A method and apparatus for measuring the growth performance and precise feeding of waterfowl

Abstract

This application discloses a method and apparatus for measuring the growth performance and precise feeding of waterfowl, belonging to the field of waterfowl farming technology. The method includes: using an RFID antenna to read the RFID tags on the waterfowl's leg bands to obtain the waterfowl's identity information; obtaining the waterfowl's weight, using the CUT-KDE algorithm to obtain the predicted daily weight, and matching it with the identity information; obtaining the feed weight and time when the waterfowl enters and leaves, obtaining a behavioral feature set; encapsulating the waterfowl's identity information, predicted daily weight, and behavioral feature set, and using an improved BP neural network model PDLR-BP to predict feed intake, obtaining the expected intake value, and then feeding; obtaining the feed conversion ratio based on the predicted daily weight and daily feed intake, classifying the waterfowl into grades, and displaying the waterfowl grades and analysis results through a visual interface for farmers' decision-making. This method significantly improves the frequency and accuracy of data collection and ensures data reliability.

Description

Technical Field

[0001] This application relates to the field of waterfowl farming technology, and in particular to a method and apparatus for measuring the growth performance of waterfowl and for precise feeding. Background Technology

[0002] Feed conversion ratio (FCR) measures the efficiency of converting feed input into body weight gain, typically expressed as the ratio of feed consumption to body weight gain. FCR is a crucial indicator for evaluating the growth performance of waterfowl; a lower value indicates higher feed utilization efficiency. Currently, FCR measurement in waterfowl primarily relies on manual operation, involving periodically weighing the waterfowl after cage rearing and recording the amount of remaining feed. This method is not only labor-intensive and time-consuming, but also causes significant stress to the waterfowl due to the small scale and low frequency of measurements, affecting production performance. Furthermore, cage rearing is detrimental to animal welfare, and manual observation is highly subjective, leading to substantial data errors and making it difficult to achieve high-precision, stable, and continuous monitoring.

[0003] Precision feeding refers to a scientific feeding method in livestock farming that precisely controls the type, ratio, feeding time, and amount of feed based on the animal's specific nutritional needs, growth stage, health status, and production goals. This aims to achieve efficient use of feed resources, improve farming efficiency, and reduce environmental pollution. However, the market currently lacks standardized and modular precision feeding systems suitable for waterfowl in different farming environments, and these systems suffer from low integration and poor compatibility between various sensors and software platforms.

[0004] To address these issues, researchers have proposed several methods and devices for measuring poultry growth performance or for precision feeding, but certain technical limitations remain. For example, existing technology discloses an automatic method for measuring waterfowl growth performance and its data management system. While this system can automatically record the feed intake and weight gain of individual waterfowl, it only has performance monitoring capabilities and fails to establish an intelligent feeding mechanism based on predictive models, making it difficult to meet the actual needs of precision feeding. Another example is an existing technology that proposes a method, system, and device for precision feeding of livestock and poultry based on image recognition. Although this analyzes the behavioral status of waterfowl through visual information, image recognition is less accurate and robust in weight measurement than physical weighing sensors. Especially in mixed-group feeding environments with multiple waterfowl, images are easily affected by occlusion, lighting, and changes in viewing angle, making it difficult to obtain high-quality, stable weight data.

[0005] Therefore, there is still a need for a waterfowl growth performance measurement and precision feeding technology solution that combines high-precision measurement capabilities with intelligent feeding control functions, in order to break through the gap between existing technologies in "monitoring" and "control" and achieve data-driven scientific feeding management and efficient resource utilization. Summary of the Invention

[0006] In view of the above-mentioned shortcomings in the prior art, this application provides a method and device for measuring the growth performance of waterfowl and for precise feeding, which solves the problems of incomplete data collection, insufficient predictive control, and lack of intelligent linkage in the prior art.

[0007] To achieve the aforementioned objectives, the technical solution adopted in this application is as follows:

[0008] First aspect:

[0009] This application provides a method for measuring the growth performance of waterfowl and for precise feeding, including:

[0010] S1: Use an RFID antenna to read the RFID tag on the waterfowl's leg band to obtain the waterfowl's identity information;

[0011] S2: Obtain the weight of the waterfowl, use the CUT-KDE algorithm to obtain the predicted daily weight of the waterfowl, and match it with the waterfowl's identity information;

[0012] S3: Obtain the feed weight and time when waterfowl enter and leave, and obtain the behavioral feature set of waterfowl;

[0013] S4: Encapsulate the waterfowl's identity information, predicted daily weight, and behavioral feature set, and use the improved BP neural network model PDLR-BP to predict feed intake to obtain the expected intake value. Feed the waterfowl based on the expected intake value. The improved BP neural network model PDLR-BP uses a multinomial learning rate decay instead of the fixed learning rate or step decay strategy in the traditional BP neural network.

[0014] S5: Based on the predicted daily body weight of waterfowl, the feed conversion rate is obtained by combining the daily feed intake. Based on the feed conversion rate, the waterfowl are classified into grades, and the waterfowl grades and analysis results are displayed through a visual interface for farmers to make decisions.

[0015] Furthermore, the method of obtaining the predicted daily body weight of waterfowl using the CUT-KDE algorithm includes:

[0016] A1: The weighing platform below the feeding station automatically reads the dynamic weight data of each waterfowl at preset time intervals, forming a dynamic weight sequence. ,in, For the first The dynamic weight data of the waterfowl collected this time. This refers to the number of data collections.

[0017] A2: Based on the age of each waterfowl, retrieve the standard weight range threshold corresponding to the age from the standard growth curve database of the same species of waterfowl, and modify the standard weight range threshold in combination with the improved Logistic growth curve model to obtain the modified weight prediction interval. Use the modified weight prediction interval to remove abnormal weight values ​​that do not conform to the physiological growth law.

[0018] The expression for removing abnormal weight values ​​that do not conform to physiological growth patterns using the modified weight prediction interval is as follows:

[0019]

[0020]

[0021]

[0022] In the formula, To obtain valid dynamic weight data after removing abnormal weight values, To calculate the minimum weight threshold using the improved Logistic growth curve model, To calculate the maximum weight threshold using the improved Logistic growth curve model, and To calculate the minimum weight threshold function and the maximum weight threshold function using the improved Logistic growth curve model, The value is the age in days. For identifying the species of waterfowl, For gender identification, The allowable percentage of weight fluctuation, for The coefficient of difference, The coefficient of difference between genders, This represents the theoretical maximum weight of the same species of waterfowl. For growth rate parameters, The inflection point age of the Logistic growth curve model;

[0023] A3: Sort the daily effective dynamic weight data of each waterfowl in ascending order, and set a lower limit and an upper limit for removal. Remove extreme values ​​from the sorted effective dynamic weight data to obtain a double-washed effective weight data set. The expression is:

[0024]

[0025] in, This is the lower limit of the exclusion ratio. This is the maximum exclusion ratio. The number of samples for effective dynamic weight data. Indicates rounding up. Indicates rounding down;

[0026] A4: The CUT-KDE algorithm is used to estimate the effective weight data set after double washing to obtain the predicted daily weight of waterfowl. The calculation formula is as follows:

[0027]

[0028] In the formula, For the predicted daily weight of waterfowl, For bandwidth parameters, For kernel function, This represents the sample size of valid weight data after double washing. To estimate the density of body weight location, The first after double cleaning One effective weight data;

[0029] Wherein, the kernel density is the Gaussian kernel function, expressed as:

[0030]

[0031] In the formula, For Gaussian kernel function, For standardized distance variables;

[0032] A5: The coefficient of determination is used to evaluate the fit of the CUT-KDE algorithm to waterfowl daily body weight data. The calculation formula is as follows:

[0033]

[0034] in, As the coefficient of determination, For the first The actual daily weight of a waterfowl This represents the average daily true body weight of all waterfowl. For the first Predicted daily weight of waterfowl For the number of waterfowl;

[0035] A6: The root mean square error (RMSE) is used to measure the average deviation between predicted daily weight and actual weight. The calculation formula is as follows:

[0036]

[0037] in, This is the root mean square error.

[0038] Furthermore, the acquisition of feed weight and time when waterfowl enter and leave, to obtain a set of behavioral characteristics of the waterfowl, includes:

[0039] S301: A feed weighing sensor installed under the feed trough monitors the change in feed weight before and after waterfowl feeding in real time.

[0040] S302: Based on the RFID tag and the weight scale platform, determine the time points when waterfowl enter and leave, extract the changes in feed weight, and obtain the amount of feed consumed per feeding.

[0041] S303: For the phenomenon of waterfowl feeding multiple times a day, record the number of times waterfowl feed and the total amount of feed consumed each day, and combine the dwell time and behavior frequency to obtain the behavioral characteristic set of waterfowl.

[0042] Furthermore, the encapsulation of the waterfowl's identity information, predicted daily weight, and behavioral feature set includes:

[0043] B1: Integrate the identity information, predicted daily weight, and behavioral characteristic set of each waterfowl using the waterfowl number as the primary key to obtain structured data entries;

[0044] B2: Based on structured data entries, establish a multi-dimensional data indexing mechanism to build a historical behavior data chain for individual waterfowl;

[0045] B3: Introduce an automatic anomaly detection mechanism. When abnormal weight, abnormal feed intake, RFID tag recognition failure, or mismatch between weight and feed intake are detected, an alarm will be triggered.

[0046] Furthermore, the step of using the improved BP neural network model PDLR-BP to predict feed intake and obtain the expected intake value, and then feeding based on the expected intake value, includes:

[0047] C1: Extract the identity information, predicted daily weight, and behavioral feature set of the encapsulated waterfowl;

[0048] C2: Feed intake is predicted based on the improved BP neural network model PDLR-BP to obtain the expected intake value;

[0049] C3: Based on the expected intake value, combined with the individual waterfowl's historical daily feeding rhythm and the group's average feeding pattern, a feeding control plan is obtained, and feeding is carried out according to the feeding control plan.

[0050] Furthermore, the improved BP neural network model PDLR-BP includes: an input layer, a hidden layer, and an output layer;

[0051] The input layer, hidden layer, and output layer all use a single-level Sigmoid activation function.

[0052]

[0053] in, It is a single-level Sigmoid activation function. The weighted sum of neurons;

[0054] The weighted sum of neurons is expressed as:

[0055]

[0056] in, For the first One trainable weight, As input features, This is a bias term.

[0057] Furthermore, the training process of the improved BP neural network model PDLR-BP includes:

[0058] In incremental mode, a weight fusion strategy is used to update the model parameters of the improved BP neural network model PDLR-BP. The weight update formula is as follows:

[0059]

[0060] In the formula, The weighted fusion coefficient, This is the weight matrix of the updated model. The weight matrix of the model before the update. This is the weight matrix obtained through forward and backward propagation training based on newly added feeding data;

[0061] During training, a multinomial learning rate is used to control model training, where the learning rate is:

[0062]

[0063] in, For the first Learning rate for each training iteration The initial learning rate, To minimize the learning rate, The total number of training rounds, indicated by the superscript. Let be the order of the polynomial.

[0064] Further, S5 includes:

[0065] S501: The feed conversion ratio is calculated based on the predicted daily body weight obtained from the CUT-KDE algorithm and the daily feed intake. The calculation formula is as follows:

[0066]

[0067] in, For feed conversion ratio, This represents the amount of food consumed that day. Today's weight. This is yesterday's weight.

[0068] S502: Based on preset The fault assessment and grading thresholds and the calculated feed conversion rate are used to classify waterfowl into three levels: high efficiency, medium efficiency, and low efficiency. The grading results, feeding efficiency, and weight gain data are then integrated and displayed in chart form on the main control platform.

[0069] The second aspect:

[0070] This application provides a waterfowl growth performance measurement and precision feeding device, including: an intelligent main control platform and multiple automatic weighing and feeding stations; the multiple automatic weighing and feeding stations are connected to the intelligent main control platform via Ethernet or wireless network;

[0071] The automatic weighing feeding station has a conical hopper at the top, an operation panel in the middle, operation buttons and an LCD screen on the operation panel, a camera at the upper right, an electric feeding unit, an electronic scale, a vibrating feeding motor and a microcontroller unit inside, and a feeding port at the bottom. Below the feeding port is a feed trough, under which a feed weighing sensor is placed. A weighing platform is located at the front, and the weighing platform and feed weighing sensor are mounted on a movable slide rail. The automatic weighing feeding station has an isolation baffle on the left side and a flat base with casters at the bottom. An RFID antenna is located on the right side of the flat base. The weighing platform, RFID antenna, feed weighing sensor, electric feeding unit, electronic scale and vibrating feeding motor are connected to the microcontroller unit via wires.

[0072] The intelligent main control platform is equipped with an LCD touch screen on the top, a lockable waterproof door panel at the front, an industrial control computer installed in the inner cavity as the core control unit of the system, and heat dissipation louvers, a handle assembly, and an air-cooling fan on the side.

[0073] The beneficial effects of this application are:

[0074] This application provides a method and apparatus for measuring the growth performance of waterfowl and for precise feeding. It achieves automatic identification of individual waterfowl and continuous, stress-free monitoring of their weight and feed intake, significantly improving the frequency and accuracy of data collection. It avoids the stress and errors caused by handling during manual weighing, ensuring data reliability. Furthermore, by utilizing an improved BP neural network model (PDLR-BP) to dynamically predict the short-term feeding behavior of individual waterfowl, it matches the optimal feeding amount based on the individual's historical behavior and real-time status. This significantly reduces feed waste and improves feed conversion rate while ensuring growth performance, thereby lowering feeding costs. Attached Figure Description

[0075] 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.

[0076] Figure 1 This is a flowchart illustrating a method for measuring the growth performance of waterfowl and for precise feeding, as provided in an embodiment of this application.

[0077] Figure 2 This is another flowchart illustrating a method for determining the growth performance of waterfowl and for precise feeding, provided in an embodiment of this application.

[0078] Figure 3 This is a schematic diagram of a waterfowl growth performance testing and precision feeding device provided in an embodiment of this application.

[0079] Figure 4 This is a three-dimensional schematic diagram of the complete structure of the automatic weighing and feeding station provided in the embodiments of this application.

[0080] Figure 5 This is a three-dimensional structural diagram of another automatic weighing and feeding station provided in an embodiment of this application.

[0081] Figure 6 This is a control logic diagram of a waterfowl growth performance measurement and precision feeding device provided in an embodiment of this application.

[0082] The components are: 1-conical hopper; 2-operation panel; 3-isolation baffle; 4-weighing scale platform; 5-swivel casters; 6-movable slide rail; 7-feeding trough; 8-camera; 9-electric feeding unit; 10-feeding port; 11-RFID antenna; 12-feed weighing sensor; 13-LCD touch screen; 14-lockable waterproof door panel; 15-heat dissipation louvers; 16-handle assembly; 17-air-cooled fan. Detailed Implementation

[0083] 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.

[0084] In existing technologies, it is difficult to collect individual data of poultry such as chickens, ducks, and geese in group-housed environments with high frequency and high precision: traditional equipment relies on manual weighing or intermittent collection of feeding information, resulting in insufficient data density and accuracy, and failing to reflect continuous changes in individual status. This is especially problematic when multiple waterfowl share a house, leading to issues such as identification confusion and data disconnection. Furthermore, it is impossible to achieve precise feeding control based on individual characteristics: existing technologies are mostly limited to performance testing, only collecting waterfowl weight or feeding data, lacking a mechanism to use historical behavioral data for predictive modeling and reverse-guided feeding, resulting in untargeted feed delivery and low resource utilization efficiency. Finally, there is a lack of a central and peripheral collaborative intelligent control system: existing feeding systems mostly adopt single-point control or closed logic, making it difficult to achieve dynamic adjustment and remote intervention, and lacking an efficient collaborative mechanism between different modules.

[0085] Example 1:

[0086] Based on this, embodiments of this application provide a method for measuring the growth performance of waterfowl and for precise feeding, which can be found in [reference needed]. Figure 1 and Figure 2 The explanation uses geese as a representative waterfowl. Figure 1 The diagram shown is a flowchart illustrating a method for determining the growth performance of waterfowl and for precise feeding, as provided in an embodiment of this application. The method includes:

[0087] S1: A set of UHF RFID antenna modules is installed below the feeding station. When geese enter the feeding station, the system continuously and actively reads the RFID electronic tag information bound to their leg bands through the antenna. The tag carries a unique identification number for each goose. The reading process is a non-contact identification. The RFID reader is connected to the main control platform through the network port.

[0088] S2: When geese enter the feeding station to feed, the RFID reader identifies each individual goose, and its weight is continuously measured by a high-precision weighing scale platform located in front of the feed trough. The weighing scale platform is connected to the main control platform via a serial port, collecting weight data of the goose standing on the scale platform once per second, and recording the weight change value throughout the entire feeding cycle. To improve accuracy, this application introduces the CUT-KDE (upper and lower bound constraints + kernel density estimation) algorithm to process the data.

[0089] In one embodiment of this application, the CUT-KDE algorithm is used to estimate daily body weight, specifically including:

[0090] A1: After the geese complete identification and enter the feeding station to eat, the weighing scale platform installed below the station will be automatically activated. The system is set to continuously collect the weight of the geese at a fixed time interval of 1 second, recording a weighing sequence containing multiple data points. ,in, For the first The dynamic weight data of the waterfowl collected this time. This represents the number of data collections. Each data entry includes a collection timestamp and a goose ID tag, forming the goose's "original weight trajectory" during that entry behavior.

[0091] A2: After collecting the weight data sequence, the system first retrieves the corresponding age value of the goose based on its identification information (such as RFID, leg band number, etc.). It also retrieves the corresponding weight range threshold for that age from the built-in standard growth curve database. This threshold is in the form of an interval, used to define the reasonable range of weight fluctuation under the current growth stage. Simultaneously, a weight prediction interval constructed using the improved Logistic growth curve model is used to jointly verify the dynamic weight data, eliminating abnormal weight values ​​that do not conform to physiological growth patterns.

[0092] The expression for removing body weight values ​​using the standard growth curve library is as follows:

[0093]

[0094] In the formula, To obtain valid dynamic weight data after removing abnormal weight values, To calculate the minimum weight threshold using the improved Logistic growth curve model, The maximum weight threshold is calculated using the improved Logistic growth curve model.

[0095] The expression for calculating the minimum weight threshold using the improved Logistic growth curve model is as follows:

[0096]

[0097] The expression for calculating the maximum weight threshold using the improved Logistic growth curve model is as follows:

[0098]

[0099] In the formula, and To calculate the minimum weight threshold function and the maximum weight threshold function using the improved Logistic growth curve model, The value is the age in days. For identifying the species of waterfowl, For gender identification, The allowable percentage of weight fluctuation, for Difference coefficient of waterfowl breeds This indicates that the variety is heavier than the basic variety, and vice versa. The coefficient of difference between genders, This represents the theoretical maximum weight of the same species of waterfowl. For growth rate parameters, This represents the inflection point age in the Logistic growth curve model.

[0100] By combining the standard library and the improved growth curve model for dual verification, outliers such as posture shift, abnormal dwell time, and sensor noise in the dynamic weight sequence can be effectively removed, resulting in a set of effective weight data after the first step of cleaning.

[0101] A3: For the remaining valid weight data in set S, due to the potential for high-frequency fluctuations caused by short-term posture changes during feeding, further optimization of data quality is needed through extreme value truncation. The daily valid dynamic weight data for each waterfowl is sorted in ascending order, and a lower and upper limit removal ratio is set. The minimum and maximum values ​​at both ends of the sequence are removed according to the preset ratios, resulting in a double-washed set of valid weight data. :

[0102]

[0103] in, This is the lower limit of the exclusion ratio. This is the maximum exclusion ratio. The number of samples for effective dynamic weight data. Indicates rounding up. This indicates rounding down to the nearest integer.

[0104] A4: After the above double cleaning, the final effective weight dataset is used as the input sample for kernel density estimation to estimate the daily weight distribution of the waterfowl and to calculate the predicted daily weight value, as shown in the following formula:

[0105]

[0106] In the formula, For the predicted daily weight of waterfowl, For bandwidth parameters, For kernel function, This represents the sample size of valid weight data after double washing. To estimate the density of body weight location, The first after double cleaning One effective weight data point.

[0107] The kernel density is chosen using a Gaussian kernel function, and its formula is as follows:

[0108]

[0109] In the formula, For Gaussian kernel function, To standardize the distance variable, the Gaussian kernel can assign higher weights to samples closer to the estimation point, thereby improving the smoothness and stability of the density estimation.

[0110] A5: To verify the accuracy of the CUT-KDE algorithm in estimating daily body weight, the coefficient of determination (COP) was used. R ²) The coefficient of determination measures how well a regression model fits the observed data; it reflects the explanatory power of the independent variables for the dependent variable. The formula for calculating the coefficient of determination is as follows:

[0111]

[0112] in, As the coefficient of determination, For the first The actual daily weight of a waterfowl This represents the average daily true body weight of all waterfowl. For the first Predicted daily weight of waterfowl To determine the number of waterfowl, the method for obtaining the actual daily weight is to select four time points on the same day, manually weigh the waterfowl, and calculate the average value.

[0113] A6: To further quantify the deviation between predicted weight and actual weight, the root mean square error is used to quantify the average deviation between the model's predicted values ​​and the actual observed values. The smaller the value, the higher the stability and accuracy of the model in predicting body weight. The calculation formula is as follows:

[0114]

[0115] in, This is the root mean square error.

[0116] S3: The system records the initial feed weight when the goose enters and the final feed weight when it leaves, and calculates the difference between the two to obtain the total amount of feed consumed during this feeding. To improve accuracy, the system supports filtering out minor vibration errors and instantaneous pecking rebounds, and eliminates invalid disturbances through sliding window averaging and dynamic threshold control.

[0117] In one embodiment of this application, S3 specifically includes:

[0118] S301: A high-precision gravity weighing sensor, i.e., a feed weighing sensor, is installed below the feed trough in the feeding station to continuously monitor the feed quality in real time. The system records the initial feed weight before the geese enter the station and records the remaining feed amount after the geese finish eating and leave the station. The difference in weight between the two points in time represents the feed consumption for that feeding session, providing data support for subsequent behavior statistics and model input.

[0119] S302: The system uses a combination of RFID identification and a weight scale platform to determine the entry and exit times of geese, and extracts feed weight changes within this time window. If the goose's stay time in the station is less than the minimum effective feeding threshold (10 seconds), the system will consider the behavior invalid data and discard it. Valid data will be recorded as a complete feeding event, and the output parameters are the single feed intake (unit: grams) and the corresponding timestamp.

[0120] S303: For geese exhibiting multiple feeding activities per day, the system automatically summarizes the effective feeding frequency and total feed intake for each individual, and combines this with the duration of each feeding session at the station to generate a daily feeding behavior feature set for the geese. This feature set is used to model and analyze individual behavioral rhythms, identify overfeeding or underfeeding, and also serves as an important reference for machine learning models to predict future intake.

[0121] S4: The system has an independent data management submodule, which is responsible for receiving and processing data uploaded from multiple feeding station terminals in real time. Whenever a goose completes a feeding process, the system organizes and encapsulates its multi-dimensional indicators, such as its identification number, sex, current age, estimated weight, single feed intake, feeding time, and cumulative number of feedings, and archives them according to timestamps.

[0122] In one embodiment of this application, S4 specifically includes:

[0123] S401: The system integrates the data collected from each step of identification (S1), weight measurement (S2), and feeding record (S3) using the goose number as the primary key to form structured data entries. Each record contains complete basic attributes (number, sex, age) and dynamic behavioral parameters (weight, feed intake, feeding time, feeding frequency, etc.). This aggregated data serves as the basic data source for analysis, modeling, and control.

[0124] S402: The system establishes a multi-dimensional data indexing mechanism, creating a timeline of behavior records for each individual goose. Users can query, analyze, and export historical data at various granularities, such as by day, week, stage, or group. Simultaneously, the front-end display module supports dynamically loading chart data through a data filtering interface, enabling flexible visualization support.

[0125] S403: To improve data quality and decision reliability, the system introduces an automatic outlier detection mechanism. When problems such as consecutive weight fluctuations (±20% or more), prolonged periods without food intake, RFID identification failure, or mismatch between weight and food intake occur, the system will automatically mark them as abnormal data and provide warnings or trigger alarms through the front-end interface, allowing users to intervene in a timely manner.

[0126] S5: The system retrieves historical behavioral data of the individual goose at different ages based on RFID identifiers. This data, combined with static features (such as sex, number, and weight) and dynamic behavioral features (such as feeding time distribution, frequency, and previous feed intake), serves as the input variables for the feed intake prediction model PDLR-BP (a BP neural network model optimized with a multinomial learning rate decay algorithm). The improved BP neural network model PDLR-BP uses multinomial learning rate decay to replace the fixed learning rate or step decay in traditional BP neural networks. This results in a larger learning rate in the early stages of training, leading to faster convergence and the ability to escape local minima, while a smaller learning rate in the later stages allows for more refined parameter updates and a more stable fit. The entire process is precisely controlled by a multinomial function.

[0127] In one embodiment of this application, S5 specifically includes:

[0128] S501: Before predicting feed intake, the system first retrieves historical information and current behavioral data of the target geese stored in the database to construct a multi-dimensional input feature set. Standardization, normalization, and missing value imputation are then used to ensure consistent feature dimensions, uniform amplitude, and controllable noise. This information includes static features such as the goose's identification number, sex, physiological age, and previous day's weight, as well as dynamic behavioral features such as the current feeding time, feeding frequency, and most recent feed intake.

[0129] S502: Based on the multidimensional input feature set constructed in step S501, the system feeds it into the PDLR-BP feed intake prediction model for inference calculation. The improved BP neural network model PDLR-BP uses multinomial learning rate decay instead of the fixed learning rate or step decay strategy in traditional BP neural networks. This allows the model to maintain a large learning rate in the early stages of training to achieve rapid convergence and the ability to escape inferior local minima; in the later stages of training, the learning rate is automatically reduced, resulting in more refined parameter updates and a more stable fitting process. The overall learning rate change is precisely controlled by a multinomial function. The model's final output target is to predict the feed intake (in grams) of the goose in the next 24 hours. This predicted value is used to characterize the system's judgment on the short-term feeding behavior trend of an individual based on historical characteristics, providing data support for subsequent intelligent feeding control.

[0130] S503: Based on the predicted intake results obtained in step S502, the system combines the individual's past daily feeding rhythm with the group's average feeding efficiency to generate a personalized feeding control plan. This plan includes recommended target feed amount (accurate to the gram), suggested feeding time periods, and whether to provide feed in segments. This plan is then sent as a parameter package to the main control platform, which drives the feeding station to perform intelligent feeding operations. This closed-loop control ensures that individuals receive nutritional support matching their growth needs, improving feed utilization efficiency.

[0131] In one embodiment of this application, the PDLR-BP prediction model specifically includes:

[0132] B1: Based on long-term, stable collection of individual goose growth and behavior data from feeding stations, a multi-dimensional input feature set covering static physiological parameters and dynamic feeding behavior is constructed, including but not limited to: individual number, sex, age, daily weight, single feed intake, feeding interval, feeding time distribution, and historical weight gain trend. Data preprocessing methods such as standardization, normalization, and missing value imputation are used to ensure that the input vector dimensions are uniform, the amplitude is consistent, and the noise is controllable, providing a stable and reliable data foundation for the neural network training and inference of the PDLR-BP model. The normalization formula is as follows:

[0133]

[0134] in, For the normalized sample, In preparation for normalization of the first One sample, The maximum value of the sample. The minimum value of the sample set is given by the normalized sample set range. .

[0135] B2: The neural network in the PDLR-BP model is built on a feedforward BP neural network framework, including an input layer, hidden layers, and an output layer. The input layer uses normalized features such as the individual goose's ID, weight, sex, and age as model input variables, mapping them to the hidden layers to achieve feature extraction and regression prediction of nonlinear feeding behavior patterns. The activation functions from the input layer to the hidden layers and from the hidden layers to the output layer both use a single-level Sigmoid function, expressed as follows:

[0136]

[0137] in, It is a single-level Sigmoid activation function. This is the weighted sum of the neurons.

[0138] The weighted sum of neuron inputs is defined as:

[0139]

[0140] in, For the first Each trainable weight determines the input features. The strength of the effect on neuron output This is a bias term used to adjust the shift amount of the activation function.

[0141] B3: A multinomial learning rate decay strategy is employed during model training. The learning rate is dynamically adjusted according to the training cycle, maintaining a larger step size in the early stages of training to accelerate convergence, and reducing the learning rate in the later stages to improve model stability and generalization ability. This strategy effectively optimizes the gradient update process of traditional BP neural networks, enabling the model to achieve higher training efficiency and predictive performance when handling high-noise, highly nonlinear data. The multinomial learning rate decay formula is:

[0142]

[0143] in, For the first Learning rate for each training iteration The initial learning rate, To minimize the learning rate, The total number of training rounds, indicated by the superscript. The order of the polynomial is used to control the slope of the learning rate decrease curve.

[0144] B4: After periodically acquiring the latest feeding records, the system automatically appends the new data, which has passed data quality verification, to the training set and triggers the model to perform incremental training or full retraining. In incremental training mode, the model adopts a weight fusion-based update strategy, proportionally fusing the weights obtained from training on the new data with the original model weights. This achieves adaptive updates of the model parameters, enabling the model to continuously absorb the latest changes in feeding behavior, weight gain trends, and environmental disturbances. The incremental weight update formula is as follows:

[0145]

[0146] In the formula, This is the weight fusion coefficient, used to adjust the fusion ratio of the old and new weights. This is the weight matrix of the updated model. The weight matrix of the model before the update. This is the weight matrix obtained through forward and backward propagation training based on newly added feeding data.

[0147] B5: After each round of feeding prediction, the main control platform performs a deviation analysis between the prediction results of the PDLR-BP model and the actual feeding data. When the system detects that the continuous prediction deviation exceeds a preset threshold, it will activate the model diagnosis and correction mechanism, including reinitializing the learning rate parameters, adjusting the network structure hyperparameters (such as the number of hidden layer nodes, activation functions, etc.), resetting the decay exponent, or triggering a manual review process. For geese exhibiting abnormal behavior (such as a sudden drop in feeding amount, abnormal feeding behavior rhythm, etc.), the system will automatically mark them and issue an alarm, realizing a data-driven prediction feedback loop, effectively improving the prediction stability, model reliability, and intelligent monitoring capabilities during long-term operation.

[0148] S6: The intelligent control platform automatically compares the geese's historical weight data daily to calculate their actual daily weight gain, and combines this with the actual feed intake data for the day to calculate the feed conversion ratio (FCR). The platform displays information such as the geese's FCR, feeding behavior frequency, and weight gain curve in chart form on the interactive interface. At the same time, the system can classify individuals according to FCR thresholds and mark high-quality or abnormal individuals.

[0149] In one embodiment of this application, S6 specifically includes:

[0150] S601: The system extracts the weight data of the geese from the daily CUT-KDE algorithm, comparing the current day's and the previous day's weight data, and calculates the individual's daily weight gain by the difference between the two. The system automatically records this indicator and archives it in conjunction with the feeding behavior data for future evaluation and model training.

[0151] S602: Substitute the model-predicted feed intake and daily weight gain into the following formula to calculate the feed conversion ratio (FCR). FCR ):

[0152]

[0153] in, For feed conversion ratio, This represents the amount of food consumed that day. Today's weight. This is yesterday's weight.

[0154] S603: The system presets the FCR evaluation grading threshold, which is usually set as follows: Level 1 is high efficiency ( FCR <1.8), Level 2 is medium (1.8≤ FCR ≤2.2), Level 3 is inefficient ( FCR >2.2). Daily system based on individual FCRThe calculation results automatically classify the geese and integrate data such as grade, feeding efficiency, and weight gain performance, displaying them in chart form on the main control platform. This allows farmers to quickly identify high-performing individuals and those to be optimized, assisting in precise breeding and adjustments to the breeding process.

[0155] S604: The main control platform's front-end interface features intuitive touch operation, allowing users to adjust settings such as recommended feeding time and feed ratio in real time. FCR The system includes core parameters such as thresholds and logic for judging abnormal feeding. It also supports exporting feeding and weight gain analysis reports by age, group, and individual, which can be used for data archiving, operational analysis, and selection of superior geese. This makes it an important tool for intelligent farming and precision breeding decision-making.

[0156] Example 2:

[0157] This application provides a device for measuring the growth performance of waterfowl and for precise feeding. This device can be found in [reference needed]. Figure 3 , Figure 4 and Figure 5 It includes: an intelligent main control platform and multiple automatic weighing and feeding stations.

[0158] The automatic weighing feeding station consists of a conical hopper 1, an operation panel 2, an isolation baffle 3, a weighing platform 4, swivel casters 5, movable slide rails 6, a feed trough 7, a camera 8, an electric feeding unit 9, a feeding port 10, an RFID antenna 11, and a feed weighing sensor 12.

[0159] In one embodiment of this application, the automatic weighing and feeding station is mounted on a flat base with casters 5, facilitating deployment and movement within the farm. An integrated conical hopper 1 is located above the device to hold dry feed, connected to an electric feeding unit 9 below via gravity flow. An integrated control box is located in the center, integrating control buttons and a digital display screen. Internally, a vibrating feeding motor and an electronic scale are connected to control the feeding action and parameter settings. The electronic scale is used to weigh the feed. A metal isolation baffle 3 is located on the left side of the station to guide the waterfowl's entry and exit, preventing interference with feeding. A weighing platform 4 and a feed trough 7 are located at the front of the base, with a weighing scale for real-time recording of waterfowl weight and a feed weighing sensor 12 for monitoring changes in feed weight installed below them. These weighing scales and feed weighing sensors can be strain gauge type weighing sensors. All sensors, motors, and RFID modules are connected to an internal microcontroller unit via wires to achieve unified data acquisition and control command response. The device communicates with the main control platform via an Ethernet interface, receives and sends instructions, and provides feedback on behavioral data such as identification, weighing, and feeding, enabling edge data acquisition and local execution.

[0160] The intelligent main control platform consists of an LCD touch screen 13, a lockable waterproof door panel 14, heat dissipation louvers 15, a handle assembly 16, and an air-cooled fan 17.

[0161] In one embodiment of this application, the intelligent main control platform adopts a vertical integrated structure design. An industrial LCD touchscreen 13 is mounted on the top sloping surface for data display, parameter setting, and alarm prompts. An openable, lockable, waterproof door 14 is located at the front, housing an industrial control computer as the core control unit of the system, running data processing and prediction algorithms. The platform's sides are equipped with heat dissipation louvers 15, a cooling fan 17, and a handle assembly 16 to ensure stable operation of the equipment in high-temperature and high-humidity environments. All feeding stations are connected to the platform via wired network ports, enabling synchronized data uploads, model prediction, feeding strategy optimization, and control command issuance. The main control platform incorporates power conversion and overvoltage protection circuits, forming a collaborative control system with the feeding stations—a "centralized analysis—edge execution" approach—enabling intelligent monitoring and precise feeding control throughout the entire breeding process.

[0162] The control logic diagram of the device provided in this application is as follows: Figure 6 As shown, it can be divided into 5 modules, including:

[0163] The automatic feeding module, featuring a large-capacity conical hopper (also known as a gravity feed hopper) at the top of the device, stores dry feed and has an anti-clogging design. The hopper is connected to the central feeding control unit via a gravity inlet, ensuring smooth feed delivery under gravity. The feeding station has a control panel in the center, equipped with an electric feeding unit containing a vibratory feed motor and electronic scale. It receives commands from the main control platform to control the feeding amount. The front panel features operation buttons and an LCD screen, allowing users to locally set feeding parameters, perform start-up tests, and check status.

[0164] The feeding and weighing module features a feeding trough located below the feed outlet. A high-precision feed weighing sensor is integrated into the lower part of the trough to monitor the weight of each feed delivery in real time. Behind the feeding trough is a waterfowl weight scale platform, supported by an independent weighing unit. Waterfowl stand on this platform while feeding, and their weight data is automatically collected. The scale platform and the feeding trough are arranged front and back to ensure that the waterfowl complete the "identification—weighing—feeding" process sequentially upon entering the station.

[0165] The RFID identification module has a UHF RFID antenna installed near the platform entrance below the device. It is used to read the electronic tags in the waterfowl leg bands to achieve automatic identification of individual identities. The identification data corresponds one-to-one with subsequent feeding and weight information and is uploaded to the central processing system as data tags.

[0166] The interactive display module is located on the tilted panel at the top of the device. It uses an industrial-grade LCD touch screen to display waterfowl feeding behavior data, weight change curves, feed conversion rate analysis charts, etc. It supports users to manually set parameters locally, such as feeding time, feeding strategy, alarm thresholds, etc.

[0167] The core processing module, namely the main control processing and data management system, integrates an industrial control computer (IPC) inside the main cavity of the intelligent main control platform. It serves as the central control unit of the feeding station network, runs the feeding management system software, and realizes data reception, algorithm processing (such as CUT-KDE, PDLR-BP), predictive model inference, and equipment control command issuance. The storage module is a solid-state drive, which is used to save age, number, weight, feed intake, prediction records, etc.

[0168] Centered on the core processing module, both the RFID identification module and the feed weighing module transmit data unidirectionally to the core processing module, providing sensing information such as waterfowl identification, weight, and feed intake. The core processing module, in turn, sends control commands unidirectionally to the automatic feeding module to adjust feeding behavior during the feeding process. The interactive display module communicates bidirectionally with the core processing module, displaying real-time data and visualization results, as well as receiving user-input control commands and parameter configurations, thus enabling human-machine interaction and system configuration functions. This structure constructs an intelligent feeding control system integrating data sensing, intelligent decision-making, and visual interaction.

[0169] In one embodiment of this application, waterfowl stand on the platform. At this time, a weighing scale measures the weight of the waterfowl, and a feed weighing sensor in the feed trough measures the remaining feed before the waterfowl enters. Then, a predictive model estimates the amount of feed the waterfowl needs to consume. An automatic feeding module uses an electronic scale to weigh the required amount of feed (the predicted amount minus the remaining feed) and dispenses it into the feed trough. The reason for adding a scale to the automatic feeding module instead of directly using the feed weighing sensor is that the waterfowl's feeding in the feed trough affects the reading of the feed weighing sensor.

[0170] This application employs an automated weighing and feeding station integrating an RFID module and a high-precision weighing scale platform. This enables automatic identification of individual waterfowl and continuous, stress-free monitoring of their weight and feed intake, significantly improving the frequency and accuracy of data collection. It avoids stress reactions and errors caused by capture during manual weighing, ensuring data reliability. Through an improved BP neural network model (PDLR-BP), this invention can dynamically predict the short-term feeding behavior of each waterfowl, matching the optimal feeding amount based on individual historical behavior and real-time status. This significantly reduces feed waste and improves feed conversion rate while ensuring growth performance, thereby lowering feeding costs. The intelligent main control platform integrates an industrial touchscreen, computing module, heat dissipation unit, and data management system. It not only enables local model deployment and visual interface operation but also provides functions such as abnormal data alarms, remote parameter configuration, and analytical decision support. The system, through real-time interconnection between the main control platform and each feeding station, constructs a closed-loop logic of "identification—measurement—prediction—control," improving system collaboration efficiency and possessing good scalability and replicability, making it suitable for large-scale farming and waterfowl breeding scenarios.

[0171] 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 waterfowl growth performance measurement and precision feeding method, characterized in that, The method comprises the following steps: S1: reading the RFID tag of the waterfowl leg ring by using the RFID antenna to obtain the identity information of the waterfowl; S2: obtaining the body weight of the waterfowl, using the CUT-KDE algorithm to obtain the predicted daily body weight of the waterfowl, and matching the identity information of the waterfowl; S3: obtaining the feed weight and time when the waterfowl enters and leaves to obtain the behavior feature set of the waterfowl; S4: encapsulating the identity information, predicted daily body weight and behavior feature set of the waterfowl, and using the improved BP neural network model PDLR-BP to predict the intake amount, obtain the expected intake value, and feed based on the expected intake value, wherein the improved BP neural network model PDLR-BP uses a polynomial learning rate decay to replace the fixed learning rate or the stepwise decay strategy in the traditional BP neural network; S5: based on the predicted daily body weight of the waterfowl, combining the daily intake amount to obtain the feed conversion rate, dividing the waterfowl into grades based on the feed conversion rate, and displaying the waterfowl grades and analysis results through a visual interface for the breeder to make decisions; The method for obtaining the predicted daily body weight of the waterfowl by using the CUT-KDE algorithm comprises the following steps: A1: The body weight scale platform under the feeding station automatically reads the dynamic body weight data of each waterfowl at preset time intervals to form a dynamic body weight sequence wherein, is the dynamic body weight data of the waterfowl collected for the first time, is the dynamic body weight data of the waterfowl collected for the nth time, is the number of times of collection; A2: according to the daily age value of each waterfowl, retrieving the standard body weight range threshold corresponding to the daily age value from the standard growth curve database of waterfowl of the same breed, and modifying the standard body weight range threshold by using the improved Logistic growth curve model to obtain a modified body weight prediction interval, and using the modified body weight prediction interval to eliminate abnormal body weight values that do not conform to the physiological growth law; Wherein, the expression for eliminating abnormal body weight values that do not conform to the physiological growth law by using the modified body weight prediction interval is: wherein, is the effective dynamic body weight data after removing abnormal body weight values, is the minimum body weight threshold calculated using the improved Logistic growth curve model, is the maximum body weight threshold calculated using the improved Logistic growth curve model, and are the minimum body weight threshold function and the maximum body weight threshold function calculated using the improved Logistic growth curve model, is the age value, is the breed identification of the waterfowl, is the gender identification, is the proportional coefficient allowing body weight fluctuation, is the difference coefficient of is the theoretical maximum body weight of the same breed of waterfowl, is the growth rate parameter, is the inflection point age of the Logistic growth curve model;​​ A3: arranging the daily effective dynamic body weight data of each waterfowl in ascending order, setting a lower limit rejection ratio and an upper limit rejection ratio, performing extreme value rejection on the effective dynamic body weight data arranged in ascending order to obtain a double-cleaned effective body weight data set, and the double-cleaned effective body weight data set The expression is: wherein, is a lower rejection ratio, is an upper rejection ratio, is a number of samples of valid dynamic weight data, denotes rounding up, denotes rounding down; A4: using the CUT-KDE algorithm to estimate the effective body weight data set after double cleaning to obtain the predicted daily body weight of the waterfowl, and the calculation formula is: wherein is the predicted daily body weight of the waterfowl, is a bandwidth parameter, is a kernel function, is the number of valid body weight data after double cleaning, is the body weight position to estimate the density, is the first valid body weight data after double cleaning, is the second valid body weight data after double cleaning. Wherein, the kernel density is a Gaussian kernel function, and the expression is: wherein is a Gaussian kernel function, is a normalized distance variable; A5: using the coefficient of determination to evaluate the fitting degree of the CUT-KDE algorithm to the daily body weight data of the waterfowl, and the calculation formula is: wherein, R2is a correlation coefficient, R2is a correlation coefficient, the real daily body weight of each waterfowl, R2is a correlation coefficient, the predicted daily body weight of each waterfowl, the predicted daily body weight of each waterfowl, the number of waterfowl; A6: using the root mean square error to measure the average deviation between the predicted daily body weight and the true body weight, and the calculation formula is: wherein is the root mean square error.

2. The waterfowl growth performance measurement and precision feeding method according to claim 1, characterized in that, The method for obtaining the behavior feature set of the waterfowl by obtaining the feed weight and time when the waterfowl enters and leaves comprises the following steps: S301: the feed weighing sensor installed below the trough monitors the change of the feed weight before and after the waterfowl feeds in real time; S302: according to the RFID tag and the body weight scale platform, the time points when the waterfowl enters and leaves are judged, the feed weight change is extracted, and the single feeding amount is obtained; S303: for the phenomenon of multiple feeding of the waterfowl within a day, the feeding frequency and total amount of the waterfowl per day are recorded, and the behavior feature set of the waterfowl is obtained in combination with the residence time and behavior frequency.

3. The waterfowl growth performance measurement and precision feeding method according to claim 1, characterized in that, The method for encapsulating the identity information, predicted daily body weight and behavior feature set of the waterfowl comprises the following steps: B1: integrating the identity information, predicted daily body weight and behavior feature set of each waterfowl as the primary key of the waterfowl number to obtain a structured data entry; B2: based on the structured data entry, a multi-dimensional data indexing mechanism is established, and a historical behavior data chain is established for each waterfowl individual. B3: An abnormal value automatic detection mechanism is introduced, and an alarm is triggered when it is monitored that abnormal weight, abnormal feeding, RFID tag identification failure or weight and feeding do not match occur continuously.

4. The waterfowl growth performance measurement and precision feeding method according to claim 1, characterized in that, The improved BP neural network model PDLR-BP is used to predict the feed intake, and an expected intake value is obtained, and feeding is performed based on the expected intake value, including: C1: Extract the identity information, predicted daily weight and behavior feature set of the packaged waterfowl; C2: Perform feed intake prediction based on the improved BP neural network model PDLR-BP to obtain an expected intake value; C3: According to the expected intake value, combine the historical daily feeding rhythm of the waterfowl individual and the average feeding rule of the group to obtain a feeding control plan, and perform feeding according to the feeding control plan.

5. The waterfowl growth performance measurement and precision feeding method according to claim 4, characterized in that, The improved BP neural network model PDLR-BP includes an input layer, a hidden layer and an output layer; The input layer, the hidden layer and the output layer all use a single-level Sigmoid activation function: wherein, is a single-stage Sigmoid activation function, is a weighted sum of neurons; The expression of the weighted sum of neurons is: wherein, is the th trainable weight, is the input feature, is the bias term.

6. The waterfowl growth performance measurement and precision feeding method according to claim 5, characterized in that, The training process of the improved BP neural network model PDLR-BP includes: In the incremental mode, the model parameters of the improved BP neural network model PDLR-BP are updated by using a weight fusion strategy, wherein the weight update formula is: In the formula, is a weight fusion coefficient, is a weight value matrix of the updated model, is a weight value matrix of the model before updating, is a weight value matrix obtained by forward propagation and back propagation training based on the newly added feeding data; In the training process, polynomial learning rate decay is used for model training control, and the expression of the learning rate decay is: wherein is the learning rate for the is the initial learning rate, is the minimum learning rate, is the total number of training epochs, the superscript is the polynomial order.

7. The waterfowl growth performance measurement and precision feeding method according to claim 1, characterized in that, The S5 includes: S501: Calculate the feed conversion rate according to the predicted daily weight obtained by the CUT-KDE algorithm and the daily feed intake, and the calculation formula is: wherein, is the feed conversion ratio, is the daily feed intake, is the today weight value, is the yesterday weight value; S502: Based on the preset Based on the preset evaluation grading threshold and the calculated feed conversion rate, the waterfowl are divided into three grades of high efficiency, medium efficiency and low efficiency, and the grade results, feed efficiency and weight gain data are integrated and displayed in the form of charts on the main control platform.

8. The device for determining the growth performance and precision feeding of waterfowls according to any one of claims 1-7, characterized in that, Including: An intelligent master control platform and a plurality of automatic weighing feeding stations; The plurality of automatic weighing feeding stations access the intelligent master control platform through Ethernet or wireless network. The top of the automatic weighing feeding station is provided with a conical hopper (1), the middle is provided with an operation panel (2), the operation panel (2) is provided with operation keys and a liquid crystal display screen, the upper right is provided with a camera (8), the inside is provided with an electric unloading unit (9), an electronic scale, a vibration type feeding motor and a microcontroller unit, and the lower part is provided with a discharging port (10); The feeding trough (7) is arranged below the discharging port (10), the feeding trough (7) is arranged below the feeding trough (7), the front is provided with a body weight scale platform (4), and the body weight scale platform (4) and the feeding weighing sensor (12) are installed on the movable slide rail (6); The left side of the automatic weighing feeding station is provided with a isolation baffle (3), and the bottom is provided with a flat base with universal casters (5), and the right side of the flat base is provided with an RFID antenna (11); The body weight scale platform (4), the RFID antenna (11), the feeding weighing sensor (12), the electric unloading unit (9), the electronic scale and the vibration type feeding motor are connected to the microcontroller unit through wires; The top of the intelligent master control platform is provided with a liquid crystal touch screen (13), the front is provided with a waterproof door plate (14) with a lock, the inner cavity is provided with an industrial computer as a system core control unit, the side is provided with a heat dissipation louver (15), a handle assembly (16) and a cooling fan (17).

Images