Method for evaluating health status of free-range poultry

Abstract

The invention relates to the technical field of poultry breeding health monitoring, in particular to a free-range poultry health state assessment method, which comprises the following steps of: constructing a three-layer acquisition network consisting of fixed sensing nodes, mobile blind compensation nodes and individual wearable equipment to realize full-range data coverage; the vegetation shielding degree is specially judged through laser ranging and scene adaptation logic; after multi-modal data such as audio, infrared and motion are preprocessed, weight fusion is adjusted according to the shielding degree on the basis of a Transform dynamic attention mechanism; an individual time sequence baseline and a group clustering baseline are established, and the health state is calculated and judged through the two-dimensional deviation degree; rFID and infrared feature codes are used for double verification and binding of data, and an iteration model is learned online. According to the method, accurate assessment and early risk identification of the health state of the free-range poultry are realized, data traceability is guaranteed, and the method is adaptive to a complex free-range environment and has relatively high practicability and popularization value.

Description

Technical Field

[0001] This invention relates to the field of poultry farming health monitoring technology, specifically to a method for assessing the health status of free-range poultry. Background Technology

[0002] As consumers' demand for higher-quality poultry products increases, the scale of free-range poultry farming is expanding year by year. Due to their wide range of activity, complex environmental dynamics, and high individual mobility, assessing the health status of free-range poultry is far more difficult than in cage-rearing. Health problems are often only discovered when obvious symptoms appear, leading to higher losses in poultry farming. Therefore, achieving real-time and accurate assessment of the health status of free-range poultry is crucial for reducing farming risks.

[0003] Existing technologies have attempted to monitor poultry health through various means. For example, Chinese invention patent application CN201711257915A discloses a method for monitoring animal health using drones. This method uses drones to collect images of the group, identify abnormal animals, and then activates internal electronic tags to obtain physiological information. While this method utilizes drones for large-scale monitoring, it has significant limitations in free-range scenarios: First, drones rely on battery power, making 24-hour continuous monitoring impossible and creating data blind spots; second, the activation distance of electronic tags is limited, and vegetation obstruction and animal movement in free-range environments lead to low activation success rates, and tag implantation is costly and easily detached by poultry pecking; third, it only collects physiological data from individuals exhibiting abnormal behavior, without establishing a health correlation model between individuals and the group, making it impossible to identify the latent risk of "individuals showing no obvious abnormalities but the group's baseline shift" at an early stage.

[0004] Another type of technology, such as the invention patent application with publication number CN202111203859.9, proposes a poultry health monitoring system that collects images and weight data through ground cameras and load-bearing structures. However, fixed cameras are limited by the shooting angle and cannot cover the vegetation-covered areas of free-range areas. Moreover, the weight data can only reflect the long-term growth status and cannot capture acute health abnormalities in real time.

[0005] In addition, the invention patent with publication number CN202210190861.5 assesses health by collecting exercise data through ankle bracelets, but it only relies on a single exercise indicator and does not integrate environmental and group data, resulting in poor anti-interference capabilities.

[0006] In summary, the existing technology has the following core defects: 1. Incomplete data collection coverage, with blind spots in fixed nodes and inefficient scheduling of mobile nodes; 2. Poor adaptability of multi-source data fusion, unable to dynamically adjust feature weights in complex environments; 3. Lack of health association assessment for individuals and groups, resulting in delayed identification of hidden risks; 4. Low accuracy of identity and data binding, easily affected by occlusion and collision.

[0007] In summary, existing technologies are difficult to adapt to the characteristics of free-range poultry farming scenarios, and there is an urgent need for a health status assessment method that can achieve full-range data collection, highly interference-resistant fusion, and accurate correlation assessment. Summary of the Invention

[0008] To solve the above problems, the present invention adopts the following technical solution.

[0009] A method for assessing the health status of free-range poultry includes the following steps:

[0010] S1. Construct a three-layer collaborative data acquisition network:

[0011] S1.1. Deploy fixed sensing nodes: N fixed nodes are evenly distributed at high-frequency activity points in the free-range area, integrating infrared thermal imaging modules, microphone arrays, and environmental sensors. Data is uploaded through LoRaWAN networking. The environmental sensors include a temperature and humidity detection unit, a vegetation shading detection unit, and a noise detection unit. The temperature and humidity detection unit is used to collect temperature and humidity; the vegetation shading detection unit is used to determine the shading degree and assign a value of 0-1.

[0012] S1.2. Scheduling mobile blind spot nodes: Deploy several drones to cruise along grid routes. When a fixed node has no data update for two consecutive times, or the occlusion degree is greater than 0.7, or the noise is greater than the preset value and lasts for a preset time, dynamic supplementary data collection is performed. The drones are equipped with high-definition cameras, RFID readers and infrared thermal imagers.

[0013] S1.3. Configure individual wearable devices: Fit lightweight leg bands for poultry. The leg bands integrate a three-axis accelerometer, a temperature sensor, and a passive RFID chip. The leg bands are printed with infrared reflective feature codes.

[0014] S2. Multimodal data preprocessing and dynamic fusion:

[0015] Preprocessing: After audio is enhanced by microphone array beamforming, an adaptive high-pass filter is triggered by the output label of the noise detection unit to extract 20-dimensional MFCC features and align them using DTW; infrared data is processed by semantic segmentation to remove background; motion data is processed by second-order difference method to detect and correct anomalies.

[0016] Dynamic fusion: Based on the Transformer model, feature weights are adjusted according to vegetation occlusion.

[0017] When the occlusion degree is <0.3, visual perception accounts for 50%, physiological function for 30%, and environmental function for 20%.

[0018] When the occlusion level is 0.3-0.7, visual perception accounts for 30%, physiological perception for 50%, and environmental perception for 20%.

[0019] When the occlusion degree > 0.7, 10% for vision, 60% for physiology, and 30% for environment), output a 128-dimensional individual health feature vector;

[0020] S3. Individual-group two-dimensional health assessment:

[0021] Build a baseline: Collect 7-day health data to establish an individual time-series baseline, cluster by age into three groups: 0 - 30 days, 31 - 60 days, and over 61 days, and establish a group baseline;

[0022] Two-dimensional reasoning: Calculate the individual abnormality degree and the group deviation degree. When the individual abnormality degree ≤ 1.5×σ t and the group deviation degree ≤ σ_group, it is determined to be healthy. When the individual abnormality degree ≤ 1.5×σ t and the group deviation degree > σ_group and the proportion of group abnormal candidates ≤ 5%, it is determined to be sub-healthy. When the individual abnormality degree > 1.5×σ t and the group deviation degree ≤ σ_group, it is determined to be mildly ill. When the individual abnormality degree > 1.5×σ t and the group deviation degree > 2×σ_group and the proportion of group abnormal candidates > 5%, it is determined to be severely ill or at risk of infectious diseases;

[0023] S4. Identity-data binding and feedback: Bind data through dual verification of RFID and infrared feature codes, push results including health status and location, and iterate the model with veterinarian-confirmed data at fixed time intervals.

[0024] Preferably, for vegetation occlusion degree determination:

[0025] The vegetation occlusion degree detection unit collects distance samples through a laser rangefinder and quantifies the occlusion degree according to the following rules:

[0026] Define the core monitoring cone area: Starting from the laser rangefinder, the maximum effective monitoring distance for the feeding trough and water source nodes is 3m, and for the rest area node is 5m;

[0027] Multi-sampling point fusion: Collect 5 samples within 1 second, remove extreme values, and take the mean as the measured distance. The measured distance for the feeding trough or water source node is A, and for the rest area is B;

[0028] Distance segmented mapping: Assign 1.0 when B ≤ 1m or A ≤ 0.5m; linearly decay to 0.0 when 1m < B ≤ 5m or 0.5 < A ≤ 3m; assign 0.0 when B > 5m or A > 3m.

[0029] Preferably, in step S1.1, the adjacent spacing of fixed nodes ≤ 30m, and the signal coverage overlap rate ≥ 10%; the sampling frequency of the laser rangefinder is 10Hz to ensure the stability of multi-sample fusion..

[0030] Preferably, in step S1.2, the drone replenishment response time is ≤2 minutes; the drone is in hibernation from 18:00 to 06:00, and the fixed node covers the rest area.

[0031] Preferably, in step S1.3, the ankle ring is made of food-grade silicone and is fixed by an elastic buckle; the infrared reflective feature code is made of anodized aluminum.

[0032] Preferably, in step S2.1, the motion data outlier detection threshold th = μ_Δm + 3σ_Δm, and the outlier correction formula is m t =0.6×m t-1 +0.4×m 众数 .

[0033] Preferably, in step S3.1, the individual time-series baseline is updated daily through a 7-day sliding window; the population is re-clustered every 15 days.

[0034] As a preferred option, if the RFID identification code and the infrared feature code are inconsistent during the double verification in step S4, the drone is triggered to descend to within 3m to re-collect the data; the attention weights are retrained every quarter during model iteration, and the feature extraction accuracy is improved by ≥2% every quarter.

[0035] As a preferred option, mobile device vibration and SMS alerts are triggered when the risk level is high.

[0036] Compared with the prior art, the beneficial effects of this invention are as follows:

[0037] This invention constructs a three-layer collaborative data acquisition network consisting of fixed sensing nodes, mobile blind spot filling nodes, and individual wearable devices to achieve full-range, blind-spot-free data acquisition in free-range areas, ensuring the complete acquisition of multi-dimensional information related to poultry health.

[0038] This invention uses a scene-adaptive vegetation occlusion determination method, combined with laser ranging and monitoring area characteristics, to dynamically quantify the degree of occlusion, providing a precise environmental adaptation basis for subsequent data fusion.

[0039] This invention fuses multi-source data through a Transformer-based dynamic attention mechanism and adaptively adjusts feature weights according to environmental occlusion, effectively filtering noise interference in complex environments and improving the reliability of health feature extraction.

[0040] This invention establishes a two-dimensional assessment model that combines individual time-series baselines and group clustering baselines, linking individual physiological abnormalities with group characteristic shifts to achieve early identification and accurate determination of latent health risks.

[0041] This invention achieves precise binding of individual identity with multi-source health data through a dual identity verification mechanism of RFID and infrared feature codes, providing reliable support for full-cycle traceability of health status.

[0042] This invention introduces an online learning iteration strategy and dynamically updates model parameters and baselines based on veterinary diagnostic results, enabling the assessment method to continuously adapt to changes in free-range environments and poultry growth cycles, thus maintaining long-term stable assessment performance. Attached Figure Description

[0043] Figure 1 This is a schematic diagram of the overall process of the present invention;

[0044] Figure 2 This is a schematic diagram of the data acquisition process of the present invention. Detailed Implementation

[0045] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0046] In the description of this invention, it should be noted that the terms "upper," "lower," "inner," "outer," "top / bottom," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0047] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installed," "equipped with," "sleeved / connected," "connected," etc., should be interpreted broadly. For example, "connection" can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium; it can be a connection within two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0048] Example

[0049] like Figure 1-2 As shown, in this embodiment, 10000m 2 The scene of free-range chickens in the forest provides a detailed description of the specific implementation of the present invention.

[0050] S1. Construct a three-layer collaborative data acquisition network:

[0051] S1.1. Deploy fixed sensing nodes:

[0052] Ten nodes are deployed (3 feeding troughs, 2 water sources, and 5 rest areas), with a node spacing of 25-30m. The LoRaWAN gateway is deployed in the regional center, and the communication radius covers the entire area. Each node integrates an infrared thermal imaging module, a microphone array, and an environmental sensor. The infrared thermal imaging module is aimed at the direction of poultry activity, and the microphone array pickup angle is adjusted to 120°. The laser rangefinder of the environmental sensor is 0.3m above the ground and aligned with the viewing axis of the infrared thermal imaging module. The maximum effective monitoring distance for the feeding trough nodes is set to 3m, and for the rest area nodes, it is set to 5m.

[0053] In this embodiment, the infrared thermal imaging module is a FLIRLEPTON 3.5 with a resolution of 160×120, a temperature measurement range of 30-45℃, an accuracy of ±0.5℃, a sampling frequency of 1 frame / second, and a monitoring angle of 60°. The microphone array is a ReSpeaker4-MicArray with a sampling frequency of 22050Hz, a pickup radius of 0-8m, and supports beamforming noise reduction—using multiple microphones to collaboratively focus on the target sound and filter environmental noise. The environmental sensor is an integrated multi-parameter sensor, including a temperature and humidity detection unit, a vegetation obstruction detection unit, and a noise detection unit. The temperature and humidity detection unit uses an SHT30 chip with a temperature measurement range of -40-125℃ and an accuracy of ±0.3℃. The humidity measurement range is 0-100%RH, with an accuracy of ±2%RH. The vegetation obstruction detection unit uses a laser rangefinder, model VL53L1X, with a ranging range of 0.1-4m, an accuracy of ±2mm, and a sampling frequency of 10Hz. Its installation angle is consistent with the central axis of the monitoring angle of the infrared thermal imaging module, and its height matches the average activity height of poultry, 0.3m above the ground, corresponding to the torso height of free-range poultry such as chickens and ducks. The noise detection unit uses a MEMS microphone, with a detection range of 30-120dB and an accuracy of ±1dB. The overall sampling frequency of the environmental sensors is once every 5 minutes. During feeding periods (such as 7:00-9:00 and 16:00-18:00), high-frequency sampling is triggered, increasing the frequency to once per minute.

[0054] In this embodiment, the occlusion degree determination logic is as follows: The vegetation occlusion degree detection unit measures the straight-line distance from the sensor to the obstacle in front using a laser rangefinder, and combines this with the monitoring range of the infrared thermal imaging module to quantify the occlusion degree according to the following rules (value range 0-1, 0 represents no occlusion, 1 represents complete occlusion):

[0055] Monitoring area definition: With the laser rangefinder as the apex, the 60° monitoring angle of the infrared thermal imaging module as the cone angle, and 5m as the maximum effective monitoring distance (obstacles beyond this distance have no impact on monitoring), a "core monitoring cone area" is constructed. This area is the key range for poultry activity and data collection.

[0056] Distance segmentation mapping:

[0057] When the measurement distance is ≤1m: it is judged as "severe occlusion", and the occlusion degree is assigned a value of 1.0;

[0058] When 1m < measurement distance ≤ 3m: it is judged as "moderate occlusion". The occlusion degree is calculated according to the linear formula: occlusion degree = 1.0 - (measurement distance - 1m) / (3m - 1m) (for example, when the distance is 2m, the occlusion degree = 1.0 - (2 - 1) / 2 = 0.5);

[0059] When 3m < measurement distance ≤ 5m: it is judged as "slight occlusion". The occlusion degree is calculated according to the linear formula: occlusion degree = 0.5 - (measurement distance - 3m) / (5m - 3m) × 0.5 (for example, when the distance is 4m, the occlusion degree = 0.5 - (4 - 3) / 2 × 0.5 = 0.25);

[0060] When the measurement distance is greater than 5m or there are no obstacles: it is judged as "unobstructed", and the occlusion degree is assigned a value of 0.0;

[0061] Multi-sampling point fusion: During each acquisition, the laser rangefinder continuously acquires 5 distance samples within 1 second, removes the maximum and minimum values, and takes the average value as the final measured distance to avoid random errors in single sampling (such as instantaneous distance jumps caused by birds flying by).

[0062] Scene adaptive correction: Adjust judgment parameters according to the deployment scene of fixed nodes:

[0063] Feed troughs and water sources: The maximum effective monitoring distance is adjusted to 3m (in areas where poultry congregate, close-range obstruction has a greater impact), with corresponding distance segments of ≤0.5m (1.0), 0.5-2m (linear attenuation), and >2m (0.0);

[0064] Rest area node: retains the default maximum effective monitoring distance of 5m to adapt to the dispersed resting characteristics of poultry;

[0065] S1.2. Scheduling mobile blind spot replacement nodes:

[0066] Deploy 3-5 small rotary-wing drones, each equipped with:

[0067] High-definition camera (1920×1080 resolution, 30fps frame rate, supports autofocus);

[0068] Ultra-high frequency RFID reader (model: R2000, identification distance 0-3m, identification speed ≥50 sheets / second);

[0069] Miniature infrared thermal imager (Model: AxisQ1922-E, field of view 60°, temperature measurement accuracy ±1℃);

[0070] Scheduling logic:

[0071] Preset patrol routes: The routes are planned according to the "grid method", with an interval of 10m between adjacent routes, a patrol altitude of 5m, and a full area coverage is completed every 30 minutes;

[0072] Dynamic blind spot triggering: When a fixed node fails to detect poultry data in the preset activity area for two consecutive times (with a 10-minute interval), or when the vegetation occlusion detection unit of the environmental sensor detects an occlusion degree ≥0.7 for more than 10 minutes, or the noise detection unit detects noise >80dB for more than 5 minutes, the server sends a supplementary data collection instruction to the nearest drone through the communication module.

[0073] Supplementary sampling process: After the drone flies to the blind area, it first scans the identification code with an RFID reader, then collects infrared images and audio data. After the supplementary sampling is completed, it returns to the cruise route. The supplementary sampling response time is ≤2 minutes.

[0074] S1.3. Configure individual wearable devices

[0075] Each poultry is fitted with a leg band made of food-grade silicone, weighing ≤4.5g, with an adjustable diameter of 2-3cm, and featuring an anti-slip texture on the surface. The leg band is secured with an elastic buckle.

[0076] The ankle bracelet integrates a three-axis accelerometer, a patch-type temperature sensor, and a passive RFID chip. The passive RFID chip is model FM1108, which stores a 16-bit unique identification code, has a read / write distance of 0-2m, is resistant to metal interference, does not require battery power, and is activated by the reader's radio frequency signal.

[0077] The ankle bracelet has an infrared reflective feature code printed on its upper surface. It is made of anodized aluminum and measures 5×5mm. It consists of a 3×3 grid of black and white blocks. Each identification code corresponds to a unique feature code. The infrared reflectivity difference is ≥30%, ensuring that the infrared camera can clearly identify it.

[0078] S2. Multimodal data preprocessing and dynamic fusion:

[0079] S2.1 Data Preprocessing

[0080] Audio data processing:

[0081] An adaptive high-pass filter is used, and the cutoff frequency is adjusted according to the real-time data of the environmental sensor noise detection unit: when the detected noise is >60dB, the cutoff frequency is set to 250Hz; when the noise is ≤60dB, the cutoff frequency is set to 200Hz, so as to filter low-frequency interference such as vegetation friction and wind and rain.

[0082] Extracting 20-dimensional frequency cepstral coefficients (MFCCs), a sound feature extracted by simulating the human ear's perception of sound, can effectively distinguish different types of calls. Through dynamic time warping (DTW), an algorithm for comparing time series of different lengths, it can stretch or compress sequence segments to allow direct comparison of poultry call features of different durations, align call segments of different lengths, and retain the effective frequency band of 1-5kHz.

[0083] Infrared temperature data processing:

[0084] Using a semantic segmentation algorithm and the U-Net model, background areas such as vegetation and ground are removed from the image, leaving only the outline of the poultry.

[0085] Calculate the mean and standard deviation of temperature within the contour, and remove outliers at the contour edges.

[0086] Motion data processing:

[0087] The second-order difference method is used to detect outliers, and the formula is as follows:

[0088] Δm t =m t+1 -2m t +m t-1 Formula explanation: Δm t Let m be the second-order difference value at time t. t Let m be the activity intensity at time t. t+1 Let m be the activity intensity at time t+1. t-1 The activity intensity at time t-1; the second-order difference can amplify abrupt changes in data and quickly identify abnormal spikes caused by collisions and sensor errors;

[0089] Threshold calculation:

[0090] th=μ_Δm+3σ_Δm

[0091] (Formula explanation: th is the anomaly detection threshold, μ_Δm is the average of the historical second-order difference values, and σ_Δm is the standard deviation of the historical second-order difference values;)

[0092] Anomaly correction: If Δm t For values ​​≥th, a linear fusion correction is applied using the individual's historical mean and the group's contemporaneous mode, as shown in the following formula:

[0093] mt=0.6×m t-1 +0.4×m 众数

[0094] Formula explanation: m 众数The mode of activity intensity for the same group during the same period is used; individual historical data is weighted at 60% and group data at 40%, which preserves the pattern of individual activity while avoiding the influence of abnormal individual data.

[0095] 2.2 Dynamic Attention Mechanism Fusion Model

[0096] Model architecture: Based on the Transformer architecture, it includes a 12-layer encoder and a 4-layer decoder, with the following input feature dimensions:

[0097] Individual physiological characteristics: 20-dimensional MFCC, 1-dimensional body temperature, and 1-dimensional activity intensity;

[0098] Environmental characteristics: 2D temperature and humidity, 1D shading degree, 1D noise;

[0099] Group characteristics: 1-dimensional cluster density, 1-dimensional activity synchronicity;

[0100] Attention weight dynamic adjustment logic: When vegetation occlusion is <0.3, visual features account for 50% of the weight, physiological features account for 30%, and environmental features account for 20%; when vegetation occlusion is between 0.3 and 0.7, visual features account for 30%, physiological features account for 50%, and environmental features account for 20%; when vegetation occlusion is >0.7, visual features account for 10%, physiological features account for 60%, and environmental features account for 30%.

[0101] When the occlusion is high, visual data is prone to distortion, so the weight of physiological features is increased; when the occlusion is low, visual data is more reliable, so the weight of visual features is increased.

[0102] Feature output: After fusion, a 128-dimensional individual health feature vector is generated, and the scale is unified by L2 normalization.

[0103] S3. Individual-Group Two-Dimensional Health Assessment Model

[0104] S3.1 Baseline Model Construction

[0105] Individual time-series baseline construction:

[0106] Collect the fusion feature vector of each poultry in a healthy state for 7 consecutive days, and divide the time period into "every 2 hours" (a total of 12 time periods, such as 00:00-02:00, 02:00-04:00...22:00-24:00);

[0107] Calculate the characteristic mean μ for each time period t With standard deviation σ t Construct individual time-series baselines:

[0108] B_individual=[μ1,σ1;μ2,σ2;…;μ12 , σ 12 ]

[0109] μ t σ is the characteristic mean of the t-th time period, representing the normal characteristic level of that time period; t Let be the characteristic standard deviation of the t-th time period, representing the normal fluctuation range;

[0110] The baseline is automatically updated every day at midnight: a sliding window is used with a window size of 7 days. That is, each time it is updated, the old data from 7 days ago is replaced with the new data of the day to ensure that the baseline adapts to changes in the poultry growth cycle (such as changes in activity intensity due to age).

[0111] Population clustering baseline construction:

[0112] Poultry were clustered into three groups based on age: G1 (0-30 days, young birds), G2 (31-60 days, medium-sized birds), and G3 (over 61 days, adult birds);

[0113] The K-means clustering algorithm (an unsupervised learning algorithm that automatically groups similar individuals into one group, where K=3 represents 3 groups) is used to calculate the characteristic cluster centers μ_group and the within-group standard deviation σ_group for each group, thus constructing the group baseline:

[0114] B_group=[μ_G1, σ_G1; μ_G2, σ_G2; μ_G3, σ_G3]

[0115] μ_group represents the cluster centers of the group's features, signifying the average level of normal features in the group; σ_group represents the standard deviation of the features within the group, signifying the range of normal differences within the group.

[0116] Every 15 days, newly grown poultry individuals are re-clustered to include them.

[0117] 3.2 Two-Dimensional Evaluation Reasoning Process

[0118] Individual abnormality calculation:

[0119] The Euclidean distance (which measures the straight-line distance between two feature vectors; the larger the distance, the greater the difference) is used for calculation, as shown in the following formula:

[0120] D_individual=||X-μ t ||2

[0121] Formula explanation: D_individual is the individual anomaly score, X is the feature vector of the current individual, and μ t This represents the individual baseline mean for the corresponding time period;

[0122] Judgment rule: If D_individual > 1.5 × σ t They are marked as individual abnormality candidates.

[0123] Group deviation calculation:

[0124] The KL divergence is used for calculation, and the formula is as follows:

[0125] D_group = KL(X||μ_group)

[0126] Formula explanation: D_group is the group deviation, X is the feature vector of the current individual, and μ_group is the cluster center of the group to which it belongs; KL divergence can reflect the degree of deviation between the individual feature distribution and the normal distribution of the group;

[0127] Judgment rules: If D_group > 2 × σ_group, and the proportion of individual anomalous candidates within the group is > 5%, it is judged as a group-related anomalous; if D_group ≤ σ_group, it is judged as an individual isolated anomalous.

[0128] Health status output:

[0129] When the individual abnormality is ≤1.5×σ t Furthermore, when the group deviation is ≤ σ_group, it is judged as "healthy" with a low risk level; when the individual abnormality is ≤ 1.5 × σt but the group deviation is > σ_group and the proportion of abnormal candidates in the group is ≤ 5%, it is judged as "sub-healthy" with a medium risk level; when the individual abnormality is > 1.5 × σt, it is judged as "sub-healthy" with a medium risk level. t However, when the group deviation is ≤ σ_group, it is judged as "mild illness" with a medium risk level; when the individual abnormality is > 1.5 × σ t When the group deviation is greater than 2×σ_group and the proportion of abnormal candidates in the group is greater than 5%, it is judged as "severe risk of disease or infectious disease", and the risk level is high.

[0130] S4. Identity-Data Dual Binding and Result Feedback

[0131] S4.1 Dual Identity Verification and Binding

[0132] Close-range verification: The RFID reader at a fixed node or on a drone reads the identification code ID1 from the ankle bracelet chip, and at the same time identifies the infrared reflection feature code through the infrared thermal imaging module to obtain the identification code ID2.

[0133] Long-distance verification: The drone's high-definition camera captures an image of the anklet, and the infrared feature code is identified through the OCR algorithm to obtain ID2, while simultaneously triggering the RFID reader to read ID1;

[0134] Binding determination:

[0135] If ID1 = ID2, bind the currently collected multi-source data with the identity code and store it in the cloud database;

[0136] If ID1 ≠ ID2 or either is missing, the drone will descend to within 3 meters to resample. If the resample fails after 3 attempts, it will be marked as pending manual verification.

[0137] S4.2 Results Feedback and Model Iteration

[0138] Real-time feedback:

[0139] The server pushes the evaluation results to the mobile device through the NB-IoT module, including: individual identification code, real-time location, health status, abnormal characteristics and environmental parameters;

[0140] When the risk level is high, the triggering of audible and visual alarms includes mobile device vibration and SMS notification, with a response time of ≤1 minute.

[0141] Model iteration:

[0142] Data accumulation: Collect veterinary diagnostic results monthly and establish a correlation dataset of assessment results, diagnostic results, and environmental parameters;

[0143] Baseline updates: Individual time-series baselines are updated daily at midnight, and population clustering baselines are updated every 15 days;

[0144] Fusion model optimization: The attention weights of the Transformer model are retrained with a new dataset every quarter, which improves the feature extraction accuracy in complex environments by ≥2% per quarter.

[0145] The above are merely preferred embodiments of the present invention; however, the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and its improved concept, should be covered within the scope of protection of the present invention.

Claims

1. A method for assessing the health status of free-range poultry, characterized in that, It includes the following steps: S1. Construct a three-layer collaborative data acquisition network: S1.

1. Deploy fixed sensing nodes: Evenly distribute N fixed nodes at high-frequency activity points in the free-range area, integrate an infrared thermal imaging module, a microphone array, and environmental sensors, and upload data through LoRaWAN networking. The environmental sensors include a temperature and humidity detection unit, a vegetation occlusion detection unit, and a noise detection unit. The temperature and humidity detection is used to collect temperature and humidity; the vegetation occlusion detection unit is used to determine the occlusion degree and assign a value of 01. S1.

2. Dispatch mobile blind spot compensation nodes: Deploy several drones to cruise along the grid route. When the fixed nodes have no data update for 2 consecutive times, or the occlusion degree > 0.7, or the noise is greater than the preset value and lasts for the preset time, dynamic supplementary collection is carried out. The drones are equipped with high-definition cameras, RFID readers, and infrared thermal imagers. S1.

3. Configure individual wearable devices: Wear lightweight ankle rings on poultry. The ankle rings integrate triaxial acceleration sensors, temperature sensors, and passive RFID chips, and the ankle rings are printed with infrared reflection feature codes. S2. Multimodal data preprocessing and dynamic fusion: Preprocessing: After the audio is enhanced by microphone array beamforming, the noise detection unit outputs a label to trigger adaptive high-pass filtering, extracts 20-dimensional MFCC features and aligns them through DTW; the infrared data removes the background through semantic segmentation; the motion data detects and corrects anomalies using the second-order difference method. Dynamic fusion: Based on the Transformer model, adjust the feature weights according to the vegetation occlusion degree. When the occlusion degree < 0.3, vision 50%, physiology 30%, environment 20%; When the occlusion degree is 0.3 - 0.7, vision 30%, physiology 50%, environment 20%; When the occlusion degree > 0.7, vision 10%, physiology 60%, environment 30%), and output a 128-dimensional individual health feature vector. S3. Individual-group two-dimensional health assessment: Construct a baseline: Collect 7-day health data to establish an individual time series baseline, cluster by age into three groups: 0 - 30 days, 31 - 60 days, and over 61 days, and establish a group baseline. Two-dimensional inference: Calculate individual anomaly and group deviation, when individual anomaly ≤ 1.5 × σ t Furthermore, individuals are considered healthy when the group deviation is ≤ σ_group, and when the individual abnormality is ≤ 1.5 × σ. t Furthermore, a group is considered sub-healthy when the group deviation is greater than σ_group and the proportion of abnormal candidates in the group is ≤5%. When the individual abnormality is greater than 1.5×σ, the sub-health status is determined. t Furthermore, a mild disease is defined as a group deviation ≤ σ_group, and an individual abnormality > 1.5 × σ_group is defined as a disease. t Furthermore, when the group deviation is greater than 2×σ__group and the proportion of abnormal candidates in the group is greater than 5%, it is judged as a risk of severe illness or infectious disease. S4. Identity-data binding and feedback: Bind data through double verification of RFID and infrared feature codes, push results including health status and location, and iterate the model with veterinarian-diagnosed data at fixed time intervals.

2. The method for assessing the health status of free-range poultry according to claim 1, characterized in that, Determination of vegetation occlusion degree: The vegetation occlusion detection unit collects distance samples through a laser rangefinder and quantifies the occlusion degree according to the following rules: Define the core monitoring cone area: Starting from the laser rangefinder, the maximum effective monitoring distance for the feeding trough and water source nodes is 3m, and for the rest area node is 5m. Multi-sampling point fusion: Collect 5 samples within 1 second, remove extreme values, and take the mean as the measured distance. The measured distance for the feeding trough or water source node is A, and for the rest area is B. Distance segmented mapping: Assign a value of 1.0 when B ≤ 1m or A ≤ 0.5m; linearly attenuate to 0.0 when 1m < B ≤ 5m or 0.5 < A ≤ 3m; assign a value of 0.0 when B > 5m or A > 3m.

3. The method for assessing the health status of free-range poultry according to claim 1, characterized in that, In step S1.1, the adjacent spacing of the fixed nodes ≤ 30m, and the signal coverage overlap rate ≥ 10%; the sampling frequency of the laser rangefinder is 10Hz to ensure the stability of multi-sample fusion..

4. The method for assessing the health status of free-range poultry according to claim 1, characterized in that, In step S1.2, the drone's response time for supplementary data collection is ≤2 minutes; from 18:00 to 06:00, the drone goes into hibernation, and fixed nodes cover the rest area.

5. A method for assessing the health status of free-range poultry according to claim 1, characterized in that, In step S1.3, the ankle ring is made of food-grade silicone and is fixed by an elastic buckle; the infrared reflective feature code is made of anodized aluminum.

6. A method for assessing the health status of free-range poultry according to claim 5, characterized in that, In step S2.1, the outlier detection threshold for motion data is th = μ_Δm + 3σ_Δm, and the outlier correction formula is m t =0.6×m t-1 +0.4×m 众数 .

7. The method for assessing the health status of free-range poultry according to claim 1, characterized in that, In step S3.1, the individual time-series baseline is updated daily through a 7-day sliding window; the population is re-clustered every 15 days.

8. The method for assessing the health status of free-range poultry according to claim 1, characterized in that, If the RFID identification code and the infrared feature code do not match during the double verification, the drone will be triggered to descend to within 3 meters to collect the data again. The model is retrained every quarter to improve the attention weights, resulting in a ≥2% improvement in feature extraction accuracy every quarter.

9. A method for assessing the health status of free-range poultry according to claim 1, characterized in that, When the risk level is high, it triggers mobile device vibration and SMS alerts.

Images