Artificial intelligence 3D sheep body size prediction system

Abstract

The invention discloses an artificial intelligence 3D sheep body size prediction system, which fuses a generative adversarial network and a Transform model, acquires and processes data through a multi-dimensional data acquisition and feature mapping unit, expands the data through the generative adversarial network, extracts advanced features through Transform, and establishes a relationship between features and sheep body sizes through a feature fusion and scale association unit. And in combination with prediction model training optimization, real-time prediction output and system monitoring adjustment, accurate prediction is realized. A generator and a discriminator in the generative adversarial network optimize sheep body features and size distribution, Transform improves a multi-head attention mechanism, feature fusion is based on a sheep body anatomical structure, training adopts a dynamic loss function, output has a growth stage calibration mechanism, system monitoring can dynamically adjust parameters, and an efficient intelligent solution is provided for sheep body size measurement.

Description

Technical Field

[0001] The present invention relates to the field of sheep growth measurement, and in particular to an artificial intelligence 3D sheep body size prediction system. Background Art

[0002] As the livestock industry transitions toward intelligent and digital technologies, accurately capturing sheep body measurement data has become a critical requirement for animal husbandry management, disease prevention and control, and breeding optimization. Traditional sheep body measurement relies on manual labor, requiring breeders to measure each sheep individually using tools like tape measures and calipers. This is not only time-consuming and labor-intensive, but also prone to data errors due to sheep's lack of cooperation during the measurement process. Furthermore, manual measurement is inefficient, making it difficult to implement large-scale, real-time monitoring, and unable to meet the modern livestock industry's demand for data timeliness and accuracy.

[0003] On a technical level, existing automated measurement technologies are mostly based on single sensors or simple algorithms. For example, some systems use ultrasonic sensors for distance measurement, but this method is significantly affected by environmental factors and prone to signal interference in complex farming environments, leading to measurement errors. Other systems based on image recognition can only process two-dimensional image information, making it difficult to obtain three-dimensional dimensional data of the sheep, and thus fail to fully reflect the sheep's true form. These technologies have limitations when processing the complex characteristics of the sheep's body and cannot effectively address issues such as the sheep's changing posture and hair obstruction.

[0004] From the perspective of model application, existing prediction models lack a deep understanding of the biological characteristics of sheep. Most models fail to fully consider the correlations between the dimensions of various body parts and the changing patterns of sheep morphology at different growth stages, resulting in significant deviations between predictions and actual conditions. Furthermore, traditional models lack data processing capabilities, making it difficult to extract the potential information from massive amounts of data, making them unable to meet the demand for accurate sheep size prediction in the modern big data environment of animal husbandry. Summary of the Invention

[0005] In order to overcome the shortcomings and deficiencies of the prior art, the present invention provides an artificial intelligence 3D sheep body size prediction system.

[0006] Artificial intelligence 3D sheep body size prediction system, the system includes:

[0007] Multi-dimensional data acquisition and feature mapping unit, which uses a specific sensor array to collect multi-angle and multi-dimensional data of the sheep's body, and converts the collected raw data into feature vectors suitable for subsequent processing according to preset mapping rules;

[0008] Generate adversarial network infrastructure unit, build an adversarial structure consisting of a generator and a discriminator. The generator receives input data and generates simulated data through multi-layer nonlinear transformations. The discriminator distinguishes the authenticity of the input data from the generated data. The two compete with each other to optimize network parameters.

[0009] The Transformer encoding and feature extraction unit inputs the processed feature vector into the Transformer architecture, uses a multi-head attention mechanism to model long-range dependencies in the data, and extracts high-level features through a multi-layer encoder;

[0010] The feature fusion and scale association unit fuses the features output by the Transformer encoding and feature extraction unit with the features generated by the generative adversarial network infrastructure unit to establish an association between the features and the sheep's body size;

[0011] The prediction model training and optimization unit uses a specific loss function combined with the back-propagation algorithm to train and optimize the fused model and adjust network parameters to improve prediction accuracy;

[0012] The real-time prediction and result output unit receives real-time input data, performs predictions based on the trained model, and outputs the prediction results in the specified format;

[0013] The system monitoring and parameter adjustment unit monitors the system operation status in real time and dynamically adjusts the parameters in the system according to the preset monitoring indicators and adjustment strategies. The units are connected and work in an orderly manner through data transmission interfaces and control signal channels.

[0014] Furthermore, the generator of the generative adversarial network infrastructure unit adopts an improved generator structure based on sheep body features, and its model formula is: Among them, x is the input sheep body feature vector, θ g is the parameter set of the generator, n is the number of layers of the generator, W i is the weight matrix of the i-th layer, f i (x;θ g ) is the activation function of the i-th layer, b i is the bias vector of the i-th layer, σ is the activation function, and the structure is optimized for the body shape and contour features of the sheep by adjusting the parameter θ g , so that the generator can generate simulated data that is more consistent with the actual characteristics of the sheep body.

[0015] Furthermore, the generative adversarial network infrastructure unit uses a discriminator model based on the sheep body size distribution, and its model formula is: Among them, y is the input data (including real data and generated data), θd is the parameter set of the discriminator, m is the number of layers of the discriminator, V j is the weight matrix of the jth layer, g j (y;θ d ) is the activation function of the jth layer, c j is the bias vector of the jth layer. The model is based on the characteristics of the actual distribution of sheep body size by adjusting the parameter θ d The training of the discriminator can enhance its ability to distinguish between real sheep data and generated data.

[0016] Furthermore, the Transformer encoding and feature extraction unit and the multi-head attention mechanism adopt an improved form for the correlation of sheep body features. When calculating the attention weight, the correlation coefficient of sheep body parts is introduced. The formula is: Among them, Q, K, V are query, key, and value matrices respectively, d k is the dimension of the key, and α is the correlation coefficient matrix of the sheep body parts. This coefficient matrix is pre-set according to the size correlation relationship between different parts of the sheep body. Through this improvement, the Transformer can more accurately extract the correlation information between the features of various parts of the sheep body.

[0017] Furthermore, the feature fusion and scale association unit adopts a feature fusion algorithm based on the anatomical structure of the sheep, and fuses the high-level features output by the Transformer encoding and feature extraction unit and the simulated features generated by the generative adversarial network infrastructure unit in a hierarchical manner according to the anatomical structure level of the sheep. The features of the sheep's head, torso, and limbs are first fused separately, and then the overall fusion is performed to establish a multi-dimensional correlation between the features and the sizes of each part of the sheep's body and the overall size.

[0018] Furthermore, the prediction model training and optimization unit adopts a dynamically adjusted loss function, which is generated by the adversarial loss L of the adversarial network. adv and the regression loss L between the predicted results and the true value reg The weights of the two are dynamically adjusted according to the training stage, and the formula is: L = β1(t)L adv +β2(t)L reg , where β1(t) and β2(t) are weight coefficients that change with the training round t. In the early stage of training, the weight of the adversarial loss is increased to enable the generative adversarial network to better learn the distribution of sheep body features. In the later stage, the weight of the regression loss is increased to improve the prediction accuracy.

[0019] Furthermore, the real-time prediction and result output unit, when outputting the prediction result, adopts a result calibration mechanism based on the sheep's growth stage, calibrates the predicted sheep size result according to the sheep's age and growth cycle information, and adjusts the prediction result before outputting it by searching a pre-established correspondence table between the sheep's growth stage and the size correction coefficient.

[0020] Furthermore, the system monitoring and parameter adjustment unit establishes a mapping relationship library between system performance indicators and parameter adjustments. When the system's prediction time and resource occupancy performance indicators are monitored to exceed the preset range, the parameters of the generative adversarial network infrastructure unit, Transformer encoding and feature extraction unit are automatically adjusted according to the rules in the mapping relationship library. The adjustment methods include adjusting the learning rate and increasing or decreasing the number of network layers.

[0021] Furthermore, the multidimensional data acquisition and feature mapping unit adopts an adaptive sampling strategy during the data acquisition process, dynamically adjusts the sampling frequency and sampling angle of the sensor according to the movement state and feature change degree of the sheep's body, ensuring that the collected data can comprehensively and accurately reflect the sheep's body characteristics. In the feature mapping process, a nonlinear mapping function based on the distribution of sheep's body characteristics is adopted to convert the original data into a more discriminative feature vector.

[0022] Artificial intelligence 3D sheep body size prediction system, the system operation includes:

[0023] S100, starting the multi-dimensional data acquisition and feature mapping unit, collecting multi-angle and multi-dimensional data of the sheep through the sensor array, and converting the collected raw data into feature vectors according to preset rules;

[0024] S200, the feature vector is input into the basic architecture unit of the generative adversarial network. The generator generates simulated data through multi-layer nonlinear transformation. The discriminator distinguishes the authenticity of the input data and the generated data. The two compete with each other to optimize the network parameters.

[0025] S300, the processed feature vector is input into the Transformer encoding and feature extraction unit, which uses a multi-head attention mechanism and a multi-layer encoder to model the long-distance dependencies in the data and extract high-level features;

[0026] S400, transmitting the features output by the Transformer encoding and feature extraction unit and the features generated by the generative adversarial network infrastructure unit to the feature fusion and scale association unit, performing feature fusion and establishing an association relationship with the sheep body size;

[0027] S500, in the prediction model training and optimization unit, uses a specific loss function combined with a back-propagation algorithm to train and optimize the fused model and adjust network parameters;

[0028] S600, the real-time prediction and result output unit receives real-time input data, performs prediction using the trained model, and outputs the prediction result in a specified format;

[0029] S700, the system monitoring and parameter adjustment unit monitors the system operating status in real time and dynamically adjusts system parameters according to preset indicators and strategies.

[0030] Beneficial effects: The present invention proposes an artificial intelligence 3D sheep body size prediction system, which deeply combines the generative adversarial network with the Transformer model and performs customized optimization around the sheep body parameters, forming a significant technical advantage. The generator and discriminator in the generative adversarial network are optimized according to the characteristics of the sheep's body shape, contour and size distribution. The generator can generate simulated data that highly restores the actual characteristics of the sheep's body, expanding the diversity of data samples; the discriminator strengthens the ability to distinguish between real sheep data and generated data, ensuring data quality. The Transformer model fully considers the size correlation between various parts of the sheep's body by introducing the correlation coefficient of sheep body parts, and improves the multi-head attention mechanism, so that it can accurately capture the long-distance dependency of the features of various parts of the sheep's body and extract more representative high-level features. In the feature fusion link, layered fusion is performed based on the anatomical structure of the sheep's body, and a multi-dimensional size correlation relationship is established from the local features of the sheep's head, trunk, limbs and other parts to the overall features, deepening the understanding of the relationship between sheep body features and size. During training, the weights of the GAN's adversarial loss and prediction regression loss are dynamically adjusted based on the characteristics of the sheep data. This helps the GAN learn the distribution of sheep characteristics in the early stages of training, while focusing on improving prediction accuracy in the later stages. Furthermore, the system calibrates predictions based on the sheep's growth stages and dynamically adjusts each unit's parameters based on system performance indicators, ensuring the system can quickly and accurately predict sheep size, providing reliable technical support for intelligent livestock management. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 A flow chart for the method operation of the present invention;

[0032] Figure 2 This is a diagram illustrating the method operation unit implementation of the present invention. DETAILED DESCRIPTION

[0033] It should be noted that, unless there is a conflict, the embodiments in this application and the features described in the embodiments can be combined with each other. The application is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0034] like Figure 1 As shown, 1. An artificial intelligence 3D sheep body size prediction system, characterized in that the system includes:

[0035] Multi-dimensional data acquisition and feature mapping unit, which uses a specific sensor array to collect multi-angle and multi-dimensional data of the sheep's body, and converts the collected raw data into feature vectors suitable for subsequent processing according to preset mapping rules;

[0036] Specifically, this unit serves as the data entry point for the AI-powered 3D sheep size prediction system, responsible for acquiring raw data and performing preliminary processing. In actual applications, the specific sensor array includes multiple sensors, such as multi-angle cameras and lidar. The multi-angle cameras can capture images of the sheep's appearance from different angles, while the lidar can accurately measure the distance between the sheep and the sensor, thereby obtaining three-dimensional spatial data of the sheep. By working together, these sensors can collect multi-angle and multi-dimensional data on the sheep's body, ensuring that the collected data fully covers the sheep's morphological characteristics, such as its outline, size, and the relative positions of various body parts.

[0037] After acquiring the raw data, the unit processes it according to preset mapping rules. These mapping rules, developed based on in-depth research and analysis of sheep body features, aim to convert the raw data into feature vectors suitable for subsequent processing. For example, for collected sheep body image data, the mapping rules extract key feature points in the image, such as key points on the head outline and limb joints, and convert the coordinate information of these feature points into elements of the feature vector. For distance data collected by the lidar, the distance information at different locations is logically integrated into a feature vector. Through this conversion, the useful information contained in the raw data is refined and structured, making it easier for subsequent units to analyze and process.

[0038] The multidimensional data collection and feature mapping unit is of great significance, providing a reliable data foundation for the entire prediction system. Accurate and comprehensive data collection and appropriate feature mapping maximize the preservation of the sheep's true body characteristics, avoiding information loss due to missing or improper data processing. This lays a solid foundation for the effective operation of subsequent generative adversarial networks and Transformer models. If the data collected by this unit is incomplete or the feature mapping is inaccurate, even the most advanced models will struggle to accurately predict sheep body size.

[0039] Generate adversarial network infrastructure unit, build an adversarial structure consisting of a generator and a discriminator. The generator receives input data and generates simulated data through multi-layer nonlinear transformations. The discriminator distinguishes the authenticity of the input data from the generated data. The two compete with each other to optimize network parameters.

[0040] Specifically, the Generative Adversarial Network infrastructure unit plays a core role in data augmentation and feature learning in the system. This unit constructs an adversarial structure consisting of a generator and a discriminator, which compete with each other and evolve co-evolved. The generator receives input data processed by the multi-dimensional data acquisition and feature mapping unit, learns and reconstructs the data through multiple layers of nonlinear transformations, and generates simulated data. For example, when processing sheep data, the generator will learn the distribution characteristics of real sheep data, including morphological features such as the sheep's body shape and outline, and then generate similar simulated sheep data. These simulated data can, to a certain extent, expand the number of samples of the original data and enrich the diversity of the data.

[0041] The discriminator's primary responsibility is to distinguish the authenticity of input data (both real and generated data). By analyzing and judging the data's features, it outputs a probability value, representing the likelihood that the input data is real. During training, the generator continuously optimizes its parameters, attempting to generate more realistic simulated data to deceive the discriminator; the discriminator, in turn, continuously improves its discriminative ability, striving to accurately distinguish between real and generated data. In this adversarial process, the two constantly adjust network parameters, resulting in increasingly higher quality data generated by the generator and stronger discriminative capabilities. For example, as training progresses, the simulated sheep data produced by the generator will become increasingly similar to real sheep data, and the discriminator will be able to more accurately distinguish between real sheep data and generated simulated data.

[0042] The introduction of the Generative Adversarial Network (GAN) infrastructure effectively addresses the issues of insufficient raw data samples and insufficient data diversity. In the sheep size prediction scenario, because the actual collected sheep data may be limited in quantity and unevenly distributed, the simulated data generated by the GAN can supplement the original data, allowing subsequent Transformer encoding and feature extraction units to learn more comprehensive sheep feature information, thereby improving the accuracy and robustness of the entire system's sheep size prediction.

[0043] The Transformer encoding and feature extraction unit inputs the processed feature vector into the Transformer architecture, uses a multi-head attention mechanism to model long-range dependencies in the data, and extracts high-level features through a multi-layer encoder;

[0044] Specifically, the Transformer encoding and feature extraction unit is a key module for high-level feature extraction in the system. Its core lies in utilizing a multi-head attention mechanism to model long-range dependencies in the data. When processing sheep data, complex size and morphological relationships exist between various parts of the sheep's body. For example, the length and height of the sheep, and the length of the limbs and the size of the torso all influence each other. The multi-head attention mechanism can analyze and capture these relationships from different perspectives. It splits the input feature vector into multiple subspaces and independently calculates attention weights in each subspace, thereby comprehensively mining the dependencies between different locations in the data. For example, when analyzing sheep data, the multi-head attention mechanism can simultaneously focus on the connection between the sheep's head and torso, the proportional relationship between the limbs and the torso, and other aspects, avoiding missing important feature association information.

[0045] Through multi-layer encoder processing, the Transformer encoding and feature extraction unit is able to gradually extract higher-level, more abstract features. Each layer of encoders further processes and refines the input features. As the number of layers increases, the extracted features gradually transition from raw surface features to more representative and discriminative high-level features. In sheep body size prediction, these high-level features can more accurately reflect the morphological characteristics and size information of the sheep. For example, after processing through multi-layer encoders, key features such as an abstract representation of the sheep's overall shape and the proportional relationship between the relative sizes of various parts can be extracted. These features are crucial for accurately predicting sheep body size.

[0046] The existence of this unit enables the system to deeply understand the complex relationships and characteristic patterns in sheep body data. Compared with traditional feature extraction methods, the Transformer encoding and feature extraction unit can more effectively process long sequence data and complex feature associations, providing high-quality feature information for subsequent feature fusion and sheep body size prediction, significantly improving the system's prediction performance and ability to express sheep body characteristics.

[0047] The feature fusion and scale association unit fuses the features output by the Transformer encoding and feature extraction unit with the features generated by the generative adversarial network infrastructure unit to establish an association between the features and the sheep's body size;

[0048] Specifically, the feature fusion and scale association unit is responsible for integrating the high-level features output by the Transformer encoding and feature extraction unit with the features generated by the generative adversarial network infrastructure unit, and establishing a correlation between them and the sheep's body dimensions. During the fusion process, a feature fusion algorithm based on the sheep's anatomical structure is used, and layered fusion is performed according to the sheep's anatomical structure. First, the features of the sheep's head, torso, limbs, and other parts are fused separately. For example, the shape features of the head and morphological features of the ears are integrated at the head level to make the features of each part more complete and accurate. Then, a holistic fusion is performed, combining the fused features of each part to form a comprehensive description of the sheep's overall morphology. This layered fusion approach fully considers the characteristics and relationships of each part of the sheep's body, avoids feature confusion, and makes the fused features more logical and structured.

[0049] To establish the relationship between features and sheep dimensions, the unit studies and analyzes a large number of samples labeled with actual sheep dimensions, uncovering potential connections between features and dimensions. For example, by comprehensively analyzing the length and width of the sheep's torso, as well as the length of its limbs, it establishes a mathematical model or mapping relationship between these features and dimensions such as length, width, and height. Once this relationship is established, the system can quickly and accurately predict various dimensions of the sheep based on these relationships when receiving new feature data.

[0050] The feature fusion and scale association unit bridges feature extraction and size prediction. It organically integrates features from different sources and establishes a direct connection with the actual application goal (sheep size prediction). In this way, the system can accurately infer sheep size information from complex feature data, improving the accuracy and reliability of predictions and providing strong support for sheep growth assessment and decision-making in livestock breeding management.

[0051] The prediction model training and optimization unit uses a specific loss function combined with the back-propagation algorithm to train and optimize the fused model and adjust network parameters to improve prediction accuracy;

[0052] Specifically, the prediction model training and optimization unit is responsible for training and adjusting the parameters of the fused model to improve the accuracy of the system's sheep body size predictions. This unit uses a specific loss function to measure the difference between the model's predictions and the actual sheep body size data. This difference information is fed back into the model's parameters via a backpropagation algorithm, thereby optimizing the model parameters. The loss function's design fully considers the characteristics of generative adversarial networks and Transformer models, as well as the practical needs of sheep body size prediction. During training, the loss function calculates a value based on the deviation between the model's predictions and the actual data, reflecting the model's current prediction error.

[0053] The backpropagation algorithm propagates the error calculated by the loss function from the output layer to the input layer along the model's network structure. During this process, the contribution of each parameter to the error is calculated and the parameter value is adjusted accordingly. For example, if the backpropagation process finds that a certain weight parameter in the Transformer encoding and feature extraction unit contributes significantly to the error, indicating that the current value of the parameter is not reasonable, the parameter will be adjusted to reduce the error. By repeating this process, the model parameters are gradually optimized, making the model's predictions increasingly close to the actual sheep body size data.

[0054] The effective operation of the prediction model training and optimization unit is crucial for the system to achieve accurate predictions. By continuously adjusting model parameters, the model can better adapt to the characteristics and patterns of sheep body data, improving the accuracy and generalization of the model's predictions of sheep body size. Only a fully trained and optimized model can accurately predict sheep body size for data from different individuals and at different growth stages, meeting the practical needs of animal husbandry.

[0055] The real-time prediction and result output unit receives real-time input data, performs predictions based on the trained model, and outputs the prediction results in the specified format;

[0056] Specifically, the real-time prediction and result output unit is a key interface for interaction between the system and the user. It is responsible for receiving real-time input data and outputting the sheep body size results predicted by the trained model in a specified format. In actual application, when new sheep body data is input into the system, this unit immediately passes the data to the trained model for prediction. Whether the sheep body data is collected in real time by the multi-dimensional data acquisition and feature mapping unit or obtained from other data sources, it can quickly enter the prediction process. The model analyzes and calculates the input data based on the relationship between previously trained features and sheep body size, and produces a predicted sheep body size result.

[0057] When outputting prediction results, the unit uses a result calibration mechanism based on the sheep's growth stage. Sheep at different growth stages have different growth rates and morphological changes, so the prediction results need to be calibrated to improve accuracy. Based on information such as the sheep's age and growth cycle, the unit searches a pre-established correspondence table between sheep growth stages and size correction coefficients and adjusts the prediction results. For example, for young sheep, which grow faster, the prediction results may need to be appropriately adjusted based on the correction coefficient to make them more consistent with actual growth conditions. The adjusted results will be output in a specified format for user viewing and use. The output format can be tabular, graphical, or other forms that are easy to understand and analyze.

[0058] The real-time prediction and output unit completes a closed-loop system from data input to output, providing users with convenient, real-time sheep size prediction services. Without complex operations, users simply input sheep data into the system to quickly obtain accurate sheep size predictions. This has important practical value for daily management, growth monitoring, and production decision-making in the livestock breeding process.

[0059] The system monitoring and parameter adjustment unit monitors the system operation status in real time and dynamically adjusts the parameters in the system according to the preset monitoring indicators and adjustment strategies. The units are connected and work in an orderly manner through data transmission interfaces and control signal channels.

[0060] Specifically, the system monitoring and parameter adjustment unit is like the "smart butler" of the entire system, monitoring the system's operating status in real time and dynamically adjusting the parameters in the system according to preset monitoring indicators and adjustment strategies. During the operation of the system, the unit will continuously monitor multiple key performance indicators, such as prediction time, resource utilization (including CPU utilization, memory utilization, etc.), model prediction error, etc. For example, when it is found that the prediction time is too long, it may mean that the system's computing efficiency is reduced, or the model complexity is high, resulting in excessive calculations; when the resource utilization rate is too high, it may affect the stability of the system and the normal operation of other tasks.

[0061] Once the system's performance indicators are monitored to be outside the preset range, the system monitoring and parameter adjustment unit will automatically adjust the parameters of the relevant units based on a pre-established library of mapping relationships between system performance indicators and parameter adjustments. There are various adjustment methods, including but not limited to adjusting the learning rate and increasing or decreasing the number of network layers. If the model prediction error is found to be large, the learning rate may be appropriately reduced to make the model more stable during training, and the parameters may be gradually optimized to improve prediction accuracy. If the system resource utilization rate is too high, the number of network layers in the Generative Adversarial Network Infrastructure Unit or the Transformer Encoding and Feature Extraction Unit may be reduced to reduce model complexity and thus reduce resource consumption.

[0062] The system monitoring and parameter adjustment unit ensures the system maintains efficient and stable operation. Through real-time monitoring and dynamic adjustment, it can promptly address any issues that arise during system operation, ensuring that the system maintains optimal performance across diverse environments and data conditions. This provides reliable technical support for sheep size prediction, extends the system's lifespan, and enhances the user experience.

[0063] Preferably, the generator of the generative adversarial network infrastructure unit adopts an improved generator structure based on sheep body features, and its model formula is: Among them, x is the input sheep body feature vector, θ g is the parameter set of the generator, n is the number of layers of the generator, W i is the weight matrix of the i-th layer, f i (x;θ g ) is the activation function of the i-th layer, b i is the bias vector of the i-th layer, σ is the activation function, and the structure is optimized for the body shape and contour features of the sheep by adjusting the parameter θ g , so that the generator can generate simulated data that is more consistent with the actual characteristics of the sheep body.

[0064] Specifically, improvements were made to the generator within the generative adversarial network infrastructure unit, and a dedicated structure based on sheep body features was constructed. This structure is closely designed around key morphological features of the sheep, such as body shape and outline, and the input sheep body feature vector is deeply processed through multiple layers of nonlinear transformations. During actual training, the generator continuously adjusts its own parameters to learn the distribution patterns of real sheep body data, such as the proportions of the sheep's torso, the length and thickness of the limbs, and other features. Through continuous optimization, the generator is able to output simulated data that is highly similar to real sheep body features, greatly expanding the original data sample. In sheep body size prediction, when the original data samples are small or the data diversity is insufficient, these simulated data can effectively supplement the training data, helping subsequent models learn more comprehensive sheep body feature information and improving the prediction system's adaptability to the characteristics of different individual sheep.

[0065] Preferably, the generative adversarial network infrastructure unit uses a discriminator model based on the sheep body size distribution, and its model formula is: Among them, y is the input data (including real data and generated data), θ d is the parameter set of the discriminator, m is the number of layers of the discriminator, V j is the weight matrix of the jth layer, g j (y;θ d ) is the activation function of the jth layer, c jis the bias vector of the jth layer. The model is based on the characteristics of the actual distribution of sheep body size by adjusting the parameter θ d The training of the discriminator can enhance its ability to distinguish between real sheep data and generated data.

[0066] Specifically, the discriminator model based on the sheep body size distribution is another important optimization of the basic architecture unit of the generative adversarial network. This model fully considers the distribution characteristics of sheep body size in actual scenarios, and performs feature analysis and authenticity judgment on the input data (including real sheep body data and simulated data output by the generator) through a multi-layer network structure. During the training process, the discriminator adjusts parameters according to the actual distribution law of sheep body size, such as the common body length and height range of different breeds of sheep, so as to enhance the ability to discriminate the authenticity of the data. When there is a difference between the simulated data output by the generator and the actual sheep body size distribution, the discriminator can quickly identify and feedback, prompting the generator to optimize the output, ensuring that the data generated by the entire generative adversarial network is more in line with the actual sheep body characteristics, and providing reliable data support for subsequent sheep body size predictions.

[0067] Preferably, the Transformer encoding and feature extraction unit and the multi-head attention mechanism adopt an improved form for the correlation of sheep body features. When calculating the attention weight, the correlation coefficient of sheep body parts is introduced. The formula is: Among them, Q, K, V are query, key, and value matrices respectively, d k is the dimension of the key, and α is the correlation coefficient matrix of the sheep body parts. This coefficient matrix is pre-set according to the size correlation relationship between different parts of the sheep body. Through this improvement, the Transformer can more accurately extract the correlation information between the features of various parts of the sheep body.

[0068] Specifically, the multi-head attention mechanism of the Transformer encoding and feature extraction unit has been innovated to introduce correlation coefficients for sheep body parts. There are close size correlations between various sheep body parts, such as head size and torso length, limb spacing and body width. These coefficient matrices are pre-set based on the sheep's anatomical structure and actual size relationships. When calculating attention weights, the mechanism comprehensively considers these correlation coefficients, capturing the long-range dependencies between the features of various sheep body parts from different perspectives. For example, when analyzing sheep image data, it not only focuses on the features of individual parts, but also accurately grasps the relative position and size ratios between each part, extracting more representative high-level features. This improvement enables the Transformer model to have a deeper understanding of sheep body characteristics, effectively improving its ability to process complex sheep morphological data and laying a solid foundation for accurately predicting sheep body size.

[0069] Preferably, the feature fusion and scale association unit adopts a feature fusion algorithm based on the anatomical structure of the sheep, and fuses the high-level features output by the Transformer encoding and feature extraction unit and the simulated features generated by the generative adversarial network infrastructure unit in layers according to the anatomical structure level of the sheep. The features of the sheep's head, trunk, and limbs are first fused separately, and then the overall fusion is performed to establish a multi-dimensional correlation relationship between the features and the sizes of each part of the sheep's body and the overall size.

[0070] Specifically, a feature fusion algorithm based on the sheep's anatomical structure is applied to the feature fusion and scale association unit. This algorithm combines the high-level features output by the Transformer encoding and feature extraction unit with the simulated features generated by the generative adversarial network infrastructure unit, performing a hierarchical fusion based on the sheep's head, torso, limbs, and other anatomical structures. First, the features of each part are integrated separately, such as the shape of the head and the form of the ear, to make the features of each part more complete. Then, a comprehensive fusion is performed to construct a complete morphological description of the sheep. Simultaneously, by analyzing a large amount of annotated data, the algorithm establishes multi-dimensional correlations between these fused features and the dimensions of each part and the overall size of the sheep. In practical applications, whether measuring sheep length and height or assessing the proportions of various body parts, this algorithm accurately predicts sheep dimensions through feature fusion and correlation analysis, providing precise data support for livestock breeding and management.

[0071] Preferably, the prediction model training and optimization unit adopts a dynamically adjusted loss function, which is generated by the adversarial loss L of the adversarial network. adv and the regression loss L between the predicted results and the true value reg The weights of the two are dynamically adjusted according to the training stage, and the formula is: L = β1(t)L adv +β2(t)L reg , where β1(t) and β2(t) are weight coefficients that change with the training round t. In the early stage of training, the weight of the adversarial loss is increased to enable the generative adversarial network to better learn the distribution of sheep body features. In the later stage, the weight of the regression loss is increased to improve the prediction accuracy.

[0072] Specifically, a dynamically adjusted loss function is applied to the prediction model training and optimization unit. It consists of the adversarial loss of the generative adversarial network and the regression loss between the predicted result and the true value, and the weights of the two change dynamically with the training stage. In the early stages of training, the weight of the adversarial loss is increased to encourage the generative adversarial network to focus on learning the distribution of sheep body features, allowing the generator to generate simulated data that is more in line with reality; as training progresses, the weight of the regression loss is gradually increased to focus the model on improving the accuracy of the prediction results. For example, at the beginning of training, by enhancing the adversarial loss, the generator is guided to quickly capture the basic morphological characteristics of the sheep body; in the later stage, the regression loss is strengthened to fine-tune the prediction results, ensuring that the model can accurately predict the sheep body size while learning the data distribution. This dynamic adjustment strategy makes model training more targeted and effectively improves the overall performance of the prediction system.

[0073] Preferably, the real-time prediction and result output unit, when outputting the prediction result, adopts a result calibration mechanism based on the sheep growth stage, calibrates the predicted sheep body size result according to the sheep's age and growth cycle information, and adjusts the prediction result before outputting it by searching a pre-established correspondence table between the sheep growth stage and the size correction coefficient.

[0074] Specifically, a result calibration mechanism based on the sheep's growth stage is the core technology of the real-time prediction and result output unit. Sheep growth rates and morphological changes vary significantly at different growth stages. Young sheep grow rapidly and their body shape varies greatly, while adult sheep remain relatively stable. This mechanism pre-establishes a correspondence table between sheep growth stages and size correction coefficients based on information such as the sheep's age and growth cycle. When outputting prediction results, the system retrieves the correction coefficient from the correspondence table based on the sheep's growth stage and calibrates the predicted sheep size results. For example, for lambs in a rapid growth period, the prediction result may be appropriately adjusted upward based on the correction coefficient to make it more consistent with actual growth trends. This calibration mechanism effectively improves the accuracy and practicality of the prediction results, providing a reliable basis for farmers to scientifically assess the growth status of sheep.

[0075] Preferably, the system monitoring and parameter adjustment unit establishes a mapping relationship library between system performance indicators and parameter adjustments. When the system's prediction time and resource occupancy performance indicators are monitored to exceed a preset range, the parameters of the generative adversarial network infrastructure unit, Transformer encoding and feature extraction unit are automatically adjusted according to the rules in the mapping relationship library. The adjustment methods include learning rate adjustment and increase or decrease in the number of network layers.

[0076] Specifically, the system monitoring and parameter adjustment unit realizes intelligent management of the system by establishing a mapping relationship library between system performance indicators and parameter adjustments. The unit monitors key performance indicators such as the system's prediction time, resource occupancy (such as CPU usage, memory usage, etc.), and model prediction error in real time. When an indicator exceeds the preset range, for example, the prediction time is too long, affecting real-time performance, or the resource occupancy is too high, causing unstable system operation, the unit will automatically adjust the parameters of related units such as the generative adversarial network infrastructure unit, Transformer encoding and feature extraction unit according to the rules in the mapping relationship library. Adjustment methods include adjusting the learning rate, increasing or decreasing the number of network layers, etc. For example, if the model training speed is found to be too slow, the learning rate can be appropriately increased to speed up the training process; if the resource occupancy is too high, the number of network layers can be reduced to reduce the complexity of the model, ensuring that the system is always in an efficient and stable operating state, providing reliable technical support for sheep body size prediction.

[0077] Preferably, the multidimensional data acquisition and feature mapping unit adopts an adaptive sampling strategy during the data acquisition process, dynamically adjusts the sampling frequency and sampling angle of the sensor according to the movement state and feature change degree of the sheep body, ensures that the collected data can fully and accurately reflect the sheep body characteristics, and adopts a nonlinear mapping function based on the sheep body feature distribution during the feature mapping process to convert the original data into a more discriminative feature vector.

[0078] Specifically, the multidimensional data acquisition and feature mapping unit was optimized, adopting an adaptive sampling strategy and a nonlinear mapping function based on the distribution of sheep body features. During the data acquisition process, the system dynamically adjusts the sensor's sampling frequency and sampling angle according to the sheep's motion state and the degree of feature change. When the sheep is actively moving, the sampling frequency is increased to capture more dynamic features; for areas with obvious feature changes, the sampling angle is adjusted to obtain clearer data. In the feature mapping process, the nonlinear mapping function based on the distribution of sheep body features can convert the raw data into a more discriminative feature vector. For example, for feature blurring caused by occlusion of sheep hair, this function can highlight key features and remove interfering information through nonlinear transformation. This adaptive sampling and mapping strategy ensures that the collected data is comprehensive and accurate, and the converted feature vectors are more conducive to subsequent model processing, providing high-quality data input for the entire sheep body size prediction system.

[0079] like Figure 2 As shown, the artificial intelligence 3D sheep body size prediction system includes:

[0080] S100, starting the multi-dimensional data acquisition and feature mapping unit, collecting multi-angle and multi-dimensional data of the sheep through the sensor array, and converting the collected raw data into feature vectors according to preset rules;

[0081] S200, the feature vector is input into the basic architecture unit of the generative adversarial network. The generator generates simulated data through multi-layer nonlinear transformation. The discriminator distinguishes the authenticity of the input data and the generated data. The two compete with each other to optimize the network parameters.

[0082] S300, the processed feature vector is input into the Transformer encoding and feature extraction unit, which uses a multi-head attention mechanism and a multi-layer encoder to model the long-distance dependencies in the data and extract high-level features;

[0083] S400, transmitting the features output by the Transformer encoding and feature extraction unit and the features generated by the generative adversarial network infrastructure unit to the feature fusion and scale association unit, performing feature fusion and establishing an association relationship with the sheep body size;

[0084] S500, in the prediction model training and optimization unit, uses a specific loss function combined with a back-propagation algorithm to train and optimize the fused model and adjust network parameters;

[0085] S600, the real-time prediction and result output unit receives real-time input data, performs prediction using the trained model, and outputs the prediction result in a specified format;

[0086] S700, the system monitoring and parameter adjustment unit monitors the system operating status in real time and dynamically adjusts system parameters according to preset indicators and strategies.

[0087] Existing sheep body measurement technologies suffer from low efficiency, large errors, and poor model adaptability. The AI-powered 3D sheep body size prediction system, with its innovative architecture and technology, specifically addresses these pain points. I will explain its breakthroughs over traditional technologies from the perspectives of data collection, model processing, and system operation.

[0088] At the data collection level, traditional manual measurement and single-sensor technology have obvious shortcomings. Manual measurement relies on farmers using tools such as tape measures and calipers to operate one by one. This is not only inefficient, but also affected by factors such as sheep's lack of cooperation and human error, making data accuracy difficult to guarantee. Single sensor technologies, such as ultrasonic sensors, are susceptible to environmental interference, and image recognition technology cannot obtain three-dimensional information, resulting in missing or inaccurate data. The multi-dimensional data acquisition and feature mapping unit of the artificial intelligence 3D sheep body size prediction system uses a sensor array composed of multi-angle cameras, lidar, etc. to achieve multi-angle and multi-dimensional data acquisition. It also uses adaptive sampling strategies and nonlinear mapping functions to dynamically adjust the acquisition method according to the sheep's body movement state, accurately converting the raw data into feature vectors, effectively solving the problems of data acquisition efficiency and accuracy, and providing a high-quality data foundation for subsequent processing.

[0089] In terms of model processing and prediction, traditional models lack deep integration of the biological characteristics of sheep and are unable to fully consider the size correlations between various parts of the sheep and the patterns of changes during growth stages, resulting in significant deviations in prediction results. Furthermore, traditional models lack data processing capabilities, making it difficult to mine the potential information within massive amounts of data. This system, however, combines a generative adversarial network (GAN) with a Transformer model. The GAN's generator and discriminator are optimized for the sheep's body shape, contour, and size distribution, generating realistic simulated data and improving data quality. The Transformer model incorporates a correlation coefficient between sheep parts to improve the multi-head attention mechanism and accurately extract information related to the characteristics of various sheep parts. The feature fusion and scale association unit establishes multi-dimensional size correlations based on hierarchical fusion features of the sheep's anatomical structure. The predictive model training and optimization unit uses a dynamically adjusted loss function to improve model prediction accuracy, comprehensively enhancing the system's understanding and prediction capabilities of sheep characteristics.

[0090] From the perspective of overall system operation, traditional technologies lack real-time monitoring and intelligent adjustment mechanisms for system operating status. When problems such as performance degradation and excessive resource usage arise, they cannot be addressed promptly, affecting system stability and measurement accuracy. The system monitoring and parameter adjustment unit of the AI-powered 3D sheep size prediction system monitors key indicators such as prediction time and resource utilization in real time. If an anomaly is detected, it automatically adjusts relevant unit parameters based on a preset mapping library, such as adjusting the learning rate and increasing or decreasing the number of network layers, to ensure the system's efficient and stable operation. At the same time, the real-time prediction and result output unit calibrates the prediction results based on the sheep's growth stage, further improving the reliability of the measurement results and effectively overcoming the shortcomings of traditional technologies in system operation and management.

[0091] While embodiments of the present invention have been shown and described, it will be understood by those skilled in the art that various equivalent changes, modifications, substitutions and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. Artificial intelligence 3D sheep body size prediction system, characterized by: The system includes: The multi-dimensional data acquisition and feature mapping unit collects data from the sheep at multiple angles and dimensions through a sensor array, and converts the collected raw data into feature vectors suitable for subsequent processing based on preset mapping rules. Generate adversarial network infrastructure unit, build an adversarial structure consisting of a generator and a discriminator. The generator receives input data and generates simulated data through multi-layer nonlinear transformations. The discriminator distinguishes the authenticity of the input data from the generated data. The two compete with each other to optimize network parameters. The Transformer encoding and feature extraction unit inputs the processed feature vector into the Transformer architecture, uses a multi-head attention mechanism to model long-range dependencies in the data, and extracts high-level features through a multi-layer encoder; The feature fusion and scale association unit fuses the features output by the Transformer encoding and feature extraction unit with the features generated by the generative adversarial network infrastructure unit to establish an association between the features and the sheep's body size; The prediction model training and optimization unit uses a specific loss function combined with the back-propagation algorithm to train and optimize the fused model and adjust network parameters to improve prediction accuracy; The real-time prediction and result output unit receives real-time input data, performs predictions based on the trained model, and outputs the prediction results in the specified format; The system monitoring and parameter adjustment unit monitors the system operation status in real time and dynamically adjusts the parameters in the system according to the preset monitoring indicators and adjustment strategies. The units are connected and work together in an orderly manner through data transmission interfaces and control signal channels.

2. The artificial intelligence 3D sheep body size prediction system according to claim 1, characterized in that: The generator of the generative adversarial network infrastructure unit adopts an improved generator structure based on sheep body features, and its model formula is: Among them, x is the input sheep body feature vector, θ g is the parameter set of the generator, n is the number of layers of the generator, W i is the weight matrix of the i-th layer, f i (x;θ g ) is the activation function of the i-th layer, b i is the bias vector of the i-th layer, σ is the activation function, and the structure is optimized for the body shape and contour features of the sheep by adjusting the parameter θ g , so that the generator can generate simulated data that is more consistent with the actual characteristics of the sheep body.

3. The artificial intelligence 3D sheep body size prediction system according to claim 2, characterized in that: The generative adversarial network infrastructure unit uses a discriminator model based on the sheep body size distribution, and its model formula is: Among them, y is the input data (including real data and generated data), θ d is the parameter set of the discriminator, m is the number of layers of the discriminator, V j is the weight matrix of the jth layer, g j (y;θ d ) is the activation function of the jth layer, c j is the bias vector of the jth layer. The model is based on the characteristics of the actual distribution of sheep body size by adjusting the parameter θ d The training of the discriminator can enhance its ability to distinguish between real sheep data and generated data.

4. The artificial intelligence 3D sheep body size prediction system according to claim 1, characterized in that: The Transformer encoding and feature extraction unit, the multi-head attention mechanism adopts an improved form for the correlation of sheep body features. When calculating the attention weight, the correlation coefficient of sheep body parts is introduced. The formula is: Among them, Q, K, V are query, key, and value matrices respectively, d k is the dimension of the key, and α is the correlation coefficient matrix of the sheep body parts. This coefficient matrix is pre-set according to the size correlation relationship between different parts of the sheep body. Through this improvement, the Transformer can more accurately extract the correlation information between the features of various parts of the sheep body.

5. The artificial intelligence 3D sheep body size prediction system according to claim 1, characterized in that: The feature fusion and scale association unit uses a feature fusion algorithm based on the sheep's anatomical structure to fuse the high-level features output by the Transformer encoding and feature extraction unit and the simulated features generated by the generative adversarial network infrastructure unit in a hierarchical manner according to the sheep's anatomical structure. The features of the sheep's head, torso, and limbs are first fused separately, and then the overall fusion is performed to establish a multi-dimensional correlation between the features and the sizes of each part of the sheep's body, as well as the overall size.

6. The artificial intelligence 3D sheep body size prediction system according to claim 1, characterized in that: The prediction model training and optimization unit adopts a dynamically adjusted loss function, which is generated by the adversarial loss L of the adversarial network. adv and the regression loss L between the predicted results and the true value reg The weights of the two are dynamically adjusted according to the training stage, and the formula is: L = β1(t)L adv +β2(t)L reg , where β1(t) and β2(t) are weight coefficients that change with the training round t. In the early stage of training, the weight of the adversarial loss is increased to enable the generative adversarial network to better learn the distribution of sheep body features. In the later stage, the weight of the regression loss is increased to improve the prediction accuracy.

7. The artificial intelligence 3D sheep body size prediction system according to claim 1, characterized in that: The real-time prediction and result output unit, when outputting the prediction result, adopts a result calibration mechanism based on the sheep's growth stage, calibrates the predicted sheep's body size result according to the sheep's age and growth cycle information, and adjusts the prediction result before outputting it by searching a pre-established correspondence table between the sheep's growth stage and the size correction coefficient.

8. The artificial intelligence 3D sheep body size prediction system according to claim 1, characterized in that: The system monitoring and parameter adjustment unit establishes a mapping relationship library between system performance indicators and parameter adjustments. When the system's prediction time and resource occupancy performance indicators are monitored to exceed a preset range, the system automatically adjusts the parameters of the generative adversarial network infrastructure unit and the Transformer encoding and feature extraction unit according to the rules in the mapping relationship library. The adjustment methods include adjusting the learning rate and increasing or decreasing the number of network layers.

9. The artificial intelligence 3D sheep body size prediction system according to claim 1, characterized in that: During the data acquisition process, the multidimensional data acquisition and feature mapping unit adopts an adaptive sampling strategy to dynamically adjust the sampling frequency and sampling angle of the sensor according to the sheep's motion state and the degree of feature change, ensuring that the collected data can fully and accurately reflect the sheep's body characteristics. During the feature mapping process, a nonlinear mapping function based on the sheep's body feature distribution is used to convert the raw data into a more discriminative feature vector.

10. The artificial intelligence 3D sheep body size prediction system according to claim 1, characterized in that: The system operation includes: S100, starting the multi-dimensional data acquisition and feature mapping unit, collecting multi-angle and multi-dimensional data of the sheep through the sensor array, and converting the collected raw data into feature vectors according to preset rules; S200, the feature vector is input into the basic architecture unit of the generative adversarial network. The generator generates simulated data through multi-layer nonlinear transformation. The discriminator distinguishes the authenticity of the input data and the generated data. The two compete with each other to optimize the network parameters. S300, the processed feature vector is input into the Transformer encoding and feature extraction unit, which uses a multi-head attention mechanism and a multi-layer encoder to model the long-distance dependencies in the data and extract high-level features; S400, transmitting the features output by the Transformer encoding and feature extraction unit and the features generated by the generative adversarial network infrastructure unit to the feature fusion and scale association unit, performing feature fusion and establishing an association relationship with the sheep body size; S500, in the prediction model training and optimization unit, uses a specific loss function combined with a back-propagation algorithm to train and optimize the fused model and adjust network parameters; S600, the real-time prediction and result output unit receives real-time input data, performs prediction using the trained model, and outputs the prediction result in a specified format; S700, the system monitoring and parameter adjustment unit monitors the system operating status in real time and dynamically adjusts system parameters according to preset indicators and strategies.

Images