Intelligent method for measuring egg-laying performance of flat-water bird individual and automatic egg collecting equipment

Abstract

This application discloses an intelligent method for measuring the egg production performance of floor-raised waterfowl and an automatic egg collection device, belonging to the field of waterfowl farming technology. The method includes: acquiring EPC data of waterfowl entering and leaving nests, obtaining individual identification information of the waterfowl and actual nest entry and exit events, identifying valid egg-laying events, and binding the individual identification information of the waterfowl with the valid egg-laying events; performing visual inspection, tracking and counting, and egg quality grading on the hatching eggs of the waterfowl, and binding the egg quality grading results with the individual identification information of the waterfowl and the valid egg-laying events; and measuring the individual egg production performance of the waterfowl based on the bound individual identification information, valid egg-laying events, and egg quality grading results. This application improves the accuracy of identification, the stability of counting, and the efficiency of egg detection, while reducing waterfowl stress and the need for manual intervention.

Description

Technical Field

[0001] This application relates to the field of waterfowl farming technology, and in particular to an intelligent method for measuring the egg production performance of floor-raised waterfowl and an automatic egg collection device. Background Technology

[0002] Currently, waterfowl production mainly includes two modes: cage rearing and free-range rearing. Cage rearing facilitates feeding control, facility cleaning, disease prevention and control, and egg production recording, and has a high land utilization rate. However, it restricts the free movement, water play, and nesting behavior of waterfowl, easily leading to animal welfare problems such as leg diseases, feather damage, and egg-laying stress. For breeds such as geese and Muscovy ducks, which have high requirements for water bathing and nesting environments, welfare-oriented free-range rearing remains an important mode for hatching egg production and breeding.

[0003] Floor-rearing is more in line with the waterfowl's natural behaviors of preferring water, living in groups, and nesting, which helps reduce stress and improve conditions for hatching egg production. However, it also weakens the natural identity constraints of "one bird, one cage" in cage-rearing. Under floor-rearing conditions, multiple waterfowl can freely enter and exit and share nests, making it difficult to automatically record breeding indicators such as individual egg production and laying rhythm. After hatching eggs are laid, they are easily exposed to bedding, feces, and damp environments, making manual egg collection labor-intensive, causing significant disturbance to the flock, and increasing the risk of dirty eggs and breakage. At the same time, factors such as waterfowl pecking at sensors, obstructed egg rolling, mechanical vibration, and foreign objects can generate a large number of interference signals, making it difficult to identify laying events, assign individual identities, and trace egg quality.

[0004] To improve the current status of measuring individual egg production performance in waterfowl, researchers have designed egg production detection methods and developed corresponding devices, but some problems still exist. For caged or small-group isolation scenarios, existing technology (CN118657824A) discloses an automatic egg production performance measurement device and system for caged geese in small family groups. This system identifies egg-laying individuals using RFID antennas and combines egg-guiding ramps with photoelectric sensors to detect egg-laying events. However, this scheme still relies on cage boundaries for identification, making it difficult to apply to floor-raised environments where multiple waterfowl freely enter and exit and share nests. It also lacks mechanisms for distinguishing multiple candidate individuals, suppressing RFID cross-reading in dense nests, identifying photoelectric impulse noise, and reverse tracing egg quality. For floor-raised breeding goose farms… The prior art (CN118278580B) discloses a method and system for measuring the egg production performance of multi-dimensional floor-raised geese. It collects data such as identity, weight, movement and vision and performs multi-source probability fusion based on timestamps. However, this scheme is still based on a synchronous time window, which makes it difficult to solve the problem of decoupling the egg-laying time from the detection trigger time caused by damp bedding, sticky feces or obstructed egg rolling. Furthermore, it does not establish physical identification rules for valid eggs, pecking noise, mechanical vibration and foreign object blockage from the photoelectric pulse of the egg guide channel itself, nor does it form a closed loop binding between egg quality results and individual identity.

[0005] In summary, while existing technologies have explored aspects such as RFID identification in caged poultry, photoelectric trigger detection, and multi-source fusion judgment in free-range poultry, they still struggle to systematically solve problems in welfare-oriented free-range waterfowl scenarios, including misassignment of multiple birds sharing nests, cross-reading of radio frequency signals in dense nests, time decoupling caused by egg rolling delays, false triggering due to photoelectric noise, lack of correlation between egg quality and individual egg-laying, and stable transmission of heterogeneous signals under large-scale deployment. Therefore, there is an urgent need for an intelligent measurement and automatic egg collection solution for individual egg-laying performance in free-range waterfowl, which, while preserving the welfare advantages of free-range farming, achieves accurate individual egg-laying identification, automatic collection of hatching eggs, reliable capture of egg-laying events, traceability of egg quality, and large-scale engineering applications. Summary of the Invention

[0006] To address the aforementioned shortcomings in existing technologies, this application provides an intelligent method for measuring the egg production performance of individual free-range waterfowl and an automatic egg collection device. This method solves a series of technical problems in free-range waterfowl scenarios, such as the difficulty in accurately identifying the egg-laying individuals in multi-nest cohabitation, RFID cross-reading in dense nests, decoupling of egg-laying time and detection trigger time due to egg rolling delay, false triggering of egg-laying events caused by photoelectric noise (waterfowl pecking, mechanical vibration, etc.), lack of online binding and reverse traceability between egg quality and individual egg-laying individuals, and difficulties in stable transmission of heterogeneous signals in large-scale deployment.

[0007] To achieve the aforementioned objectives, the technical solution adopted in this application is as follows: First aspect: This application provides an intelligent method for determining the egg production performance of floor-raised waterfowl, including: S1: Obtain EPC data of waterfowl entering and leaving nests, obtain the individual identity information of waterfowl and the actual access events of waterfowl entering and leaving nests, determine the valid egg-laying events, and bind the individual identity information of waterfowl with the valid egg-laying events; S2: Visually inspect, track and count, and grade the quality of hatching eggs from waterfowl, and link the egg quality grading results with the individual identification information of the waterfowl and valid egg-laying events; S3: Based on the bound individual identity information, valid egg-laying events, and egg quality grading results, determine the egg-laying performance of individual waterfowl.

[0008] Further, S1 includes: S101: Obtain EPC data on waterfowl entering and leaving nests, obtain individual identification information and time series data of waterfowl, and identify actual access events of waterfowl entering and leaving nests based on EPC data; S102: Calculate the pulse width of a single occlusion event of hatching eggs, divide the single occlusion event of hatching eggs into electronic jitter events, pecking noise events, valid egg candidate events and foreign object blockage events, and update the upper and lower limits of the valid egg candidate interval when the event is a valid egg candidate event. S103: Initially bind valid egg candidate events with the individual identity information of waterfowl.

[0009] Further, S101 includes: A1: The RFID reader continuously reads the EPC data within the range of each egg-shaped antenna according to a preset polling cycle, and performs anti-crosstalk processing by combining power constraints and physical shielding with metal signal partitions. EPC data on non-target egg-shaped antennas is marked as crosstalk candidates. The expression for the EPC data is:

[0010] In the formula, For the first Candidate waterfowl leg band tag codes, For the time of reading, To count the number of reads within the window, Number the antenna. For receiving signal strength indication; Among them, the power constraint is based on the stable reading distance within the target nest and the cross-read suppression distance between adjacent nests; if the number of times the antenna of a non-target nest reads the same EPC is less than the cross-read reading number threshold, and the average RSSI reading of the antenna of a non-target nest is less than the cross-read RSSI threshold, then it is marked as a cross-read candidate. A2: For the same Adjacent read records are time-merged. When the interval between two adjacent reads is less than the continuity threshold, they are merged into the same continuous read sequence. The dwell time is calculated by marking the dwell time with the first and last read times. The calculation formula is as follows:

[0011] in, For the first The EPC in the first The dwell time of a continuous reading sequence. This refers to the first read time in a continuous card reading sequence. This is the last read time in the continuous card reading sequence; A3: Based on dwell time, determine the actual entry and exit events of waterfowl from the nest, and obtain candidate EPCs. The judgment criteria are:

[0012] In the formula, The minimum dwell time threshold, For the first The cumulative number of reads within a continuous card reading sequence. The minimum number of reads threshold, For the first The arithmetic mean of all RSSI values ​​within a continuous reading sequence. The minimum average RSSI threshold.

[0013] The beneficial effects of the above scheme are as follows: Addressing the issues of RFID cross-reading between adjacent nests, short-term nest exploration, lingering and backtracking, and identity confusion caused by multiple waterfowl approaching simultaneously under dense nesting conditions, this application proposes a hardware-software co-operational nest-end identity recognition method. By continuously polling EPC data using a near-field RFID antenna, and combining read / write power constraints, physical shielding with metal partitions, continuous card reading sequence merging, and dwelling threshold determination, a comprehensive analysis is performed on the reading frequency, RSSI strength, dwelling duration, target antenna ratio, and entry-dwelling-departure sequence of the same tag. This method can automatically filter weak cross-reading signals at long distances and invalid nest exploration records, accurately outputting the true nesting status, nesting status, nesting status, and abnormal lingering status of waterfowl, solving the problem of stable individual identity collection in multi-bird nesting environments.

[0014] Further, S102 includes: B1: Calculate the pulse width of a single occlusion event of the hatching egg. The calculation formula is as follows:

[0015] In the formula, The duration for which the infrared light path is blocked from the hatching eggs. This represents the starting moment when the photoelectric signal switches from an unobstructed state to an obstructed state. This is the end time when the photoelectric signal returns from an obstructed state to an unobstructed state; B2: Based on the pulse width of a single occlusion event of a hatching egg, and according to the geometric dimensions of the waterfowl egg and the slope speed, the single occlusion event of a hatching egg is divided into electronic jitter events, pecking noise events, valid egg candidate events, and foreign object blockage events. The calculation formula is as follows:

[0016] In the formula, The classification results are for a single photoelectric pulse width event. This is the upper limit threshold for electronic jitter. This is the lower threshold for valid candidate egg events. The upper threshold for valid candidate egg events; B3: Statistically analyze the pulse width set of valid egg candidate events, calculate the mean and standard deviation, and update the upper and lower limits of the interval for valid egg candidate events to obtain valid egg-laying events. The calculation formula is as follows:

[0017]

[0018] In the formula, and These are the lower and upper thresholds for the updated valid candidate egg events, respectively. and These are the initial calibration values. For the effective candidate pulse width sample set, The average pulse width of the set. The standard deviation of the pulse width for this set. is the confidence coefficient.

[0019] The beneficial effects of the above scheme are as follows: Addressing the problem of false infrared triggering caused by waterfowl pecking sensors, mechanical vibration after egg rolling, foreign object obstruction, and bedding interference, this application analyzes the pulse width, trigger interval, duration, and repetition frequency of the photoelectric signal, classifying them into types such as short-pulse noise, effective egg-rolling pulses, long-term obstruction, and mechanical vibration. Furthermore, by combining biological refractory period constraints and an adaptive threshold update mechanism, genuine egg-laying events are confirmed, and non-egg-laying triggers are suppressed, thereby improving the purity and reliability of egg-laying event capture from the signal source.

[0020] Further, S103 includes: C1: Extracting symmetrical time windows Feature vectors of each candidate EPC And calculate the target antenna ratio, the calculation formula is:

[0021] In the formula, The target antenna percentage, This represents the number of times a candidate EPC is read by the target antenna within the current valid egg-laying event time window. This represents the total number of times a candidate EPC is read by the target nest and adjacent related antennas within the current time window. For the first One valid egg-laying event The infrared triggering time, and These are the candidate read time window lengths before and after the trigger time, respectively. For read frequency, For the average RSSI, For the length of stay, For time proximity, The most recent valid read time of the candidate EPC. The time proximity decay constant; C2: Based on the target antenna ratio, the DBSCAN clustering method is used to group candidate EPCs into three categories: target-stayed, interference-passed, and crosstalk-readout. For the candidate EPC regions in the target-stayed category, a weighted confidence score is calculated using the following formula:

[0022] in, In order to achieve effective egg production events Under the conditions of occurrence, the first The confidence level of each candidate EPC. For feature vectors The first in Each feature component The corresponding parameters are read frequency, average RSSI, dwell time, target antenna ratio, and time proximity. For the first The weighting coefficients of each characteristic component, This is the Min-Max normalization function; C3: When and At that time, the effective egg-laying event will be Initially bind with the corresponding candidate EPC; otherwise proceed to C4. In the formula, The highest attribution confidence among candidate EPCs. It has the second highest confidence level of attribution. The minimum binding confidence threshold. This is the minimum distinguishing threshold between the highest confidence level and the second highest confidence level; C4: Calculate the time series variance of the received signal strength indication for the continuous card reading sequence of candidate waterfowl, and extract the nesting segment of the time series variance below the attitude variance threshold as the candidate egg-laying period. The calculation formula is as follows:

[0023] In the formula, For the first The candidate EPC in the first The variance of the RSSI time series of consecutive card reading sequences. For the first The total number of records read in a continuous card reader sequence. For a moment The instantaneous value of RSSI, For the first The arithmetic mean of the RSSI of a continuous sequence of card readings; C5: In a symmetrical time window Based on the infrared triggering time of the egg-laying event, As the right endpoint, towards Back to the previous moment Duration, obtaining candidate backtracking windows The time decay factor is calculated for each candidate EPC, using the following formula:

[0024] In the formula, The time decay factor, The moment when the egg-laying event is triggered. EPC departure time The time difference between them The attenuation constant is The one-way backtracking duration for the candidate backtracking window; C6: Calculate the extended attribution confidence based on the time decay factor, and for fuzzy sets with insufficient extended confidence differences, complete the probabilistic data association and attribution between valid egg-laying events and individual identities according to the principle of priority matching based on the time of departure from the nest. The calculation formula is as follows:

[0025] In the formula, For the time decoupling condition, the first Each candidate EPC corresponds to a valid egg-laying event. Extended attribution confidence, To normalize the variance of the RSSI time series, To normalize low-variance nesting time, For biological hard constraints, , , These are variance weights, nesting time weights, and biological constraint weights, respectively.

[0026] The beneficial effects of the above scheme are as follows: Addressing the discrepancy between the actual egg-laying time and the infrared detection time caused by damp bedding, sticky feces, changes in nest floor friction, or obstructed egg rolling in waterfowl floor-raised nests, this application proposes a retrospective attribution algorithm for delayed egg-laying events. Using the infrared trigger time as a reference, it retrieves historical nesting records of candidate waterfowl and calculates the attribution probability by comprehensively considering dwell time, nest departure time difference, RSSI stability, and time decay weights. This method can accurately retrospectively attribute the actual egg-laying individual even when the laying individual has left the nest, the next waterfowl has entered, or the trigger signal appears late, thus improving the accuracy of egg-laying individual attribution in time-decoupled scenarios.

[0027] Further, S2 includes: S201: Use YOLOv8-nano lightweight target detector to detect hatching eggs at the end of the conveyor belt. If the cross-union ratio between detection frames is greater than the cross-union ratio deduplication threshold and the centroid distance is less than the centroid distance threshold, then merge the corresponding detection frames into the same egg. S202: The target tracking algorithm is used to establish the movement trajectory of the detected hatching eggs, and the camera field of view is divided into two coding regions. The region coding sequence of each hatching egg in different frames is tracked. When the coding sequence changes in accordance with the transport direction, the count is incremented by one. S203: For hatching eggs that have completed visual line crossing count, the region of interest image of the corresponding hatching egg is extracted according to the detection box coordinates, and the trajectory identifier, line crossing time, detection confidence and conveyor belt position code are recorded. The region of interest image is used as the input for egg quality grading. S204: Construct a lightweight egg quality grading neural network model, and use the lightweight egg quality grading neural network model to grade the quality of hatching eggs based on the region of interest image; S205: Based on the quality grading of hatching eggs, the infrared trigger count at the egg nest end and the visual line-crossing count at the end of the conveyor belt are time-aligned according to the conveyor delay. The difference between the infrared count and the visual count for the same batch is calculated using the following formula:

[0028] In the formula, For the first The difference between batch infrared counts and visual counts, For the first Batch at time Infrared triggering cumulative counting at the nest end of the egg, For the first Batch at time The visual line crossing cumulative count at the end of the conveyor belt. This is the time offset used for counting alignment between the infrared channel and the visual channel; S206: Calculate the anomaly intensity based on the count difference. When the anomaly intensity exceeds the preset alarm threshold, output an anomaly alarm. The calculation formula is as follows:

[0029] In the formula, For the first Batch infrared-visual counting anomaly intensity; S207: Align the visual tracking ID with the infrared event according to the transmission delay. When the matching judgment condition is met, bind the egg quality grading result with the individual identity information of the waterfowl and the valid egg-laying event. The matching judgment condition is:

[0030] In the formula, The moment when the visual trajectory crosses the counting line. For the first One valid egg-laying event The infrared triggering time, To compensate for the time delay in the transport from the egg nest to the final visual inspection area, This represents the maximum permissible matching time error between infrared events and visual trajectories.

[0031] Furthermore, the lightweight egg product grading neural network model includes: a backbone network, a direction-aware attention module, a multi-scale feature fusion module, and a classification head; The backbone network adopts the MobileNetV2 inverse residual structure. The orientation-aware attention module is placed after the depthwise separable convolution of the inverse residual unit of the backbone network and before the residual summation, and adopts the EfficientRCAM structure. The multi-scale feature fusion module adopts the CascadedRCAM three-layer architecture and performs feature fusion in the order of single module enhancement, residual fusion, and channel adaptation. The classification head consists of global average pooling, a fully connected layer, and a Softmax function; The training loss function of the lightweight egg grading neural network model adopts a general knowledge distillation form, which is a weighted combination of cross-entropy loss and KL divergence loss:

[0032] in, Let be the total training loss function. For cross-entropy loss based on true class labels, The KL divergence loss is used to compare the soft label distributions of the teacher and student models. For distillation weight, This refers to the distillation temperature.

[0033] The beneficial effects of the above scheme are as follows: Addressing the challenges of complex egg sources, continuous transport, significant eggshell reflectivity, and diverse defect types such as broken shells, dirt, deformities, soft shells, and blood spots during centralized egg collection in free-range waterfowl, an online egg quality detection method deployed at the end of the egg collection conveyor belt is proposed. The method involves acquiring images from the end-of-line transport, first completing egg target localization, deduplication of detection boxes, trajectory confirmation, and effective counting, and then performing quality grading identification on images of eggs confirmed to have passed through the counted area. The quality detection model employs a lightweight CA-MobileNetV2-MSFF network, introducing Coordinate Attention into the MobileNetV2 inverse residual structure, and fusing semantic information from crack edges, eggshell texture, and dirt areas using MSFF multi-scale features to achieve online discrimination of categories such as qualified eggs, broken eggs, dirty eggs, deformed eggs, soft-shelled eggs, and blood-spotted eggs. This method not only protects the specific lightweight model structure, but also covers the complete online detection process, including end-point image acquisition, target confirmation, quality identification, real-time inference, and writing back detection results to the egg-laying event, providing a basis for screening abnormal hens, nutritional adjustment, and disease early warning.

[0034] Further, S3 includes: S301: Based on valid egg-laying events and egg quality grading results, using individual leg band codes as the primary key, valid egg-laying events, nesting sequences, and egg quality grading are aggregated into an individual event stream within a preset evaluation window. Multiple behavioral indicators are quantified, and after normalization, an individual behavioral profile vector is generated. An adaptive baseline is established for the behavioral indicators, calculated using the following formula:

[0035]

[0036] In the formula, for Baseline mean The baseline standard deviation, For individuals No. Daily Indicators The observed values, Baseline smoothing coefficient; S302: Statistically analyze nest utilization rate, flock laying time distribution, batch abnormality rate, and egg quality grading distribution. Employ spatial distribution monitoring methods and quantitatively assess the balance of nest usage using the Gini coefficient of each nest's time utilization rate. The calculation formula is as follows:

[0037] In the formula, The total number of nests. The Gini coefficient represents the utilization rate of the nesting time. For the first The time utilization rate of each egg nest within the evaluation window. For the first Time utilization rate of each nest within the same evaluation window; S303: Compare each individual's daily behavioral indicators with the adaptive baseline, calculate the standardized deviation of each indicator, and only sum the deviations in the unfavorable direction to obtain the individual's comprehensive abnormality index. The calculation formula is as follows:

[0038] In the formula, For individual comprehensive abnormality index, As an indicator The weight, The reliability coefficient is the indicator. As an indicator Unfavorable deviation direction sign, It is a positive number; S304: The results of individual egg production performance measurement for waterfowl are obtained based on the individual comprehensive abnormality index and the Gini coefficient of nest time utilization rate.

[0039] Further, S304 includes: When at least one of the following conditions is met: the comprehensive abnormality index is greater than the abnormality index threshold, an individual has not laid eggs for several consecutive days, or the abnormality rate is greater than the abnormality rate warning threshold, an abnormal individual warning will be output. When at least one of the following conditions is met: the nest utilization rate is lower than the lower limit threshold, or the Gini coefficient is greater than the preset coefficient, nest optimization suggestions are output.

[0040] The second aspect: This application provides an automatic egg collection device for realizing an intelligent method for measuring the egg production performance of floor-raised waterfowl, including: an egg-laying nest unit, an RFID identification module, an infrared triggering module, a signal grading acquisition module, an egg collection and conveying module, an end-visual counting and grading module, a main control processing module, and a data management module. The egg-laying nest unit includes multiple egg-laying nests arranged linearly and at equal intervals along the longitudinal direction, metal signal partitions vertically arranged between adjacent egg-laying nests, side baffles at both ends of the egg-laying nest array, and a V-shaped egg-guiding plate at the rear outlet of each egg-laying nest. The RFID identification module includes an RFID antenna and an RFID reader at the bottom of the egg-laying nest. The RFID reader is connected to a signal aggregation router via an Ethernet interface. The signal aggregation router aggregates the EPC data uploaded by multiple RFID readers and uploads it to the main control processing module in real time using the UDP protocol. The infrared triggering module includes a reflective infrared photoelectric sensor located at the narrow opening of the V-shaped egg guide plate; the signal hierarchical acquisition module includes a signal aggregation router, an infrared sub-acquisition box, an infrared master acquisition box, and a PLC; the signal from the reflective infrared photoelectric sensor is input to an infrared sub-acquisition box, the infrared sub-acquisition boxes converge to an infrared master acquisition box, and the infrared master acquisition box is connected to the PLC via an RS-485 bus; the PLC cyclically scans the entire field of photoelectric channels, timestamps each valid trigger event, and then transmits the photoelectric trigger data to the main control processing module via TCP protocol; The egg collection and conveying module includes an egg collection conveyor belt and a conveyor belt motor. The egg collection conveyor belt is a flat belt made of food-grade nylon material, which is laid horizontally along the longitudinal direction of the egg-laying nest array and directly below the rear outlet of all egg-laying nests. The conveyor belt motor is installed at the conveying end of the egg collection conveyor belt on the side facing the end visual inspection area and is fixed to the end of the support structure. The output shaft of the conveyor belt motor is connected to the drive roller of the egg collection conveyor belt, driving the egg collection conveyor belt to unidirectionally and uniformly convey hatching eggs along the longitudinal direction of the egg-laying nest array. The end-of-line visual counting and grading module is arranged above the end of the egg collection conveyor belt and is used to complete visual inspection, tracking counting and egg quality grading. It includes an embedded vision module, a support structure, a fixed background plate and a supplementary lighting unit. The embedded vision module includes an industrial-grade embedded AI camera, which is fixedly installed on the top crossbeam of the support structure in a top view and uploads the end-of-line visual inspection results to the main control processing module. The fixed background plate is horizontally set directly below the embedded vision module and is close to the imaging area of ​​the egg collection conveyor belt. The supplementary lighting unit uses a strip LED diffused light source, which is symmetrically arranged on both sides of the top crossbeam of the support structure along the width direction of the egg collection conveyor belt and faces the imaging area of ​​the fixed background plate. The main control processing module is fixedly placed on the ground at the end of the egg collection conveyor belt of the egg nest array. The main control processing module integrates a PLC, power supply, communication and interface units in its cabinet. The data management module communicates with the main control processing module via a local area network to receive and store data.

[0041] The beneficial effects of this application are: This application provides an intelligent method for measuring the egg production performance of free-range waterfowl and an automatic egg collection device. Through the organic integration of three-level collaborative sensing of "identity recognition - event capture - egg quality detection" with RFID anti-cross-reading, photoelectric pulse width physical identification, multi-feature confidence binding, hysteresis probability attribution, end-point visual tracking counting and lightweight egg quality grading model, it realizes the full-process automation of accurate attribution of egg production of free-range waterfowl, reliable capture of egg production events, online grading of egg quality and reverse traceability. It significantly improves the accuracy of identity attribution, counting stability and egg quality detection efficiency, while reducing waterfowl stress and the need for manual intervention. Attached Figure Description

[0042] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other embodiments can be obtained based on these drawings.

[0043] Figure 1 This is a flowchart illustrating an intelligent method for determining the egg production performance of floor-raised waterfowl, as provided in an embodiment of this application.

[0044] Figure 2 This is a flowchart of an anti-tampering identity recognition and nest entry / exit determination method provided in an embodiment of this application.

[0045] Figure 3 This is a schematic diagram of a photoelectric pulse width four-zone classification and biological refractory period locking window provided in an embodiment of this application.

[0046] Figure 4 This is a flowchart illustrating an event-identity binding process for card reader feature clustering and confidence scoring, provided as an embodiment of this application.

[0047] Figure 5 This application provides a flowchart for the probabilistic data association and attribution of RSSI variance and time decay in an embodiment of the present application.

[0048] Figure 6 This is a structural diagram of a CA-MobileNetV2-MSFF egg quality classification model provided in an embodiment of this application.

[0049] Figure 7 This application provides a flowchart for a dual-channel closed-loop verification and anomaly diagnosis.

[0050] Figure 8 This application provides a flowchart for constructing an individual nest behavior profile, an adaptive baseline, and a production auxiliary decision-making process.

[0051] Figure 9This is a schematic diagram of the overall structure of an intelligent measurement and automatic egg collection device for the egg production performance of free-range waterfowl, provided in an embodiment of this application.

[0052] Figure 10 This is a schematic diagram of a nesting structure provided in an embodiment of this application.

[0053] The components include: 1. Egg-laying nest; 2-c. Metal signal partition; 2-b. Side baffle; 4. V-shaped egg-gathering guide plate; 3-a. RFID antenna; 3-b. RFID reader; 5. Reflective infrared photoelectric sensor; 6. Signal aggregation router; 7-a. Egg-collecting conveyor belt; 7-b. Conveyor belt motor; 8. Embedded vision module; 9. Main control processing module; 10-a. Infrared sub-acquisition box; 10-b. Infrared mother acquisition box; 11. PLC. Detailed Implementation

[0054] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art based on this application are within the scope of protection of this application.

[0055] Example 1: This application provides an intelligent method for determining the egg production performance of individually raised waterfowl. This method can be found in [reference needed]. Figure 1 ,include: S1: Continuously poll and collect EPC data throughout the entire process of waterfowl entering, staying, and leaving the nest to obtain individual identification information of waterfowl and actual access events of waterfowl entering and leaving the nest; when waterfowl lay hatching eggs and the hatching eggs trigger the infrared photoelectric sensor through the egg guide channel, complete the RFID identification and anti-cross-reading processing at the nest end, infrared photoelectric triggering event capture, photoelectric pulse width physical identification, binding of valid egg laying events with candidate individual identities, and probability data association and attribution under time decoupling conditions.

[0056] Further, S1 includes: S101: Obtain EPC data on waterfowl entering and leaving nests, obtain individual identification information and time series data of waterfowl, and identify actual access events of waterfowl entering and leaving nests based on EPC data.

[0057] In one embodiment of this application, such as Figure 2 As shown, the RFID reader operates according to a preset polling cycle. Continuously read EPC data within the range of each egg-shaped antenna, recording a quintuple each time. ;in, For RFID polling cycle, For the first Candidate waterfowl leg band tag codes, For the time of reading, To count the number of reads within the window, Number the antenna. This is to receive signal strength indication. The five-tuple is used to reconstruct the presence status of waterfowl in the nesting resource area, including entry time, departure time, frequency of visits, and duration of visits. The RFID reader operates in a preset frequency band. It supports the EPC Gen 2 anti-collision protocol dense mode to improve the identification reliability under the condition of multiple tag concurrent reading; RFID data is aggregated through Cat5 UTP cabling via GigabitEthernet switch and then uploaded to the main control processing module in real time via UDP protocol.

[0058] The main control processing module limits the RFID read / write power to The following limits the metal interference area. Adaptive adjustment, combined with a metal signal partition, forms a collaborative anti-crosstalk mechanism of "power constraint - physical shielding"; power calibration is based on the stable reading distance within the target egg nest. Distance between adjacent nests and read suppression To constrain the readings, the target nest tag is kept reading stably while weak readings from adjacent nests are suppressed. For readings recorded on antennas of non-target nests, if the reading frequency... And the average RSSI If it is, then it is marked as a candidate for serial reading; among them, This represents the number of times the non-target antenna reads the same EPC. This is the threshold for the number of serial reads. Read the mean RSSI value for the non-target antenna. This is the RSSI threshold for serial reading.

[0059] The main control processing module is for the same Adjacent read records are time-merged, when the interval between two adjacent reads is... The time is merged into the same continuous card reading sequence. ; at the first reading time With the last reading time The process of establishing a nest and the duration of stay are calibrated:

[0060] in, For the first The EPC in the first The dwell time of a continuous reading sequence. This refers to the first read time in a continuous card reading sequence. This refers to the last read time in the continuous card reading sequence. For the first The first EPC A continuous sequence of card readers, The time interval between two consecutive record reads. This is the continuity threshold.

[0061] Only if both conditions are met 、 、 Under three conditions, it is determined that the EPC has truly entered the current nest; among them, The minimum dwell time threshold, This represents the cumulative number of reads within this sequence segment. The minimum number of reads threshold, It is the arithmetic mean of all RSSI values ​​within this sequence segment. The minimum average RSSI threshold is used. This step employs the Bout-merging time threshold merging algorithm, using the first timestamp as the entry time and the last timestamp as the exit time, and uses double conversion logic to confirm the complete "entry-stay-leave" sequence. Records that do not form a complete access sequence, have too short a duration, or have insufficient signal strength are treated as nest exploration, passing through, or cross-reading interference and are not included in the egg-laying identity candidate set.

[0062] S102: Calculate the pulse width of a single occlusion event of hatching eggs, classify the single occlusion event of hatching eggs into four types of events: electronic jitter event, pecking noise event, valid egg candidate event and foreign object blockage event, and update the upper and lower limits of the valid egg candidate interval when the event is a valid egg candidate event.

[0063] In one embodiment of this application, when the hatching egg passes through the infrared photoelectric sensor at the narrow opening of the V-shaped egg-converging guide plate, the PLC collects the falling edge time. With rising edge time Calculate the pulse width of a single occlusion event. .in, The duration for which the infrared light path is blocked from the hatching eggs. This represents the starting moment when the photoelectric signal switches from an unobstructed state to an obstructed state. This is the end time when the photoelectric signal returns from an obstructed state to an unobstructed state. This step is based on the fundamental idea of ​​using infrared or laser counters to record egg passage events via photoelectric pulses, and is expressed with pulse width scaling in conjunction with the waterfowl egg size, egg guide ramp speed, and sensor installation position in the embodiments of this application.

[0064] Based on the geometric dimensions of waterfowl eggs and the slope speed, the main control processing module classifies single pulse events into four categories: electronic jitter, pecking noise, valid egg candidates, and foreign object obstruction. The photoelectric pulse width is further classified into four zones as follows: Figure 3As shown, the expression is:

[0065] in, Let be the equivalent geometric length of the waterfowl egg along the direction of passage. The equivalent rolling speed of the hatching egg as it passes the infrared detection position along the guide slope. The classification results are for a single photoelectric pulse width event. This is the upper limit threshold for electronic jitter. The lower limit threshold for effective egg candidates. This is the upper limit threshold for valid egg candidates. Only when... The system only generates an egg-laying event when the egg falls within the valid egg candidate range. And enter the identity binding process; when When the channel is suspected of being blocked by a foreign object, stuck egg, or obstructed by bedding, the system will mark it and output a corresponding nest maintenance alarm. Once a valid egg candidate is identified, the system will activate the biological refractory period locking window. Subsequent pulses on the same channel are discarded within the locking window to prevent mechanical jitter from triggering secondary triggers or repeated counting. The lockout duration for preventing repeated counting of the same photoelectric channel after a valid egg-laying event.

[0066] The system references the processing concept of statistically calibrating thresholds based on access duration or event duration, and performs processing on recent... Secondary effective egg candidate pulse width set Perform statistics and calculate the mean. with standard deviation And update the upper and lower limits of the valid candidate egg range online:

[0067]

[0068] in, and These are the updated lower and upper thresholds for valid egg candidates, respectively. and This is the initial calibration value, used to prevent the threshold from drifting beyond a physically reasonable range. For the effective candidate pulse width sample set, The number of sliding window samples used in the statistics. The average pulse width of this set. The standard deviation of the pulse width for this set. is the confidence coefficient. This adaptive update is used to compensate for the slow changes in the rolling pulse width caused by factors such as wear of the nest bedding, slight changes in the nest bottom angle, and slight drift in the conveyor speed, without changing the basic boundaries of the four-zone physical classification.

[0069] S103: Initially bind valid egg candidate events with the individual identity information of waterfowl.

[0070] In one embodiment of this application, such as Figure 4 As shown, when a valid egg-laying event is output... Subsequently, the main control processing module draws on the multimodal association concept of visual trajectory and RFID identity binding in group-housed animal scenarios to extract... Feature vectors of each candidate EPC within the time window ;in, For effective egg production events The infrared triggering time, and These are the candidate read time window lengths before and after the trigger time, respectively. For read frequency, For the average RSSI, For the length of stay, The target antenna percentage, For time proximity, The most recent valid read time of the candidate EPC. This represents the time proximity attenuation constant. Target antenna proportion. Calculate as follows:

[0071] in, This represents the number of times a candidate EPC is read by the target nest antenna within the current egg-laying event time window. This represents the total number of times the EPC is read by the target nest and adjacent related antennas within the current time window.

[0072] The system uses the DBSCAN clustering method to group candidate EPCs, with a neighborhood radius of [missing value]. The minimum sample size is The clustering results categorize candidate EPCs into three classes: target-residing, perturbed, and cross-read. The target-residing class proceeds to the subsequent confidence ranking, while the perturbed and cross-read classes are retained only as low-priority candidates or anomalous records. A weighted confidence score is calculated for each candidate EPC.

[0073] in, For the egg-laying event Under the conditions of occurrence, the first The confidence level of the attribution of each candidate EPC; For feature vectors The first in Each feature component The corresponding parameters are read frequency, average RSSI, dwell time, target antenna ratio, and time proximity. For the first The weighting coefficients of each feature component, the weight vector These correspond to the reading frequency, average RSSI, dwell time, target antenna ratio, and time proximity, respectively, and satisfy the following conditions: ; This is the Min-Max normalization function. When... and At that time, the effective egg-laying event will be Initial binding with the corresponding EPC; among which, The highest attribution confidence among candidate EPCs. It has the second highest confidence level of attribution. The minimum binding confidence threshold. The minimum threshold for distinguishing between the highest confidence level and the second highest confidence level; otherwise, it is marked as an event to be verified and corrected in conjunction with the end-visual closed-loop verification results or manual review results.

[0074] When the initial binding result output meets the confidence threshold and there are no risks of egg rolling delay, egg jamming, foreign object blockage, nest contamination, or multiple candidate conflicts, the system directly adopts the binding result; when , If, within the same nest, multiple candidate nest departure sequences, abnormal infrared pulse widths, egg jam alarms, or abnormal transport delays exist within the retrospective time window, the system enters the time decoupling probability attribution process, such as... Figure 5 As shown. The system borrows the soft association weighting concept of "multiple candidate targets - multiple observations" from probabilistic data association Kalman filtering, treating data association as a latent variable, and the association weights satisfy... The measurement likelihood satisfies .in, For the first The candidate target and the first The correlation weight between individual observations To match the quality matrix elements to the corresponding candidate-observation, To remove the first line (number) (the first candidate target) and the first Column (number) The submatrix following the observation results For matrix permanents, For state Lower observation The measurement likelihood, The likelihood decay coefficient is... This refers to the complement distance of the Intersection over Union (IoU) in visual tracking. This application does not directly replicate the spatial IoU association in visual tracking, but rather simplifies it under the temporal decoupling condition of RFID group rearing into a comprehensive attribution calculation based on RSSI variance, time decay, and biological constraints.

[0075] The system calculates the RSSI time series variance for each continuous card reading sequence:

[0076] in, For the first The first EPC The variance of the RSSI time series of consecutive card reading sequences. This represents the total number of records read from this sequence segment. For a moment The instantaneous value of RSSI, This represents the arithmetic mean of the RSSI values ​​for this sequence. RFID access event studies can provide basic variables such as entry, exit, access frequency, and access duration. This application further uses RSSI fluctuation as an engineering proxy for waterfowl posture stability: below the posture variance threshold... The variance of a certain value corresponds to a quiet lying down or laying state, while a high variance corresponds to a state of frequent movement, turning, or exploring the nest. Based on this, the low variance nesting segment is extracted from the continuous card reading sequence as a candidate laying period.

[0077] To address the decoupling of egg-laying timing from infrared triggering timing caused by delayed egg rolling, egg jamming, or nest contamination, a symmetrical time window is used. Based on the infrared trigger time of the egg-laying event, a candidate backtracking window is constructed by extending unidirectionally forward; the candidate backtracking window refers to the window that is based on the infrared trigger time of the egg-laying event. As the right endpoint, backtrack to a point before that time. One-way time interval formed by duration This is used to include candidate EPCs (Egg Collectors) that left the nest earlier than the trigger time but may still be genuine egg-laying individuals in the attribution assessment, thus covering the decoupling situation where egg roll-off lags behind individual nest departure time. The candidate backtracking window is extended to include individuals before the trigger time. The interval is defined, and the time decay factor is calculated for each candidate EPC:

[0078] in, The time decay factor, The moment when the egg-laying event is triggered. The EPC departure time The time difference between them The attenuation constant is The one-way backtracking duration of the candidate backtracking window, i.e., the candidate backtracking window. The length of the interval; The time decay factor for this candidate EPC is used to quantify the temporal proximity between its lodging time and the triggering time of the egg-laying event: where the natural exponent is used. Time difference Perform attenuation mapping when (When the candidate individual leaves the nest near the trigger time) The corresponding attribution weight is the largest. The larger The more exponentially it decays, the smaller the corresponding weight. The system can be adjusted based on the contamination level of the nest, the frequency of recent egg jamming alarms, or the worst-case egg rolling delay manually calibrated. The more severe the contamination or the more frequent the egg jamming, the greater the time tolerance for candidates who leave the nest earlier. The system can be adjusted based on the contamination level of the nest, the frequency of recent egg jamming alarms, or the worst-case egg rolling delay manually calibrated. The more severe the contamination or the more frequent the egg jamming, the greater the time tolerance for candidates who leave the nest earlier.

[0079] Calculate the overall confidence level of extended egg production attribution:

[0080] in, For the time decoupling condition, the first One candidate EPC for egg-laying events Extended attribution confidence, To normalize the RSSI variance, To normalize low-variance nesting time, and All were normalized to Min-Max or threshold truncation. interval, This is a biological hard constraint, where the time interval between a candidate waterfowl's last effective egg production and its current egg production is less than the physiological refractory period threshold. hour, Set to zero, otherwise set to one or assign a value according to the physiological feasibility score; , , These are the variance weights, nesting time weights, and biological constraint weights, respectively. The system only applies to... For fuzzy sets with insufficient discrepancies, probability assignment is performed; for other events, they are directly bound according to the maximum confidence level. When there are multiple stranded eggs or multiple triggers within the backtracking window, a FIFO state machine with priority matching based on the nest departure time is adopted and combined with... Joint pairing.

[0081] S2: After hatching eggs enter the egg collection conveyor belt through the egg guide channel, the system completes visual inspection, tracking and counting, egg quality grading, infrared trigger counting and visual counting closed-loop verification at the end of the conveyor belt, and reverse-binds the egg quality results to the egg laying event and the egg laying individual.

[0082] In one embodiment of this application, the embedded vision module uses a YOLOv8-nano lightweight target detector to detect hatching eggs at the end of the conveyor belt; if the IoU between detection frames is greater than 1, the detection module will detect hatching eggs at the end of the conveyor belt. And the distance between the centroids is less than If the corresponding detection boxes are not found, they will be merged into the same egg to suppress multiple boxes per egg. The detector uses a multi-scale feature fusion structure to output detection boxes, class confidence scores, and target center points, with an input size of [size missing]. The initial learning rate is Momentum is Batch size is The number of training rounds is The optimizer is ;in, To determine the threshold for deduplication of detection bounding boxes, To detect the threshold distance between the centroid and the frame, Input image size to the model. To detect the initial learning rate of the model, The momentum coefficient, To detect the training batch size of the model, To detect the number of training epochs of the model, To detect the model optimizer type.

[0083] The embedded vision module uses centroid tracking, Kalman state prediction, and Hungarian algorithm frame matching to establish trajectories. The matching distance is determined by combining the target centroid distance, IoU distance, and motion prediction residual; when short-term occlusion or missed detection occurs, the trajectory is determined within the maximum number of lost frames. The system remains in a pending matching state. The system sets a counter. With lateral hysteresis bands The hysteresis band width is The count is incremented only when the same trajectory completes the full process of "entering from side A - passing through L - leaving from side B"; rollbacks are not counted. For the first The trajectory of the hatching egg. The maximum number of consecutive frames that the trajectory is allowed to be lost. For virtual counting lines, The hysteresis band regions on both sides of the counting line. This represents the hysteresis band width.

[0084] The system adopts the spatial coding concept from egg visual counting research, encoding the camera's field of view (FOV) into zones 0 and 1, with the coding region size being [size missing]. Tracking the encoded sequence of each egg The count is incremented by one when the encoded sequence undergoes a 0→1 or 1→0 transition that conforms to the transmission direction; where, The spatial size of a single coded region. For the first The region-coded sequence of each hatching egg in consecutive frames. For this type of egg in frame 1 to frame 2 The region coding value corresponding to the frame, The number of consecutive frames involved in the judgment. Simultaneously, drawing on ByteTrack's low-confidence detection box secondary association strategy and OC-SORT's observation center technology, high-confidence detection boxes are prioritized for matching, while low-confidence detection boxes are used for trajectory completion, reducing the risk of missed tracking caused by brief pauses, local occlusion, or nonlinear motion of hatching eggs. For hatching egg trajectories that have completed effective line crossing counts or are located within stable detection areas... The system extracts the corresponding ROI image of the hatching egg based on the coordinates of the detection box. It also records the trajectory ID and the time of crossing the line. The confidence level and conveyor belt position encoding are detected, and the ROI image is input into the lightweight egg grading neural network model CA-MobileNetV2-MSFF. Figure 6 As shown, where, For the first Tracking trajectory of hatching eggs For from the first The trajectory corresponds to the region of interest image of the hatching egg captured by the detection box. The moment when the trajectory successfully crosses the line.

[0085] The system constructs a lightweight egg product grading neural network model CA-MobileNetV2-MSFF, with the backbone network being a MobileNetV2 inverse residual structure, including... The model employs the lightweight agricultural product defect classification network architecture, incorporating MobileNetV2, Coordinate Attention, EfficientRCAM, and CascadedRCAM. It introduces a direction-aware attention module, Coordinate Attention (CoordAtt), into the inverted residual backbone, using an EfficientRCAM structure. This EfficientRCAM architecture is named for its ability to aggregate features along both the horizontal and vertical directions while preserving spatial orientation information. This direction-aware attention module is not located as an independent branch outside the backbone network but is embedded within the inverted residual units of the backbone network. Specifically, the CoordAtt submodule is encapsulated within an EfficientRCAM unit and placed within a depthwise separable convolution of this inverted residual unit. Following DWConv and before residual summation, it, along with residual connections and H-Swish activation, constitutes EfficientRCAM. It is deployed only at the last Bottleneck layer of each stage, thus maintaining the lightweight backbone structure of MobileNetV2 while providing orientation-aware adaptive enhancement to the multi-scale defect features output by each stage. This balances defect localization capabilities and embedded inference efficiency. For the number of layers in the Bottleneck module, is the edge length of the kernel for depthwise separable convolution.

[0086] The CA-MobileNetV2-MSFF egg grading model uses direction-aware attention calculations. ;in, For the first The spatial location of each channel Increased output due to increased attention For the corresponding input feature values, It is the Sigmoid activation function. and These represent the directional attention responses in the horizontal and vertical directions, respectively. Global pooling is applied along both the horizontal and vertical directions. , ,in, and The first Each channel is at a height position and width position The results of directional pooling on the up direction, For the first Input feature maps of each channel, and These represent the height and width of the input feature map, respectively. The two sets of directional features are then concatenated and shared. After convolution, batch normalization, and nonlinear activation, the mixture is split according to spatial orientation, and horizontal and vertical attention weights are generated respectively to preserve spatial orientation information and adapt to the ellipsoidal geometric features of the egg.

[0087] The CascadedRCAM three-layer architecture is used to construct a multi-scale feature fusion MSFF module, which fuses features in the order of single-module enhancement, residual fusion, and channel adaptation. Layer and First Layer features:

[0088] in, The output features after cross-layer fusion These are low-level features with high resolution and rich edge and texture details. These are high-level features with strong semantic expression but low spatial resolution. and These are the lower-level feature layer numbers and the higher-level feature layer numbers, respectively. For pointwise convolution used for channel alignment, For upsampling operation, The kernel size is Depth-separable convolution, To fuse convolutional kernel sizes, the channel compression ratio is: This module applies the concept of cross-layer residual fusion and channel adaptation from CascadedRCAM to egg quality grading, enabling the model to utilize both fine-grained information such as crack edges and high-level semantic information such as dirty areas.

[0089] The classification head is constructed using GAP, FC, and Softmax, with a total of [number] classification categories. The classification categories include one or more of the following: qualified eggs, cracked eggs, dirty eggs, deformed eggs, soft-shelled eggs, and blood-spotted eggs; among them, This represents the total number of egg product quality classification categories. During the training phase, the pre-training distillation concept from lightweight visual models is adopted. The teacher model, ConvNeXt-Base, outputs logits and pre-computes them, storing them on disk. The student model loads these logits directly during training to reduce the overhead of repetitive forward computations. Distillation optimization employs a general knowledge distillation form using a weighted combination of cross-entropy loss and KL divergence loss.

[0090] in, For total training losses, For cross-entropy loss based on true class labels, The KL divergence loss is used to compare the soft label distributions of the teacher and student models. For distillation weight, The distillation temperature; the training strategy includes the initial learning rate. Cosine decay, number of preheating cycles AdamW optimizer, momentum= weight_decay= batch_size= Number of training rounds and input dimensions ;in, The initial learning rate for the classification model. Number of warm-up rounds for learning rate For the optimizer momentum parameter, This is the weight decay coefficient. For the training batch size of the classification model, The number of training rounds for the classification model. Input image dimensions to the classification model.

[0091] like Figure 7 As shown, according to the delivery delay Aligned infrared counting Visual counting Calculate the first Batch infrared-visual counting difference; where the infrared channel provides the starting point of the nest-end event, and the visual channel references the end-counting paradigm in poultry target detection-tracking-counting research, and the two together constitute a dual-channel verification:

[0092] in, For the first The difference between batch infrared counts and visual counts, For the first Batch at time Infrared triggering cumulative counting at the nest end of the egg, For the first Batch at time The visual line crossing cumulative count at the end of the conveyor belt. This refers to the transport delay required for hatching eggs to be conveyed from the infrared trigger position at the end of the nest to the visual detection area at the end via the egg collection conveyor belt. In other words, it is the time offset between the infrared channel and the visual channel used for counting alignment, and its value is determined by the operating speed of the egg collection conveyor belt. The transport distance from the infrared detection position at the egg nest end to the visual detection area at the end. according to Estimates can also be obtained through on-site calibration to compensate for actual retention errors of hatching eggs in the egg-guiding channel and conveyor belt; among which, The distance along the egg collection conveyor belt from the infrared detection position at the egg nest end to the visual detection area at the end.

[0093] Calculate the anomaly intensity based on the count difference:

[0094] in, For the first The batch's infrared-visual count anomaly intensity, with a constant of 1 in the denominator to avoid a division-by-zero error when both the infrared and visual counts are zero. Output abnormal alarms, among which This is the threshold for abnormal intensity alarms; If the problem is suspected to be a lost egg, stuck egg, transport delay, or visual miss, If the cause is suspected to be an imported egg, an infrared leak trigger, or a rollback recount, If the count is consistent, the system will determine that the faulty nest or faulty group is located by combining the RFID module number, infrared mother / daughter acquisition box number, nest number and conveyor section number.

[0095] Visual tracking IDs and infrared events Aligned according to delivery delay, when A successful pairing is determined at this time; among which, The moment when the visual trajectory crosses the counting line. For egg-laying events The infrared triggering time, To compensate for the transport delay from the egg nest to the final visual inspection area, This represents the maximum permissible time error for matching infrared events and visual trajectories. The alignment process borrows from the concept of associating trajectories with RFID identities based on a unified time axis in multimodal identity binding. The system uses... The identity binding result is used to obtain the EPC of the laying individual, and the egg category, counting trajectory and abnormal mark of the terminal visual output are written back to the event to establish a reverse traceability link of "egg quality - laying event - laying individual".

[0096] Alarm records are generated according to the priority of sensor anomalies, transportation anomalies, identification and attribution anomalies, and egg quality anomalies. When multiple anomalies are triggered by the same event, the anomaly type that affects the authenticity of the egg production event and the accuracy of identification and attribution is retained first, and the original RFID sequence, infrared pulse width, visual trajectory and classification results are saved simultaneously for manual review.

[0097] To address the issues of visual undercounting and double counting caused by reflection, rolling, brief pauses, backflips, or target occlusion of hatching eggs at the end of the conveyor belt, this application employs a top-down visual acquisition structure for target detection, trajectory tracking, and line-crossing counting of hatching eggs. The system features a unidirectional counting line and a hysteresis band, confirming a valid count only when the target completely crosses the counting area in a specified direction. Combined with a target trajectory continuity and short-term loss recovery mechanism, double counting is avoided, improving the stability of the end-of-line visual counting results.

[0098] S3: Write the obtained valid egg-laying events, individual identity attribution results, and obtained egg quality, count verification and abnormal alarm results into the floor-raised waterfowl individual egg-laying data management system to generate reports such as individual egg-laying records, group egg-laying performance, egg quality grading, abnormal individual warnings, equipment status and nest utilization rate, so as to realize intelligent measurement of individual egg-laying performance, automatic egg collection and production auxiliary decision-making.

[0099] In one embodiment of this application, such as Figure 8 As shown, based on the processing results of S1 and S2, the individual multidimensional behavioral profile construction method of this application is used to form an individual nest behavior profile: using the individual leg band code as the primary key, within an evaluation window of one day, the egg-laying events obtained from identity attribution in S1 and the effective nesting sequence, and the egg quality categories output in S2 are aggregated into an individual event stream. Multidimensional behavioral indicators such as daily egg-laying frequency, average egg-laying interval, nesting duration, nest entry and nest-exploration frequency, nest location preference, brooding tendency, egg quality qualification rate, and egg-laying rhythm are quantified and then normalized and concatenated into an individual behavioral profile vector. To eliminate natural differences between individuals and avoid misjudging individuals based on group mean, the system performs each indicator for each individual... Establish a dedicated adaptive statistical baseline and update the baseline mean online using an exponentially weighted moving average (EWMA). Compared with baseline standard deviation As the core benchmark for subsequent deviation scoring:

[0100]

[0101] in, For individuals No. Daily Indicators The observed values, The baseline smoothing coefficient is the initial baseline, which is statistically determined by observations over the previous few days. This adaptive baseline slowly drifts with the individual's physiological stages (start of laying, peak egg production, and rest period) to avoid misjudging normal changes as abnormal.

[0102] Construct a population-dimensional dataset and perform population behavior statistics and spatial distribution monitoring: Within the evaluation window, statistically analyze the utilization rate of each nest, the distribution of egg-laying time in the population, the batch anomaly rate, and the distribution of egg quality grading; and employ spatial distribution monitoring methods, focusing on the time utilization rate of each nest. Gini coefficient Quantitatively assessing the balance of hatch site utilization provides a data foundation for hatch site optimization, population health assessment, and batch hatching quality prediction.

[0103] in, The total number of nests. For the first The time utilization rate of a nest within the evaluation window is equal to the ratio of the cumulative time the nest was effectively occupied (nesting or laying eggs) to the total evaluation window time. For the first Time utilization rate of each nest within the same evaluation window and All are nest number indices, and the value range is [missing information]. ,molecular The summation of the absolute differences in time utilization rates for all pairwise combinations of nests characterizes the dispersion of utilization rates. The denominator... As a normalization factor, make The value falls within interval, The Gini coefficient represents the nest time utilization rate and is used to quantitatively characterize the degree of balance in nest usage. The closer to 0, the more evenly the nests are used. The larger the value, the more unbalanced the distribution, indicating that some nests are overcrowded while others are idle. Based on this, the system weights and integrates the group egg quality distribution with behavioral statistical indicators to output a predicted value for batch hatching quality.

[0104] Based on individual behavioral profiles, individual and group statistical baselines, egg quality grading, abnormal events, and equipment status, the system generates production support decision-making results. The system compares each individual's daily behavioral indicators with the obtained adaptive baseline, calculates the standardized deviation (Z-score) of each indicator, and only weights and sums the deviations in unfavorable directions (such as decreased egg production frequency, decreased pass rate, increased broodiness tendency, and a sudden increase in broodiness frequency) to obtain the individual's comprehensive abnormality index. As a core criterion for health early warning:

[0105] in, As an indicator weight and , The reliability coefficient is the indicator. As an indicator Unfavorable deviation direction sign, To prevent small positive numbers from being divided by zero, an explicit production rule is superimposed on this: when Greater than the abnormal index threshold, individual continuity No eggs laid in a day, abnormal rate greater than If any of the three conditions is met, an alert for abnormal individuals (suspected disease, cessation of egg production, brooding) will be issued; when the nest utilization rate is less than Gini coefficient If either of the two conditions is met, a nesting site optimization suggestion is output; and a comprehensive breeding selection score is calculated by combining individual egg production performance and abnormality index. Decision records are generated according to categories such as health warning, broodiness warning, nesting site optimization, and batch hatching quality prediction. Simultaneously, corresponding behavioral profiles, baselines, and original event sequences are saved for manual review. The threshold for the number of consecutive days without egg production. The abnormality rate warning threshold, This is the lower limit threshold for space utilization.

[0106] The system forms a data management platform for individual egg production of free-range waterfowl, providing functions such as individual egg production record reports, group egg production performance statistics reports, egg quality grading reports, abnormal individual early warning reports, equipment status and alarm reports, nest utilization analysis, and batch hatching quality prediction. It also supports multi-dimensional querying and export by individual, group, time period, and nest number.

[0107] Example 2: This application provides an automatic egg collection device to implement the intelligent method for determining the egg production performance of floor-raised waterfowl provided in Embodiment 1. (See also...) Figure 9 It includes: egg-laying nest unit, RFID identification module, infrared triggering module, signal grading acquisition module, egg collection and conveying module, end-point visual counting and grading module, main control processing module 9 and data management module.

[0108] like Figure 10 As shown, the nesting unit consists of multiple nests 1 arranged linearly and at equal intervals along the longitudinal direction, metal signal partitions 2-c vertically arranged between adjacent nests, side baffles 2-b at both ends of the nest array, and a V-shaped egg-guiding plate 4 at the rear exit of each nest, forming a row of nests. The bottom plate of each nest is at a preset angle to the horizontal plane. The slope is directed towards the rear of the nest, allowing the eggs to roll backward under gravity after being laid. A pair of V-shaped egg-guiding plates 4 are installed at the rear exit of the nest to gather and guide the eggs rolling down the entire width of the nest to a narrow central opening. The width of the narrow opening is... This ensures that hatching eggs fall in a single row through a narrow opening into the egg-collecting conveyor belt below, while also providing a fixed detection position for the reflective infrared photoelectric sensor 5; the front of the laying nest is an open entrance without a front baffle; the metal signal partition 2-c uses a material with a thickness of not less than The aluminum alloy plate extends from the bottom of the nest to the top of the nest and upwards. It also extends rearward to the edge of the exit at the rear of the nest to physically shield the coupling of RFID radio frequency signals between adjacent nests; the internal dimensions of each nest are [length missing]. ,Width ,high It can accommodate Only waterfowl nest at the same time; among them, The angle of the egg-laying nest floor. The narrow opening of the egg guide plate is narrowed by a V-shape. The thickness of the metal signal separator. The metal signal partition extends upwards beyond the top of the laying nest. , and These are the internal length, width, and height of the nesting chamber. This refers to the number of waterfowl that a single nest can accommodate at the same time.

[0109] The RFID identification module consists of an RFID antenna 3-a at the bottom of the egg-laying nest 1 and an RFID reader 3-b; the RFID antenna 3-a is a near-field UHF antenna, operating at a frequency of [frequency missing]. Installed below or embedded inside the base of the nesting box, the antenna's effective reading range is limited to the interior space of a single nesting box, with a reading distance not exceeding [a certain value]. The RFID reader 3-b supports multi-channel concurrent reading, each... Each nesting brood shares one reader, which is connected to the signal aggregation router 6 via an Ethernet interface; the EPC tags worn by the waterfowl are passive leg band tags, with each waterfowl corresponding to a unique EPC code; the read / write power of the RFID reader 3-b is uniformly limited to a preset threshold. The following, combined with the physical shielding of the metal signal partition 2-c, ensures that the EPC reading success rate within the target nesting nest is no less than The EPC cross-read rate of adjacent nests is not higher than ;in, The operating frequency of the RFID antenna. This represents the upper limit of the effective reading distance of the antenna. This refers to the number of nests connected to a single RFID reader. This refers to the upper limit of RFID read / write power. The lower limit of the success rate of EPC reading of the target laying nest. This represents the upper limit of EPC read rate between adjacent nests.

[0110] The infrared trigger module consists of a reflective infrared photoelectric sensor 5 located at the narrow opening of the V-shaped egg-gathering guide plate 4. The reflective infrared photoelectric sensor is installed on the exit side of the narrow opening of the V-shaped egg-gathering guide plate, with a sensing distance of [missing information]. When the hatching egg passes through the fixed narrow channel, it generates a complete "blocking-recovery" binary pulse signal. The signal hierarchical acquisition module consists of a signal aggregation router 6, an infrared sub-acquisition box 10-a, an infrared mother acquisition box 10-b, and a PLC 11; each Each RFID reader in a laying kennel connects to the nearest signal aggregation router 6 via Ethernet. The signal aggregation router aggregates data from multiple readers and uploads it in real time to the main control processing module 9 using the UDP protocol. The signal from the road-reflective infrared photoelectric sensor 5 is input to an infrared sub-acquisition box 10-a, each Each infrared sub-acquisition box 10-a converges to an infrared master acquisition box 10-b, and the infrared master acquisition box 10-b is connected to PLC 11 via an RS-485 bus; PLC 11 connects to PLC 11 at a frequency of not less than The scanning frequency is used to cyclically scan the entire field of photoelectric channels, perform hardware-level de-jitter filtering, and timestamp each valid trigger event. The photoelectric trigger data is then transmitted to the main control processing module 9 via TCP protocol. The sensing distance of a reflective infrared photoelectric sensor. This refers to the number of reflective infrared photoelectric sensor channels connected to a single infrared sub-acquisition box. This refers to the number of infrared sub-acquisition boxes aggregated by a single infrared master acquisition box. This is the lower limit of the scanning frequency of the PLC for the photoelectric channel.

[0111] The egg collection and conveying module consists of an egg collection conveyor belt 7-a and a conveyor belt motor 7-b. The egg collection conveyor belt 7-a is a flat belt made of food-grade nylon, with a width of [missing information]. They are laid horizontally along the longitudinal direction of the nesting array, directly below the rear exit of all nests, and with... The conveyor belt motor 7-b is installed at the end of the egg collection conveyor belt 7-a on the side facing the end visual detection area and is fixed to the end of the support structure. Its output shaft is connected to the active roller drive of the egg collection conveyor belt 7-a, driving the egg collection conveyor belt 7-a to transport hatching eggs unidirectionally and uniformly along the longitudinal direction of the egg-laying nest array.

[0112] The end-of-line visual counting and grading module consists of an embedded vision module 8, a support structure, a fixed background plate, and a supplementary lighting unit, and is arranged above the end of the egg collection conveyor belt 7-a. The embedded vision module 8 uses an industrial-grade embedded AI camera, which is fixedly installed on the top beam of the support in a top-down view, at a height of [height missing]. The field of view covers the entire width of the conveyor belt; the fixed background plate is horizontally positioned directly below the embedded vision module 8 and closely attached to the imaging area of ​​the egg collection conveyor belt 7-a, serving as a uniform background for egg imaging to enhance the contrast between the foreground and background; the supplementary lighting unit uses a strip-shaped LED diffused light source with a color temperature of [insert color temperature here]. The image acquisition frame rate is no less than Inference delay is no greater than They are symmetrically arranged on both sides of the top crossbeam of the support structure along the width of the egg-collecting conveyor belt, and are located around the camera lens of the embedded vision module 8 and facing the imaging area of ​​the fixed background plate. For the width of the egg-collecting conveyor belt, To control the speed of the egg-collecting conveyor belt Installation height of the vision module, To compensate for the color temperature of the light source, For image acquisition frame rate, This represents the upper limit of visual inference latency per frame.

[0113] The main control processing module 9 is an independent rack-mounted control unit, fixedly placed on the field end ground on one side of the egg collection conveyor belt 7-a of the egg-laying nest array. The rack integrates PLC11 and power supply, communication and interface units. The main control processing module 9 receives EPC data uploaded by each RFID reader 3-b via signal aggregation router 6 using UDP protocol, receives infrared photoelectric trigger data via infrared mother acquisition box 10-b and PLC11 using TCP protocol, and connects to the end-point visual detection results output by embedded vision module 8. It is responsible for real-time data processing and calculation such as RFID anti-cross-reading processing, egg-laying event capture, binding of egg-laying events and candidate individual identities, probability data association and attribution under time decoupling conditions, and infrared-visual dual-channel closed-loop verification.

[0114] The data management module is a software unit running on the host computer of the main control processing module 9. It communicates with the main control processing module 9 via the local area network, receives and persistently stores data such as egg production events, individual identification results, egg quality grading and abnormal alarms. It is responsible for generating reports such as individual egg production records, group egg production performance, egg quality grading, abnormal individual warnings, equipment status and nest utilization rate, and provides users with data management and application functions such as data storage, statistical analysis, multi-dimensional query and production auxiliary decision-making.

[0115] To address the challenges of large-scale floor-rearing scenarios involving numerous nests, high concurrency of RFID identification signals, and stringent stability requirements for infrared trigger signals, a hierarchical networked transmission architecture is proposed. RFID identification data is transmitted directly to the main control unit with low latency via UDP. Infrared photoelectric trigger data is aggregated by a PLC or edge acquisition unit and then uploaded uniformly. The main control unit then synchronizes, verifies, and stores data from different sources and at different time scales. This device innovatively constructs a three-level collaborative sensing link of "identity recognition—event capture—egg quality detection," integrating nest entry / exit identification, egg-laying event triggering, automatic egg delivery, egg quality detection, and individual identification traceability into one system. The overall structure eliminates the need for mechanically closing or forcibly restraining the waterfowl, enabling low-stress, continuous, and automated individual egg-laying performance measurement under floor-rearing conditions with free access to nests.

[0116] This application achieves full-process automation by organically integrating three-level collaborative sensing of "identity recognition - event capture - egg detection" with RFID anti-cross-reading, photoelectric pulse width physical identification, multi-feature confidence binding, hysteresis probability attribution, end-point visual tracking and counting, and a lightweight egg grading model. This enables accurate attribution of individual egg-laying status for free-range waterfowl, reliable capture of egg-laying events, online grading of egg quality, and reverse traceability. It significantly improves the accuracy of identity attribution, counting stability, and egg detection efficiency, while reducing waterfowl stress and the need for manual intervention.

[0117] It should be noted that those skilled in the art will recognize that the embodiments described herein are for the purpose of helping readers understand the principles of this application, and should be understood as not limiting the scope of protection of this application to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this application without departing from the essence of this application, and these modifications and combinations are still within the scope of protection of this application.

Claims

1. A method for intelligently determining the egg production performance of floor-raised waterfowl, characterized in that, include: S1: Obtain EPC data of waterfowl entering and leaving nests, obtain the individual identity information of waterfowl and the actual access events of waterfowl entering and leaving nests, determine the valid egg-laying events, and bind the individual identity information of waterfowl with the valid egg-laying events; S2: Visually inspect, track and count, and grade the quality of hatching eggs from waterfowl, and link the egg quality grading results with the individual identification information of the waterfowl and valid egg-laying events; S3: Based on the bound individual identity information, valid egg-laying events, and egg quality grading results, determine the egg-laying performance of individual waterfowl; S1 includes: S101: Obtain EPC data on waterfowl entering and leaving nests, obtain individual identification information and time series data of waterfowl, and identify actual access events of waterfowl entering and leaving nests based on EPC data; S102: Calculate the pulse width of a single occlusion event of hatching eggs, divide the single occlusion event of hatching eggs into electronic jitter events, pecking noise events, valid egg candidate events and foreign object blockage events, and update the upper and lower limits of the valid egg candidate interval when the event is a valid egg candidate event. S103: Initially bind valid egg candidate events with the individual identity information of waterfowl; S101 includes: A1: The RFID reader continuously reads the EPC data within the range of each egg-shaped antenna according to a preset polling cycle, and performs anti-crosstalk processing by combining power constraints and physical shielding with metal signal partitions. EPC data on non-target egg-shaped antennas is marked as crosstalk candidates. The expression for the EPC data is: In the formula, For the first Candidate waterfowl leg band tag codes, For the time of reading, To count the number of reads within the window, Number the antenna. For receiving signal strength indication; Among them, the power constraint is based on the stable reading distance within the target nest and the cross-read suppression distance between adjacent nests; if the number of times the antenna of a non-target nest reads the same EPC is less than the cross-read reading number threshold, and the average RSSI reading of the antenna of a non-target nest is less than the cross-read RSSI threshold, then it is marked as a cross-read candidate. A2: For the same Adjacent read records are time-merged. When the interval between two adjacent reads is less than the continuity threshold, they are merged into the same continuous read sequence. The dwell time is calculated by marking the dwell time with the first and last read times. The calculation formula is as follows: in, For the first The EPC in the first The dwell time of a continuous reading sequence. This refers to the first read time in a continuous card reading sequence. This is the last read time in the continuous card reading sequence; A3: Based on dwell time, determine the actual entry and exit events of waterfowl from the nest, and obtain candidate EPCs. The judgment criteria are: In the formula, The minimum dwell time threshold, For the first The cumulative number of reads within a continuous card reading sequence. The minimum number of reads threshold, For the first The arithmetic mean of all RSSI values ​​within a continuous reading sequence. The minimum average RSSI threshold; S102 includes: B1: Calculate the pulse width of a single occlusion event of the hatching egg. The calculation formula is as follows: In the formula, The duration for which the infrared light path is blocked from the hatching eggs. This represents the starting moment when the photoelectric signal switches from an unobstructed state to an obstructed state. This is the end time when the photoelectric signal returns to an unblocked state from an obstructed state; B2: Based on the pulse width of a single occlusion event of a hatching egg, and according to the geometric dimensions of the waterfowl egg and the slope speed, the single occlusion event of a hatching egg is divided into electronic jitter events, pecking noise events, valid egg candidate events, and foreign object blockage events. The calculation formula is as follows: In the formula, The classification results are for a single photoelectric pulse width event. This is the upper limit threshold for electronic jitter. This is the lower threshold for valid candidate egg events. The upper threshold for valid candidate egg events; B3: Statistically analyze the pulse width set of valid egg candidate events, calculate the mean and standard deviation, and update the upper and lower limits of the interval for valid egg candidate events to obtain valid egg-laying events. The calculation formula is as follows: In the formula, and These are the lower and upper thresholds for the updated valid candidate egg events, respectively. and The initial calibration value, For the effective candidate pulse width sample set, The average pulse width of the set. The standard deviation of the pulse width for this set. Confidence coefficient; S103 includes: C1: Extracting symmetrical time windows Feature vectors of each candidate EPC And calculate the target antenna ratio, the calculation formula is: In the formula, The target antenna percentage, This represents the number of times a candidate EPC is read by the target antenna within the current valid egg-laying event time window. This represents the total number of times a candidate EPC is read by the target nest and adjacent related antennas within the current time window. For the first One valid egg-laying event The infrared triggering time, and These are the candidate read time window lengths before and after the trigger time, respectively. For read frequency, For the average RSSI, For the length of stay, For time proximity, The most recent valid read time of the candidate EPC. The time proximity decay constant; C2: Based on the target antenna ratio, the DBSCAN clustering method is used to group candidate EPCs into three categories: target-stayed, interference-passed, and crosstalk-readout. For the candidate EPC regions in the target-stayed category, a weighted confidence score is calculated using the following formula: in, In order to achieve effective egg production events Under the conditions of occurrence, the first The confidence level of each candidate EPC. For feature vectors The first in Each feature component The corresponding parameters are read frequency, average RSSI, dwell time, target antenna ratio, and time proximity. For the first The weighting coefficients of each characteristic component, This is the Min-Max normalization function; C3: When and At that time, the effective egg-laying event will be Initially bind with the corresponding candidate EPC; otherwise proceed to C4. In the formula, The highest attribution confidence among candidate EPCs. It has the second highest confidence level of attribution. The minimum binding confidence threshold. This is the minimum threshold for distinguishing between the highest confidence level and the second-highest confidence level. C4: Calculate the time series variance of the received signal strength indication for the continuous card reading sequence of candidate waterfowl, and extract the nesting segment of the time series variance below the attitude variance threshold as the candidate egg-laying period. The calculation formula is as follows: In the formula, For the first The candidate EPC in the first The variance of the RSSI time series of consecutive card reading sequences. For the first The total number of records read in a continuous card reader sequence. For a moment The instantaneous value of RSSI, For the first The arithmetic mean of the RSSI of a continuous sequence of card readings; C5: In a symmetrical time window Based on the infrared triggering time of the egg-laying event, As the right endpoint, towards Back to the previous moment Duration, obtaining candidate backtracking windows The time decay factor is calculated for each candidate EPC, using the following formula: In the formula, The time decay factor, The moment when the egg-laying event is triggered. EPC departure time The time difference between them The attenuation constant is The one-way backtracking duration for the candidate backtracking window; C6: Calculate the extended attribution confidence based on the time decay factor, and for fuzzy sets with insufficient extended confidence differences, complete the probabilistic data association and attribution between valid egg-laying events and individual identities according to the principle of priority matching based on the time of departure from the nest. The calculation formula is as follows: In the formula, For the time decoupling condition, the first Each candidate EPC corresponds to a valid egg-laying event. Extended attribution confidence, To normalize the variance of the RSSI time series, To normalize low-variance nesting time, This is a biological hard constraint. , , These are variance weights, nesting time weights, and biological constraint weights, respectively.

2. The intelligent method for determining the egg production performance of floor-raised waterfowl according to claim 1, characterized in that, The S2 includes: S201: Use YOLOv8-nano lightweight target detector to detect hatching eggs at the end of the conveyor belt. If the cross-union ratio between detection frames is greater than the cross-union ratio deduplication threshold and the centroid distance is less than the centroid distance threshold, then merge the corresponding detection frames into the same egg. S202: The target tracking algorithm is used to establish the movement trajectory of the detected hatching eggs, and the camera field of view is divided into two coding regions. The region coding sequence of each hatching egg in different frames is tracked. When the coding sequence changes in accordance with the transport direction, the count is incremented by one. S203: For hatching eggs that have completed visual line crossing count, the region of interest image of the corresponding hatching egg is extracted according to the detection box coordinates, and the trajectory identifier, line crossing time, detection confidence and conveyor belt position code are recorded. The region of interest image is used as the input for egg quality grading. S204: Construct a lightweight egg quality grading neural network model, and use the lightweight egg quality grading neural network model to grade the quality of hatching eggs based on the region of interest image; S205: Based on the quality grading of hatching eggs, the infrared trigger count at the egg nest end and the visual line-crossing count at the end of the conveyor belt are time-aligned according to the conveyor delay. The difference between the infrared count and the visual count for the same batch is calculated using the following formula: In the formula, For the first The difference between batch infrared counts and visual counts, For the first Batch at time Infrared triggering cumulative counting at the nest end of the egg, For the first Batch at time The visual line crossing cumulative count at the end of the conveyor belt. This is the time offset used for counting alignment between the infrared channel and the visual channel; S206: Calculate the anomaly intensity based on the count difference. When the anomaly intensity exceeds the preset alarm threshold, output an anomaly alarm. The calculation formula is as follows: In the formula, For the first Batch infrared-visual counting anomaly intensity; S207: Align the visual tracking ID with the infrared event according to the transmission delay. When the matching judgment condition is met, bind the egg quality grading result with the individual identity information of the waterfowl and the valid egg-laying event. The matching judgment condition is: In the formula, The moment when the visual trajectory crosses the counting line. For the first One valid egg-laying event The infrared triggering time, To compensate for the time delay in the transport from the egg nest to the final visual inspection area, This represents the maximum permissible matching time error between infrared events and visual trajectories.

3. The intelligent method for determining the egg production performance of floor-raised waterfowl according to claim 2, characterized in that, The lightweight egg product grading neural network model includes: a backbone network, a direction-aware attention module, a multi-scale feature fusion module, and a classification head; The backbone network adopts the MobileNetV2 inverse residual structure. The orientation-aware attention module is placed after the depthwise separable convolution of the inverse residual unit of the backbone network and before the residual summation, and adopts the EfficientRCAM structure. The multi-scale feature fusion module adopts the CascadedRCAM three-layer architecture and performs feature fusion in the order of single module enhancement, residual fusion, and channel adaptation. The classification head consists of global average pooling, a fully connected layer, and a Softmax function; The training loss function of the lightweight egg grading neural network model adopts a general knowledge distillation form, which is a weighted combination of cross-entropy loss and KL divergence loss: in, Let be the total training loss function. For cross-entropy loss based on true class labels, The KL divergence loss is used to compare the soft label distributions of the teacher and student models. For distillation weight, This refers to the distillation temperature.

4. The intelligent method for determining the egg production performance of floor-raised waterfowl according to claim 1, characterized in that, The S3 includes: S301: Based on valid egg-laying events and egg quality grading results, using individual leg band codes as the primary key, valid egg-laying events, nesting sequences, and egg quality grading are aggregated into an individual event stream within a preset evaluation window. Multiple behavioral indicators are quantified, and after normalization, an individual behavioral profile vector is generated. An adaptive baseline is established for the behavioral indicators, calculated using the following formula: In the formula, for Baseline mean The baseline standard deviation, For individuals No. Daily Indicators The observed values, Baseline smoothing coefficient; S302: Statistically analyze nest utilization rate, flock laying time distribution, batch abnormality rate, and egg quality grading distribution. Employ spatial distribution monitoring methods and quantitatively assess the balance of nest usage using the Gini coefficient of each nest's time utilization rate. The calculation formula is as follows: In the formula, The total number of nests. The Gini coefficient represents the utilization rate of the nesting time. For the first The time utilization rate of each egg nest within the evaluation window. For the first Time utilization rate of each nest within the same evaluation window; S303: Compare each individual's daily behavioral indicators with the adaptive baseline, calculate the standardized deviation of each indicator, and only sum the deviations in the unfavorable direction to obtain the individual's comprehensive abnormality index. The calculation formula is as follows: In the formula, For individual comprehensive abnormality index, As an indicator The weight, The reliability coefficient is the indicator. As an indicator Unfavorable deviation direction sign, It is a positive number; S304: The results of individual egg production performance measurement for waterfowl are obtained based on the individual comprehensive abnormality index and the Gini coefficient of nest time utilization rate.

5. The intelligent method for determining the egg production performance of floor-raised waterfowl according to claim 4, characterized in that, S304 includes: When at least one of the following conditions is met: the comprehensive abnormality index is greater than the abnormality index threshold, an individual has not laid eggs for several consecutive days, or the abnormality rate is greater than the abnormality rate warning threshold, an abnormal individual warning will be output. When at least one of the following conditions is met: the nest utilization rate is lower than the lower limit threshold, or the Gini coefficient is greater than the preset coefficient, nest optimization suggestions are output.

6. An automatic egg collection device for implementing the intelligent method for determining the egg production performance of floor-raised waterfowl as described in any one of claims 1-5, characterized in that, include: Egg-laying nest unit, RFID identification module, infrared triggering module, signal grading acquisition module, egg collection and conveying module, end-point visual counting and grading module, main control processing module (9) and data management module; The egg-laying nest unit includes multiple egg-laying nests (1) arranged linearly and equally along the longitudinal direction, metal signal partitions (2-c) vertically arranged between adjacent egg-laying nests (1), side baffles (2-b) at both ends of the egg-laying nest (1) array, and V-shaped egg-gathering guide plates (4) at the rear outlet of each egg-laying nest (1). The RFID identification module includes an RFID antenna (3-a) and an RFID reader (3-b) at the bottom of the egg-laying nest (1). The RFID reader (3-b) is connected to the signal aggregation router (6) via an Ethernet interface. The signal aggregation router (6) aggregates the EPC data uploaded by multiple RFID readers (3-b) and uploads it to the main control processing module (9) in real time using the UDP protocol. The infrared triggering module includes a reflective infrared photoelectric sensor (5) located at the narrow opening of the V-shaped egg guide plate (4); the signal hierarchical acquisition module includes a signal aggregation router (6), an infrared sub-acquisition box (10-a), an infrared mother acquisition box (10-b), and a PLC (11); the signal of the reflective infrared photoelectric sensor (5) is connected to an infrared sub-acquisition box (10-a), the infrared sub-acquisition box (10-a) is aggregated to an infrared mother acquisition box (10-b), and the infrared mother acquisition box (10-b) is connected to the PLC (11) via an RS-485 bus; the PLC (11) performs a cyclic scan of the entire field photoelectric channel, timestamps each valid trigger event, and then transmits the photoelectric trigger data to the main control processing module (9) via the TCP protocol. The egg collection and conveying module includes an egg collection conveyor belt (7-a) and a conveyor belt motor (7-b). The egg collection conveyor belt (7-a) is a flat belt made of food-grade nylon material, which is laid horizontally along the longitudinal direction of the egg-laying nest (1) array and directly below the rear outlet of all egg-laying nests (1). The conveyor belt motor (7-b) is installed at the conveying end of the egg collection conveyor belt (7-a) on the side facing the end visual detection area and is fixed to the end of the support structure. The output shaft of the conveyor belt motor (7-b) is connected to the active roller drive of the egg collection conveyor belt (7-a) to drive the egg collection conveyor belt (7-a) to convey hatching eggs unidirectionally and uniformly along the longitudinal direction of the egg-laying nest (1) array. The end-visual counting and grading module is arranged above the end of the egg collection conveyor belt (7-a) and is used to complete visual inspection, tracking counting and egg quality grading. The system includes an embedded vision module (8), a support structure, a fixed background plate, and a supplementary lighting unit. The embedded vision module (8) includes an industrial-grade embedded AI camera, which is fixedly installed on the top beam of the support structure in a top-down manner and uploads the end vision detection results to the main control processing module (9). The fixed background plate is horizontally set directly below the embedded vision module (8) and is close to the imaging area of ​​the egg collection conveyor belt (7-a). The supplementary lighting unit uses a strip LED diffuse light source, which is symmetrically arranged on both sides of the top beam of the support structure along the width direction of the egg collection conveyor belt (7-a) and faces the imaging area of ​​the fixed background plate. The main control processing module (9) is fixedly placed on the field end ground on one side of the egg collection conveyor belt (7-a) of the egg nest array. The main control processing module (9) integrates a PLC (11) and power supply, communication and interface units in its cabinet. The data management module communicates with the main control processing module (9) via a local area network to receive and store data.

Images