Intelligent feed proportioning system based on deep learning

Abstract

The present invention discloses a feed intelligent proportioning system based on deep learning, which includes a data acquisition module that generates synchronous preprocessing data; a conditional diffusion semantic segmentation module that outputs semantic segmentation mask data during the back-diffusion iterative process; a behavior recognition module that inputs into a poultry behavior recognition network built based on an improved ResNeXt architecture to obtain feeding behavior recognition result data; an individual trajectory association module that outputs feeding behavior time series data; a feeding-feed prediction module that generates feeding rate prediction data and feed remaining amount prediction data respectively; a dynamic feeding optimization decision module that obtains target feed amount data and nutrient ratio adjustment coefficient data; and an execution control module that adjusts the intelligent feed proportioning in real time. The present invention adjusts the target feed amount and the nutrient ratio of protein, energy, and trace elements in real time in stages, achieving closed-loop control of the entire process. The feeding accuracy and economy are superior to existing automatic feeding solutions.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent feed proportioning, and in particular to an intelligent feed proportioning system based on deep learning. Background Art

[0002] With the continuous improvement of the scale and density of modern poultry farming, poultry behavior monitoring and dynamic feed control technologies based on computer vision and intelligent analysis have gradually become important means to improve farming efficiency and the level of automated management. Existing technologies generally use traditional convolutional neural networks or the YOLO series of algorithms based on target detection to identify and count individual poultry behaviors. Existing technologies can achieve basic behavior detection in low-density scenarios, but in actual factory environments with dense poultry numbers, severe group occlusion, and frequent changes in ambient light, individual missed detections, misjudgments, and inaccurate behavior recognition are prone to occur, resulting in the inability to obtain highly reliable individual feeding behavior statistics, which limits the accuracy and automation level of subsequent intelligent feeding.

[0003] On the other hand, existing farms generally adopt a timed or quantitative feed delivery system, which makes it difficult to dynamically adjust the feed amount and formula according to the actual feeding status of the poultry flock, individual differences and growth stage, which can easily lead to feed waste or insufficient nutrition supply for individual poultry, affecting the overall breeding efficiency and poultry flock health. Summary of the Invention

[0004] One purpose of the present invention is to propose an intelligent feed proportioning system based on deep learning. The present invention adjusts the target feed amount and the nutritional ratio of protein, energy and trace elements in real time in stages, realizing closed-loop control of the entire process. The feeding accuracy and economy are better than existing automatic feeding solutions.

[0005] According to an embodiment of the present invention, a deep learning-based intelligent feed proportioning system includes:

[0006] The data acquisition module collects poultry house video image data and feed status data, performs preprocessing, and generates synchronous preprocessed data;

[0007] The conditional diffusion semantic segmentation module uses the conditional diffusion semantic segmentation model to perform pixel-level semantic segmentation on the poultry house video image data in the synchronous preprocessed data, and outputs semantic segmentation mask data in the reverse diffusion iterative process;

[0008] The behavior recognition module performs pixel-aligned fusion of semantic segmentation mask data and corresponding poultry house video image data, and inputs them into the poultry behavior recognition network built based on the improved ResNeXt architecture to obtain feeding behavior recognition result data;

[0009] The individual trajectory association module eliminates inter-frame jitter of the feeding behavior recognition result data and associates individual trajectories to output feeding behavior time series data;

[0010] The feeding-feed prediction module inputs the feeding behavior time series data and the feed status data in the synchronous preprocessed data into the feeding rate prediction model and the feed remaining amount prediction model to generate feeding rate prediction data and feed remaining amount prediction data respectively;

[0011] The dynamic feeding optimization decision module solves the multi-objective optimization function based on the bird age, feeding rate prediction data and feed remaining amount prediction data to obtain the target feed amount data and nutrient ratio adjustment coefficient data;

[0012] The execution control module drives the variable-speed screw feeder and multi-bin formula selector to adjust the intelligent feed ratio in real time based on the target feed amount data and the nutrient ratio adjustment coefficient data.

[0013] Optionally, the data acquisition module includes:

[0014] The poultry house video image data is collected by a multispectral camera device arranged on the top of the poultry house and above the feed trough, and the feed status data is synchronously collected through a weight flow sensor device. The poultry house video image data and feed status data are time-stamp aligned, format standardized and noise suppressed.

[0015] Optionally, the construction of the conditional diffusion semantic segmentation model includes:

[0016] Calculate the adaptive noise scheduling sequence based on the ambient light intensity data in the synchronous preprocessed data;

[0017] The conditional input tensor is constructed based on the spatial heat map of the bird's head, the prior information of the feed trough outline, the historical segmentation mask and the feed spectral energy distribution map;

[0018] At the start of the back diffusion, the poultry house video image data and the initial noise tensor are weightedly added to obtain the initial noise image tensor;

[0019] At each back-diffusion time step, the conditional diffusion network parameterized function is input with the initial noise image tensor, the conditional input tensor and the current time step number, and the noise residual is output;

[0020] Based on the noise residual and the adaptive noise scheduling sequence, the denoised image prediction is updated using a weighted residual method;

[0021] In the anti-diffusion iterative process, the multi-objective loss function is optimized;

[0022] After the iteration is completed, the final noise-free image is obtained. The energy-threshold segmentation function is used to multiply the final noise-free image pixel by pixel by the energy weight matrix estimated based on the feed spectral energy distribution map. The final semantic segmentation mask data is obtained through normalized probability mapping.

[0023] Optionally, the behavior recognition module includes:

[0024] The final semantic segmentation mask data is aligned pixel by pixel with the corresponding poultry house video image data, and fused into multi-channel fusion input tensor data;

[0025] The multi-channel fusion input tensor data is input into a multi-granularity attention feature extraction module based on an improved ResNeXt structure and customized in combination with poultry feeding behavior perception and dynamic feed regulation scenarios. The input is preliminarily encoded with spatial features using grouped convolutional units to obtain preliminary spatial feature map data.

[0026] The preliminary spatial feature map data is input in parallel to the card-type parallel feature extraction modules of the posture perception path, the head fine movement perception path, and the head-trough interaction perception path;

[0027] The posture perception path captures the global body posture and spatial motion characteristics of the bird through the residual module and outputs posture feature map data;

[0028] In the head fine movement perception path, an adaptive head attention tensor is constructed using binary mask data of the bird's head region to output feature map data of the head fine movement;

[0029] In the head-trough interaction perception path, the interactive attention tensor is constructed based on the intersection of the binary mask data of the bird's head region and the binary mask data of the trough region, and the head-trough interaction feature map data is output;

[0030] The posture feature map data, the head fine movement feature map data and the head-trough interaction feature map data are fused at the feature level to obtain the comprehensive behavior feature data;

[0031] The comprehensive behavioral feature data is input into the fully connected classification layer, and after probability normalization mapping, the poultry feeding behavior recognition result data is output. The poultry feeding behavior recognition result data includes the probability distribution of the four behavioral states of poultry in the poultry house scene: feeding, drinking, leaving the trough and resting.

[0032] Optionally, the improved ResNeXt structure includes:

[0033] The multi-channel fusion input tensor data is input into the multi-granularity attention feature extraction module built based on the improved ResNeXt structure, and the spatial feature map data of the grouped convolution path is output;

[0034] For the output group spatial feature map data of each group convolution subpath, the behavioral feature perception regulation factor is calculated based on the average response intensity of the binary mask data of the bird's head area and the binary mask data of the feeding trough area. The spatial feature map data and the corresponding behavioral feature perception regulation factor are weighted and summed to obtain the fused feature map data.

[0035] Perform three convolution operations on the fused feature map data to obtain three sets of residual feature map data with different receptive fields. Add the three sets of residual feature map data and the fused feature map data at the pixel level to obtain multi-scale enhanced feature map data.

[0036] The spatial position attention weight tensor is calculated based on the joint response map of the binary mask data of the bird's head area and the binary mask data of the feeding trough area. The spatial position attention weight tensor is multiplied point by point with the multi-scale enhanced feature map data at each pixel and channel to obtain the preliminary spatial feature map data.

[0037] Optionally, the process of eliminating inter-frame jitter of the feeding behavior recognition result data and associating individual trajectories to output feeding behavior time series data includes:

[0038] The probability sequence tensor of bird feeding behavior is formed by the four types of behavior probability values ​​of each frame of feeding behavior recognition result data;

[0039] The Kalman filter algorithm is used to suppress the inter-frame fluctuation of the probability sequence tensor of poultry feeding behavior, and a smooth output sequence of poultry feeding behavior is obtained;

[0040] Perform time series splicing on the smoothed output sequence of each bird's feeding behavior in the video image to generate a time series matrix of the bird's individual behavior state;

[0041] Through individual tracking identification information and spatial position features, the inter-frame motion coherence rule is used to perform temporal tracking and matching of individual poultry identities:

[0042] If the distance is lower than the preset position matching threshold, the two individuals are considered to be the same identity, otherwise they are given new individual numbers;

[0043] The time series matrix of individual behavioral states of birds with stable inter-frame tracking is aggregated according to individual numbers to obtain the time series data of bird feeding behavior.

[0044] Optionally, the food intake-feed prediction module includes:

[0045] Align the poultry feeding behavior time series data with the feed status data in the synchronous preprocessed data in chronological order to construct a feeding behavior-feed status joint sequence tensor;

[0046] Based on the feeding behavior-feed state joint sequence tensor, the behavioral state of each individual bird at each time step is statistically analyzed to obtain the maximum probability behavior category and the set of individual bird behavior states. The transition statistics of the behavioral states of any two consecutive time steps are also performed to obtain the behavioral state transition probability matrix.

[0047] The feeding behavior-feed state joint sequence tensor and the behavior state transition probability matrix are input into the dynamic feeding rate prediction model. At each time step, the dynamic feeding rate prediction model outputs the predicted feeding rate of the current time step based on the current behavior state of the bird, the behavior state at the previous moment, and the feed remaining at the previous moment.

[0048] The predicted feeding rate under continuous time steps is time-integrated, and a feed remaining amount prediction model is established in combination with the unit time length. At each time step, the feed remaining amount prediction model outputs the predicted feed remaining amount of the current time step based on the actual feed remaining amount of the previous time step and the predicted feeding rate and time step of the current time step.

[0049] Optionally, the dynamic feeding optimization decision module includes:

[0050] Collect age parameters and environmental data of the bird flock;

[0051] The time proportion and duration of the four behavioral states of feeding, drinking, leaving the trough, and resting in the feeding behavior time series data, together with the feeding rate prediction data, feed remaining prediction data, and bird age parameters, are input into the dynamic feeding optimization module. The dynamic feeding optimization module synchronously determines the target feed amount data and the nutrient ratio adjustment coefficient data based on the multi-objective optimization function;

[0052] The dynamic feeding optimization module prioritizes meeting the preset feed nutritional balance goals during the multi-objective optimization function solution process. It also dynamically adjusts feed and formula based on real-time statistical results of four behavioral states: eating, drinking, leaving the trough, and resting, minimizing the cost per unit weight gain.

[0053] Combined with the constraints of the multi-objective optimization function, the final target feed amount data and nutrient ratio adjustment coefficient data are output.

[0054] Optionally, the multi-objective optimization function takes the following optimization objectives as constraints:

[0055] The target feed rate is determined by multiplying the predicted feeding rate by the proportion of time spent in the feeding behavior state. If the proportion of time spent in the feeding behavior state is lower than the historical average, the target feed rate is adjusted downwards, and vice versa.

[0056] If the proportion of drinking behavior increases significantly, the system will determine it is a high temperature or stress state based on environmental data, and increase the electrolyte and energy components in the nutritional ratio accordingly;

[0057] If the proportion of away from the trough or resting behavior state increases, it indicates that the feeding activity has decreased. The target feed amount will be lowered based on the weighted comprehensive result of the feeding rate prediction data and the proportion of feeding behavior state time. At the same time, the nutrient ratio adjustment coefficient will increase the protein ratio and reduce the energy ratio.

[0058] Optionally, the nutrient ratio adjustment coefficient data is set as:

[0059] When the feeding behavior state accounts for a high proportion and the feeding rate is higher than the recommended standard, the nutrient ratio adjustment coefficient data is set according to the high-protein, high-energy formula standard to increase the protein and energy coefficients;

[0060] When the proportion of drinking behavior is higher than the historical average, the nutritional ratio adjustment coefficient data increases the proportion of electrolytes and vitamins;

[0061] When the proportion of off-trough and resting behavior states is higher than the threshold, the nutrient ratio adjustment coefficient data reduces energy and increases the vitamin ratio.

[0062] The beneficial effects of the present invention are:

[0063] (1) The present invention proposes a semantic segmentation method based on the fusion of conditional diffusion model and spatial prior, which effectively improves the accuracy and robustness of individual feeding behavior detection in dense scenes. In the poultry house video image processing link, the conditional diffusion semantic segmentation model is used to dynamically superimpose multimodal prior information, combined with the ambient light adaptive noise scheduling and energy-weighted segmentation strategy. In scenes with severe occlusion, large lighting changes and dense groups, it can still generate pixel-level masks with clear boundaries and strong category discrimination. The risk of false detection and missed detection is significantly reduced in the actual poultry house test set.

[0064] (2) The present invention introduces a ResNeXt behavior recognition network with multi-granularity feature fusion and structure-guided attention mechanism to achieve accurate recognition of individual behavior-region interaction. In view of the characteristics of dense occlusion and complex movements of birds, the conditional diffusion segmentation results are fused with the original video image at the pixel level and input into the multi-granularity attention-enhanced ResNeXt network. Grouped convolution and card-type parallel feature paths respectively extract the overall posture of the birds, fine head movements and head-trough interaction features. The structure-guided spatial attention mechanism dynamically improves the response of the key feeding area. The channel-space joint weighting fully solves the problem of identifying the behavioral details of multiple birds occluding, turning their heads quickly, and temporarily leaving the trough in the same frame.

[0065] (3) At the back end of behavior recognition, the present invention realizes a dynamic feeding control closed loop based on behavioral causal chain and multi-objective optimization by jointly modeling the feeding behavior time series and feed status data, supports intelligent adaptive adjustment of target feed amount and nutrient ratio, and the multi-objective optimization module combines the nutritional requirement reference standard to adjust the target feed amount and the nutrient ratio of protein, energy and trace elements in real time in stages, realizing closed-loop control of the entire process. The feeding accuracy and economy are better than those of existing automatic feeding solutions. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0067] Figure 1 This is a system block diagram of a deep learning-based intelligent feed proportioning system proposed by the present invention;

[0068] Figure 2 This is a schematic diagram of the feature extraction structure in the deep learning-based intelligent feed proportioning system proposed by the present invention. DETAILED DESCRIPTION

[0069] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0070] refer to Figure 1-Figure 2 , a feed intelligent proportioning system based on deep learning, including:

[0071] The data acquisition module collects poultry house video image data and feed status data, performs preprocessing, and generates synchronous preprocessed data;

[0072] In this embodiment, the data acquisition module includes:

[0073] The poultry house video image data is collected by a multispectral camera device arranged on the top of the poultry house and above the feed trough, and the feed status data is synchronously collected through a weight flow sensor device. The poultry house video image data and feed status data are time-stamp aligned, format standardized and noise suppressed.

[0074] The conditional diffusion semantic segmentation module uses the conditional diffusion semantic segmentation model to perform pixel-level semantic segmentation on the poultry house video image data in the synchronous preprocessed data, and outputs semantic segmentation mask data in the reverse diffusion iterative process;

[0075] In this embodiment, the construction of the conditional diffusion semantic segmentation model includes:

[0076] Calculate the adaptive noise scheduling sequence based on the ambient light intensity data in the synchronous preprocessed data;

[0077] The ambient light intensity data is exponentially related to the preset minimum noise scale coefficient, maximum noise scale coefficient, and light attenuation coefficient to generate a noise scale coefficient sequence corresponding to each diffusion time step. Each item in the noise scale coefficient sequence is adaptively adjusted as the ambient light intensity data changes. It is used to dynamically set the noise amplitude of each step in the back-diffusion process, automatically increasing the noise removal step size in low-light environments and automatically reducing the noise removal step size in high-light environments.

[0078] The adaptive noise scheduling sequence is used to initialize the step size configuration of the inverse diffusion process. Each item of the adaptive noise scheduling sequence is adaptively adjusted by the ambient light intensity data.

[0079] The conditional input tensor is constructed based on the spatial heat map of the bird's head, the prior information of the feed trough outline, the historical segmentation mask and the feed spectral energy distribution map;

[0080] The conditional input tensor is composed of the spatial heat map of the bird's head, the prior information of the feed trough outline, the historical segmentation mask, and the feed spectral energy distribution map superimposed in a fixed order.

[0081] The feed spectral energy distribution diagram is obtained by sampling the feed area in the visible-near infrared band through a multispectral camera device and then reducing the dimension, which is used to distinguish between feed particles and poultry body areas.

[0082] The spatial heat map of poultry heads is obtained by performing key point detection and spatial distribution statistics on the poultry head area in each frame of the poultry house video image data. The specific process is to use a pre-trained head detection model to detect the candidate head area of ​​the poultry on a pixel-by-pixel basis in each frame of the image, and then aggregate the center points of the detected candidate head areas in the current frame and the most recent frames to obtain the probability distribution of each pixel being a head, forming a spatial heat map of the poultry head.

[0083] The trough outline prior information is obtained by performing image segmentation and edge detection on the poultry house video image data. This process involves segmenting the video image based on the shape, color, and spatial layout characteristics of the trough structure. An edge detection algorithm is then used to extract the trough area's boundaries. This is then geometrically corrected based on the known physical dimensions of the poultry house structure to obtain the precise trough outline position in each frame, forming the trough outline prior information.

[0084] At the start of the back diffusion, the poultry house video image data and the initial noise tensor are weightedly added to obtain the initial noise image tensor;

[0085] The weights of the initial noise tensor are given by the first term of the adaptive noise scheduling sequence.

[0086] At each back-diffusion time step, the conditional diffusion network parameterized function is input with the initial noise image tensor, the conditional input tensor and the current time step number, and the noise residual is output;

[0087] The noise residual is controlled by the feed attention gating block inside the conditional diffusion network, highlighting the feature expressions of the head area and the feeding trough area.

[0088] Based on the noise residual and the adaptive noise scheduling sequence, the denoised image prediction is updated using a weighted residual method;

[0089] In the anti-diffusion iterative process, the multi-objective loss function is optimized;

[0090] The multi-objective loss function includes the weighted Dice loss of semantic labels, which expresses the segmentation accuracy of the four categories of semantic labels; the Euclidean distance preservation loss between the bird's head area and the feeding trough area, which constrains the geometric consistency of the feeding posture; and the temporal consistency intersection-over-union loss between the mask and the historical segmentation mask.

[0091] After the iteration is completed, the final noise-free image is obtained. The energy-threshold segmentation function is used to multiply the final noise-free image pixel by pixel by the energy weight matrix estimated based on the feed spectral energy distribution map. The final semantic segmentation mask data is obtained through normalized probability mapping.

[0092] The energy-threshold segmentation function uses the feed spectral energy distribution map to perform pixel-by-pixel dynamic weighting and threshold judgment on the segmentation output according to the perception of poultry feeding behavior and dynamic feed control scenarios. When the energy-threshold segmentation function is used to process the final noise-free image, each pixel value of the final noise-free image is multiplied pixel by pixel by the weight value in the energy weight matrix at the corresponding position. The energy weight matrix is ​​obtained by normalizing the feed spectral energy distribution map and is used to give a higher response to pixels in the key feeding area. The product results of all pixels are sent to the normalized probability mapping module. The probability value of each category is adjusted to the same probability distribution range through pixel-by-pixel normalization. The category with the largest probability value is selected at each pixel position as the final classification label of the pixel, thereby generating the final semantic segmentation mask data, realizing dynamic improvement of category confidence and precise segmentation of the key feeding area.

[0093] The behavior recognition module performs pixel-aligned fusion of semantic segmentation mask data and corresponding poultry house video image data, and inputs them into the poultry behavior recognition network built based on the improved ResNeXt architecture to obtain feeding behavior recognition result data;

[0094] In this embodiment, the behavior recognition module includes:

[0095] The final semantic segmentation mask data is aligned pixel by pixel with the corresponding poultry house video image data, and fused into multi-channel fusion input tensor data;

[0096] The multi-channel fusion input tensor data includes the red channel data, green channel data, blue channel data of the poultry house video image data, and the binary mask data of the poultry head area and the binary mask data of the feeding trough area extracted from the final semantic segmentation mask data.

[0097] The red channel data, green channel data, and blue channel data of the poultry house video image data are used to maintain the original visual texture and color information of the poultry house, respectively. The binary mask data of the poultry head area is used to highlight the spatial position and morphological characteristics of the poultry head in the video image. The binary mask data of the feeding trough area is used to enhance the spatial contextual information of the interaction between the feeding trough and the head area.

[0098] The multi-channel fusion input tensor data is input into a multi-granularity attention feature extraction module based on an improved ResNeXt structure and customized in combination with poultry feeding behavior perception and dynamic feed regulation scenarios. The input is preliminarily encoded with spatial features using grouped convolutional units to obtain preliminary spatial feature map data.

[0099] The preliminary spatial feature map data is input in parallel to the card-type parallel feature extraction modules of the posture perception path, the head fine movement perception path, and the head-trough interaction perception path;

[0100] The posture perception path captures the global body posture and spatial motion characteristics of the bird through the residual module and outputs posture feature map data;

[0101] In the head fine movement perception path, an adaptive head attention tensor is constructed using binary mask data of the bird's head region to output feature map data of the head fine movement;

[0102] The adaptive head attention tensor weights the preliminary spatial feature map data in a pixel-by-pixel manner to highlight the key motion features of the avian head.

[0103] In the head-trough interaction perception path, the interactive attention tensor is constructed based on the intersection of the binary mask data of the bird's head region and the binary mask data of the trough region, and the head-trough interaction feature map data is output;

[0104] The interactive attention tensor weights the preliminary spatial feature map data in a pixel-by-pixel manner to highlight the pixel-level correlation features of the dynamic interaction between the bird's head and the feeding trough area.

[0105] The posture feature map data, the head fine movement feature map data and the head-trough interaction feature map data are fused at the feature level to obtain the comprehensive behavior feature data;

[0106] During the feature fusion process, the channel-spatial attention mechanism is used to dynamically enhance the response intensity of each feature channel to the perception of poultry feeding behavior.

[0107] The channel-space attention mechanism is a deep learning feature enhancement method. During the feature extraction process, different weights are assigned to data from different channels and data from different spatial regions by automatically analyzing the importance of each channel feature and spatial position feature. In this embodiment, the channel attention mechanism dynamically adjusts the contribution of each channel during feature fusion according to the overall response strength of each channel in the feature tensor, and the spatial attention mechanism dynamically adjusts the feature responses of different spatial regions according to the importance of each pixel or position in the feature tensor. After combination, it can not only highlight the semantic information channels related to the task but also enhance the expression ability of key spatial regions, thereby improving the recognition accuracy and robustness of complex scenes of poultry feeding behavior.

[0108] The comprehensive behavioral feature data is input into the fully connected classification layer, and after probability normalization mapping, the poultry feeding behavior recognition result data is output. The poultry feeding behavior recognition result data includes the probability distribution of the four behavioral states of poultry in the poultry house scene: feeding, drinking water, leaving the trough and resting. Each probability value is used for the refined analysis of poultry feeding behavior and the basis for dynamic feed regulation decision-making.

[0109] In this embodiment, the ResNeXt structure is improved, including:

[0110] The multi-channel fusion input tensor data is input into the multi-granularity attention feature extraction module built based on the improved ResNeXt structure, and the spatial feature map data of the grouped convolution path is output;

[0111] The multi-granularity attention feature extraction module includes multiple grouped convolution subpaths. Each grouped convolution subpath takes the same multi-channel fused input tensor data as input and outputs the spatial feature map data of the grouped convolution path. The total number of grouped convolution paths is represented by G. Each grouped convolution path models different feature subspaces of the multi-channel fused input tensor data through convolution operations. The spatial dimension and channel dimension of the output feature map data maintain the same dimension.

[0112] For the output group spatial feature map data of each group convolution subpath, the behavioral feature perception regulation factor is calculated based on the average response intensity of the binary mask data of the bird's head area and the binary mask data of the feeding trough area. The spatial feature map data and the corresponding behavioral feature perception regulation factor are weighted and summed to obtain the fused feature map data.

[0113] The behavioral feature perception regulation factor is used to measure the response sensitivity of the grouped convolution subpath to the key areas of poultry feeding behavior. The sum of all behavioral feature perception regulation factors is equal to 1. The fused feature map data spatially enhances the response characteristics of the significant areas of feeding behavior.

[0114] Perform three convolution operations on the fused feature map data to obtain three sets of residual feature map data with different receptive fields. Add the three sets of residual feature map data and the fused feature map data at the pixel level to obtain multi-scale enhanced feature map data.

[0115] The convolution kernel sizes are one by one, three by three, and five by five, respectively. The multi-scale enhanced feature map data retains the semantic detail information at different spatial scales, improving the stable recognition ability of the feeding behavior of poultry at different posture scales.

[0116] The spatial position attention weight tensor is calculated based on the joint response map of the binary mask data of the bird's head area and the binary mask data of the feeding trough area. The spatial position attention weight tensor is multiplied point by point with the multi-scale enhanced feature map data at each pixel and channel to obtain the preliminary spatial feature map data.

[0117] The spatial position attention weight tensor is defined point by point on the pixel coordinates and channel indices, and normalized by the Sigmoid function. The preliminary spatial feature map data gives higher response priority to the key areas and feature channels of poultry feeding actions, suppresses the response of the background area, and realizes the sensitive feature expression of the target area.

[0118] The individual trajectory association module eliminates inter-frame jitter of the feeding behavior recognition result data and associates individual trajectories to output feeding behavior time series data;

[0119] In this embodiment, the process of eliminating inter-frame jitter of the feeding behavior recognition result data and associating individual trajectories to output feeding behavior time series data includes:

[0120] The probability sequence tensor of bird feeding behavior is formed by the four types of behavior probability values ​​of each frame of feeding behavior recognition result data;

[0121] The four types of behavior probability values ​​are feeding behavior probability value, drinking behavior probability value, resting behavior probability value and leaving the trough behavior probability value. The poultry feeding behavior probability sequence tensor is used to measure the probability of the poultry belonging to the four behavioral states at each time step.

[0122] The Kalman filter algorithm is used to suppress the inter-frame fluctuation of the probability sequence tensor of poultry feeding behavior, and a smooth output sequence of poultry feeding behavior is obtained;

[0123] The Kalman filter algorithm dynamically balances instantaneous recognition and historical estimation results by weighting the behavior probability value of the current frame and the smoothed behavior state vector of the previous frame. The smoothed output sequence of poultry feeding behavior is used to eliminate jitter and occasional recognition errors between frames.

[0124] Perform time series splicing on the smoothed output sequence of each bird's feeding behavior in the video image to generate a time series matrix of the bird's individual behavior state;

[0125] The time series matrix of the behavior state of individual birds is composed of the smoothed behavior state vector of each frame in a continuous time period of the individual birds, which is used to describe the behavioral evolution trajectory of the individual birds in a continuous observation period.

[0126] Through individual tracking identification information and spatial position features, the inter-frame motion coherence rule is used to perform temporal tracking and matching of individual poultry identities:

[0127] If the distance is lower than the preset position matching threshold, the two individuals are considered to be the same identity, otherwise they are given new individual numbers;

[0128] The inter-frame motion continuity rule is based on matching the spatial center coordinate distance of each individual in the current frame and the previous frame. The inter-frame motion continuity rule is used to eliminate trajectory interruptions caused by mismatching and occlusion re-identification, and ensure the continuity of individual identity.

[0129] The time series matrix of individual behavioral states of birds with stable inter-frame tracking is aggregated according to individual numbers to obtain the time series data of bird feeding behavior.

[0130] The time series data of poultry feeding behavior consists of the smoothed behavioral state vector of individual poultry, individual identity information, spatial trajectory points and timestamps.

[0131] The feeding-feed prediction module inputs the feeding behavior time series data and the feed status data in the synchronous preprocessed data into the feeding rate prediction model and the feed remaining amount prediction model to generate feeding rate prediction data and feed remaining amount prediction data respectively;

[0132] In this embodiment, the feeding-feed prediction module includes:

[0133] Align the poultry feeding behavior time series data with the feed status data in the synchronous preprocessed data in chronological order to construct a feeding behavior-feed status joint sequence tensor;

[0134] The feeding behavior-feed state joint sequence tensor consists of the behavior state vector of each individual bird at each time step and the feed remaining at the corresponding time step. The behavior state vector contains four types of behavior confidence.

[0135] Based on the feeding behavior-feed state joint sequence tensor, the behavioral state of each individual bird at each time step is statistically analyzed to obtain the maximum probability behavior category and the set of individual bird behavior states. The transition statistics of the behavioral states of any two consecutive time steps are also performed to obtain the behavioral state transition probability matrix.

[0136] The behavioral state transition probability matrix is ​​used to represent the temporal evolution pattern between eating, drinking, resting and leaving the trough. Each element of the behavioral state transition probability matrix represents the probability of transitioning from the current behavioral state to the next behavioral state.

[0137] The feeding behavior-feed state joint sequence tensor and the behavior state transition probability matrix are input into the dynamic feeding rate prediction model. At each time step, the dynamic feeding rate prediction model outputs the predicted feeding rate of the current time step based on the current behavior state of the bird, the behavior state at the previous moment, and the feed remaining at the previous moment.

[0138] The dynamic feeding rate prediction model uses a tensor representing the joint sequence of feeding behavior and feed status and a behavioral state transition probability matrix as input. The model reads the behavioral state of each individual bird at the current time step and combines it with the previous behavioral state and feed remaining at the previous time step. When predicting the feeding rate at the current time step, the model not only considers the current behavioral state and historical states themselves but also infers the probability trend of feeding behavior over the next period of time based on the behavioral state transition probability matrix. The dynamic feeding rate prediction model retrieves the probability of transitioning from the current behavioral state to the feeding behavior in the behavioral state transition probability matrix and applies this probability as a dynamic weight to the fundamental impact of the current behavioral state and feed remaining on feeding rate. By training on historical behavioral sequences and feed status data, the model learns the quantitative mapping between different behavioral patterns, changes in feed remaining, and actual feeding rate. The dynamic feeding rate prediction model integrates the current bird behavior state, historical behavior state, feed remaining amount and behavior trend weight based on the behavior state transition probability matrix, and outputs the predicted feeding rate of the bird individual at the current time step.

[0139] The entire dynamic feeding rate prediction model uses supervised learning to train parameters on a labeled dataset containing historical behavior sequences, feed remaining and actual feeding rate to achieve high-precision rate prediction based on the temporal behavior causal chain and feed status changes.

[0140] The predicted feed intake rate under continuous time steps is time-integrated, and a feed remaining amount prediction model is established based on the unit time length. At each time step, the feed remaining amount prediction model outputs the predicted feed remaining amount value of the current time step based on the actual feed remaining amount in the previous time step and the predicted feed intake rate and time step length of the current time step;

[0141] The feed remaining amount prediction model is based on the feeding behavior-feed state joint sequence tensor as input, and is gradually recursively deduced in combination with the output results of the dynamic feeding rate prediction model. In this embodiment, the feed remaining amount prediction model obtains the actual feed remaining amount of the previous time step and the feeding rate prediction value of the current time step at each time step, and dynamically updates the feed remaining amount of the previous time step using the feeding rate prediction value of the current time step according to the set time step. The updating process is achieved by subtracting the feed remaining amount of the previous time step from the feed remaining amount of the previous time step from the feed amount expected to be consumed in the current time step. The expected feed amount to be consumed is equal to the product of the feeding rate prediction value of the current time step and the time step. In this way, the recursive feed remaining amount prediction model outputs the latest feed remaining amount prediction value in real time at each time step, realizing continuous dynamic tracking and prediction of the feed consumption process.

[0142] In another embodiment, a feeding behavior causal reasoning graph is established by taking the feeding behavior-feed state joint sequence tensor and the behavior state transition probability matrix as the global information entry, combining the behavior state transition trend, the historical feeding rate prediction results and the behavior state transition probability matrix;

[0143] The causal reasoning graph of feeding behavior consists of a node set and an edge set. The node set includes the behavioral state, feed remaining amount and feeding rate prediction value at each time step. The weight of each edge in the edge set is given by the corresponding element of the behavioral state transition probability matrix, which directly reflects the temporal causal relationship between behavioral evolution and feed consumption.

[0144] Based on the feeding behavior causal reasoning graph, real-time traversal reasoning is performed to determine whether the current cycle is in a behavior-intensive - feed shortage warning state. If the feeding rate prediction value at the current time step is higher than the feeding rate upper threshold, and the feed remaining amount prediction value is lower than the remaining feed safety threshold, the compensation control logic is triggered;

[0145] The upper threshold of feeding rate and the safety threshold of remaining feed are both adjustable parameters set by experience.

[0146] The dynamic feeding optimization decision module solves the multi-objective optimization function based on the bird age, feeding rate prediction data and feed remaining amount prediction data to obtain the target feed amount data and nutrient ratio adjustment coefficient data;

[0147] In this embodiment, the dynamic feeding optimization decision module includes:

[0148] Collect age parameters and environmental data of the bird flock;

[0149] The poultry age parameter is the actual age of the current poultry individual or group, which is used to describe the specific stage of the poultry's growth cycle.

[0150] The time proportion and duration of the four behavioral states of feeding, drinking, leaving the trough, and resting in the feeding behavior time series data, together with the feeding rate prediction data, feed remaining prediction data, and bird age parameters, are input into the dynamic feeding optimization module. The dynamic feeding optimization module synchronously determines the target feed amount data and the nutrient ratio adjustment coefficient data based on the multi-objective optimization function;

[0151] The multi-objective optimization function is constrained by the following optimization objectives and behavior patterns:

[0152] The target feed rate should meet the feed consumption requirements of the forecast period. It is determined by the product of the predicted feeding rate data and the proportion of time in the feeding behavior state. If the proportion of feeding behavior state is lower than the historical average, the target feed rate will be adjusted downward, otherwise it will be adjusted upward.

[0153] If the proportion of drinking behavior increases significantly, the system will determine that it is a high temperature or stress state based on environmental data. Based on the target feed amount meeting the feed consumption demand within the forecast period, the electrolyte and energy components in the nutrient ratio will be increased proportionally;

[0154] If the proportion of time spent away from the trough or resting increases, indicating a decrease in feeding activity, the target feed rate should be adjusted downward based on the weighted comprehensive result of the feeding rate prediction data and the proportion of time spent in the feeding behavior state. At the same time, the nutrient ratio adjustment coefficient should be increased to increase the protein ratio and reduce the energy ratio to promote the resumption of feeding willingness after entering the group.

[0155] The upper limit of the target feed amount is set according to the age parameters and growth stage of the poultry. The maximum feed amount does not exceed the maximum safe intake standard of the poultry at the current stage. The lower limit of the target feed amount is determined jointly based on the feed remaining amount prediction data and behavioral status analysis. If the feed remaining amount is predicted to be lower than the safety threshold and the proportion of feeding behavior status continues to be high, the target feed amount will be increased. If the remaining amount is sufficient or the behavioral activity decreases, the target feed amount will be reduced.

[0156] The nutrient ratio adjustment coefficient data is set as follows, and the nutrient ratio adjustment coefficient data is composed of protein, increased energy and trace element composition ratio:

[0157] When the feeding behavior state accounts for a high proportion and the feeding rate is higher than the recommended standard, the nutrient ratio adjustment coefficient data is set according to the high-protein, high-energy formula standard to increase the protein and energy coefficients;

[0158] When the proportion of drinking behavior is higher than the historical average, the nutrient ratio adjustment coefficient data increases the proportion of electrolytes and vitamins, and enhances heat resistance and stress resistance;

[0159] When the proportion of off-trough and resting behavior states is high or fluctuates violently, the nutrient ratio adjustment coefficient data reduces energy and increases the vitamin ratio;

[0160] The upper and lower limits of the nutrient ratio adjustment coefficient are dynamically adjusted according to the bird age parameters and standard feeding recommendations to prevent over- or under-nutrition;

[0161] The dynamic feeding optimization module prioritizes meeting the preset nutritional balance goals of feed supplementation during the multi-objective optimization function solution process. It also dynamically adjusts feed and formula based on real-time statistical results of four behavioral states: eating, drinking, leaving the trough, and resting. This minimizes the cost per unit weight gain and achieves economically optimal regulation across the entire behavior-data-formula chain.

[0162] Combined with the constraints of the multi-objective optimization function, the final target feed amount data and nutrient ratio adjustment coefficient data are output. The target feed amount data is used to drive the precise supply of the feeding equipment, and the nutrient ratio adjustment coefficient data is used for real-time ratio adjustment of the multi-bin feeding device.

[0163] The execution control module drives the variable-speed screw feeder and multi-bin formula selector to adjust the intelligent feed ratio in real time based on the target feed amount data and the nutrient ratio adjustment coefficient data.

[0164] Example 1: The system detected a bird numbered "21587." Its head heatmap highly overlapped with the trough outline, and the feeding probability output was 0.93. At this point, the multispectral image signal successfully segmented the feeding action region using the conditional diffusion model. The temporal consistency intersection-over-union ratio between the feeding mask and the historical mask for the region containing bird number "21587" was 0.87, exceeding the set threshold, indicating stable recognition of continuous feeding behavior.

[0165] The average probability of continuous feeding state output by the feeding behavior recognition network is 0.91, and the switching frequency of feeding-drinking-feeding state is once every 13 minutes. The feed decline rate of the feeding trough where the bird numbered "21587" is located during the corresponding time is measured to be 12.6g / minute. During the same period, the feeding probability of five birds numbered "20981", "21734" and others in the same area are all above 0.85. Through the behavioral time series, the standard deviation of the feeding state error after Kalman filtering is less than 0.06, and no individual trajectory is lost.

[0166] Subsequent monitoring found that the remaining feed amount in the feeding trough had dropped to 420g, which was lower than the safety threshold set by the system (500g). The system automatically integrated the feeding rate prediction model with the remaining amount prediction data, and combined the behavioral state sequence to conclude that the expected consumption in the next 30 minutes was 410g. According to the optimization rules, the target feed amount was set to 600g, and the nutritional ratio adjustment coefficient automatically increased the protein ratio to 1.08 (based on the historical standard of 1.00), and the energy adjustment coefficient was maintained at 1.00.

[0167] The intelligent feeding equipment receives the target feeding amount data and the nutrient ratio adjustment coefficient, the feeding process starts, the equipment outputs the actual protein content increase in the compartment, and the metering device provides real-time feedback. The actual feeding amount is 601g, and the protein and energy group ratio errors are +0.2% and +0.1% respectively, which is much lower than the 2% error of the traditional manual adjustment scheme.

[0168] Through behavioral causal chain analysis, the system found that the feeding behavior maintained a continuous high proportion (>0.92) for 22 minutes, the predicted feeding rate gradually increased from 9.8g / minute to 12.3g / minute, and the difference between the feed remaining decline curve and the actual measurement was 38g, which was much better than the average error of 120g in historical data. The alarm rate of abnormal feeding behavior of chickens before and after feeding dropped to 0.5%, while the abnormal alarm rate was often above 2% when the timed feeding mode was used in the past.

[0169] Affected by the external high temperature environment, the proportion of drinking behavior increased significantly. The system automatically determined the heat stress trend, and the nutrient ratio adjustment coefficient immediately increased the electrolyte ratio to 1.12, the vitamin ratio to 1.05, and the feeding behavior ratio to 3.1%. The target feed rate was automatically reduced by 11%, and the nutrient ratio adjustment took effect in real time. The equipment feedback showed that the actual formula adjustment error was less than 0.3%.

[0170] Individual numbered "21865" exhibited a high frequency switching state of feeding-resting-feeding-drinking. The behavioral state transition probability matrix showed that the transition probability from feeding to resting increased. Based on this, the system warned that the feeding activity in this section decreased, and the target feed amount was lowered to 220g per hour. The protein coefficient was simultaneously increased to 1.12, and the energy was reduced to 0.98. The frequency of the multi-bin module's feeding action was highly consistent with actual demand, and manual inspections found no feed waste or delayed feeding.

[0171] During the 72-hour test run, the system detected 1,289 feeding events, 375 drinking events, 213 trough-leaving events, and 841 resting events. Segmentation accuracy remained consistently above 96.5%, and the individual behavior tracking breakpoint rate was below 1.2%. The F1 score of conventional data collection methods in the same field was only 0.89, while the proposed method achieved 0.97.

[0172] Comparative experimental samples: The inventive group, using a combined sequence tensor model of feeding behavior and feed status and dynamic nutrition control, had an average feed intake of 60.3 kg per 24 hours, an abnormal feeding alarm rate of 0.6%, an eggshell breakage rate of 1.7%, and a daily weight gain of 0.042 kg per bird. The traditional group, on the other hand, had a daily feed intake of 64.2 kg, an alarm rate of 2.5%, an eggshell breakage rate of 2.9%, and a daily weight gain of 0.037 kg per bird. Training set example: The inventive method had a total of 3,245 training sets and 816 test sets for feeding behavior classification, achieving a feeding recognition accuracy of 97.2%. The traditional method had a training set of 2,980 sets and a test set of 780 sets, achieving an accuracy of 88.9%.

[0173] The results of Example 1 show that the present invention can continuously and automatically sense changes in feeding behavior in dense and complex scenarios, and realize dynamic feed amount and nutrient ratio regulation driven by behavioral data and feed status data, which can effectively reduce feed waste and improve eggshell quality and weight gain efficiency.

[0174] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A feed intelligent proportioning system based on deep learning, characterized in that: include: The data acquisition module collects poultry house video image data and feed status data, performs preprocessing, and generates synchronous preprocessed data; The conditional diffusion semantic segmentation module uses the conditional diffusion semantic segmentation model to perform pixel-level semantic segmentation on the poultry house video image data in the synchronous preprocessed data, and outputs semantic segmentation mask data in the reverse diffusion iterative process; The behavior recognition module performs pixel-aligned fusion of semantic segmentation mask data and corresponding poultry house video image data, and inputs them into the poultry behavior recognition network built based on the improved ResNeXt architecture to obtain feeding behavior recognition result data; The individual trajectory association module eliminates inter-frame jitter of the feeding behavior recognition result data and associates individual trajectories to output feeding behavior time series data; The feeding-feed prediction module inputs the feeding behavior time series data and the feed status data in the synchronous preprocessed data into the feeding rate prediction model and the feed remaining amount prediction model to generate feeding rate prediction data and feed remaining amount prediction data respectively; The dynamic feeding optimization decision module solves the multi-objective optimization function based on the bird age, feeding rate prediction data and feed remaining amount prediction data to obtain the target feed amount data and nutrient ratio adjustment coefficient data; The execution control module drives the variable-speed screw feeder and multi-bin formula selector to adjust the intelligent feed ratio in real time based on the target feed amount data and the nutrient ratio adjustment coefficient data; The construction of the conditional diffusion semantic segmentation model includes: Calculate the adaptive noise scheduling sequence based on the ambient light intensity data in the synchronous preprocessed data; The conditional input tensor is constructed based on the spatial heat map of the bird's head, the prior information of the feed trough outline, the historical segmentation mask and the feed spectral energy distribution map; At the start of the back diffusion, the poultry house video image data and the initial noise tensor are weightedly added to obtain the initial noise image tensor; At each back-diffusion time step, the conditional diffusion network parameterized function is input with the initial noise image tensor, the conditional input tensor and the current time step number, and the noise residual is output; Based on the noise residual and the adaptive noise scheduling sequence, the denoised image prediction is updated using a weighted residual method; In the anti-diffusion iterative process, the multi-objective loss function is optimized; After the iteration is completed, the final noise-free image is obtained. The energy-threshold segmentation function is used to multiply the final noise-free image pixel by pixel by the energy weight matrix estimated based on the feed spectral energy distribution map. The final semantic segmentation mask data is obtained through normalized probability mapping. The improved ResNeXt structure includes: The multi-channel fusion input tensor data is input into the multi-granularity attention feature extraction module built based on the improved ResNeXt structure, and the spatial feature map data of the grouped convolution path is output; For the output group spatial feature map data of each group convolution subpath, the behavioral feature perception regulation factor is calculated based on the average response intensity of the binary mask data of the bird's head area and the binary mask data of the feeding trough area. The spatial feature map data and the corresponding behavioral feature perception regulation factor are weighted and summed to obtain the fused feature map data. Perform three convolution operations on the fused feature map data to obtain three sets of residual feature map data with different receptive fields. Add the three sets of residual feature map data and the fused feature map data at the pixel level to obtain multi-scale enhanced feature map data. The spatial position attention weight tensor is calculated based on the joint response map of the binary mask data of the bird's head area and the binary mask data of the feeding trough area. The spatial position attention weight tensor is multiplied point by point with the multi-scale enhanced feature map data at each pixel and channel to obtain the preliminary spatial feature map data.

2. The intelligent feed proportioning system based on deep learning according to claim 1, characterized in that: The data acquisition module includes: The poultry house video image data is collected by a multispectral camera device arranged on the top of the poultry house and above the feed trough, and the feed status data is synchronously collected through a weight flow sensor device. The poultry house video image data and feed status data are time-stamp aligned, format standardized and noise suppressed.

3. The intelligent feed proportioning system based on deep learning according to claim 1, characterized in that: The behavior recognition module includes: The final semantic segmentation mask data is aligned pixel by pixel with the corresponding poultry house video image data, and fused into multi-channel fusion input tensor data; The multi-channel fusion input tensor data is input into a multi-granularity attention feature extraction module based on an improved ResNeXt structure and customized in combination with poultry feeding behavior perception and dynamic feed regulation scenarios. The input is preliminarily encoded with spatial features using grouped convolutional units to obtain preliminary spatial feature map data. The preliminary spatial feature map data is input in parallel to the card-type parallel feature extraction modules of the posture perception path, the head fine movement perception path, and the head-trough interaction perception path; The posture perception path captures the global body posture and spatial motion characteristics of the bird through the residual module and outputs posture feature map data; In the head fine movement perception path, an adaptive head attention tensor is constructed based on the binary mask data of the bird's head area, and the head fine movement feature map data is output; In the head-trough interaction perception path, the interactive attention tensor is constructed based on the intersection of the binary mask data of the bird's head region and the binary mask data of the trough region, and the head-trough interaction feature map data is output; The posture feature map data, the head fine movement feature map data and the head-trough interaction feature map data are fused at the feature level to obtain the comprehensive behavior feature data; The comprehensive behavioral feature data is input into the fully connected classification layer, and after probability normalization mapping, the poultry feeding behavior recognition result data is output. The poultry feeding behavior recognition result data includes the probability distribution of the four behavioral states of poultry in the poultry house scene: feeding, drinking, leaving the trough and resting.

4. The intelligent feed proportioning system based on deep learning according to claim 1, characterized in that: The process of eliminating inter-frame jitter of the feeding behavior recognition result data and associating individual trajectories to output feeding behavior time series data includes: The probability sequence tensor of bird feeding behavior is formed by the four types of behavior probability values ​​of each frame of feeding behavior recognition result data; The Kalman filter algorithm is used to suppress the inter-frame fluctuation of the probability sequence tensor of poultry feeding behavior, and a smooth output sequence of poultry feeding behavior is obtained; Perform time series splicing on the smoothed output sequence of each bird's feeding behavior in the video image to generate a time series matrix of the bird's individual behavior state; Through individual tracking identification information and spatial position features, the inter-frame motion coherence rule is used to perform temporal tracking and matching of individual poultry identities: If the distance is lower than the preset position matching threshold, the two individuals are considered to be the same identity, otherwise they are given new individual numbers; The time series matrix of individual behavioral states of birds with stable inter-frame tracking is aggregated according to individual numbers to obtain the time series data of bird feeding behavior.

5. The intelligent feed proportioning system based on deep learning according to claim 1, characterized in that: The feeding-feed prediction module includes: Align the poultry feeding behavior time series data with the feed status data in the synchronous preprocessed data in chronological order to construct a feeding behavior-feed status joint sequence tensor; Based on the feeding behavior-feed state joint sequence tensor, the behavioral state of each individual bird at each time step is statistically analyzed to obtain the maximum probability behavior category and the set of individual bird behavior states. The transition statistics of the behavioral states of any two consecutive time steps are also performed to obtain the behavioral state transition probability matrix. The feeding behavior-feed state joint sequence tensor and the behavior state transition probability matrix are input into the dynamic feeding rate prediction model. At each time step, the dynamic feeding rate prediction model outputs the predicted feeding rate of the current time step based on the current behavior state of the bird, the behavior state at the previous moment, and the feed remaining at the previous moment. The predicted feeding rate under continuous time steps is time-integrated, and a feed remaining amount prediction model is established in combination with the unit time length. At each time step, the feed remaining amount prediction model outputs the predicted feed remaining amount of the current time step based on the actual feed remaining amount of the previous time step and the predicted feeding rate and time step of the current time step.

6. The intelligent feed proportioning system based on deep learning according to claim 1, characterized in that: The dynamic feeding optimization decision module includes: Collect age parameters and environmental data of the bird flock; The time proportion and duration of the four behavioral states of feeding, drinking, leaving the trough, and resting in the feeding behavior time series data, together with the feeding rate prediction data, feed remaining prediction data, and bird age parameters, are input into the dynamic feeding optimization module. The dynamic feeding optimization module synchronously determines the target feed amount data and the nutrient ratio adjustment coefficient data based on the multi-objective optimization function; The dynamic feeding optimization module prioritizes meeting the preset feed nutritional balance goals during the multi-objective optimization function solution process. It also dynamically adjusts feed and formula based on real-time statistical results of four behavioral states: eating, drinking, leaving the trough, and resting, minimizing the cost per unit weight gain. Combined with the constraints of the multi-objective optimization function, the final target feed amount data and nutrient ratio adjustment coefficient data are output.

7. The deep learning-based intelligent feed proportioning system according to claim 6, characterized in that: The multi-objective optimization function is constrained by the following optimization objectives: The target feed rate is determined by multiplying the predicted feeding rate by the proportion of time spent in the feeding behavior state. If the proportion of time spent in the feeding behavior state is lower than the historical average, the target feed rate is adjusted downwards, and vice versa. If the proportion of drinking behavior increases significantly, the system will determine it is a high temperature or stress state based on environmental data, and increase the electrolyte and energy components in the nutritional ratio accordingly; If the proportion of away from the trough or resting behavior state increases, it indicates that the feeding activity has decreased. The target feed amount will be lowered based on the weighted comprehensive result of the feeding rate prediction data and the proportion of feeding behavior state time. At the same time, the nutrient ratio adjustment coefficient will increase the protein ratio and reduce the energy ratio.

8. The intelligent feed proportioning system based on deep learning according to claim 7, characterized in that: The nutrient ratio adjustment coefficient data is set as: When the proportion of feeding behavior is high and the feeding rate is higher than the recommended standard, the nutrient ratio adjustment coefficient data is to increase the protein and energy coefficients; When the proportion of drinking behavior is higher than the historical average, the nutritional ratio adjustment coefficient data is to increase the proportion of electrolytes and vitamins; When the proportion of off-trough and resting behavior states is higher than the threshold, the nutritional ratio adjustment coefficient data is to reduce energy and increase the vitamin ratio.

Images