Intelligent feed delivery method and system for automated rabbit farm

Abstract

This invention relates to the field of intelligent animal husbandry technology, and discloses an intelligent feed delivery method and system for automated rabbit farms. The intelligent feed delivery method for automated rabbit farms includes: collecting image and radio frequency data from the rabbit farm to confirm the identity of the doe; matching the doe's associated records to assess the current lactation performance of the doe; assessing feeding requirements; calculating feeding amounts and checking inventory; monitoring the behavior data of does nursing in each cage in real time and identifying nursing behavior, analyzing nursing patterns and predicting feeding times; planning and dynamically adjusting the path of the feeding cart; controlling the feeding cart to perform feeding operations, and diagnosing and handling abnormal situations during the execution process. This invention achieves accurate assessment of the nutritional needs of lactating does and dynamic adjustment of individualized feeding amounts, solving the problem that traditional feeding methods using fixed feeding amounts cannot adapt to individual differences and dynamic changes in the lactation stage.

Description

Technical Field

[0001] This invention relates to the field of intelligent farming technology, and more specifically, to an intelligent feed delivery method and system for automated rabbit farms. Background Technology

[0002] With the accelerated modernization of animal husbandry, automated feeding and management in large-scale rabbit farms has become an inevitable trend in the industry. In breeding rabbit farms, lactating does are key targets for nutritional management, as their nutritional supply during lactation directly affects the growth and development of their kits and the recovery of the does's reproductive performance. However, the nutritional needs of lactating does are highly dynamic, with milk production initially increasing and then decreasing with postpartum age, and is influenced by a combination of factors such as parity, litter size, and ambient temperature. Traditional fixed-amount and timed-quantity feeding strategies are insufficient to meet the individualized and precise nutritional needs of these does.

[0003] While existing automated feeding systems have achieved mechanized feeding, they still have significant shortcomings in terms of intelligence. Most systems can only feed rabbits at preset times according to a pre-set feeding program, lacking the ability to perceive the individual rabbit's condition and assess its nutritional needs. Although some systems have introduced individual recognition technology, the recognition method is simplistic and lacks stability in complex breeding environments such as changes in light intensity and dust. Regarding nutritional needs assessment, existing methods have failed to establish a correlation model between the mother rabbit's lactation performance and the growth status of the kits, and cannot accurately capture the changes in nutritional needs linked between mother and kits.

[0004] Therefore, there is an urgent need for an intelligent feed delivery method that can accurately identify individual rabbits, dynamically assess lactation performance, accurately calculate nutritional requirements, intelligently adjust feeding amounts, intelligently decide on feeding timing, and intelligently plan delivery paths. This would solve the technical problems existing in current technologies, such as insufficient stability in individual identification, inaccurate assessment of nutritional requirements, and fixed feeding timing, and achieve precise nutritional management of lactating rabbit populations. Summary of the Invention

[0005] This invention provides an intelligent feed delivery method and system for automated rabbit farms, solving the technical problems of inaccurate nutritional requirement assessment and fixed feeding timing in related technologies.

[0006] This invention provides an intelligent feed delivery method for automated rabbit farms, comprising:

[0007] Image and radio frequency data from the rabbit farm were collected to identify the female rabbits and obtain the identification results.

[0008] Based on the identification results of the female rabbit, the associated files of the female rabbit are matched to assess the current lactation performance of the female rabbit and obtain the current lactation stage label and nutritional requirement level.

[0009] Based on the current lactation stage label and nutritional requirement level, the feeding requirement is assessed to obtain the standard feeding parameters for individual does;

[0010] The target feeding amount is obtained by calculating the feeding amount and checking the inventory based on the standard feeding parameters of individual female rabbits.

[0011] Real-time monitoring of the behavior data of the mother rabbits nursing in each cage and identification of nursing behavior, analysis of nursing patterns and prediction of feeding time, to obtain the best feeding time and feeding urgency score for each cage.

[0012] Based on the optimal feeding time and feeding urgency score for each rabbit cage, the path of the feeding cart is planned and dynamically adjusted to obtain the feeding cart driving instructions.

[0013] Based on the target feeding amount and the feeding trolley's driving instructions, the feeding trolley is controlled to perform feeding operations, and abnormal situations during the execution process are diagnosed and handled to obtain feeding execution logs and optimization suggestions.

[0014] In a preferred embodiment, the step of collecting image data and radio frequency data from the rabbit farm to verify the identity of the female rabbits includes:

[0015] An RGB-D depth camera is used to simultaneously acquire color and depth images of the mother rabbit, detect and locate the mother rabbit, obtain the bounding box and key point positions of the mother rabbit, and output visual feature vectors.

[0016] An array of ultra-high frequency RFID readers is deployed in the rabbit cage area to collect signals from the electronic ear tags worn by the mother rabbits. Based on the phase difference of the signals received by multiple antennas, the three-dimensional position coordinates of the tags are estimated and the radio frequency characteristics are spliced ​​and output.

[0017] A weighted fusion strategy is adopted to integrate visual features and radio frequency features, and the results are mapped to a standardized embedding space through a pre-trained metric learning network to obtain an identity verification feature vector.

[0018] In a preferred embodiment, mapping the pre-trained metric learning network to a normalized embedding space includes:

[0019] Training is performed using the triplet loss function;

[0020] Maintain a feature library to store the reference embedding vectors of all registered female rabbits;

[0021] Cosine similarity is used as the distance metric to calculate the similarity score between the current embedded vector and all reference vectors in the feature library, and candidate matching scores are selected based on the similarity scores.

[0022] Based on the matching score, select the identity confirmation result to obtain the mother rabbit's identity confirmation result.

[0023] In a preferred embodiment, the process of matching the doe's associated profile and assessing the current doe's lactation performance includes:

[0024] Based on the confirmation of the mother rabbit's identity, the corresponding mother rabbit's file information is retrieved from the breeding management database, the integrity of the file information is verified, and the postpartum age is calculated based on the current date and the most recent birth date.

[0025] Obtain daily weight data of the rabbit litter, calculate the daily weight gain of the rabbit litter, and use an energy balance model to back-calculate the daily milk production of the mother rabbit.

[0026] The temperature data of the rabbit hutch was used to correct the estimated milk production value. The temperature correction coefficient was calculated using a piecewise linear function.

[0027] Based on the time series of continuously collected body shape feature parameters of female rabbits, a one-dimensional convolutional neural network is used to extract the temporal features of body condition changes and automatically predict the body condition score sequence.

[0028] Based on the estimated lactation volume and body condition score sequence, a time-series prediction network with a multi-task learning architecture is used to simultaneously predict changes in lactation volume and body condition score over several future days.

[0029] Based on the lactation volume prediction curve, a change point detection algorithm is used to identify the lactation stage transition point, and obtain the current lactation stage label and the stage transition prediction time.

[0030] In a preferred embodiment, the feeding requirement assessment includes:

[0031] Based on the current lactation stage label, combined with the parity and body condition score of the does, the does are classified into the preset nutritional requirement level.

[0032] Based on the predicted milk production value, the correction coefficient of milk production load on feeding amount is calculated. When the milk production of the doe exceeds the grade standard value, the feeding amount is increased to meet the additional nutrient output.

[0033] Based on the predicted body condition score, the body condition score is compared with the preset target body condition score to calculate the body condition deviation value and obtain the feeding amount compensation coefficient for body condition correction.

[0034] The weight of the doe was estimated based on the doe's body size parameters, and the correction coefficient for the feeding amount was calculated by combining the rabbit house temperature data obtained from the environmental monitoring system.

[0035] Based on the baseline feeding amount, lactation load correction factor, body condition compensation factor, and environmental correction factor, the correction factors are comprehensively calculated, and the final feeding amount is adjusted upwards using a safety factor.

[0036] In a preferred embodiment, the feeding amount calculation and inventory check include:

[0037] Based on the rabbit feed formula and standard feeding parameters, the feeding amount for the female rabbits was calculated, and the rationality of the feeding amount was verified.

[0038] Check the current inventory level of rabbit feed. If the current inventory level is lower than the preset threshold, activate the inventory shortage handling strategy, including priority allocation strategy, delayed feeding strategy and emergency replenishment strategy.

[0039] Based on the target feeding amount and rabbit feed formulation, a pre-trained palatability prediction model was used to evaluate the expected feeding effect of the feeding amount.

[0040] When the expected feed intake score is lower than the preset palatability threshold, the feeding amount is adjusted accordingly. The adjustment strategies include feeding in multiple sessions, moderately reducing the feeding amount, and adjusting the feeding timing.

[0041] In a preferred embodiment, the analysis of breastfeeding patterns and prediction of feeding timing includes:

[0042] The results of breastfeeding posture recognition, breastfeeding activity pattern recognition, and breastfeeding environment acoustic feature recognition are obtained. Multimodal fusion decision is performed using the evidence theory framework to obtain a comprehensive judgment result of the breastfeeding status.

[0043] Based on the historical lactation time data of female rabbits recorded by the system, combined with the lactation management mode information of the rabbit farm, a multi-modal adaptive modeling method is used to establish the individual lactation time pattern and obtain the predicted distribution of lactation time.

[0044] Based on the predicted distribution of lactation time under different management models, combined with the digestive physiological rhythm of the mother rabbits and the constraints of actual management operations, a utility function is used to score the candidate feeding times, and the time with the largest utility function value is selected as the optimal feeding time.

[0045] In a preferred embodiment, the planning and dynamic adjustment of the feeding trolley path includes:

[0046] Based on the optimal feeding time and feeding urgency score for each rabbit cage, a feeding time window constraint is constructed for each rabbit cage.

[0047] Based on the spatial layout data of the rabbit farm and the topology data of the track network, a matrix of travel distances between rabbit cages is constructed.

[0048] Based on time window constraints and the travel distance matrix, the material feeding path planning problem is formalized into a vehicle path problem with time window, and an adaptive large neighborhood search algorithm is used to solve it to obtain the initial feasible path;

[0049] Based on the initial feasible path, dynamic replanning is performed during actual execution according to the changes in breastfeeding management status monitored in real time. A local adjustment strategy is adopted to obtain a dynamic adjustment path.

[0050] Based on dynamically adjusted paths, driving instructions for the material feeding trolley are generated.

[0051] In a preferred embodiment, controlling the feeding trolley to perform the feeding operation and diagnosing and handling abnormal situations during the execution process includes:

[0052] Based on driving commands, control the feeding trolley to move along the planned path;

[0053] Once the feeding trolley reaches the target rabbit cage, the screw conveyor is controlled to feed the rabbits quantitatively according to the target feeding amount. The feeding control adopts a strategy that combines open-loop preset and closed-loop correction.

[0054] After feeding is completed, continuously monitor the data of the weighing sensor in the feed trough, record the change curve of feed balance over time, analyze and process the feeding curve, and calculate the first derivative of the curve to obtain the feeding rate curve.

[0055] Based on the actual feed intake, the expected feed intake output by the palatability prediction model is compared to calculate the feed intake deviation rate. When the absolute value of the feed intake deviation rate exceeds the preset normal fluctuation threshold, it is judged as an abnormal feed intake.

[0056] When abnormal feeding is detected, the knowledge graph-based abnormal cause diagnosis module is activated to perform reasoning and query, and output an abnormal diagnosis report and corresponding treatment suggestions.

[0057] Based on the anomaly diagnosis report and handling suggestions, an adjustment plan for system parameters is generated, and continuous learning and optimization are carried out through a closed-loop feedback mechanism.

[0058] In a preferred embodiment, an intelligent feed delivery system for an automated rabbit farm is used to execute the aforementioned intelligent feed delivery method for an automated rabbit farm, comprising:

[0059] The data acquisition module is used to collect image data and radio frequency data from the rabbit farm to identify the female rabbits and obtain the identification results.

[0060] The lactation assessment module matches the doe's profile with the doe's identity confirmation results to assess the doe's current lactation performance and obtain the current lactation stage label and nutritional requirement level.

[0061] The nutrition calculation module assesses the feeding requirements based on the current lactation stage label and nutritional requirement level, and obtains the standard feeding parameters for each doe.

[0062] The feeding amount calculation module calculates the feeding amount and checks the inventory based on the standard feeding amount parameters of individual female rabbits to obtain the target feeding amount;

[0063] The behavior monitoring module is used to monitor the behavior data of the mother rabbits in each cage in real time during the nursing process, identify the nursing behavior, analyze the nursing pattern and predict the feeding time, and obtain the best feeding time and feeding stress score for each cage.

[0064] The path planning module plans the path of the feeding cart and makes dynamic adjustments based on the optimal feeding time and feeding urgency score of each rabbit cage, and obtains the driving instructions of the feeding cart.

[0065] The feeding execution module controls the feeding trolley to perform feeding operations based on the target feeding amount and the feeding trolley's driving instructions. It also diagnoses and handles abnormal situations during the execution process, and obtains feeding execution logs and optimization suggestions.

[0066] The beneficial effects of this invention are as follows:

[0067] This invention achieves accurate assessment of the nutritional needs of lactating does and dynamic adjustment of individual feeding amounts through a multimodal feature fusion-based individual identification method and a multi-source data fusion-based lactation performance evaluation model. It solves the problem that traditional feeding methods with fixed feeding amounts cannot adapt to individual differences and dynamic changes in the lactation stage. This allows each does to receive nutritional supply that matches their current physiological state, effectively improving the nutritional utilization efficiency of feed and reducing feed waste and production performance loss caused by over- or under-nutrition.

[0068] This invention identifies the optimal feeding time by recognizing multimodal lactation activities and analyzing the adaptive temporal patterns of management modes. It provides corresponding feeding strategies for different management modes such as natural lactation, artificial assisted lactation, and partitioned management. Combined with a path optimization algorithm with time window constraints, it plans the driving path of the feeding cart, enabling precise feeding during the peak feeding period of the mother rabbits. This solves the problem that traditional timed feeding methods cannot adapt to individual differences in lactation rhythms and actual management constraints, effectively improving feed intake and conversion efficiency.

[0069] This invention enhances the stability and reliability of the system in complex rabbitry environments through sensor protection and self-cleaning mechanisms, environmental noise suppression technology, and other measures. At the same time, it achieves continuous optimization of the system through a closed-loop feedback mechanism, thereby improving the overall intelligence level and practical application value of breeding management. Attached Figure Description

[0070] Figure 1 This is a main flowchart of an intelligent feed delivery method for an automated rabbit farm in this invention;

[0071] Figure 2This is a detailed flowchart of an intelligent feed delivery method for an automated rabbit farm according to the present invention;

[0072] Figure 3 This is a block diagram of an intelligent feed delivery method for an automated rabbit farm according to the present invention. Detailed Implementation

[0073] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0074] At least one embodiment of the present invention discloses an intelligent feed delivery method for automated rabbit farms, such as... Figures 1 to 2 As shown, it includes the following steps:

[0075] Step 1: Collect image data and radio frequency data from the rabbit farm to verify the identity of the female rabbits and obtain the identification results.

[0076] Step 1.1, Image data acquisition and preprocessing;

[0077] Given the variable lighting conditions and frequent dust interference in the rabbit farm environment, the image data includes color and depth images of the female rabbits. An RGB-D depth camera was used to simultaneously acquire both color and corresponding depth images of the female rabbits. The color images contain appearance features such as fur color texture and spot distribution, while the depth images contain the three-dimensional morphological contour information of the female rabbits.

[0078] For color images, a contrast-limited adaptive histogram equalization method is used for enhancement processing. This method divides the image into several sub-regions and performs histogram equalization independently in each sub-region. At the same time, a contrast-limiting parameter is used to prevent excessive amplification of noise, and bilinear interpolation is used to eliminate the discontinuity of the sub-region boundaries, resulting in an enhanced color image with uniform illumination and moderate contrast.

[0079] For depth images, hole regions caused by sensor measurement errors are detected. A hole-filling algorithm based on image inpainting is adopted. The effective depth value of the hole boundary is used to iteratively fill the pixels inside the hole through the diffusion equation. Then, a bilateral filter is used for smoothing. The bilateral filter considers both the spatial distance of pixels and the difference in depth value. While smoothing noise, it maintains the sharpness of the depth edge and obtains a complete and smooth depth image.

[0080] Step 1.2, Detection and key point localization of the mother rabbit;

[0081] Based on enhanced color images, an improved single-stage object detection network is used to detect and locate female rabbits in the images. The improved single-stage object detection network is based on a lightweight feature extraction backbone network and fuses feature maps of different scales through a feature pyramid structure to achieve effective detection of female rabbits of different sizes.

[0082] The improved single-stage object detection network comprises two main components: a feature extraction layer and a head detection layer. The feature extraction layer is responsible for extracting deep feature representations containing rabbit appearance information from the input image. The head detection layer simultaneously outputs bounding box regression results and keypoint heatmaps. The bounding box is used to determine the rabbit's location range in the image, and keypoints include anatomically significant locations such as the head center, ear roots, shoulders, and rump. Keypoint localization uses heatmap regression; the network outputs a heatmap for each keypoint, and the location with the highest response value in the heatmap is the predicted coordinate of the keypoint. By performing sub-pixel-level refinement of the heatmap, a quadratic surface fitting method is used to interpolate and calculate more accurate keypoint coordinates within the neighborhood of the response peak, obtaining the bounding box and keypoint locations of the rabbit. Simultaneously, the feature extraction layer of the improved single-stage object detection network outputs a visual feature vector containing rabbit appearance information. This feature vector encodes discriminative features such as fur texture and body shape contours.

[0083] Equipped with sensor protection and self-cleaning mechanisms, including a timed air-blowing cleaning device and a dustproof and moisture-proof protective cover, it ensures long-term stable operation of the image sensor in dusty and rabbit fur environments. A sensor status monitoring mechanism is established to monitor image quality and equipment operating status in real time.

[0084] Step 1.3, 3D point cloud reconstruction and volume measurement;

[0085] Based on the complete and smooth depth image and the keypoint locations obtained in step 1.2, a depth image to point cloud conversion algorithm is used to convert the depth pixels of the mother rabbit region into 3D point cloud data. The conversion process utilizes the intrinsic parameter matrix of the depth camera to map the image coordinates and depth values ​​of each depth pixel to 3D spatial coordinates in the camera coordinate system. The point cloud is cropped based on the mother rabbit's bounding box, retaining a subset of points belonging to the mother rabbit's body. A region-growing-based point cloud segmentation algorithm is used to remove background noise. This algorithm starts from the seed points corresponding to the keypoints, determines whether they belong to the same object surface based on the Euclidean distance and the angle between the normal vectors of the points, and iteratively expands until the entire mother rabbit's body region is covered. Based on the clean point cloud after segmentation, combined with the keypoint locations, the mother rabbit's body shape feature parameters are calculated: body length is defined as the Euclidean distance from the center point of the head to the center point of the rump; body height is defined as the vertical distance from the shoulder keypoints to the ground reference plane; and body width is defined as the length of the projection of the line connecting the left and right shoulder keypoints onto the horizontal plane. By applying median filtering to the measurement results of multiple consecutive frames, random errors in single-frame measurements are eliminated, resulting in stable 3D body shape measurements.

[0086] Step 1.4, Radio frequency signal acquisition and spatial positioning;

[0087] An array of UHF RFID readers was deployed in the rabbit cage area to collect signals from the electronic ear tags worn by the female rabbits. Each reader was equipped with multiple antennas arranged in a specific geometric configuration to achieve spatial diversity. The readers sequentially activated each electronic ear tag via an anti-collision protocol, acquiring the tag's unique identification code and the signal parameters received by each antenna, including the received signal strength indicator and signal phase value. Based on the phase difference of the signals received by multiple antennas, a spatial positioning algorithm based on the phase difference of arrival was used to estimate the tag's azimuth angle relative to the antenna array. This algorithm utilizes the geometric relationship between the phase difference between adjacent antennas and the signal angle of arrival, solving a system of phase difference equations to obtain the estimated azimuth angle of the tag. Combining the positioning results from multiple readers, a weighted least squares method was used to fuse the azimuth angle estimates from each reader, calculating the tag's three-dimensional position coordinates in the rabbit farm coordinate system. The weights of the position estimates were determined based on the received signal strength of each reader; higher signal strength indicated a closer distance between the reader and the tag, resulting in higher positioning accuracy and a larger weight. Through the above processing, the unique identification code of the electronic ear tag and its estimated position in space were obtained, and the radio frequency characteristics were spliced ​​to output the RFID feature.

[0088] Step 1.5, Multimodal Identity Verification;

[0089] Based on the visual features output in step 1.2 and the radio frequency (RF) features output in step 1.4, environmental sensors are deployed within the rabbit farm to collect current environmental parameters such as light intensity, temperature, and humidity. Due to the uniqueness and stability advantages of RFID electronic ear tag identification, RF identification is used as the primary means of identity verification, while visual identification serves as an auxiliary verification method. The reliability of each sensing modality is evaluated based on environmental parameters: when the light intensity is below a preset first light threshold, the imaging quality of the visual modality decreases, correspondingly reducing the reliability coefficient of the visual features; when there are numerous metal devices or electromagnetic interference sources in the environment, the signal quality of the RF modality decreases, correspondingly reducing the reliability coefficient of the RF features. The reliability coefficient is calculated using a sigmoid function mapping, mapping the environmental parameter values ​​to the range of zero to one. Based on the calculated reliability coefficients, a weighted fusion strategy is used to integrate visual and RF features, with the weight of RF features set to 0.7 and the weight of visual features set to 0.3, reflecting the dominant role of RFID identification. The visual feature vector is output by the feature extraction layer of the detection network, while the RF feature vector is constructed by concatenating the embedded representation of the tag ID and the location coordinates. The identity verification feature vector is obtained by weighted fusion, which integrates information from both modalities.

[0090] Step 1.6, Identity Verification and Confidence Assessment;

[0091] Based on the identity verification feature vector obtained in step 1.5, a pre-trained metric learning network is used to map it to a standardized embedding space. The metric learning network is trained using a triplet loss function, ensuring that features collected from the same female rabbit at different times and under different conditions are close in the embedding space, while features from different female rabbits are far apart. The network outputs a fixed-dimensional embedding vector, which is then L2 normalized for subsequent matching. The system maintains a feature library storing reference embedding vectors for all registered female rabbits, with each rabbit corresponding to multiple reference vectors collected under different conditions to improve matching robustness. Cosine similarity is used as the distance metric to calculate the similarity score between the current embedding vector and all reference vectors in the feature library. For each candidate female rabbit, the maximum similarity of all its reference vectors is taken as the matching score for that candidate. The candidate with the highest matching score is selected as the identity verification result, and this highest score is the verification confidence score. If the highest confidence score is lower than a preset verification threshold, the individual is determined to be unknown, triggering a re-verification process or using an alternative identification scheme. Through the above processing, the female rabbit identity verification result is obtained, including the female rabbit's individual ID and verification confidence score.

[0092] Furthermore, since standard metric learning networks may confuse cases with high similarity in appearance among female rabbits, a fine-grained feature extraction network based on an attention mechanism can be used to replace the standard metric learning network. The aim is to enhance the network's ability to perceive subtle differences between individual female rabbits and improve the accuracy of distinguishing similar individuals. Specifically, this fine-grained feature extraction network introduces a multi-scale attention module on top of the backbone network. This module divides the feature map into multiple groups along the channel dimension, and each group of feature maps is processed through an independent attention branch. Each attention branch contains a global average pooling layer, a fully connected layer, and a sigmoid activation function, outputting the channel attention weights of the features in that group. Different groups use different receptive field sizes, and multi-scale feature extraction is achieved in parallel through dilated convolution. The weighted features from each group are concatenated, and then the spatial attention module further highlights the discriminative regions. The spatial attention module performs max pooling and average pooling on the feature map along the channel dimension, concatenates the two pooling results, and generates a spatial attention map through a convolutional layer. This attention map is then multiplied element-wise with the original feature map to obtain the spatially weighted features. The final embedding vector is obtained by fusing multi-scale channel attention features and spatial attention features, which can better capture fine-grained discriminative features such as the shape of the mother rabbit's ears and the spots on its fur.

[0093] Step 2: Based on the confirmation of the mother rabbit's identity, match the mother rabbit's associated files, assess the current lactation performance of the mother rabbit, and obtain the current lactation stage label and nutritional requirement level;

[0094] Step 2.1, Historical Archives Retrieval and Data Verification;

[0095] Based on the identification result of the female rabbit output in step 1, retrieve the complete file information of the female rabbit from the breeding management database. The file information includes the female rabbit's breed, date of birth, parity, mating dates, farrowing dates, litter size, number of kits per litter, number of kits at each weaning, and other reproductive records, as well as production performance data such as historical weight change curves, historical feed intake records, and historical health events. In rabbit farming, kits typically refer to young rabbits from birth to weaning, during which they mainly rely on their mother's milk for nutrition.

[0096] The retrieved data undergoes integrity verification to check for missing fields or outliers. Missing historical data is filled using the average value of the same population or breed standard values; outliers significantly exceeding reasonable ranges are corrected using interpolation from adjacent time points. Postpartum age is calculated based on the current date and the most recent farrowing date; this parameter is a key input for determining the subsequent lactation stage. Through data association processing, the mother rabbit's file is linked to the litter of kits she is currently nursing, establishing a mother-kitten relationship.

[0097] Step 2.2, reverse calculation of energy balance lactation volume;

[0098] The daily weight gain of the rabbits in their nests is calculated by deploying weighing sensors within the nests and acquiring daily weight data. Simultaneously, weighing sensors in the rabbit feed troughs monitor the amount of solid feed consumed by the baby rabbits. Before the baby rabbits begin consuming solid feed, their weight gain is entirely due to maternal milk intake; after they begin consuming solid feed, their weight gain is a combination of maternal milk and solid feed contributions.

[0099] The daily milk production of the doe was estimated using an energy balance model: The total energy intake of the kits was calculated based on their daily weight gain and maintenance metabolic requirements. Maintenance metabolic requirements were calculated using the metabolic weight formula based on the kits' average weight and ambient temperature. Energy intake from solid feed was calculated based on the amount of solid feed consumed and its metabolizable energy content. Subtracting the energy intake from solid feed from the total energy intake yielded the energy intake from the doe's milk. The doe's milk energy intake was then converted into milk volume based on the standard energy concentration of rabbit milk, resulting in an estimated daily milk production. The rabbit milk energy concentration parameter used in this estimation process adopted breed standard values, with default values ​​determined based on literature analysis of the milk composition of that breed of rabbit.

[0100] Step 2.3, Ambient temperature and milk production correction calibration;

[0101] Based on rabbit hutch temperature data obtained from the environmental monitoring system, the estimated milk production value obtained in step 2.2 is corrected for temperature. Ambient temperature indirectly affects the accuracy of milk production estimation by influencing the metabolic rate and growth efficiency of baby rabbits. When the ambient temperature is below the lower limit of the baby rabbit's thermoneutral zone, the baby rabbits need to consume extra energy to maintain their body temperature, resulting in a higher energy intake required for the same weight gain; without correction, milk production will be overestimated. When the ambient temperature is above the upper limit of the thermoneutral zone, the baby rabbits' feed intake and growth rate decrease; without correction, milk production will be underestimated. The temperature correction coefficient is calculated using a piecewise linear function. Within the thermoneutral zone, the correction coefficient is one; below the lower limit, the correction coefficient increases as the temperature decreases; above the upper limit, the correction coefficient decreases as the temperature increases. The upper and lower limits of the thermoneutral zone are dynamically adjusted according to the baby rabbits' age. Newborn baby rabbits have a narrower thermoneutral zone and higher temperature requirements; as they age, the thermoneutral zone gradually expands and shifts towards lower temperatures. Dividing the estimated milk production value from step 2.2 by the temperature correction coefficient yields the corrected estimated milk production value.

[0102] Step 2.4, Body shape sequence body condition score prediction;

[0103] Based on the time series of three-dimensional body measurements of doe rabbits collected over multiple consecutive days in step 1, a one-dimensional convolutional neural network was used to extract the temporal features of body condition changes. Body condition scores reflect the nutritional reserve status of doe rabbits and are an important indicator for assessing nutritional balance during lactation.

[0104] This system automatically predicts body condition scores by analyzing trends in body shape parameters. The one-dimensional convolutional neural network (CNN) takes as input a sequence of measurements of body length, height, and width over several days. The network contains multiple convolutional and pooling layers. The convolutional layers slide along the time dimension to extract local change patterns, while the pooling layers reduce feature dimensionality and enhance translation invariance. The output layer is a fully connected layer that outputs a sequence of body condition scores, ranging from underweight to obese. The network's training data comes from historical samples labeled with body condition scores by professionals. Supervised learning is used to establish a mapping relationship between changes in body shape parameters and body condition scores.

[0105] Step 2.5, Multi-task time-series lactation performance prediction;

[0106] Based on the estimated lactation volume obtained in steps 2.2 and 2.3, and the body condition score sequence obtained in step 2.4, a temporal prediction network using a multi-task learning architecture simultaneously predicts changes in lactation volume and body condition score over several future days, resulting in lactation volume prediction curves and body condition score prediction curves. This network employs an encoder-decoder structure. The encoder is a bidirectional long short-term memory network that processes the lactation volume sequence and body condition score sequence separately, extracting their respective temporal dependency features. The hidden states of the two encoders interact through a cross-attention mechanism, enabling lactation volume prediction to perceive trends in body condition changes, and body condition prediction to perceive changes in lactation burden. The decoder also uses a long short-term memory network structure, using the final hidden state of the encoder as the initial state to autoregressively generate prediction values ​​for future time steps. The advantage of multi-task learning lies in the inherent correlation between lactation volume and body condition; high lactation volume is usually accompanied by a decline in body condition. Joint modeling can leverage this correlation to improve the prediction accuracy of both tasks. The network outputs daily predicted lactation volume and body condition score values ​​for a preset number of future days, forming the prediction curves.

[0107] Step 2.6, detection and identification of change points during lactation;

[0108] Based on the predicted milk production curve obtained in step 2.5, a change point detection algorithm is used to identify lactation stage transition points, obtaining the current lactation stage label and the predicted transition time. The lactation process in does typically consists of three stages: the rising lactation period from postpartum to peak milk production, the peak lactation period where milk production remains at a high level, and the declining lactation period from the end of the peak period to weaning. The change point detection algorithm models the lactation curve as a piecewise linear function, where milk production changes linearly within each stage, with a sudden change in slope at the stage transition points. The algorithm uses a Bayesian inference framework to calculate the posterior probability of each possible change point location given the observed data, selecting the location with the highest posterior probability as the change point estimate. By detecting changes that have occurred in historical lactation curves, the current lactation stage of the doe is determined. By analyzing potential changes in the prediction curve, the future stage transition time is predicted, outputting the current lactation stage label and the predicted date of the next stage transition.

[0109] This step ultimately outputs the lactation performance assessment results, including the current lactation stage label, the doe's nutritional requirements level, and the predicted date of the next stage transition.

[0110] Furthermore, since standard Long Short-Term Memory (LSTM) networks may suffer from gradient vanishing when processing long-term sequences, leading to insufficient capture of long-term dependencies, a Transformer-based temporal prediction network can be used to replace LSM networks. The aim is to directly model the dependencies between arbitrary time steps through a self-attention mechanism, enhancing the learning ability of the long-term variation patterns of the lactation curve. Specifically, this Transformer temporal prediction network performs positional encoding on the input lactation volume sequence and body condition score sequence, embedding time step information into the feature representation. The encoder consists of multiple stacked self-attention modules, each layer containing a multi-head self-attention sublayer and a feedforward neural network sublayer. The self-attention sublayer calculates the attention weights of each time step in the sequence with all other time steps, achieving global information aggregation. To adapt to the temporal prediction task, a causal mask is introduced in the self-attention calculation to ensure that each time step only focuses on its preceding time steps, preventing information leakage. The decoder adopts a similar structure, adding an additional encoder-decoder attention sublayer, enabling the decoder to focus on the output features of the encoder. By stacking multiple attention modules, the network can learn complex periodic patterns and long-term trends in the lactation curve, improving prediction accuracy.

[0111] Step 3: Based on the current lactation stage label and nutritional requirement level, conduct a feeding requirement assessment to obtain the standard feeding parameters for individual does;

[0112] Step 3.1, Configure the baseline values ​​for nutritional requirement levels;

[0113] Based on the current lactation stage label output in step 2, combined with the doe's parity and body condition score, the doe is categorized into a preset nutritional requirement level, yielding the corresponding standard feeding baseline value. The nutritional requirement level rules are as follows:

[0114] Grade A consists of primiparous does (first parity) that are in the rising lactation stage and require a higher nutritional supply to support mammary gland development.

[0115] Grade B consists of multiparous does (≥2 parities) that are at their peak lactation period, with their nutritional needs at their peak levels.

[0116] Grade C refers to does of any parity that are in a period of declining lactation and require a balance between maintaining lactation and restoring body condition.

[0117] Grade D refers to female rabbits with a body condition score of less than 4 (regardless of the lactation stage) that require additional nutritional supplementation.

[0118] Grade E refers to female rabbits with a body condition score of ≥4 who are in the maintenance phase (non-lactation period or with extremely low milk production).

[0119] Grading is conducted according to priority: first priority is assessing body condition, second priority is determining lactation stage, and third priority is considering parity to determine the final grade. Each nutritional requirement grade corresponds to a pre-set baseline feeding amount, which is determined based on the average weight and standard nutritional requirements of does at that grade. The specific values ​​for the grade standards are determined based on breed nutritional standards and breeding experience, with default values ​​set according to industry-recommended breeding rabbit feeding standards.

[0120] Step 3.2, Adjustment of feeding amount for lactation load;

[0121] Based on the predicted milk production output from step 2.5, the correction coefficient for feeding amount based on milk production load is calculated to obtain the feeding amount after correction. When the milk production of the doe is higher than the standard value, the feeding amount needs to be increased to meet the additional nutritional output; when the milk production is lower than the standard value, the feeding amount can be appropriately reduced to avoid nutritional excess. The calculation method for the milk production load correction coefficient is as follows: calculate the difference between the actual milk production and the standard milk production, then divide the difference by the standard milk production to obtain the relative deviation, then multiply the relative deviation by the adjustment factor, and finally add the baseline value to obtain the correction coefficient. The adjustment factor is set according to the doe's parity and body condition. Primiparous doe and doe in poor body condition have a larger adjustment factor, requiring more nutritional reserves; multiparous doe and doe in good body condition have a smaller adjustment factor, with higher nutritional utilization efficiency. The corrected feeding amount considering the difference in milk production load is obtained through the above calculation.

[0122] Step 3.3, Body condition correction feeding amount compensation;

[0123] Based on the predicted body condition score output in step 2.5, it is compared with the preset target body condition score to calculate the body condition deviation value, thus obtaining the feeding compensation coefficient for body condition correction. The target body condition score is dynamically set according to the lactation stage: during the lactation rise and peak periods, a moderate decline in body condition is allowed to support lactation, and the target body condition can be slightly lower than the body condition at mating; during the lactation decline period, body condition needs to be gradually restored, and the target body condition should be close to or reach the standard body condition at mating. The calculation of the body condition compensation coefficient adopts a simplified proportional adjustment strategy: calculate the difference between the target body condition score and the current body condition score, multiply the difference by the body condition adjustment factor, and add the benchmark value to obtain the compensation coefficient. The body condition adjustment factor is determined according to the lactation stage; the adjustment factor is larger during the lactation decline period to promote body condition recovery, and smaller during the lactation peak period to avoid overcompensation affecting lactation. By adjusting the feeding amount through the body condition compensation coefficient, a balance between body condition recovery and lactation maintenance is achieved.

[0124] Step 3.4, Environmental correction and feeding amount adjustment;

[0125] Based on the body shape parameters of the doe obtained in step 1, the doe's weight is estimated. Combined with the rabbitry temperature data obtained from the environmental monitoring system, a correction coefficient for environmental factors on the feeding amount is calculated, resulting in the adjusted feeding amount after environmental correction. Environmental temperature has a certain impact on the doe's feed intake and nutritional requirements: when the environmental temperature is below the lower limit of the thermoneutral zone, the doe needs to consume extra energy to maintain body temperature, and the feeding amount should be increased; when the environmental temperature is above the upper limit of the thermoneutral zone, the doe's feed intake decreases, and the feeding amount should be appropriately reduced while increasing the nutrient concentration of the feed. The environmental correction coefficient is calculated using a piecewise linear function: the correction coefficient is 1 within the thermoneutral zone; below the lower limit, the correction coefficient increases as the temperature decreases; above the upper limit, the correction coefficient decreases as the temperature increases. The upper and lower limits of the thermoneutral zone are dynamically adjusted according to the doe's weight and lactation stage; does with larger weight and higher milk production have relatively wider thermoneutral zones. Multiplying the baseline feeding amount by the environmental correction coefficient yields the adjusted feeding amount adapted to the current environmental conditions.

[0126] Step 3.5: Determine the standard feeding parameters;

[0127] Based on the baseline feeding amount determined in step 3.1, the lactation load correction coefficient calculated in step 3.2, the body condition compensation coefficient calculated in step 3.3, and the environmental correction coefficient calculated in step 3.4, the final standard feeding amount is obtained by comprehensively calculating each correction coefficient. The comprehensive correction coefficient is obtained by multiplying the lactation load correction coefficient, the body condition compensation coefficient, and the environmental correction coefficient. The final feeding amount is obtained by multiplying the baseline feeding amount by the comprehensive correction coefficient. Considering various uncertainties in actual breeding processes, including individual differences, environmental fluctuations, and deviations in feed nutrient composition, a safety factor is applied upwards to the final feeding amount. The value of the safety factor is determined based on the importance of the does and the stability of feed supply: for high-producing does and key breeding does, the safety factor is relatively large to ensure sufficient nutrient supply; for general does, the safety factor is relatively small to avoid waste caused by overfeeding. The default value of the safety factor is set to 1.05 to 1.15, which can be adjusted according to the actual situation of the rabbit farm. The above processing yields the standard feeding parameters for individual female rabbits, including the baseline feeding amount, the comprehensive correction coefficient, and the final recommended feeding amount.

[0128] Furthermore, since a fixed safety factor cannot adapt to the risk differences among different individuals and under different environmental conditions, a dynamic safety factor calculation method based on uncertainty quantification can be used instead of a fixed safety factor. The aim is to dynamically adjust the safety margin based on the confidence level of the prediction model and environmental stability, thereby ensuring sufficient feeding while reducing unnecessary overfeeding. Specifically, this dynamic safety factor calculation method assesses the uncertainty of milk production prediction in step 2. By introducing a dropout layer into the prediction network and performing multiple forward propagations, the mean and variance of the prediction results are calculated, with the variance reflecting the degree of uncertainty in the prediction. Simultaneously, the stability of environmental parameters is assessed by calculating the coefficients of variation of parameters such as temperature and humidity over the past several days. Using prediction uncertainty and environmental stability as inputs, a pre-trained safety factor mapping network outputs a dynamic safety factor. The training data for this mapping network comes from historical aquaculture records, using actual cases of underfeeding or overfeeding as labels, to learn the optimal safety factor values ​​under different uncertainty levels. By dynamically adjusting the safety factor, the safety margin is increased under high uncertainty conditions, and overfeeding is reduced under low uncertainty conditions.

[0129] Step 4: Calculate the feeding amount and check the inventory based on the standard feeding parameters for individual female rabbits to obtain the target feeding amount;

[0130] Step 4.1, accurately calculate the feeding amount;

[0131] Based on a rabbit feed formula designed and validated by nutrition experts, and combining this formula with the standard feeding parameters output in step 3.5, the precise feeding amount for the doe is calculated to obtain the target feeding amount. The feeding amount calculation uses a direct mapping method: the target feeding amount equals the final recommended feeding amount output in step 3.5. Simultaneously, a feeding amount rationality check is performed to ensure the calculation result is within an acceptable range: It is checked whether the feeding amount exceeds the doe's maximum feed intake capacity, which is estimated based on the doe's weight and physiological state, generally 3% to 5% of body weight; it is checked whether the feeding amount is lower than the minimum required for maintenance, which is 1.5% to 2% of body weight; if the calculated feeding amount exceeds the reasonable range, boundary constraint adjustments are made to limit the feeding amount between the minimum and maximum values. Through the above processing, a rationally validated target feeding amount is obtained.

[0132] Step 4.2, Check feed inventory adequacy;

[0133] Check the current rabbit feed inventory to ensure it can meet the feeding needs of all does. Calculate the total feeding requirement for all does in the current batch and compare it with the inventory. If the inventory is sufficient, use the target feeding amount for each does; if the inventory is insufficient, activate the inventory shortage handling strategy. The handling strategies include: a priority allocation strategy, prioritizing the feeding needs of high-producing does and does in poor condition, and appropriately reducing the feeding amount for other does; a delayed feeding strategy, recording the feeding deficit when the inventory is severely insufficient and scheduling it for priority feeding after the next batch is replenished; and an emergency replenishment strategy, triggering emergency feed procurement or remixing. During the adjustment process, it is necessary to ensure that the basic nutritional needs of the does are met, prioritizing maintenance needs and basic lactation needs. Through the above processing, a practically executable feeding plan considering inventory constraints is obtained.

[0134] Step 4.3, verify the palatability of the feeding amount;

[0135] Based on the target feeding amount and rabbit feed formula obtained in step 4.1, a pre-trained palatability prediction model is used to evaluate the expected feed intake effect under this feeding amount, obtaining the expected feed intake rate score. Feed palatability directly affects the feed intake of the does, thus affecting the actual nutrient intake. The palatability prediction model takes the feeding amount and individual characteristics of the does as input, and outputs the expected feed intake rate as a score, defined as the ratio of actual feed intake to the feeding amount. The model is constructed using a gradient boosting decision tree algorithm, and the training data comes from historical actual feeding records of various feeding amount combinations, including feeding amount, does's weight, lactation stage, and nutritional requirement level. The model can learn feeding patterns under different feeding amounts; for example, the feed intake rate decreases when overfeeding, and is higher when feeding moderately. If the expected feed intake rate score is lower than the preset palatability threshold (usually set to 0.85), the feeding amount needs to be adjusted.

[0136] Step 4.4, optimize and adjust the feeding amount;

[0137] When the expected feed intake score in step 4.3 is lower than the palatability threshold, the feeding amount is adjusted accordingly to improve feeding effectiveness. The palatability threshold is usually set at 0.85. An expected feed intake lower than the threshold indicates a potential risk of insufficient feed intake, requiring adjustment of the feeding strategy; a feed intake higher than the threshold indicates acceptable palatability, and the target feeding amount can be used directly.

[0138] The specific strategies for adjusting the feeding amount are as follows: A staggered feeding strategy, dividing the target feeding amount into 2 to 3 feedings, reducing the amount of food each time but increasing the frequency to improve the completion rate of each feeding; a moderate reduction in feeding amount strategy, appropriately reducing the feeding amount while ensuring basic nutritional needs are met, avoiding decreased feed intake and feed waste due to overfeeding; and a feeding timing adjustment strategy, adjusting the feeding time according to the does's feeding habits, selecting periods when the does are more actively eating. After adjustments, the expected feed intake rate is recalculated. If the feed intake rate increases above the threshold, the final feeding plan is determined; if it is still below the threshold, the plan with the highest feed intake rate is selected as a compromise, and the final target feeding amount is output.

[0139] Furthermore, since palatability prediction models trained on historical data may not accurately predict the feeding effects of new feeding combinations, an online palatability model update mechanism based on active learning can be used to replace the static prediction model. The aim is to continuously collect actual feeding feedback from new feeding schemes, dynamically update model parameters, and improve the accuracy and timeliness of predictions. Specifically, after each new feeding scheme is generated, the active learning mechanism marks it as a sample to be validated. When the scheme is actually fed, the system collects the actual feed intake data of the does and calculates the deviation between the actual feed intake rate and the predicted feed intake rate. If the deviation exceeds a preset threshold, it indicates that the sample has high information value, and it is added to the training set, triggering an incremental model update. The model update uses an online learning algorithm, incorporating information from new samples while retaining existing knowledge, avoiding the computational overhead of full retraining. Simultaneously, the system maintains an uncertainty sample pool, storing feeding schemes with low model prediction confidence, prioritizing the actual validation of these schemes to accelerate the model's learning in uncertain regions. Through the active learning mechanism, the palatability prediction model can continuously adapt to the influence of factors such as changes in raw material batches and seasonal variations.

[0140] Step 5: Monitor the behavior data of the mother rabbits nursing in each cage in real time and identify nursing behavior, analyze nursing patterns and predict feeding time, and obtain the best feeding time and feeding urgency score for each cage.

[0141] Step 5.1, Infrared thermal imaging breastfeeding posture recognition:

[0142] Infrared thermal imaging sensors were deployed above the rabbit cages to collect thermal image sequences of the cage area. Thermal region detection and posture analysis algorithms were used to obtain nursing posture recognition results. Infrared thermal imaging can operate in low-light or even complete darkness, overcoming the limitations of visible light vision in nighttime monitoring. When nursing, the mother rabbit enters the birthing box to contact her kits, and the kits gather around the mother's abdomen to suckle, forming a characteristic thermal distribution pattern. The thermal images were preprocessed, including temperature calibration, background subtraction, and noise filtering, to obtain thermal images highlighting the foreground target. A temperature gradient-based region segmentation algorithm was used to detect high-temperature regions in the thermal images. The mother rabbit's body temperature is higher than the ambient temperature, and the area where the kits gather exhibits an even higher temperature due to the superposition of body heat. The shape characteristics and spatial distribution of high-temperature regions were analyzed. When a large area of ​​high temperature was detected where the mother rabbit was in a prone position and multiple small high-temperature clusters were present on the abdomen, it was determined to be a nursing posture. The posture recognition results include whether the rabbit is in a nursing state and a confidence score for the nursing posture.

[0143] Step 5.2, vibration sensing breastfeeding activity pattern recognition;

[0144] A triaxial accelerometer was installed on the bottom plate of the rabbit cage to collect vibration data. Vibration pattern feature extraction and classification algorithms were used to obtain the results of lactation activity pattern recognition. The mother rabbit exhibits characteristic activity patterns when entering the whelping box to nurse: at the beginning of lactation, the mother rabbit enters the whelping box, causing positional movement vibrations; during lactation, the mother rabbit remains relatively still while the kits gather and move, producing continuous slight vibrations; after lactation ends, the mother rabbit leaves the whelping box, producing obvious positional change vibrations. The acceleration data was preprocessed, including removing DC components, bandpass filtering, and frame segmentation. Temporal and frequency domain features were extracted from each frame: temporal features included mean, variance, peak value, and zero-crossing rate; frequency domain features were obtained through Fast Fourier Transform, including dominant frequency, spectral energy distribution, and spectral entropy. The extracted feature vectors were input into a pre-trained support vector machine classifier, which outputs the behavior category at the current moment, including stillness, normal activity, and lactation activity. The characteristic of lactation activity is the presence of periodic vibration components within a specific frequency range, which corresponds to the overall activity pattern frequency of the mother rabbit during lactation. The classification results are fused with the thermal imaging recognition results from step 5.1 to improve the accuracy of breastfeeding behavior recognition.

[0145] Step 5.3, Acoustic feature identification of the breastfeeding environment using audio sensing;

[0146] Directional audio sensors were deployed near the rabbit cages to collect environmental sound data. Spectrum analysis and acoustic environment change detection algorithms were used to obtain the acoustic feature recognition results of the nursing environment. When the mother rabbit enters the birthing box to nurse, characteristic environmental acoustic changes occur, including the sound of the mother rabbit moving, the sound of the kits gathering, and the overall pattern of changes in environmental acoustic activity.

[0147] To address the complex background noise environment within the rabbit hutch, including fan noise, rabbit activity noise, and equipment operation noise, a multi-level adaptive noise suppression technique is employed. The audio data undergoes preprocessing, including pre-emphasis, framing, windowing, and Fast Fourier Transform, to obtain a short-time spectral representation. Spectral subtraction is used to suppress steady-state background noise, while Wiener filtering is employed to suppress non-steady-state interference noise.

[0148] Environmental acoustic features are extracted, including sound pressure level change rate, spectral centroid shift, and acoustic activity index—parameters reflecting changes in the acoustic environment during breastfeeding. An acoustic environment change detection model based on a long short-term memory network is used to analyze the audio stream in real time, and the model outputs the probability of occurrence of breastfeeding environment acoustic patterns. When the probability of a breastfeeding environment acoustic feature exceeds a preset threshold, a breastfeeding-related acoustic environment change is determined to have been detected.

[0149] Meanwhile, the system is equipped with sensor protection and self-cleaning mechanisms, including a timed air-blowing cleaning device and a dustproof and moisture-proof protective cover, ensuring long-term stable operation of the audio sensor in dusty and rabbit fur environments. A sensor status monitoring mechanism is established to monitor signal quality and equipment operating status in real time.

[0150] Step 5.4, Multimodal evidence fusion decision;

[0151] Based on the thermal imaging recognition results of step 5.1, the breastfeeding activity pattern recognition results of step 5.2, and the acoustic feature recognition results of the breastfeeding environment of step 5.3, an evidence theory framework is used for fusion decision-making to obtain a comprehensive judgment result of the breastfeeding state. Evidence theory can handle uncertainty and conflicting information and is suitable for multi-sensor data fusion scenarios. The recognition results of each modality are converted into basic probability assignment functions, which are defined on the hypothesis space of the breastfeeding state, including three hypotheses: breastfeeding, non-breastfeeding, and uncertainty. The value of the basic probability assignment is determined according to the confidence score of each modality recognition; a higher confidence score corresponds to a larger probability assignment value for the hypothesis. The Dempster combination rule is used to fuse the basic probability assignments of the three modalities. This rule calculates the intersection of the evidence of each modality and normalizes to handle conflicts, resulting in the fused basic probability assignment. The confidence and likelihood of the breastfeeding hypothesis are extracted from the fusion result. The confidence represents the minimum strength of evidence supporting the breastfeeding hypothesis, and the likelihood represents the maximum strength of evidence not opposing the breastfeeding hypothesis. When the confidence level exceeds the preset judgment threshold, the mother rabbit is confirmed to be in a nursing state; when the likelihood level is lower than the threshold, the mother rabbit is confirmed to be in a non-nursing state; when it is in between, the rabbit remains in an uncertain state and awaits more evidence.

[0152] Step 5.5, Modeling the Adaptability of Breastfeeding Management Models;

[0153] Based on the historical lactation time data of female rabbits recorded by the system, combined with the lactation management mode information of the rabbit farm, a multi-modal adaptive modeling method is used to establish the individual lactation time pattern and obtain the predicted distribution of lactation time.

[0154] The system first identifies the nursing management mode type for each mother rabbit:

[0155] Natural lactation pattern: Multiparous and maternally healthy female rabbits are able to complete lactation behavior autonomously. The Gaussian process regression method is used to model the natural lactation time pattern. The kernel function adopts a combination of periodic kernel and radial basis kernel to capture daily periodicity and time correlation.

[0156] Artificial Assisted Lactation Mode: Primiparous does or does with poor maternal instincts require artificial intervention to assist lactation. The system is integrated with the rabbit farm management system to obtain artificial lactation plans and execution records, and to establish a time prediction model based on the artificial management plan. This model uses the management plan as the main input and is corrected by combining historical execution deviations.

[0157] Partition Management Mode: Rabbit cages that separate mother and baby rabbits are used. The contact time between the mother rabbit and the baby rabbit is controlled by the partition. The system is linked with the partition control system to record the opening and closing time of the partition and establish a nursing time prediction model based on the partition operation cycle.

[0158] Management pattern recognition employs a decision tree algorithm. Input features include the doe's parity, historical lactation success rate, frequency of human intervention, and partition operation records. For does whose management patterns change, a change point detection algorithm is used to identify the timing of the pattern switch and update the corresponding prediction model in a timely manner. The prediction models for each pattern output the probability distribution of future lactation times, providing input for optimizing subsequent feeding timing.

[0159] Step 5.6, Optimize decision on the best feeding time in multiple modes;

[0160] Based on the predicted distribution of lactation time under different management modes obtained in step 5.5, combined with the digestive physiological rhythm of the mother rabbits and the constraints of actual management operations, the feeding time is calculated to obtain the optimal feeding time and feeding stress score for each rabbit cage.

[0161] Optimize feeding timing strategies for different breastfeeding management models:

[0162] In the natural breastfeeding mode, the optimal feeding time is set at the peak feeding period after the predicted end time of breastfeeding. The starting point of the time window is determined by the physiological recovery time after breastfeeding, and the ending point is determined by the duration of the peak feeding period.

[0163] In artificial feeding mode, the timing of feeding needs to be coordinated with the artificial feeding plan to avoid interference caused by feeding during artificial operation. At the same time, the stress recovery time of the mother rabbit after artificial feeding should be considered, and the feeding time should be appropriately delayed.

[0164] Under the partition management model, feeding should be scheduled during the period when the partition is closed to avoid feeding during the nursing period when the partition is open, thereby reducing interference with the mother rabbit's nursing behavior. At the same time, the impact of the partition operation cycle on the mother rabbit's feeding rhythm should be considered.

[0165] A utility function is used to score candidate feeding times, taking into account four factors: the timeliness of the feeding time matching peak feeding times, the feasibility of the feeding system scheduling, the risks arising from forecast uncertainty, and the coordination of management operations. The coordination score is determined based on the degree of conflict between the feeding time and the manual operation plan and the partition operation plan. The system is synchronized in real time with the rabbit farm management information system to obtain the latest management operation arrangements, ensuring that the feeding plan is coordinated with actual management operations. The time with the highest utility function value is selected as the optimal feeding time. Simultaneously, a feeding urgency score is calculated, reflecting the urgency of the current time relative to the optimal feeding time, which is used for priority ranking in subsequent path planning.

[0166] Furthermore, since adaptive modeling of lactation management patterns based on the historical data of a single doe has low prediction accuracy when data is insufficient, a group-based lactation pattern modeling method based on transfer learning can be used instead of single-individual modeling. The aim is to utilize the lactation data of other does in the group to assist the learning of patterns in new individuals or individuals with sparse data, thereby improving prediction stability. Specifically, this transfer learning method trains a basic lactation pattern model at the group level. This model takes the doe's feature vector as conditional input, including parity, number of litters, postpartum age, body condition score, and management pattern type, and outputs the lactation time distribution parameters under those conditions. The basic model learns the general relationship between lactation patterns and individual characteristics from the historical data of all does. For the prediction of a specific doe, the prior distribution is obtained from the basic model based on its feature vector. The prior distribution is then updated using the doe's own historical lactation data to obtain the posterior distribution as the final prediction. When individual historical data is abundant, the posterior distribution is mainly determined by the individual data; when individual historical data is sparse, the posterior distribution relies more on the group prior, achieving effective knowledge transfer.

[0167] Step 6: Based on the optimal feeding time and feeding urgency score for each rabbit cage, plan the path of the feeding cart and make dynamic adjustments to obtain the driving instructions for the feeding cart.

[0168] Step 6.1, Constructing the feeding time window constraint;

[0169] Based on the optimal feeding time and feeding urgency score output from step 5 for each rabbit cage, a feeding time window constraint is constructed for each cage, resulting in the feeding time window constraint for each cage. The time window defines the allowed time range for the feeding cart to access the cage, including the earliest and latest arrival times. The earliest arrival time is set as the optimal feeding time minus an advance allowance. The advance allowance is determined based on the physiological recovery time of the doe after lactation; feeding too early may reduce feed intake because the doe has not yet entered a feeding state. The latest arrival time is set as the optimal feeding time plus a delay tolerance. The delay tolerance is determined based on the duration of the feeding peak and the feeding urgency score. A higher urgency score indicates a more urgent feeding situation for the cage, and a smaller delay tolerance. For cages with urgency scores exceeding a preset urgency threshold, they are marked as having a hard time window constraint, requiring feeding to be completed within the time window. For cages with lower urgency scores, they are marked as having a soft time window constraint, allowing violation of the time window when necessary but incurring penalty costs. Through the above processing, the feeding requirements of each rabbit cage are converted into access constraints with time windows.

[0170] Step 6.2, construct the orbital network distance matrix;

[0171] Based on the spatial layout data of the rabbit farm and the topology data of the track network, a distance matrix between rabbit cages is constructed. The track network consists of track segments and intersection nodes. The feeding trolley can only travel along the tracks, and different travel directions can be selected at intersection nodes. The track network is modeled as a directed graph, where the nodes include the positions of each rabbit cage and the track intersections. The edges of the graph represent the track segments between adjacent nodes, and the weight of the edge is the length of the track segment. For unidirectional track segments, only unidirectional edges are added; for bidirectional track segments, both bidirectional edges are added. The Floyd-Warshall algorithm is used to pre-calculate the shortest path distance between any two nodes to obtain the complete distance matrix. This pre-calculation process is performed once during system initialization, and subsequent path planning directly looks up the distance values ​​to avoid redundant calculations. The distance matrix also records the specific sequence of nodes traversed by the shortest path, which is used to generate detailed driving instructions later.

[0172] Step 6.3, Solving the vehicle routing problem within the time window;

[0173] Based on the time window constraints constructed in step 6.1 and the travel distance matrix calculated in step 6.2, the feeding path planning problem is formalized as a vehicle path problem with time windows and solved using an adaptive large neighborhood search algorithm to obtain an initial feasible path. The goal of this problem is to find a path that starts from the starting point, visits all cages to be fed, and finally returns to the starting point, minimizing the total travel distance while satisfying the time window constraints of each cage. The adaptive large neighborhood search algorithm is a metaheuristic algorithm that explores the solution space by iteratively breaking and repairing the current solution. The algorithm maintains a set of breaking and repair operators. Breaking operators remove visits to some cages from the current solution, and repair operators re-insert the removed cages into the path. Breaking operators include random removal, worst-case removal, and relevant removal, while repair operators include greedy insertion and regret insertion. The algorithm adaptively adjusts the selection probability of each operator based on its performance in historical iterations, with operators that perform well being selected more frequently. After each iteration, if the new solution is better than the current optimal solution, the optimal solution is updated; if the new solution satisfies the acceptance criterion of simulated annealing, the new solution is accepted as the starting point for the next iteration. Iteration continues until the termination condition is met, and the optimal path found during the search process is output as the initial feasible path.

[0174] Step 6.4, Dynamic path replanning adjustment;

[0175] Based on the initial feasible path obtained in step 6.3, dynamic replanning is performed during actual execution based on real-time monitoring of changes in lactation management status, resulting in a dynamically adjusted path. Due to the uncertainty in lactation time prediction, the actual lactation time may deviate from the predicted time. Simultaneously, the artificial lactation plan and partition operation plan may be temporarily adjusted, causing changes in the time window constraints of the original planned path. The system continuously monitors the lactation status and management operation status of each rabbit cage. When it detects that the actual lactation end time of a mother rabbit in a cage deviates from the predicted time by more than a threshold, or when the management operation plan changes, dynamic replanning is triggered. The replanning adopts a local adjustment strategy, optimizing only the affected path segments while keeping other path segments unchanged to reduce computational overhead and execution interruptions. Local adjustment methods include: adjusting the access order of affected rabbit cages in the path, adjusting the waiting time of the feeding cart at certain nodes, and skipping some low-priority rabbit cages for later supplementary feeding when necessary. The replanning algorithm uses a fast insertion heuristic, seeking the solution with the minimum adjustment cost while ensuring the time window constraints of high-priority rabbit cages.

[0176] Step 6.5: Detailed driving instructions are generated;

[0177] Based on the dynamically adjusted path obtained in step 6.4, a feeding trolley driving instruction is generated. This instruction includes the visit order of each node on the path, the driving direction selection at each node, the dwell time at each rabbit cage, and a complete feeding schedule. The complete feeding schedule includes the planned arrival time, planned feeding amount, and planned dwell time for each rabbit cage. The feeding trolley carries rabbit feed and delivers it precisely according to the target feeding amount.

[0178] The planned arrival time is calculated based on the kinematic model of the feeding trolley, considering the time consumption of acceleration, constant speed travel, deceleration, and stopping. For each cage on the path, the cumulative time from the starting point to the planned arrival time is calculated and added to the starting time. A buffer time allocation mechanism is introduced into the time estimation, distributing the total time margin of the path according to the urgency score of each cage. Cages with high urgency are allocated less buffer to ensure on-time arrival, while cages with low urgency are allocated more buffer to absorb the time fluctuations of previous execution.

[0179] Furthermore, since a single feeding cart may not be able to complete all feeding tasks within all time windows in a large-scale rabbit farm, a multi-car cooperative scheduling algorithm can be used to replace single-car path planning. The aim is to improve system throughput through the parallel operation of multiple feeding carts and ensure that the time window constraints of all rabbit cages are met. Specifically, this multi-car cooperative scheduling algorithm divides the feeding task into several sub-regions based on the spatial layout and time window distribution of the rabbit farm, with one feeding cart assigned to each sub-region. The region division uses a clustering algorithm, clustering based on the spatial location of the rabbit cages and the center of their time windows, ensuring that rabbit cages within the same region are spatially close and their time windows are compatible. For rabbit cages with conflicting time windows at the region boundaries, an auction mechanism is used for allocation. Each cart bids based on the marginal cost of accessing the cage, and the cart with the lowest cost obtains the service right for that cage. Each cart independently plans its path within its respective region, while a central scheduler coordinates the driving order on the shared track segment to avoid collisions and deadlocks. The scheduler adopts a priority-based conflict resolution strategy. When two vehicles need to use the same track segment at the same time, the vehicle with higher priority goes first. The priority is determined according to the urgency of the vehicle's current task.

[0180] Step 7: Based on the target feeding amount and the feeding trolley's driving instructions, control the feeding trolley to perform feeding operations, and diagnose and handle any abnormal situations during the execution process to obtain feeding execution logs and optimization suggestions;

[0181] Step 7.1, Track tracking control of the feeding trolley;

[0182] The movement of the feeding trolley is controlled based on path planning results, using a trajectory tracking control algorithm.

[0183] Based on the driving command output in step 6, the feeding trolley is controlled to move along the planned path, obtaining the trolley's real-time position and motion status. The feeding trolley is equipped with a drive motor, encoder, and position sensor, enabling precise position control. The trajectory tracking control employs a proportional-integral-derivative (PID) control algorithm. The controller input is the deviation between the target position and the actual position, and the output is the control signal for the drive motor. The proportional term generates an instantaneous correction based on the current position deviation, the integral term eliminates steady-state error, and the derivative term suppresses overshoot and oscillation. The controller parameters are tuned based on the feeding trolley's dynamic characteristics and track conditions. The tuning method uses an empirical formula based on step response, and the default parameter values ​​are determined based on the optimal response under typical operating conditions. During operation, the trolley continuously reports its real-time position and motion status, including current coordinates, speed, acceleration, and motor current. The system monitors the trolley's operating status based on the reported data. When a position deviation exceeds a threshold or an abnormal motor current is detected, a protection mechanism is triggered to suspend operation and issue an alarm.

[0184] Step 7.2, Precise feeding control of the screw conveyor;

[0185] Once the feeding trolley reaches the target rabbit cage, the screw conveyor mechanism is controlled to quantitatively feed the rabbits according to the target feeding amount output in step 4. The screw conveyor pushes the rabbit feed from the hopper to the discharge port through the rotation of the screw; the feeding amount is proportional to the number of screw rotations. The feeding control employs a strategy combining open-loop preset and closed-loop calibration: the preset number of rotations is calculated based on the target feeding amount and the screw's single-rotation conveying capacity, and the screw is controlled to rotate the preset number of rotations to complete the main feeding; the actual feeding amount is measured by the trough weighing sensor, and the deviation from the target amount is calculated; if the deviation exceeds the allowable range, supplementary feeding is performed or underfeeding is recorded. The screw's single-rotation conveying capacity is affected by feed density and flowability, and the system periodically calibrates and updates this parameter. During the feeding process, process data such as screw speed, feeding time, and weighing sensor reading changes are recorded for subsequent analysis and system optimization.

[0186] Step 7.3: Continuous monitoring and analysis of the foraging process;

[0187] After feeding, the system continuously monitors the data from the weighing sensor in the feed trough, recording the change curve of the remaining feed over time, i.e., the feeding curve. The feeding curve reflects the feeding behavior pattern of the doe, including information such as feeding speed, feeding rhythm, and total feed intake. The feeding curve is analyzed and processed: smoothing filters are used to remove weighing noise, and a moving average filter is used to smooth the raw data; the first derivative of the curve is calculated to obtain the feeding rate curve, identifying the peak time and duration of the feeding rate; the total feed intake within a preset time period after feeding is calculated, which is equal to the initial feed amount minus the remaining feed in the trough at the end of that period. Feeding behavior features are extracted from the feeding curve, including feeding initiation delay time, duration of peak feeding, peak feeding rate, and number of feeding intervals. These features are used to assess the rationality of feeding timing and feed palatability, and also serve as input data for subsequent nutritional requirement assessments.

[0188] Step 7.4, Detection and judgment of abnormal feeding deviations;

[0189] Based on the actual feed intake obtained in step 7.3, the expected feed intake output by the palatability prediction model in step 4 is compared to calculate the feed intake deviation rate. The feed intake deviation rate is defined as the difference between the actual feed intake and the expected feed intake divided by the expected feed intake. When the absolute value of the feed intake deviation rate exceeds the preset normal fluctuation threshold, it is judged as an abnormal feed intake. Abnormal feed intake is divided into two categories: insufficient feed intake and excessive feed intake. Insufficient feed intake is manifested as actual feed intake being lower than expected, and possible causes include feed palatability issues, doe health issues, environmental stress, etc. Excessive feed intake is manifested as actual feed intake being higher than expected, and possible causes include compensatory feeding due to insufficient feeding in the previous period, low nutritional requirement assessment, etc. For detected abnormal feed intake, the system records the abnormality type, abnormality degree, and occurrence time, and triggers subsequent abnormality diagnosis processes. At the same time, the feed intake deviation data is fed back to the palatability prediction model for online updating and calibration of the model.

[0190] Step 7.5, Diagnosis of anomalies in the knowledge graph;

[0191] When abnormal feeding is detected in step 7.4, the knowledge graph-based abnormal cause diagnosis module is activated to obtain an abnormal diagnosis report and treatment suggestions. The knowledge graph integrates expert knowledge in the field of rabbit breeding and includes three types of elements: entities, relationships, and attributes. Entities include diseases, symptoms, nutrients, feed ingredients, and environmental factors; relationships include causes, mitigation, inclusion, and impact; attributes include severity, probability of occurrence, and duration. The diagnosis process collects contextual information related to the abnormal rabbit, including recent changes in weight, body condition score, and environmental parameters; it performs reasoning queries in the knowledge graph, starting from the observed fact of abnormal feeding and tracing back along the causal relationship chain to find possible cause nodes. The reasoning engine uses a combination of rule-based forward chain reasoning and embedding-based similarity retrieval: rule-based reasoning performs logical deduction based on predefined diagnostic rules, while embedding retrieval searches for similar historical cases in the knowledge graph embedding space based on the current symptom vector. The results of both methods are combined to output an abnormal diagnosis report and corresponding treatment suggestions.

[0192] Step 7.6: System parameter feedback optimization and adjustment;

[0193] Based on the anomaly diagnosis report and handling suggestions output in step 7.5, an adjustment plan for system parameters is generated, yielding optimization recommendations. The adjustment plan is determined according to the type of anomaly diagnosed: if the cause is a deviation in feeding amount assessment, the safety factor or correction factor parameters in the feeding amount calculation module are adjusted; if the cause is inappropriate feeding timing, the parameters of the nursing pattern prediction module or the buffer parameters of the time window construction are adjusted; if the cause is environmental stress, environmental control suggestions are sent to management personnel. Parameter adjustments employ a gradual update strategy, with each adjustment not exceeding the preset maximum step size to avoid drastic parameter changes that could lead to system instability. The adjusted parameters take effect in subsequent feeding cycles. The system continuously monitors the adjustment effect; if the anomaly improves, the parameter adjustment is confirmed as effective; if the anomaly persists, diagnosis and adjustment continue. Through this closed-loop feedback mechanism, the system can continuously learn and optimize, gradually improving feeding accuracy and efficiency.

[0194] Step 7.7: Generate log statistics report;

[0195] Based on various data collected throughout the feeding process, feeding execution logs and statistical reports are generated. The feeding execution log records detailed information for each feeding operation in chronological order, including fields such as target rabbit cage number, planned arrival time, actual arrival time, planned feed amount, actual feed amount, feed intake monitoring results, and abnormal events. Statistical reports summarize system performance indicators over a period of time, including feeding on-time rate, feeding accuracy, average feed intake rate, anomaly rate, and path execution efficiency. Trend analysis is performed on the statistical data to identify system performance trends and potential problems. Visual charts are generated, including feed distribution histograms, feed intake rate time-series curves, anomaly type pie charts, and path heatmaps, allowing managers to intuitively understand the system's operating status. Feeding execution logs and reports are stored in a database, supporting historical queries and data export, providing data support for system optimization and management decisions.

[0196] Furthermore, since rule-based knowledge graph reasoning may fail to provide accurate diagnoses when faced with complex or rare anomalies, an end-to-end anomaly diagnosis model based on graph neural networks can be used to replace the rule-based reasoning engine. The aim is to automatically discover complex correlations between anomaly patterns and their causes through deep learning, thereby improving the accuracy and coverage of diagnoses. Specifically, this graph neural network diagnosis model models the anomaly diagnosis problem as a link prediction task on a knowledge graph. The model input is a subgraph consisting of the current anomaly event node and its associated context nodes, and the output is the probability of a causal relationship between the anomaly event and each candidate cause node. The model adopts a relational graph attention network architecture, which considers the relationship type of the edges during message passing, and different attention weights are calculated for different types of relationships. The initial embedding of a node is obtained from the node attributes through an embedding layer, and after multi-layer graph attention propagation, a node representation incorporating neighborhood information is obtained. The final representation of the anomaly event node interacts with the representations of each candidate cause node for computation, and the causal relationship probability is output through a scoring function. The model is trained on historical anomaly diagnosis cases, learning the mapping relationship from anomaly manifestation to cause, and is able to handle novel anomaly patterns not covered by the rule base.

[0197] An intelligent feed delivery system for automated rabbit farms, such as Figure 3 As shown, a smart feed delivery method for an automated rabbit farm, as described above, includes:

[0198] The data acquisition module is used to collect image data and radio frequency data from the rabbit farm to identify the female rabbits and obtain the identification results.

[0199] The lactation assessment module matches the doe's profile with the doe's identity confirmation results to assess the doe's current lactation performance and obtain the current lactation stage label and nutritional requirement level.

[0200] The nutrition calculation module assesses the feeding requirements based on the current lactation stage label and nutritional requirement level, and obtains the standard feeding parameters for each doe.

[0201] The feeding amount calculation module calculates the feeding amount and checks the inventory based on the standard feeding amount parameters of individual female rabbits to obtain the target feeding amount;

[0202] The behavior monitoring module is used to monitor the behavior data of the mother rabbits in each cage in real time during the nursing process, identify the nursing behavior, analyze the nursing pattern and predict the feeding time, and obtain the best feeding time and feeding stress score for each cage.

[0203] The path planning module plans the path of the feeding cart and makes dynamic adjustments based on the optimal feeding time and feeding urgency score of each rabbit cage, and obtains the driving instructions of the feeding cart.

[0204] The feeding execution module controls the feeding trolley to perform feeding operations based on the target feeding amount and the feeding trolley's driving instructions. It also diagnoses and handles abnormal situations during the execution process, and obtains feeding execution logs and optimization suggestions.

[0205] In one embodiment of the present invention, a specific example is provided:

[0206] This invention focuses on the application of intelligent feeding and management in large-scale rabbit breeding farms. The following is a specific application example to illustrate the practical application effect of the method of this invention.

[0207] A modern rabbit breeding base employs the intelligent feed delivery method and system described in this invention for precise nutritional management of its lactating does. The base houses a number of lactating does in a three-tiered cage system equipped with a track-mounted automatic feeding system. The system utilizes various sensing devices, including RGB-D depth cameras, UHF RFID readers, infrared thermal imaging sensors, vibration sensors, audio sensors, weighing sensors, and environmental monitoring sensors. A central server runs the intelligent feed delivery software system described in this invention, achieving fully automated management from individual rabbit identification to feed delivery execution.

[0208] Table 1 shows some examples of data collected during system operation:

[0209] Table 1: Examples of some data collected during the operation;

[0210]

[0211] Based on the above data, the system calculates individualized nutritional requirements for each does and determines differentiated feeding amounts. Does in the rising and peak lactation phase receive higher feeding amounts to meet their lactation needs, while does in the declining lactation phase receive appropriately lower feeding amounts while ensuring their body condition recovers.

[0212] The system uses multimodal lactation behavior recognition and temporal pattern analysis to determine the optimal feeding time for each does. Table 2 shows examples of lactation patterns and feeding timing data for some does.

[0213] Table 2: Examples of nursing patterns and feeding timing data for some female rabbits;

[0214]

[0215] By feeding the rabbits during their peak feeding period after lactation, the system effectively improves feed intake and utilization efficiency.

[0216] The intelligent feed delivery method described in this invention, applied in this rabbit breeding base, enables precise nutritional management of lactating does. The system accurately identifies each does and links it to their physiological status profile, dynamically adjusting the feed amount based on their lactation performance and body condition. Feed is delivered at the optimal time to improve feed intake, and system parameters are continuously optimized through a closed-loop feedback mechanism. This method effectively solves the problem that traditional feeding methods cannot adapt to individual differences and dynamic changes, providing a feasible technical solution for the intelligent upgrading of large-scale rabbit farms.

[0217] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for intelligent feed delivery for automated rabbitries, comprising: Includes the following steps: Image and radio frequency data from the rabbit farm were collected to identify the female rabbits and obtain the identification results. Based on the identification results of the female rabbit, the associated files of the female rabbit are matched to assess the current lactation performance of the female rabbit and obtain the current lactation stage label and nutritional requirement level. Based on the current lactation stage label and nutritional requirement level, the feeding requirement is assessed to obtain the standard feeding parameters for individual does; The target feeding amount is obtained by calculating the feeding amount and checking the inventory based on the standard feeding parameters of individual female rabbits. Real-time monitoring of the behavior data of the mother rabbits nursing in each cage and identification of nursing behavior, analysis of nursing patterns and prediction of feeding time, to obtain the best feeding time and feeding urgency score for each cage. Based on the optimal feeding time and feeding urgency score for each rabbit cage, the path of the feeding cart is planned and dynamically adjusted to obtain the feeding cart driving instructions. Based on the target feeding amount and the feeding trolley's driving instructions, the feeding trolley is controlled to perform feeding operations, and abnormal situations during the execution process are diagnosed and handled to obtain feeding execution logs and optimization suggestions; The evaluation of the current lactation performance of the doe includes: based on the doe's identity confirmation, retrieving the corresponding doe's file information from the breeding management database, verifying the integrity of the file information, and calculating the postpartum age based on the current date and the most recent farrowing date; obtaining daily weight data of the kit litter, calculating the daily weight gain of the kit litter, and using an energy balance model to back-calculate the doe's daily milk production; obtaining rabbit hutch temperature data, applying temperature correction to the estimated milk production value, with the temperature correction coefficient calculated using a piecewise linear function; based on the time series of continuously collected doe body shape characteristic parameters, using a one-dimensional convolutional neural network to extract the temporal features of body condition changes, and automatically predicting the body condition score sequence; based on the estimated milk production value and the body condition score sequence, using a multi-task learning architecture temporal prediction network to simultaneously predict changes in milk production and body condition scores for the next few days; and based on the milk production prediction curve, using a change point detection algorithm to identify lactation stage transition points, obtaining the current lactation stage label and the predicted stage transition time. The feeding requirement assessment includes: classifying does into preset nutritional requirement levels based on the current lactation stage label, combined with the does's parity and body condition score; calculating the correction coefficient for feeding amount based on predicted milk production, increasing feeding amount when the does's milk production exceeds the level standard value to meet the additional nutritional output; comparing the predicted body condition score with the preset target body condition score to calculate the body condition deviation value and obtain the feeding amount compensation coefficient for body condition correction; estimating the does's weight based on body size parameters and combining it with rabbitry temperature data obtained from the environmental monitoring system to calculate the correction coefficient for feeding amount based on environmental factors; and comprehensively calculating each correction coefficient based on the baseline feeding amount, lactation load correction coefficient, body condition compensation coefficient, and environmental correction coefficient, and applying a safety factor upwards to the final feeding amount. The analysis of lactation patterns and prediction of feeding times includes: obtaining lactation posture recognition results, lactation activity pattern recognition results, and lactation environment acoustic feature recognition results; using an evidence theory framework for multimodal fusion decision-making to obtain a comprehensive judgment result of lactation status; based on the historical lactation time data of the doe recorded by the system, combined with the lactation management mode information of the rabbit farm, using a multimodal adaptive modeling method to establish the individual lactation time pattern and obtain the predicted distribution of lactation time; based on the predicted distribution of lactation time under different management modes, combined with the digestive physiological rhythm of the doe and the constraints of actual management operations, using a utility function to score the candidate feeding times, and selecting the time with the largest utility function value as the optimal feeding time; The process of planning and dynamically adjusting the feeding trolley route includes: constructing a feeding time window constraint for each rabbit cage based on the optimal feeding time and feeding urgency score; constructing a travel distance matrix between rabbit cages based on the spatial layout data of the rabbit farm and the topology data of the track network; formalizing the feeding path planning problem into a vehicle path problem with a time window based on the time window constraint and the travel distance matrix, and solving it using an adaptive large neighborhood search algorithm to obtain an initial feasible path; based on the initial feasible path, dynamically replanning according to the changes in the lactation management status monitored in real time during actual execution, and using a local adjustment strategy to obtain a dynamically adjusted path; and generating the feeding trolley driving instructions based on the dynamically adjusted path.

2. The intelligent feed delivery method for an automated rabbit farm according to claim 1, characterized in that, The collection of image and radio frequency data from the rabbit farm for female rabbit identification includes: An RGB-D depth camera is used to simultaneously acquire color and depth images of the mother rabbit, detect and locate the mother rabbit, obtain the bounding box and key point positions of the mother rabbit, and output visual feature vectors. An array of ultra-high frequency RFID readers is deployed in the rabbit cage area to collect signals from the electronic ear tags worn by the mother rabbits. Based on the phase difference of the signals received by multiple antennas, the three-dimensional position coordinates of the tags are estimated and the radio frequency characteristics are spliced ​​and output. A weighted fusion strategy is adopted to integrate visual features and radio frequency features, and the results are mapped to a standardized embedding space through a pre-trained metric learning network to obtain an identity verification feature vector.

3. The intelligent feed delivery method for an automated rabbit farm according to claim 2, characterized in that, The mapping from the pre-trained metric learning network to the standardized embedding space includes: Training is performed using the triplet loss function; Maintain a feature library to store the reference embedding vectors of all registered female rabbits; Cosine similarity is used as the distance metric to calculate the similarity score between the current embedded vector and all reference vectors in the feature library, and candidate matching scores are selected based on the similarity scores. Based on the matching score, select the identity confirmation result to obtain the mother rabbit's identity confirmation result.

4. The intelligent feed delivery method for an automated rabbit farm according to claim 1, characterized in that, The calculation of feeding amount and inventory check include: Based on the rabbit feed formula and standard feeding parameters, the feeding amount for the female rabbits was calculated, and the rationality of the feeding amount was verified. Check the current inventory level of rabbit feed. If the current inventory level is lower than the preset threshold, activate the inventory shortage handling strategy, including priority allocation strategy, delayed feeding strategy and emergency replenishment strategy. Based on the target feeding amount and rabbit feed formulation, a pre-trained palatability prediction model was used to evaluate the expected feeding effect of the feeding amount. When the expected feed intake score is lower than the preset palatability threshold, the feeding amount is adjusted accordingly. The adjustment strategies include feeding in multiple sessions, moderately reducing the feeding amount, and adjusting the feeding timing.

5. The intelligent feed delivery method for an automated rabbit farm according to claim 1, characterized in that, The control of the feeding trolley to perform the feeding operation, and the diagnosis and handling of abnormal situations during the execution process, include: Based on driving commands, control the feeding trolley to move along the planned path; Once the feeding trolley reaches the target rabbit cage, the screw conveyor is controlled to feed the rabbits quantitatively according to the target feeding amount. The feeding control adopts a strategy that combines open-loop preset and closed-loop correction. After feeding is completed, continuously monitor the data of the weighing sensor in the feed trough, record the change curve of feed balance over time, analyze and process the feeding curve, and calculate the first derivative of the curve to obtain the feeding rate curve. Based on the actual feed intake, the expected feed intake output by the palatability prediction model is compared to calculate the feed intake deviation rate. When the absolute value of the feed intake deviation rate exceeds the preset normal fluctuation threshold, it is judged as an abnormal feed intake. When abnormal feeding is detected, the knowledge graph-based abnormal cause diagnosis module is activated to perform reasoning and query, and output an abnormal diagnosis report and corresponding treatment suggestions. Based on the anomaly diagnosis report and handling suggestions, an adjustment plan for system parameters is generated, and continuous learning and optimization are carried out through a closed-loop feedback mechanism.

6. An intelligent feed delivery system for automated rabbit farms, characterized in that, A method for performing an intelligent feed delivery system for an automated rabbit farm as described in any one of claims 1-5, comprising: The data acquisition module is used to collect image data and radio frequency data from the rabbit farm to identify the female rabbits and obtain the identification results. The lactation assessment module matches the doe's profile with the doe's identity confirmation results to assess the doe's current lactation performance and obtain the current lactation stage label and nutritional requirement level. The nutrition calculation module assesses the feeding requirements based on the current lactation stage label and nutritional requirement level, and obtains the standard feeding parameters for each doe. The feeding amount calculation module calculates the feeding amount and checks the inventory based on the standard feeding amount parameters of individual female rabbits to obtain the target feeding amount; The behavior monitoring module is used to monitor the behavior data of the mother rabbits in each cage in real time during the nursing process, identify the nursing behavior, analyze the nursing pattern and predict the feeding time, and obtain the best feeding time and feeding stress score for each cage. The path planning module plans the path of the feeding cart and makes dynamic adjustments based on the optimal feeding time and feeding urgency score of each rabbit cage, and obtains the driving instructions of the feeding cart. The feeding execution module controls the feeding trolley to perform feeding operations based on the target feeding amount and the feeding trolley's driving instructions. It also diagnoses and handles abnormal situations during the execution process, and obtains feeding execution logs and optimization suggestions.

Images