to the fence area, path changes frequently but range is stable.

[0046] b. Active range contraction / aggregation Features: Trajectory points are densely distributed in a corner or local area, and the activity radius decreases. Criteria: Trajectory path length decreases per unit time, area coverage decreases, and spatial variance decreases.

[0047] c. Path back-and-forth / repeated walking Features: Trajectory shows obvious path overlap in a short time. Criteria: Trajectory sub-section similarity is high (e.g. DTW distance < threshold), and back-and-forth times ≥ 3.

[0048] The system outputs corresponding score indicators for each type of feature as structured output for trajectory recognition channels.

[0049] 3、Edge and cloud collaborative analysis mechanism To ensure system efficiency and energy consumption control, this module uses a "edge preliminary screening + cloud fusion" collaborative recognition mechanism: Edge gateway local processing: Perform basic trajectory point calculation, noise removal, and preliminary feature extraction. Cloud trajectory model analysis: Perform high-order trajectory modeling, behavior trend classification, and recognition probability calculation. Edge-cloud interaction mechanism: Edge devices only upload trajectory summary data and abnormal markers, saving communication resources.

[0050] 4、Trajectory data structure definition and interface format To ensure that virtual trajectory point strings have good system readability and recognition algorithm compatibility, this module defines the trajectory data structure uniformly. Each trajectory record contains the following key fields: Device identification field (Device_ID): Used to uniquely identify the identity number of the ear tag terminal, ensuring that the trajectory data of different individuals can be independently tracked and managed. Timestamp: records the specific time information when the trajectory point is generated, which is used to build time series and analyze behavior rhythm; Virtual coordinate (X / Y coordinate): two-dimensional relative coordinates calculated based on RSSI signal strength inversion, representing the estimated position of the trajectory point in the pasture plane coordinate system; RSSI Set: contains the Bluetooth signal strength values received by multiple gateway nodes at the time of trajectory point generation, which is used for multi-source calculation and credibility analysis of trajectory points; Validity Score: a score value calculated based on RSSI data distribution, gateway geometric distribution and algorithm confidence, reflecting the availability and accuracy of the trajectory point; Zone Label: used to label the functional area to which the trajectory point belongs, such as "south corner of the shed", "drinking area", "rest area", etc., to support area heat analysis and behavior spatial aggregation degree judgment; Behavior Flag: a structured result output by the trajectory algorithm recognition module, indicating whether the point matches certain abnormal path behavior characteristics (such as returning, detouring, gathering, etc.), which is an important input basis for subsequent behavior recognition channels.

[0051] The above data structure can be sent to the recognition module or cloud platform by the edge gateway after preliminary packaging, realizing behavior trend modeling, trajectory clustering analysis, estrus path early warning, etc., while supporting historical data backtracking and multi-cow behavior comparison analysis. This structure has good scalability, and can add extension fields such as path speed, weighted direction, and interaction area in subsequent versions to adapt to different algorithm model calling needs.

[0052] 5. Deployment and scalability This module does not rely on high-cost positioning means such as GPS and UWB; The number of required gateways can be flexibly adjusted according to the size of the site, with good deployment adaptability; Supporting AoA, ToF and other higher precision Bluetooth positioning schemes; Suitable for open pasture and shed scenes, with day and night running ability.

[0053] In summary, the virtual positioning trajectory module is the basic component of the second recognition channel of the system, which can construct parallel recognition paths with attitude angle data, independently realize abnormal detection of trajectory behavior, and also participate in multi-channel fusion judgment mechanism. It is a key spatial behavior perception module in the estrus recognition process.

[0054] (Three) Edge recognition module

[0055] This module serves as an intermediate recognition unit in the system, deployed in the intelligent ear tag terminal or pasture gateway edge device, responsible for the preliminary identification of attitude angle and trajectory data and the local execution of high-frequency trigger logic. This module has the characteristics of low delay, low power consumption, and high response capability, and is a key component of the recognition process for "near-source judgment and on-site response".

[0056] 1. Deployment location and data input source

[0057] The edge recognition module can have the following two physical deployment forms according to the system deployment strategy:

[0058] Ear tag built-in: The recognition algorithm is embedded in the ear tag main control chip, and the pre-processing of attitude angle data and behavior scoring logic are directly executed locally in the terminal. It is suitable for ear tag structures with computing power;

[0059] Gateway processing: The recognition logic is deployed in the edge processor of the receiving gateway, which processes broadcast data from multiple ear tags in a centralized manner to achieve centralized concurrent behavior recognition scheduling.

[0060] The data sources received by this module include: Low-frequency sampling data of attitude angle (Yaw, Pitch, Roll); Virtual trajectory point string and signal strength data; High-frequency sampling trigger conditions and sampling window configuration instructions.

[0061] 2. Behavior feature recognition function

[0062] The edge recognition module is responsible for the first round of analysis and risk scoring of attitude angle data and trajectory features. Its typical functions include but are not limited to: Circle behavior recognition: By identifying the periodic fluctuations and amplitude changes of the Yaw angle, it determines whether there is a behavior of turning around in place or repeating the path; Frequent head turning / lying down recognition: Statistics on high-frequency small fluctuations of Yaw and Pitch axes in attitude angle to determine whether there is a state of accelerated behavior rhythm, frequent lying down, etc. Local aggregation and path repetition judgment: Combining parameters such as spatial aggregation degree, path coincidence degree, and area stay time of trajectory points, it identifies whether there is an abnormal path behavior.

[0063] The above behaviors can be realized through algorithms such as sliding window standard deviation calculation, extreme point density analysis, and trajectory similarity evaluation, and finally output the confidence score of each behavior type.

[0064] 3. High-frequency sampling trigger logic and response mechanism

[0065] The edge recognition module has the ability to monitor abnormal behaviors in real time and supports the following high-frequency sampling window trigger mechanism:

[0066] Passive trigger: When the low-frequency recognition score or trajectory channel recognition score exceeds the suspected threshold (such as score > 0.6), the edge module immediately sends an open high-frequency sampling window instruction to the ear tag.

[0067] Active sampling trigger: According to the platform scheduling strategy, random sampling of individuals who do not show obvious abnormalities is performed at fixed time periods or uncertain points defined by the system every day, ensuring that silent or intermittent individuals are not missed.

[0068] The module records the trigger results and synchronously uploads the sampling instruction response status to the cloud to form a complete closed loop of the recognition chain.

[0069] 4. Score output and collaborative interface

[0070] The edge recognition module outputs various behavior recognition results in the form of structured scores, including the following main fields: Behavior type label (such as turning head, rotating in place, path return, etc.); Behavior confidence score (0-1 floating point value); Data source label (pose angle / trajectory); Whether to suggest opening the high-frequency window mark; Current recognition window identification (start and end time, recognition round).

[0071] The output score can be used for local fusion judgment, or sent to the cloud through the communication module to participate in the full-channel fusion recognition process and historical behavior modeling.

[0072] 5. Module capabilities and technical advantage explanation Supports near-end rapid processing, reducing data transmission volume and platform load; Can support new recognition models and new behavior patterns through firmware upgrades; Flexible architecture, adaptable to centralized deployment (gateway processing) or distributed deployment (terminal embedded); Excellent energy consumption control, does not affect terminal endurance; Fault tolerance and edge caching mechanism, supports local judgment in offline mode.

[0073] In summary, the edge recognition module plays a key role in the system as a "local recognition, near-source response" unit, not only reducing recognition latency and communication burden, but also providing support for high-frequency sampling scheduling execution and risk score advancement. It is an indispensable central execution unit in the three-channel fusion structure.

[0074] (Four) Cloud fusion judgment module

[0075] This module serves as the centralized decision-making center of the system, deployed in the cloud platform or local private cloud server, mainly responsible for receiving scoring results and behavior feature data from multiple recognition channels (low-frequency attitude angle, trajectory analysis, high-frequency sampling), and making unified decisions through multi-channel fusion algorithms to generate comprehensive recognition results of estrus state. This module has key capabilities such as recognition logic aggregation, scoring strategy management, model upgrade, and behavior trend tracking, and is the core platform component for large-scale, intelligent estrus early warning.

[0076] 1. Input structure and data sources The data input of the cloud fusion decision module mainly includes the following three types of structured results: Low-frequency attitude angle recognition channel output: behavior scoring results from edge recognition modules or terminal processors, such as head turning frequency, circle rhythm, and standing-up frequency; Trajectory analysis channel output: path feature labels and confidence scores identified by virtual trajectory point strings, such as fence wandering, activity range contraction, and path return; High-frequency sampling channel output: high-time sequence recognition results generated after passive or active high-frequency sampling, extracting detailed feature scores such as short-time perturbation, reactive sharp turn, and continuous disturbance.

[0077] All channel outputs will be standardized, unified into behavior labels + confidence structure form, and attached with timestamp and device identification information.

[0078] 2. Multi-channel fusion recognition mechanism The cloud module integrates multiple channel output results through fusion algorithms to improve overall recognition accuracy and robustness. Fusion strategies include but are not limited to:

[0079] a. Threshold priority decision mechanism (Early Trigger) If any channel output score exceeds the set strong confidence threshold (such as ≥ 0.95), it is considered as an independent decision of estrus event; At this time, the system directly enters the estrus state output without waiting for the results of other channels.

[0080] b. Weighted voting fusion mechanism (Majority Fusion) If no single channel meets the strong confidence decision, the output results of each channel are combined by weighting according to channel weights; The typical strategy is "any two channel scores ≥ 0.6" to trigger estrus warning, or set the trajectory channel + high-frequency channel combination as the priority.

[0081] c. Behavior trend enhancement mechanism The system continuously records the behavior score change trend of each cow; If the individual score continues to rise or repeatedly reaches the medium intensity score over a period of time, the system will trigger the "trend enhancement" decision mechanism; It can effectively identify intermittent, fluctuating, and delayed estrus performance.

[0082] 3. Identification tag output and state management The cloud fusion module generates a structured output tag for each identification round, including the following fields: Individual ID (corresponding to ear tag device); Current fusion score (0-1); Decision result (estrus / non-estrus / suspicious); Dominant channel (the highest scoring identification channel); Decision reason summary (e.g., "trajectory aggregation + high-frequency disturbance"); Trigger time and identification window duration; Current behavior state tag update record.

[0083] The system synchronously pushes the decision result to the user end and records it to the individual behavior archive for long-term trend analysis and historical behavior correlation modeling.

[0084] 4. Model flexibility and platform expansion mechanism This module supports the following capabilities to meet long-term iteration and scene adaptation needs:

[0085] a. Multi-model scheduling support Support for rule engines, statistical scoring models, machine learning classifiers, or neural network models, and other fusion paths; Support model switching and parameter optimization based on identification accuracy feedback.

[0086] b. Identification channel scalability In addition to the existing three-channel structure, a fourth channel (such as body temperature channel, voice channel) can be added in the future; The system supports the addition of new channels participating in weighted fusion and voting mechanisms together with existing paths.

[0087] c. Edge feedback and self-learning mechanism The cloud platform can receive user behavior annotations or false positive feedback information; Support for training self-learning models by continuously accumulating sample data to improve scene adaptation capabilities.

[0088] 5. Identification closed loop and platform linkage mechanism After the fusion decision result is output, the system will trigger the following linkage operations: Update the individual state to "estrus"; Push warning notifications to the user interface or APP end; Record the relevant data (scoring structure, feature summary, channel path) for this identification into the behavior log database; Control the subsequent sampling rhythm, such as limiting repeated warnings, activating the status tracking channel, or reducing the sampling frequency.

[0089] In summary, the cloud-based fusion judgment module constitutes the core decision-making center of this system. Through the unified fusion of multi-channel recognition paths, trend modeling, and state management mechanisms, it achieves high-confidence, low-latency, and structured intelligent judgment of dairy cow estrus behavior. It is the core hub for this system to achieve the design goals of "early recognition, wide coverage, and low power consumption".

[0090] (v) Early warning output module

[0091] This module, serving as the final execution unit in the system's identification chain, is primarily responsible for synchronously outputting estrus identification results to ranch managers, the platform system, or third-party interfaces in a visual and operable manner, forming a timely and intuitive early warning and response mechanism. This module supports multiple output formats, multi-terminal synchronization mechanisms, and multi-level information access control, ensuring that early warning information is accurately and efficiently delivered to key management personnel or system interfaces in the first instance, thus forming a closed loop of "identification-response-intervention."

[0092] 1. Warning Triggering Conditions The startup mechanism of this module is based on the three-channel recognition path proposed in this invention. The three channels include: a low-frequency sampling recognition channel for attitude angles, a trajectory recognition channel, and a high-frequency sampling recognition channel for attitude angles. The three channels maintain a parallel and collaborative processing architecture during system operation. The recognition results can be used independently for early warning judgment, or they can be fused and enhanced to form a multi-path comprehensive early warning judgment.

[0093] To improve response sensitivity while maintaining judgment accuracy, this module introduces a multi-source independent early warning triggering mechanism and a fusion-enhanced judgment mechanism that are executed in parallel. The specific early warning triggering logic is as follows:

[0094] a. Single-channel strong confidence triggering mechanism: When any identification channel (including the low-frequency sampling identification channel of attitude angle, the trajectory identification channel, or the high-frequency sampling identification channel of attitude angle) outputs an estrus identification score value that reaches the preset high confidence threshold (e.g., ≥ 0.95) in the current identification window period, it is determined that the channel has the stability and reliability to independently complete the judgment. The system can directly trigger the estrus warning output without waiting for other channels to participate, ensuring the response timeliness under high intensity signals.

[0095] b. Multi-channel fusion voting trigger mechanism: If no single-channel score reaches the strong confidence threshold, the system will collect the recognition outputs of the three channels in parallel and perform fusion judgment based on channel weight settings and score values, which specifically includes: Any two-channel score simultaneously exceeds the medium confidence threshold (e.g., ≥ 0.6); The weighted fusion score result exceeds the fusion warning threshold (e.g., Σ(score_i × weight_i) ≥ 0.85); Synergistic estrus behavior characteristics are detected between recognition channels (e.g., trajectory return and posture disturbance appear synchronously); Such mechanisms take into account the advantages of multi-source synergy, improving recognition stability and accuracy.

[0096] c. Behavior trend enhancement trigger mechanism: For a channel recognition score that shows stable increase or repeated fluctuations near the threshold in consecutive periods, the system will invoke the trend enhancement model for trend curve evaluation. If the score growth trend is significant, it can also trigger an early warning to avoid late estrus signal missed recognition.

[0097] d. Administrator attention trigger mechanism (optional configuration): The system supports manual setting of "key attention" individuals. Even if the recognition score has not reached the regular threshold, it can also trigger level one or two warnings in advance according to the strategy setting, which is used for reproductive tracking, breeding evaluation, and other targeted management scenarios.

[0098] Through the above four-level trigger logic design, the module ensures dual capabilities of "highly reliable single-channel immediate response" and "multi-channel fusion enhanced judgment", ensuring low false negative rate while achieving high timeliness and adaptability of warning response capability.

[0099] 2. Output content structure Each warning information contains the following core fields: Individual identification (ear tag number or cow number); Estrus judgment result (estrus / suspected / normal); Fusion recognition score (confidence 0-1); Dominant behavior characteristics (e.g., "enclosure + high-frequency disturbance"); Warning generation timestamp; Current state suggestion (e.g., "reproductive management intervention can be performed"); Source channel combination (posture angle / trajectory / high-frequency sampling); Recognition round and log tracking number.

[0100] This structure supports standardized platform interface calls and can also be rendered locally as human-readable content.

[0101] 3. Multi-terminal early warning synchronization mechanism To adapt to the management mode and hardware conditions of pastures of different sizes, this module supports a multi-terminal synchronization mechanism, and the output forms include but are not limited to:

[0102] a. Mobile terminal push: Push instant early warning messages to the pasture manager APP or WeChat applet; Support click to view behavior details, score trend chart and historical records; Can set up group attention individuals and customize notification rules (such as only high score triggers).

[0103] b. Platform pop-up and Web interface display: Pop up identification prompts on the Web management platform; Mark the spatial position and state color coding of estrus cows in the group view; Support "unhandled early warning" list management and event confirmation.

[0104] c. SMS and voice notification (optional): Send SMS to the preset mobile phone; For users who do not often use the platform system, you can set up a voice broadcast information (need to be configured).

[0105] d. Pasture large screen synchronization display (optional): Display early warning information in the form of icons on the pasture monitoring large screen; Mark the estrus cow position with the camera to assist manual intervention.

[0106] e. Third-party system interface: Support pushing early warning information through RESTful API, MQTT, Modbus, etc. Standard protocols; Can be connected with pasture ERP system, reproductive management system or feeding control system to realize cross-system linkage.

[0107] 4. Multi-level permission and information control strategy This module has a built-in information grading mechanism that supports different roles and permission receiving strategy configurations: Ordinary operator: Receive basic early warning information (individual + time); Reproductive manager: Receive detailed behavior summary, historical score and judgment path; System administrator: Can view identification algorithm channel score, log number and misjudgment rate trend; Third-party interface: Can receive desensitized structured data for secondary analysis.

[0108] The permission control mechanism ensures that different users are managed and matched while protecting data privacy.

[0109] 5. Early warning management and response mechanism To support the subsequent management closed loop, the module also includes the following functions: Early warning status marking (viewed / handled / ignored); Support one-key allocation intervention task (such as "arrange breeding check"); All early warning information is recorded to the platform log module, supporting tracking and statistical analysis; Individual priority, response strategy, reminder frequency and other personalized configurations can be set.

[0110] In summary, the early warning output module constitutes the result transmission core of the system, ensuring that the estrus recognition results with high confidence can be delivered to the terminal system of different roles in an efficient, clear and controllable manner, supporting the pasture personnel to respond and take intervention measures at the first time. It is an indispensable execution link for the estrus monitoring system to build an intelligent farming scene.

[0111] (Six) Power management and low-power operation guarantee module This module is the basic guarantee unit for the long-term stable operation of the system, and builds a system-level closed loop around "low-power sampling mechanism + energy-saving communication path + signal blocking strategy", ensuring that the continuous estrus behavior recognition and early warning function of dairy cows can be realized without relying on complex perception fusion path.

[0112] To achieve ultra-long endurance and stable recognition ability, the system is designed collaboratively at the hardware and algorithm levels, forming the following operation guarantee strategies:

[0113] 1. Intermittent sampling driving mechanism The system implements dynamic regulation on the posture angle data collection process based on the "low-frequency trigger + high-frequency short window" sampling mechanism. In the daily recognition process, only three-axis posture angle data (Yaw, Pitch, Roll) is periodically sampled, and when the behavior trend period is found, the high-frequency window sampling mode is entered, thereby greatly reducing the overall energy consumption of the system and improving the behavior sensitivity and resource utilization efficiency of data collection.

[0114] 2. Low-power Bluetooth communication architecture The system integrates the short-range broadcast mechanism of BLE protocol, and uses the low-rate, short-message periodic broadcast mode to complete the device identity identification and part of the recognition data upload. In the non-recognition period, the terminal device maintains a sleep state, further prolonging the battery usage period. The platform also supports breakpoint supplement and edge cache strategy to avoid data waste and energy consumption growth caused by unstable communication.

[0115] 3. Long-cycle endurance capability guarantee Through the above energy-saving mechanism and ultra-low power hardware design, the system supports the terminal device to run continuously for more than 36 months without battery replacement, covering most of the dairy cow breeding management cycle, and significantly reducing the terminal operation frequency and labor cost.

[0116] 4. Identification signal path definition statement The identification mechanism used by the system is based only on three-axis attitude angle signals (i.e., yaw angle, pitch angle, and roll angle), and does not rely on six-axis, nine-axis, or other fusion sensor paths. The system does not integrate acceleration data, temperature, heart rate, and other sensing signal channels, nor does it introduce external images, sounds, or physiological indicators as identification factors, thereby avoiding energy consumption, delay, and misidentification problems caused by multi-signal fusion from the source.

[0117] 5. System deployment adaptability and anti-circumvention path blocking The system structure does not depend on ear tag shell shape, material structure, or wearing form, and does not require a specific appearance. It can adapt to different types of ear tag shells for structural bearing. The system blocks all possibilities of circumventing the identification of the invention based on temperature, heart rate, multi-axis fusion, etc., forming a closed identification signal path and enhancing the strength of patent protection.

[0118] In summary, based on the protection of terminal ultra-long endurance, the module provides system-level support for the actual deployment and long-term stable operation of the system in large-scale pastures through identification signal closure design, communication architecture low-power optimization, and operation cycle reduction strategies, which is an important capability foundation for the landing and transformation of the invention.

[0119] III. System structure and module function description

[0120] The system aims to realize intelligent identification and accurate early warning of dairy cow estrus behavior, and to build a low-power, scalable, and high-credibility fusion identification platform. To this end, the overall structure of the system is designed as a "six-module collaborative operation" architecture, with each module having clear division of labor and close collaboration in identification data collection, behavior trend determination, trajectory reconstruction, early warning output, and system endurance, etc., to build a complete identification and response closed loop from the front-end wearable device to the back-end management platform.

[0121] 1. The ear tag type attitude angle acquisition module, as the front-end sensing unit of the system, is worn on the ear of the dairy cow and is responsible for real-time acquisition of the attitude angle information (Yaw, Pitch, Roll) of the individual in three-dimensional space. The module supports low-frequency intermittent sampling mechanism and triggerable high-frequency short-time sampling window, and integrates BLE low-power Bluetooth communication capability, providing accurate and efficient time series data source for subsequent identification.

[0122] 2、Virtual positioning trajectory module based on Bluetooth broadcast and multi-gateway cooperative receiving mechanism, to build a relative two-dimensional space trajectory point string, through the analysis of the path form and regional characteristics (such as wandering, returning, fence gathering, etc.), auxiliary identification of typical activity mode during estrus. This module does not rely on GPS or UWB and other high-cost positioning methods, with flexible deployment, strong adaptability.

[0123] 3、Edge recognition module is deployed in ear tag terminal or local relay box, with the ability to independently process attitude angle data, which can locally and real-time identify estrus trend behaviors such as frequent head turning and circling. The module is equipped with a high-frequency triggering mechanism and a simplified model structure, supporting the completion of edge preliminary judgment in a low-power environment, improving the overall system response speed and network resource utilization.

[0124] 4、Cloud fusion judgment module is responsible for centralized processing of the results output by multiple channels (attitude angle low-frequency recognition channel, trajectory recognition channel, high-frequency detail behavior recognition channel). This module outputs a unified recognition probability and determines whether to trigger estrus warning through fusion strategy and probability weighting mechanism, which is the core decision unit of the system's recognition and decision.

[0125] 5、Early warning output module based on platform configuration logic, synchronously pushes the recognition results in the form of multiple terminals, including but not limited to SMS notification, APP pop-up window, pasture management large screen display, and platform background interface. The module supports targeted push of warning content according to roles (such as breeders, veterinarians, and farm owners), achieving multi-level improvement of management efficiency.

[0126] 6、Power management and low-power operation guarantee module ensures the system to maintain low maintenance and high stability in long-term operation. This module realizes stable operation of terminal devices for more than 180 days without battery replacement through intermittent sampling mechanism, BLE communication optimization strategy, and recognition signal path limitation mechanism. At the same time, this module blocks unnecessary sensing paths such as six-axis, nine-axis, temperature, and heart rate, effectively preventing the system from being bypassed by low-cost pseudo-fusion methods, ensuring the exclusivity and patent stability of the recognition path.

[0127] In summary, the system structure revolves around the key identification elements of estrus behavior, fuses attitude angle and trajectory information, adopts a dual-layer mechanism of terminal recognition and cloud fusion, and forms an intelligent estrus recognition platform with low power consumption, high precision, strong expansion, safety control, and other multiple advantages, especially suitable for the breeding management and digital upgrading needs of large-scale dairy farms.

[0128] It is worth mentioning that the identification algorithm logic, sampling mechanism configuration and determination path in the system can preferably call the technical solution of the application for "a dairy cow estrus recognition method based on posture angle and trajectory fusion" as the algorithm basis. The system is the specific application of the system implementation and deployment path of the method in the engineering layer, and the two can form a complementary technical protection system.

[0129] IV. System architecture closedness explanation and protection boundary declaration In order to protect the deployment structure, module coordination path and core identification mechanism of the system, prevent others from bypassing the system structure characteristics of the application through redundant fusion or non-key module addition, and specifically set this section to clearly explain the identification structure composition, core channel boundary and closed protection range of the system.

[0131] (I) Input path limitation of main identification system structure

[0132] The system described in the application is constructed based on double main input paths:

[0133] 1. Three-axis attitude angle data channel Yaw, Pitch and Roll data from the ear tag terminal enter the edge and cloud identification module through intermittent low-frequency sampling or high-frequency short-time sampling, constituting one of the main behavior characteristic channels of the system.

[0134] 2. Virtual trajectory data channel The trajectory point string is formed by the cooperation of Bluetooth broadcast and multi-gateway coordinated reception, and the relative activity path is output to construct, and the spatial behavior is analyzed in combination with the area label. The trajectory signal does not constitute a positioning service and does not depend on GPS, UWB and other systems, and its identification path is completely constructed by RSSI.

[0135] The above two channels constitute the core identification structure of the system, and all identification mechanisms are generated around them.

[0136] (II) Closed logic of high-frequency identification mechanism of attitude angle The high-frequency identification module in the system is an independent identification mechanism, which has the following limited boundaries: Only attitude angle signals are collected, without introducing acceleration, temperature, sound and other signals; The starting mechanism is based on suspected state triggering or platform sampling strategy; High-frequency sampling results can independently participate in identification and determination, and directly trigger an early warning when the confidence reaches a preset threshold (such as 95%); The module is deployed on the ear tag terminal or relay terminal and has independent edge running capability without cloud support.

[0137] This path serves as an enhanced module supplement, although it is an additional mechanism, but it still constitutes a complete closed structure.

[0138] (Three) System-level pseudo-fusion path exclusion declaration In order to prevent the sensor fusion mode without independent identification value from bypassing the threat to the system, it is declared as follows: the path is not considered as an innovative path, and belongs to the protection range of the system of the application: Six-axis fusion path (attitude angle + acceleration): high power consumption, heavy algorithm, and overlapping features; Nine-axis fusion path (attitude angle + acceleration + geomagnetic): redundant information, high interference; Temperature, heart rate and other environmental signal auxiliary path: no behavior trend modeling capability, only for health monitoring data.

[0139] Therefore, any system that claims "multi-source fusion" but still uses attitude angle as the main axis of identification is considered a technical equivalent solution and falls within the protection boundary of the application.

[0140] (Four) Non-positioning attribute of trajectory path The "trajectory point string" used by the system is based on RSSI and only has behavior trend reference significance, and its characteristics are as follows: Non-geometric coordinates, no spatial absolute positioning provided; Tolerate error range 1.5-3 meters, identify behavior patterns through fuzzy matching; Used to identify behaviors such as staying, wandering, and area returning; Can independently reach the identification confidence threshold to trigger an early warning.

[0141] Therefore, the system does not rely on traditional positioning systems and only uses trajectory trends as identification elements.

[0142] (Five) Exclusion boundary of non-wearable and visual path The system does not involve the following identification paths: Image recognition, video analysis systems; Infrared, temperature sensing non-contact systems; Identification systems based on voice, sound, and electrical signals; Delayed trigger systems such as climbing frequency identification mechanisms.

[0143] The above paths do not constitute the technical basis of the system and are not applicable to the architecture of the system.

[0144] (Six) Closed system structure declaration The system claims the following structure combinations as protected ranges: Three-axis attitude angle data identification path; RSSI virtual trajectory point string construction path; Multi-module linked three-channel structure; Closed trigger mechanism of high-frequency recognition module BLE broadcast and lightweight configuration of edge deployment.

[0145] Any system that claims to identify the estrus state of a cow based on behavioral trends, and whose core signal source is the three-axis attitude angle or RSSI trajectory, is considered to fall within the scope of the patent protection of the present system if it does not explicitly construct an innovative mechanism outside the system structure path.

[0146] V. Definition of terms and explanation of path boundaries

[0147] To ensure the accuracy of the expression of the system architecture, recognition mechanism, and path boundaries of the present invention, and to prevent related recognition mechanisms from being confused and replaced by non-inventive paths, the key terms used in the present invention and their applicable boundaries are defined and explained as follows:

[0148] (1) Definition of "estrus recognition behavior chain"

[0149] The "estrus recognition behavior chain" described in the present invention refers to the entire process of collecting, analyzing, modeling, and warning output of various behavioral signals generated by a cow individual within a specific time period, covering multiple parallel recognition paths such as attitude angle trend channels, virtual trajectory recognition channels, and high-frequency behavior sampling channels, and forming a complete recognition closed loop through edge and cloud modules.

[0150] This behavior chain emphasizes the fusion expression of behavioral characteristics in time, space, and frequency dimensions, and is the basis for scheduling the core recognition mechanism of the present system.

[0151] (2) Definition and functional boundaries of "three-axis attitude angle"

[0152] "Attitude angle" refers to the three-axis spatial direction angle data obtained by the terminal collection module of the present system, including: Yaw (yaw angle): describes the left and right turning or turning actions; Pitch (pitch angle): reflects the up and down inclination such as lifting the head, lowering the head, or lying down; Roll (roll angle): represents left and right tilting, lying on the side, or lying on the side.

[0153] Attitude angle is the only perception source used to construct a closed behavior recognition model in the present system, and all trend recognition, rhythm modeling, and perturbation analysis in the channels are based on it.

[0154] (3) Definition of "virtual trajectory point string"

[0155] "Virtual trajectory point string" refers to the relative position sequence point set reconstructed based on RSSI strength after the low-power Bluetooth ear tag broadcast signal of the present system is received by multiple fixed gateways, which is used to: predict the path trend of a cow individual; extract trajectory features such as fence wandering, path returning, and area staying; trigger the suspected early warning logic of the trajectory identification channel.

[0156] It is characterized by not relying on GPS or UWB positioning, not generating real coordinates, and only being used for behavior trend identification. The invention protects the system that uses this mechanism for trajectory identification, but does not claim the right to the geometric positioning accuracy or coordinate expression form.

[0157] (Four) "attitude angle trend identification path" boundary description

[0158] The invention defines all behavior identification models constructed based on three-axis attitude angle time series trends as "attitude angle trend identification path". Including but not limited to: attitude angle fluctuation amplitude analysis; behavior rhythm identification; direction switching mode extraction; attitude stability offset detection, etc.

[0159] If any system identification mechanism uses attitude angle as the only input signal and constructs a behavior warning mechanism, it is considered to fall within the scope of the invention.

[0160] (Five) "high-frequency sampling window" mechanism definition

[0161] "High-frequency sampling window" refers to a short period of high-frequency attitude angle data collection triggered automatically by the system in the identification channel when the score of a certain channel reaches the suspected state, which is used to identify the following short-time difficult-to-catch behaviors: passive estrus or silent estrus; climbing and crossing reaction individuals; local kicking, rotating and other micro-disturbance actions.

[0162] This mechanism belongs to the reinforced identification channel in the system architecture and is controlled by algorithm threshold to start and stop, which does not constitute the regular energy consumption burden of the system.

[0163] (Six) "lock-type identification mechanism" definition and exclusion declaration

[0164] The attitude angle identification path in the system is a lock-type main channel mechanism, which specifically refers to the identification path based on three-axis attitude angle trend modeling. To clarify the scope of patent protection, it is declared as follows:

[0165] 1. Prohibit pseudo-fusion path avoidance: Any system that uses posture angle as the main input, but integrates other signals (such as acceleration, temperature, humidity, image, sound, etc.) and does not provide effective gain to the behavior criterion, only making redundant superposition, is also considered a pseudo-fusion avoidance method within the scope of this channel blockade, and its independence is not recognized.

[0166] 2. Exclusion of six-axis, nine-axis and multi-modal sensing alternative paths: Any recognition mechanism based on six-axis (posture angle + acceleration), nine-axis (including geomagnetic) or introducing image, microphone and other modal paths, which are not the main channel of the system, still fall within the protection boundary if the posture angle recognition model essence is not changed.

[0167] 3. Explanation of the independent main channel status of the trajectory channel: The virtual trajectory recognition mechanism in the system is a parallel main channel with the posture angle recognition, and is not a subordinate module of the posture angle recognition channel. Both are modeled in parallel and cooperatively determined.

[0168] 4. Explanation of the path recognition right protection range: Any system built on the posture angle time series recognition model, regardless of the sampling method (high frequency, low frequency or mixed mode), falls within the protection range of the blocking recognition mechanism of the present invention.

[0169] Six, beneficial effects

[0170] The present application provides a kind of based on low-power Bluetooth ear mark's cow estrus recognition system, system constructs the three-channel parallel recognition architecture consisting of posture angle low-frequency sampling module, virtual trajectory point string analysis module and high-frequency behavior recognition module, with high recognition sensitivity, low power adaptability and multi-mode deployment flexibility, embodied in the following aspects:

[0171] 1. System-level recognition channel design realizes comprehensive coverage of complex estrus behavior The present system breaks through the traditional method of relying only on acceleration or passive behavior triggering in the recognition path, and independently constructs three recognition channels of "posture angle trend recognition", "virtual trajectory deduction" and "high-frequency behavior reinforcement". Through system-level modular design, comprehensive judgment of behavior trend, spatial activity mode and micro-disturbance is realized, effectively covering the edge state of "silent estrus", "passive estrus", "intermittent estrus" and other traditional methods that are difficult to identify, and improving the response range and early warning depth of the system to complex behavior types.

[0172] 2. Multi-path parallel and fusion judgment strategy significantly enhances system recognition stability The system supports the synchronous operation and independent early warning output of three channels, combines the "score merging + weight voting" mechanism, strengthens the collaborative judgment ability between multiple source data, effectively reduces the system misjudgment risk caused by single channel recognition error, and guarantees the stability and accuracy of the output result. The mechanism is controllable at the software and hardware level, supports adaptive parameter adjustment, and adapts to different pasture recognition needs.

[0173] 3. Structure distributed sampling mechanism balances response ability and terminal energy consumption The system introduces an intermittent trigger high-frequency sampling mechanism in the architecture, triggered by low-frequency attitude angle recognition or trajectory anomaly point string, which activates the high-energy consumption module only when it enters the suspected stage, greatly reducing the terminal power consumption. The system as a whole maintains high recognition sensitivity while supporting continuous operation for more than 6 months, meeting the dual requirements of power consumption and service life for large-scale deployment.

[0174] 4. Edge-cloud collaborative architecture supports flexible deployment and scene adaptation Each recognition module of the system has a clear division of labor and deployment adaptability: The low-frequency attitude angle recognition module can be deployed on the terminal or platform edge node; The trajectory point string analysis module is deployed on the cloud platform, suitable for handling large amounts of device behavior paths and hot zone aggregation modeling; The high-frequency sampling recognition module runs on the ear tag terminal or near-end edge computing node, realizing local rapid response and behavior capture.

[0175] This architecture realizes the system design of "edge recognition and transmission, cloud judgment collaboration", which adapts to different network conditions, hardware capabilities and operation and maintenance strategies, and has wide landing feasibility.

[0176] 5. Recognition mechanism block path definition, system structure controllable, technical boundary clear The recognition mechanism used by the system is completely based on three-axis attitude angle and virtual trajectory point string signal source construction, excluding the calculation load and copyright dispute risk brought by multi-modal fusion path, image processing path or six-axis pseudo fusion path, with the advantages of strong controllability of system design, clear recognition path boundary, providing a solid foundation for subsequent industrialization promotion and intellectual property protection.

[0177] 6. System structure has good universality and can be applied to other ruminant behavior recognition needs The system platform supports quick replacement of recognition templates and parameter models, and can widely adapt to the behavior feature recognition needs of goats, buffaloes and other ruminants on the basis of maintaining the three-channel structure, with good horizontal expansion ability and industrial universality.

[0178] In summary, the system has significant technical advantages in structural design, recognition mechanism, module cooperation, energy consumption management, deployment adaptability and scalability, fully breaks through the technical bottlenecks of traditional solutions in recognition accuracy, power consumption control and deployment adaptability, and constructs a standardized deployment, industry replication and patent circumvention-proof dairy cow estrus recognition system solution. BRIEF DESCRIPTION OF DRAWINGS

[0179] Figure 1 : The overall structure diagram of the system of the application; Figure 2 : The structure diagram of the posture angle acquisition terminal module; Figure 3 : The flow chart of trajectory reconstruction and trajectory behavior recognition; Figure 4 : The flow chart of high-frequency sampling trigger mechanism and behavior recognition; Figure 5 : The structure diagram of edge recognition and cloud fusion determination module; Figure 6 : The system early warning output flow chart. DETAILED DESCRIPTION

[0180] In order to make the purpose, technical scheme and beneficial effects of the present application more clear and definite, the present application will be further described in detail below in combination with the drawings and examples. The present application is not limited to the following specific examples, and any equivalent replacement or improvement within the spirit and principles of the present application shall be included within the protection scope of the present application.

[0181] Example one: posture angle low-frequency sampling module in actual pasture recognition application

[0182] After deploying the system in a large pasture in North China, the posture angle low-frequency recognition module is configured to operate in an intermittent mode with a 1-minute sampling period and 5-second sampling time. The three-axis posture angle data (Yaw, Pitch, Roll) of the cow are collected by the ear tag built-in gyroscope. The cow numbered #0146 showed the following trends in the continuous 7-day low-frequency data: The Yaw angle fluctuation amplitude continued to increase, and the standard deviation increased from 4.2° to 9.8°; The Roll angle frequently switched, indicating repeated turning behavior; The high-frequency short-period signal was concentrated in the 0.12-0.25Hz interval, consistent with the fence wandering behavior model.

[0183] The posture angle low-frequency recognition module of the system determined the recognition probability to be 93.5% in this cycle, triggered the behavior label and high-frequency sampling module preparation state, and waited for the judgment of other channels.

[0184] Example two: trajectory recognition module independent trigger estrus early warning

[0185] The trajectory recognition module of the system reconstructs the virtual trajectory point string based on the RSSI signals between the ear tags and the multi-gateway, and updates the relative path every 5 minutes. The path reconstruction is performed on the behavior of cow #0146 at night (22:00-02:00), and the system finds the following trajectory characteristics: The in-pen back-and-forth path exceeds 15 times; The return rate is significantly higher than the group average; The direction switching angle is concentrated at 90°±15°, and the path linearization is significant.

[0186] The trajectory recognition module calculates the recognition probability of 98.1% according to the trajectory behavior model, which exceeds the set single-channel warning threshold, and directly issues the estrus warning by the fusion judgment and warning module, and synchronously uploads the behavior chain record.

[0187] Example Three: High-frequency sampling module assisted identification of estrus micro-disturbance action

[0188] After the trajectory module determines that cow #0146 is a suspected individual, the high-frequency sampling trigger module automatically starts a short high-frequency window (30Hz x 10s) to capture its micro-disturbance action. The data shows: Yaw angle appears rapid swing and large fluctuation; Pitch angle shows short-term rapid rise and fall; The disturbance frequency of Roll angle is as high as 2.8Hz, which is consistent with the characteristics of passive avoidance behavior.

[0189] The high-frequency recognition module of the system outputs the estrus probability of 90.6%, and the trajectory module forms a "double-channel consistent" high-confidence result. The fusion judgment module generates a formal warning accordingly, and records all behavior parameters and identification basis in the high-frequency sampling window.

[0190] Example Four: Three-channel fusion and continuous monitoring mechanism system linkage application

[0191] Under the continuous monitoring of the system, the recognition probability of cow #0221 is 68% for the low-frequency posture angle module, 73% for the trajectory module, and 94.3% for the high-frequency sampling module. According to the logic rule that "any channel recognition rate exceeding the set threshold triggers the warning", the system automatically triggers the warning mechanism.

[0192] To prevent "edge value early stop" misjudgment, the system synchronously starts the "high-frequency continuous recognition mechanism", maintains the high-frequency module running for 60 minutes, and monitors the change trajectory of the subsequent behavior chain. The system platform uploads the recognition results and feature logs at the set period for the breeder to review and intervene.

[0193] Example Five: Adaptive system behavior of dynamic adjustment of recognition threshold and fusion mechanism

[0194] In the long-term operation of the pasture, the system platform dynamically adjusts its single-channel threshold setting according to the historical recognition effect and false positive rate of each channel: the trajectory channel is set to 92%, and the high-frequency channel is adjusted to 96%. In the process of recognizing the behavior of cow #0385, the system recognition output is as follows: Posture angle low-frequency channel recognition rate: 91.5%; Trajectory recognition module recognition rate: 89.2%; High-frequency perturbation recognition module recognition rate: 87.4%.

[0195] The fusion judgment module uses a "three-channel voting mechanism" to calculate the comprehensive confidence as 91.3%, which is higher than the set system warning threshold 90%, successfully generating estrus warning. The system automatically labels the fusion path as "low frequency + trajectory", and records the threshold dynamic adjustment trajectory.

[0196] Example six: low-power deployment optimization and system endurance verification scheme

[0197] Deploy the system in a large-scale dairy farm in Inner Mongolia, select a low-power Bluetooth ear tag terminal that supports three-axis attitude angle collection, and run the modules described in the system according to the following strategy: The posture angle recognition module is configured to use an intermittent sampling method with a 60-second cycle and a 6-second sampling duration, and the edge processing module completes preliminary recognition and storage; The trajectory recognition module reconstructs the trajectory point string through the gateway at intervals of 15 minutes and uploads it to the cloud for processing; The high-frequency sampling trigger module is only started when there is a suspected state or regular sampling, with no more than 8 times per day and no more than 30 seconds each time.

[0198] The system platform sets data upload to a batch mode at the hourly level to reduce communication energy consumption. Running statistics show that under the standard battery capacity of 850mAh, the terminal device can run continuously for 783 days with more than 20% of the battery remaining. This deployment scheme verifies that the system has ultra-long endurance and extremely low power consumption characteristics under the premise of ensuring recognition sensitivity, making it suitable for large-scale dairy farm deployment.

[0199] Example seven: verification of system adaptation mechanism under environmental changes and breed differences

[0200] Deploy the system in an extreme low-temperature environment in a pasture in North China in winter, and the system automatically adapts the strategies of each module according to environmental parameters: The cloud platform prolongs the data reconstruction interval of the trajectory recognition module (from 15 minutes to 40 minutes) to reduce non-behavioral interference; The posture angle recognition module dynamically reduces the motion recognition threshold to enhance the perception ability of small movements; In the module coordination strategy, different channels and model parameters are set for different breeds of cows (such as Holstein and Jersey) to automatically adjust the recognition threshold and classification logic in edge computing.

[0201] The test results show that the estrus recognition accuracy of the system exceeds 93% under different individuals and environments, demonstrating the high adaptability and breed behavior feature perception ability of the system.

[0202] Example Eight: Verification of the Migration and Adaptability of the System in Other Ruminants

[0203] To verify the cross-species versatility of the system, the core modules of the system were deployed at a goat breeding farm in the southwest region: The same structure of Bluetooth ear tag devices was used, and only the parameters of the posture angle recognition module and the trajectory recognition module were adjusted; The low-frequency sampling period was set to 40 seconds, and the high-frequency trigger threshold was lowered to adapt to the size and activity frequency of goats; The cloud platform loaded trajectory behavior templates specific to goats, supporting path compression and lightweight modeling.

[0204] The test period was 60 days, and the estrus recognition results output by the system were compared with the on-site manual records, with an accuracy rate of 89% and an early warning rate of 74%. In addition, without replacing the terminal hardware, the system remained stable for 90 days, confirming that the system has good migration ability and behavior recognition ability for goats and other ruminants.

[0205] Example Nine: Application Example of Channel Closing and Intelligent Recovery Strategy in Three-Channel System

[0206] During the daily operation of the system, individual cows frequently triggered the high-frequency sampling module due to abnormal individual activity. To improve recognition efficiency and energy-saving ability, the following strategies were deployed: The edge recognition module sets a "continuous non-compliance trigger shutdown mechanism": if any recognition channel outputs less than 50% recognition probability for three consecutive rounds, the system automatically closes the sampling task for that channel; The platform fusion module also sets a "regular forced sampling mechanism", which resumes sampling for health verification twice every 48 hours even if the channel is closed; If any channel outputs more than 70% recognition probability, the system immediately restores the three-channel parallel recognition mechanism.

[0207] This strategy was verified in a group of 120 cows, with overall power consumption reduced by 34%, but recognition accuracy maintained above 95%, demonstrating the system's good channel dynamic control ability and recognition stability.

[0208] Example Ten: Verification of System-Level Misjudgment Filtering and Recognition Correction Mechanism

[0209] In daily operation, some cows have behavior pattern disturbances due to disturbances, changes in feeding rhythm, or the influence of barn structure. This system constructs three false alarm filtering mechanisms: Behavior chain backtracking mechanism: The cloud fusion module automatically compares the recognition results with the behavior chain within 6 hours. If the high recognition value is an isolated event, it is marked as "low confidence warning"; Group state comparison mechanism: When 3 or more individuals in the group output similar trends at the same time, the system suspends the warning and enters the "environmental disturbance evaluation mode"; Rhythmic anomaly filtering mechanism: If the individual recognition fluctuates frequently within a single day, the system automatically reduces the weight of the recognition result and delays the next round of judgment.

[0210] In the joint pilot of five farms in three provinces, the false alarm rate decreased by 42%, and 128 ineffective warning prompts were withdrawn. The overall recognition results of the system have significantly improved in stability, and the system has been verified to have high fault tolerance capability in complex environments.

Claims

1. A cow estrus recognition system based on low-power Bluetooth ear tags, characterized in that, The system includes: 1) Attitude angle acquisition module, which is installed in the Bluetooth ear tag terminal worn by the cow, is used to periodically collect the cow's three-axis attitude angle data, including yaw, pitch and roll. 2) The trajectory point string construction module receives the ear tag signal strength (RSSI) through multiple Bluetooth receiving gateways deployed in the ranch, and periodically reconstructs the virtual trajectory point string of individual cows to reflect their relative movement trend; 3) Posture angle trend recognition module, used to identify behavioral trends of dairy cows such as circling, frequent turning, and turning in place based on the time series features of posture angle data; 4) Trajectory behavior recognition module, used to identify trajectory behavior characteristics of dairy cows such as path retracing, pen wandering, and concentration of activity hotspots; 5) High-frequency sampling trigger module, used to control the ear tag to enter the short-term high-frequency attitude angle sampling window under the condition that the suspected estrus trend signal is output by any identification channel or the periodic self-trigger condition set by the system, so as to capture subtle behavioral features; 6) Fusion judgment and early warning module, used to independently output estrus warning based on the judgment result of any one of the three recognition channels, or to perform fusion enhancement judgment when multiple channels output at the same time, so as to improve the recognition confidence and fault tolerance.

2. The system according to claim 1, wherein, The attitude angle acquisition module adopts a low-power intermittent sampling mechanism, and its sampling period and duration are configurable to adapt to different deployment energy consumption requirements and recognition sensitivity requirements.

3. The system according to claim 1, wherein, The trajectory point string construction module infers the cow's activity direction and movement area by creating signal strength differences between multiple fixed Bluetooth gateways through ear tag broadcast signals, without relying on GPS, UWB or other absolute positioning systems.

4. The system according to claim 1, wherein, The attitude angle trend recognition module is deployed on the edge device to improve real-time performance and terminal response speed; the trajectory behavior recognition module is deployed on the cloud server to process large-scale data and match trajectory models.

5. The system according to claim 1, wherein, The high-frequency sampling triggering module supports a dual triggering mechanism, namely: 1) Triggered by suspected behavioral signals output by the attitude angle recognition or trajectory recognition module; 2) The system automatically triggers periodically according to the time window set by the system, which is used for sampling and detection of potentially silent estrus individuals.

6. The system according to claim 1, wherein, The high-frequency sampling window collects short-term attitude angle data at a higher frequency than the conventional sampling frequency after being triggered. The sampling interval and duration are configurable to adapt to the timing accuracy requirements of different recognition algorithms.

7. The system according to claim 1, wherein, When the confidence level of the identification reaches a preset threshold, the fusion judgment and early warning module can directly output the early warning result from a single identification channel; when multiple channels output simultaneously, a fusion weighted judgment mechanism is adopted to improve the identification accuracy.

8. The system according to claim 1, wherein, The fusion judgment and early warning module supports three modes: "parallel operation + independent triggering + collaborative enhancement". The system can automatically switch judgment strategies according to the real-time status of the identification channel.

9. The system according to claim 1, wherein, The ear tag terminal integrates only a three-axis attitude angle sensor. The main recognition path of the system does not rely on accelerometers, geomagnetic sensors, temperature sensors or image recognition devices, thus eliminating the main recognition capability of six-axis, nine-axis and other pseudo-fusion paths.

10. The system according to claim 1, wherein, This system can be adapted to the estrus recognition task of various ruminants such as dairy cows, buffalo, and goats through an algorithm template parameter switching mechanism, and has good species adaptability and scene transferability.

Images