AI-driven predictive regulatory methods and systems for animal nutrition

Abstract

This invention provides an AI-driven method and system for predicting and monitoring animal nutrition, comprising: acquiring basic and physiological information of each animal; collecting the animal's dwell time at the feeding trough and obtaining dietary information corresponding to the dwell time by recording changes in the weight of the feeding trough; predicting the animal's nutrient intake information at the current time using a first AI model based on the basic, physiological, and dietary information; acquiring growth targets; predicting required nutrient intake using a second AI model based on the growth targets and the animal's nutrient intake information; and obtaining a nutrient intake plan based on the prediction results. This method enables non-invasive, intelligent, and accurate prediction and monitoring of animal nutrition.

Description

Technical Field

[0001] This invention relates to the field of data technology, and more specifically, to an AI-driven method and system for predicting and monitoring animal nutrition. Background Technology

[0002] In animal husbandry, the nutritional status of animals has a significant impact on their growth, reproduction, and health.

[0003] Existing animal nutrition status monitoring systems typically include modules for managing basic animal information, feed formulation, feed administration, and nutritional indicators. These modules help farmers and relevant departments better manage and monitor animal nutrition status, thereby improving farming efficiency and animal welfare.

[0004] However, existing animal nutrition status monitoring systems typically rely on manual observation and experience-based judgment, which suffers from subjectivity and inaccuracy. Moreover, manual assessment is inefficient and cannot monitor animal nutrition status in real time; it lacks dynamic nutritional requirements analysis and cannot provide targeted nutritional optimization solutions. Therefore, a system that can accurately assess animal nutrition status is needed to improve the efficiency and quality of animal husbandry. Summary of the Invention

[0005] In view of this, the present invention proposes an AI-driven method and system for predicting and monitoring animal nutrition to solve the problems existing in the prior art.

[0006] To achieve the above objectives, this invention proposes an AI-driven method for predictive supervision of animal nutrition, comprising: Each animal is tagged with an RFID tag, and the RFID tag and the corresponding basic and physiological information of each animal are stored. The RFID tag reader reads the RFID tags to collect the time the animals spend at the feeding trough, and obtains the dietary information corresponding to the time spent at the feeding trough by recording the weight changes of the feeding trough. The first AI model is used to predict basic information, physiological information and dietary information to obtain the animal's nutritional intake information at the current time. Obtain growth targets, and based on growth targets and animal nutrition intake information, use a second AI model to predict required nutrient intake, and obtain a nutrition intake plan based on the prediction results.

[0007] Optionally, the basic information includes the animal's species, breed, age, and sex; the physiological information includes weight, height, body length, heart rate, respiratory rate, growth status, and performance status; and the dietary information includes the type of feed, the proportion of nutrients, feed consumption, and water consumption.

[0008] Optionally, the first AI model is a deep learning model, wherein the first AI model is used to represent the mapping relationship between the basic information, physiological information, residence time, dietary information and animal nutritional intake information.

[0009] Optionally, the first AI model employs a dynamic dual-path adversarial correction method for prediction. The first AI model includes a prediction network and a correction network connected sequentially. The prediction network extracts deep features from basic information, physiological information, and dietary information, and decodes the extracted deep features to obtain predicted nutrient intake information. The correction network judges the predicted nutrient intake information and physiological information to obtain the predicted animal growth rate. The deviation between the predicted animal growth rate and the actual animal growth rate is calculated, and the output of the prediction network is corrected based on the deviation. The prediction network and the correction network undergo adversarial training.

[0010] Optional, adversarial training processes include: The first and second phases are executed repeatedly to complete adversarial training. The first phase includes acquiring sample data, fixing the model parameters of the prediction network, and training the calibration network using the sample data. The second phase includes fixing the model parameters of the calibration network and training the prediction network using the sample data. The comprehensive loss in the training of the prediction network is calculated. The comprehensive loss is a weighted sum of the supervision loss and the adversarial loss, where the supervision loss is the bias loss of the prediction network's output data and the adversarial loss is the bias loss of the calibration network's output data. The prediction network is then trained based on the comprehensive loss.

[0011] Optionally, in the prediction network, a temporal processing method is used to process dietary information to obtain behavioral embedding vectors. Basic information and physiological information are processed by introducing noise through latent variable encoding to obtain individual embedding vectors. The behavioral embedding vectors and individual embedding vectors are concatenated and nonlinearly transformed to obtain high-level features. The high-level features are then processed through a linear processing layer to obtain the animal's nutritional intake information at the current time.

[0012] Optionally, in the second AI model, based on auxiliary variables and historical data, a deep learning model is used to predict physiological indicators for future time series. The auxiliary variables include the required nutrient intake; the historical data includes basic information, physiological information, residence time, and dietary information at historical times; the auxiliary variables are adjusted according to the physiological indicators through an optimization method, and the execution of the deep learning model and optimization method is iterated to obtain the predicted nutrient intake result.

[0013] On the other hand, the present invention provides an AI-driven animal nutrition prediction and monitoring system for performing the above-described method.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention indirectly and accurately derives the true nutrient absorption that is difficult to measure directly by utilizing readily available animal behavior and growth data. This overcomes the problems of high cost, poor timeliness, and invasiveness associated with traditional methods that rely on manual sampling and experimental analysis. By employing AI models for multi-objective optimization, this invention can develop future nutritional needs plans for each individual animal that match its real-time physiological state, genetic potential, and health trends. This represents a leap from uniform group feeding to precise individual nutrition, and an upgrade from static nutrient supply to dynamic target tracking. It can automatically monitor, analyze, and generate feeding plans, greatly reducing reliance on human experience, improving management efficiency, and providing core technological support for unmanned and intelligent farming. Attached Figure Description

[0015] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. In the drawings: Figure 1 This is a schematic diagram of the method flow in an embodiment of the present invention. Detailed Implementation

[0016] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0017] This embodiment proposes an AI-driven predictive regulatory method for animal nutrition, such as... Figure 1 As shown, it includes: RFID tags are set for the animals, and each RFID tag is associated with a single animal. Basic information and physiological information of each animal are recorded. Install corresponding RFID tag reading devices at the feeding troughs to collect the animal's stay time at the feeding troughs and the corresponding dietary information; The first AI model processes basic information, physiological information, dwell time and dietary information to obtain the animal's nutritional intake information at the current time. Obtain growth targets, and based on growth targets and animal nutrition intake information, use a second AI model to predict required nutrient intake, and obtain a nutrition intake plan based on the prediction results.

[0018] The basic information mentioned above includes the animal's species, breed, age, and sex; the physiological information includes weight, height, body length, heart rate, respiratory rate, growth status, and performance. Dietary information includes the type of feed, its nutritional composition, consumption amount, and water consumption. Basic information is pre-entered. Physiological information is monitored by setting up corresponding devices at specific locations in the enclosure or by setting up corresponding sensors on the animals, or by setting up more intelligent images or other types of detection devices. Dietary information is collected by devices such as feed trough weight detection or water flow detection. The collection method is not fixed and can be achieved by existing technologies, so it will not be described in detail.

[0019] This invention uses a first AI model to predict animal nutrient intake at the current time. This first AI model employs a deep learning model and a dynamic dual-path adversarial correction method. Conventional methods typically calculate nutrient consumption by manually or experimentally analyzing the difference between food intake and excretion. However, this calculation requires significant manual sampling and experimentation, incurring substantial labor and time costs. To address this issue, since animal nutrient intake is reflected in their growth, a mapping relationship exists between nutrition and animal growth. This invention utilizes a specific deep learning model to deduce nutrient intake using readily available dietary and growth information, even when direct and precise measurement of each animal's daily absorption is not possible.

[0020] This deep learning model introduces an adversarial correction mechanism. Instead of treating intake prediction as a simple regression problem, it constructs a game-theoretic system consisting of two neural networks. The first AI model comprises a prediction network and a correction network. The prediction network guesses the intake based on daily behavioral data, while the correction network evaluates the reasonableness of the prediction based on the animal's actual growth performance. By allowing these two networks to compete and learn from each other, a highly accurate intake estimate is ultimately output that aligns with both daily behavioral observations and the final growth results.

[0021] The predictive network is responsible for extracting features from the raw data and making preliminary predictions about intake. It employs a dual-path design to process different types of data separately, which includes a first path and a second path: the first path processes relevant time-series dietary information through the backbone network, and its input data includes high-frequency time-series data, such as the total daily residence time in the feed trough, the nutrient composition ratio of the feed (the proportion of nutrient cost evenly mixed in the feed), feed consumption, water consumption, and ambient temperature.

[0022] In the first approach, a one-dimensional convolutional neural network is used to extract local feeding pattern features in the short term. Based on these feeding pattern features, subsequent temporal pattern features are extracted; that is, a long short-term memory (LSTM) network is then used to capture long-term feeding rhythms and behavioral patterns, such as feeding frequency characteristics at different time points. The final behavioral embedding vector, i.e., the temporal behavioral feature, is generated through the LSM network, encoding the animal's feeding behavior characteristics.

[0023] In the second path, the encoding focuses on the static and dynamic individual characteristics of the animal. The main data processed includes basic information and physiological information over a certain time interval, such as the animal's species, breed, age, and sex. Dynamic individual information mainly includes weight, growth status, and performance scores. The basic information is encoded using an embedding layer, transforming it into corresponding numerical vectors. Growth status is represented by the rate of change in weight, height, and body length. Other information is represented by corresponding numerical values ​​to construct the input data. Simultaneously, relevant normalization processing can be performed on the above information, introducing the numerical values ​​into the [0, 1] interval for easier subsequent computation. The encoder encodes the above encoding and basic information. In the encoder: The primary task of the encoder is to map the aforementioned physiological and basic information into latent health state variables, which are comprehensive latent health state variables it transforms. This aims to capture intrinsic, difficult-to-quantify health states in animals, such as digestive efficiency, immune load, or chronic stress levels.

[0024] The encoder employs a fully connected network. The input layer of the fully connected network receives the input data, which is then further processed by the hidden layers of the encoder for low-dimensional feature extraction. Each hidden layer consists of several (e.g., 2-3) fully connected layers. Taking two fully connected layers as an example, the first hidden layer, acting as a feature abstraction layer, has a relatively large number of neurons (e.g., 64) and uses the ReLU activation function. This hidden layer learns the complex nonlinear interactions between input features. For example, it might identify a subtle pattern that foreshadows potential problems, such as a mismatch between the rate of aging and weight gain. After this layer, a Dropout operation is introduced, randomly ignoring a portion of neurons to force the network to learn more robust features. Next, the output data from the first hidden layer enters the second hidden layer, the feature compression layer. This layer has fewer neurons (e.g., 32) and acts as an information bottleneck, refining and condensing the rich features extracted by the previous layer, discarding redundant information and retaining the core signals most relevant to health status, thus forming a higher-level composite feature representation. If there are more hidden layers, the design is adjusted according to the actual needs of the number of neurons.

[0025] After passing through several hidden layers, the composite features are processed using a parallel dual-head output structure. In this structure, one output head calculates the mean vector (z_mean) of the latent variables, while the other calculates the log-variance vector (z_log_var). The log-variance is used to ensure that the variance is always positive and to make the training process more stable. Based on the mean and log-variance of the latent variables, a probability distribution for the corresponding latent variables can be generated. This probability distribution is a multivariate normal distribution. The final latent variables are obtained from this probability distribution using a reparameterization method. In this reparameterization method, a random noise vector (epsilon) is first sampled from the standard normal distribution, and then the final latent variable z is generated using a specific formula: z = z_mean + exp(z_log_var * 0.5) * epsilon. This process introduces necessary randomness into the model to model the uncertainty of nutrition regulation prediction, while ensuring that the entire network is differentiable from beginning to end, so that the error gradient can be smoothly backpropagated, thereby training the encoder to learn how to accurately infer health status from physiological data.

[0026] Finally, the sampled latent variable z, along with its distribution parameters z_mean and z_log_var, is used as the encoder output. This final latent variable is then fed into the next stage of path two, where it is fused with other features. z_mean and z_log_var are primarily used to calculate the KL divergence loss, which, as part of the model's total loss function, constrains the distribution of the latent variable, preventing it from degenerating into meaningless noise and ensuring it becomes a structured and interpretable latent space of health status. The final latent variable serves as a latent variable for health status, aiming to quantify intrinsic factors that are difficult to measure directly but affect nutrient absorption, such as digestive efficiency, immune system activity, or subclinical health status. The final output, the final latent variable, serves as an individual embedding vector, encoding the animal's inherent characteristics and physiological state.

[0027] The behavior embedding vector and the individual embedding vector are concatenated and fused into a comprehensive feature vector. This comprehensive feature vector is then nonlinearly transformed through several fully connected layers to obtain high-level features. These high-level features are then processed by a linear output layer to output the final predicted nutrient intake. Optionally, multiple forward propagation predictions can be performed based on the same data, and the prediction results are averaged to determine the final predicted nutrient intake.

[0028] The calibration network, an independent neural network, is crucial for implementing the model's inversion logic. Its input data consists of the intake predicted by the prediction network. And the individual embedding vector generated by the individual attribute paths in the main regression network. Based on this, the predicted nutrient intake... Based on the individual's characteristics, it predicts the extent of growth deviation over a future period. Its correction network primarily learns to determine whether the predicted intake will lead to poor growth.

[0029] For the correction network, its core task is not to directly predict intake, but to assess whether the intake predicted by the master regression network is reasonable. It achieves this supervisory function by determining what growth outcomes the animal would exhibit if fed at this intake level.

[0030] The calibration network receives predicted intakes from the master regression network. The first input is a scalar value representing the main network's best guess about how much the animal ate based on its behavior. The second input is an individual embedding vector from the second path, the individual attribute path. This is an information-rich, dense vector that encodes all individualized features such as the animal's breed, age, and health status. These two inputs are concatenated along the channel dimension to form a novel concatenated feature vector. This concatenated feature vector links the guessed behavioral outcome to the individual's physiological state.

[0031] The concatenated feature vectors are fed into the discriminative network, a deep neural network consisting of several fully connected layers. The first hidden layer, serving as the feature interaction layer, has a moderate number of neurons (e.g., 50) and introduces nonlinearity using activation functions such as ReLU. The primary responsibility of this layer is to deeply mine the complex constraints between the predicted intake and individual attributes. The data then flows into the second hidden layer, the decision abstraction layer. This layer has even fewer neurons (e.g., 25), further refining and condensing the complex features extracted by the previous layer, synthesizing various nonlinear relationships into a higher-order abstract representation to prepare for the final decision. Deeper hidden layers can be selected for further feature extraction to generate the final decision features.

[0032] The final decision features are processed through the output layer of the calibration network. The output layer uses a linear activation function and contains only one neuron. The output is a continuous scalar value, representing the predicted growth rate Δ. The sign and magnitude of this value are crucial: a negative value indicates that, based on the discriminator's experience, the intake predicted by the main regression network is negative. If the animal is fed, its growth will be less than expected (actual weight gain is less than theoretical weight gain); a positive value means that the predicted intake may be sufficient to exceed expectations; while a value close to zero means that the intake is just right.

[0033] Throughout the model training process, the weight updates of the calibrator discriminator depend on a very clear objective: maximizing its predicted growth rate Δ. The discriminator maintains consistency with the actual measured rate of growth change Δg (the true value of the rate of change in weight and height). In other words, the discriminator strives to become an accurate predictor. Meanwhile, the prediction network engages in a dynamic game with the correction network: the prediction network adjusts its parameters to ensure that its predicted intake... After being fed into this trained discriminator, the predicted growth rate Δ can be obtained. The difference between the predicted and actual values ​​should be as small as possible. This means that the master regression network is constantly learning how to propose an intake prediction that the discriminator will find perfect and without any problems. Through this adversarial training mechanism, the discriminator successfully back-injects the global true signal about growth outcomes obtained from regular weight measurements into the master regression network's daily, behavior-based intake predictions, thereby achieving a high-level, biophysiologically based closed-loop correction.

[0034] The model training process includes training a calibration network and a prediction network. First, the calibration network is trained, while the master regression network is kept stationary. The goal is to make the discriminator an accurate predictor. The training process then incorporates the input predicted by the master regression network. Individual features are input into the discriminator. The discriminator outputs a predicted growth rate Δ. Compare the predicted growth rate Δ The difference between the actual measured growth rate Δg and the mean squared error (MSE) is considered. During this process, only the network weights of the discriminator are updated, making its predictions increasingly accurate.

[0035] Then, the master regression network is trained, with the calibration discriminator fixed during this process. The goal is to make the master regression network a predictor that can fool the discriminator. The master regression network then predicts the intake again based on behavior and data. The random noise vector (epsilon) generated during the intake generation process is randomly sampled at different stages. The discriminator uses this... Determine the extent of growth bias it will cause. Calculate the total loss of the main regression network, which includes supervised loss and adversarial loss. The supervised loss is calculated on a small number of data points with actual intake labels, yielding predicted values. The error between the predicted and actual values. The adversarial loss is the magnitude of the loss that corrects for the rate of growth change predicted by the network. The goal of the master regression network is to minimize this bias loss, which means it must adjust its predictions to make its output more accurate. The discriminator considers this an intake level that perfectly supports the expected growth and will not cause bias loss. During this process, only the weights of the main regression network are updated.

[0036] Through iterative processes of these two stages, the predictive network, or principal regression network, driven by the need to counteract loss, learns to incorporate feedback from the animal's growth performance into its intake prediction. Ultimately, even without actual intake data, it can output a physiologically corrected and highly accurate intake estimate that closely matches the animal's actual growth trajectory.

[0037] After obtaining the growth target, a second AI model is used to predict the required nutrient intake based on the growth target and animal nutrient intake information. A corresponding nutrient intake plan is generated based on the prediction results. Traditional animal nutrition plans often focus on planning nutrition for a fixed nutrient intake target set at the beginning of the rearing process (e.g., how much nutrition to consume at a specific age). However, this static paradigm ignores two crucial dynamic variables in the rearing process: animals are not precisely functioning machines; their growth trajectory fluctuates due to health, genetic potential, and early performance, and their performance varies at different stages. This method abandons these static elements and introduces the concept of a dynamic nutrient intake target. This target is no longer static but an optimal nutrient requirement plan that can be adjusted based on the animal's real-time performance. The second AI model primarily continuously re-predicts and calibrates the dynamic target, and subsequently plans the most efficient nutrient path to achieve that target.

[0038] To achieve the above-mentioned approach, real-time and historical data of individual animals are continuously acquired. These data mainly include two categories: first, nutritional intake data, such as feed intake, animal nutritional intake information, water consumption, and feed composition; second, physiological status data, such as weight, body size, daily weight gain, heart rate, respiratory rate, and the rate of change of weight, height, and body length.

[0039] The aforementioned data is processed in real time by a second AI model, which operates as a dynamic growth target optimizer. First, based on the input current and historical physiological data, the second AI model uses a time-series prediction algorithm to simulate the animal's future physiological development trajectory under various intervention scenarios while maintaining its current state, generating a set of prediction curves covering different physiological information and health states. Subsequently, in the second stage, the second AI model evaluates these predicted trajectories using a multi-objective optimization algorithm. The optimization objectives of this algorithm mainly include: 1) growth performance, referring to the optimal weight gain and growth within the breed's genetic potential; 2) health risk, minimizing the predicted probability of metabolic diseases or nutritional deficiencies; and 3) feed conversion efficiency, pursuing the efficient utilization of feed and nutrients. Through optimization calculations, a trajectory that achieves the best balance among the above biological objectives is selected from numerous possible paths; this trajectory is defined as the individual's dynamic physiological target for the next stage.

[0040] The second AI model includes a predictor and an optimizer. The predictor uses a long short-term memory network or a gated recurrent unit. It uses historical data from a previous period and data from the current period as basic data, and feed consumption and nutrient consumption for possible feeding in the future as auxiliary variables. The basic data and auxiliary variables are concatenated as the input of the predictor. The predictor generates time-series data on changes in physiological information (all quantified by existing scoring methods) and possible abnormal data in subsequent periods. Possible abnormal data include the probability of occurrence of metabolic diseases or nutritional deficiencies.

[0041] Using the output data of the aforementioned predictor as the fitness calculation data to guide the optimization algorithm, and setting the aforementioned auxiliary variables as design variables, after obtaining a certain amount of predictor output data, the corresponding optimization objectives are calculated using the output data of the prediction period and the aforementioned auxiliary variables. The optimization objectives are: growth performance objective, health robustness objective, and nutritional efficiency objective. The growth performance objective uses normalized weight, height, body length, and growth status as indicators; the higher the objective, the better. The health robustness objective uses the trend of changes in physiological information obtained from statistical prediction to maintain a stable or positive trend. Specifically, the quantitative values ​​of weight, height, body length, growth status, and performance status must ensure an upward trend, and the higher the value, the better; heart rate and respiratory rate should maintain a stable trend, with smaller variance values ​​being better, and the number of predicted abnormal data should be minimized. Nutritional efficiency targets aim to achieve better growth and health outcomes with less nutrient intake, i.e., higher predicted feed conversion efficiency. This is calculated as the ratio of information such as body weight, length, and growth status to feed and nutrient composition; higher is better. Different optimization targets are integrated using a weighted sum method, with higher values ​​receiving positive weights and lower values ​​receiving negative weights. Different weighted calculations are performed to generate corresponding optimization targets. Simultaneously, this optimization target (corresponding to fitness) guides the adjustment of design variables in the next iteration. The optimization algorithm employs swarm intelligence, genetic algorithms, or particle swarm optimization, using existing algorithmic processes to adjust the optimization target and design variables under fitness conditions.

[0042] Simultaneously, relevant constraints are added, including that values ​​such as weight and length, and their corresponding rates of change, cannot exceed the known biological limits for the breed. Quantitative scores for growth status, performance, and other information must be maintained within a preset health range. Predicted requirements for any feed and nutrients must be within safe and normal intake ranges, with feed consumption selected within a certain range, and the average consumption over a previous historical period serving as the center of that range.

[0043] After several iterations of the optimization algorithm, the auxiliary variables corresponding to the optimization results are used as the required nutrient intake. This forms a specific dynamic nutrient requirement set for the animal's daily needs for core nutrients such as energy, protein, minerals, and vitamins during the subsequent planning period (e.g., the next few days to a week). With ensuring all nutrient needs are met as the primary premise, the proportion of nutrients in the total feed is adjusted to generate one or more feasible personalized nutrient supply plans. The goal is to achieve a precise match between nutrient intake and the animal's physiological goals.

[0044] The entire process forms a closed loop. The new round of physiological and intake data generated after the implementation of the nutrition plan will be captured by the system again and used as new input to start the next decision-making cycle, thereby realizing continuous calibration and optimization of dynamic nutritional needs and nutritional supply plans, forming a precise management cycle that adapts to individual needs.

[0045] This invention achieves high-precision, non-invasive inversion monitoring of actual nutrient intake based on animal behavior data and growth performance by constructing a dynamic dual-path adversarial correction network. This overcomes the shortcomings of traditional methods, such as reliance on human experience, poor timeliness, and high cost. Furthermore, through a dynamic physiological target optimizer, it realizes forward-looking and precise nutrient regulation based on the real-time status of individuals, forming a complete intelligent closed loop from perception and decision-making to execution, which significantly improves the automation level of aquaculture management and nutrient utilization efficiency.

[0046] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. An AI-driven method for predictive supervision of animal nutrition, characterized in that, include: Each animal is tagged with an RFID tag, and the RFID tag and the corresponding basic and physiological information of each animal are stored. The RFID tag reader reads the RFID tags to collect the time the animals spend at the feeding trough, and obtains the dietary information corresponding to the time spent at the feeding trough by recording the weight changes of the feeding trough. The first AI model is used to predict basic information, physiological information and dietary information to obtain the animal's nutritional intake information at the current time. Obtain growth targets, and based on growth targets and animal nutrition intake information, use a second AI model to predict required nutrient intake, and obtain a nutrition intake plan based on the prediction results. The first AI model employs a dynamic dual-path adversarial correction method for prediction. This first AI model includes a prediction network and a correction network connected sequentially. The prediction network extracts deep features from basic, physiological, and dietary information, and decodes these features to obtain predicted nutrient intake information. The correction network then evaluates the predicted nutrient intake and physiological information to obtain the predicted animal growth rate. The deviation between the predicted and actual animal growth rates is calculated, and the output of the prediction network is corrected based on this deviation. The prediction network and the correction network undergo adversarial training. In the prediction network, a temporal processing method is used to process dietary information to obtain behavioral embedding vectors. Basic information and physiological information are processed by introducing noise through latent variable encoding to obtain individual embedding vectors. The behavioral embedding vectors and individual embedding vectors are concatenated and nonlinearly transformed to obtain high-level features. The high-level features are processed through a linear processing layer to obtain the animal nutrition intake information at the current time. The temporal characteristics of the behavior encode the feeding behavior characteristics of the animal; Individual embedding vectors encode the animal’s inherent characteristics and physiological state.

2. The method according to claim 1, characterized in that, Basic information includes the animal's species, breed, age, and sex; physiological information includes weight, height, body length, heart rate, respiratory rate, growth status, and performance; dietary information includes the type of feed, the proportion of nutrients, feed consumption, and water consumption.

3. The method according to claim 1, characterized in that, The first AI model employs a deep learning model, wherein the first AI model is used to represent the mapping relationship between the basic information, physiological information, residence time, dietary information, and animal nutrient intake information.

4. The method according to claim 1, characterized in that, The process of adversarial training includes: The first and second phases are executed repeatedly to complete adversarial training. The first phase includes acquiring sample data, fixing the model parameters of the prediction network, and training the calibration network using the sample data. The second phase includes fixing the model parameters of the calibration network and training the prediction network using the sample data. The comprehensive loss in the training of the prediction network is calculated. The comprehensive loss is a weighted sum of the supervision loss and the adversarial loss, where the supervision loss is the bias loss of the prediction network's output data and the adversarial loss is the bias loss of the calibration network's output data. The prediction network is then trained based on the comprehensive loss.

5. The method according to claim 1, characterized in that, In the second AI model, based on auxiliary variables and historical data, a deep learning model is used to predict physiological indicators for future time series. The auxiliary variables include the required nutrient intake; the historical data includes basic information, physiological information, residence time, and dietary information at historical times; the auxiliary variables are adjusted according to the physiological indicators through an optimization method, and the execution of the deep learning model and optimization method is iterated to obtain the predicted nutrient intake results.

6. An AI-driven animal nutrition prediction and monitoring system, characterized in that, Used to perform the method described in any one of claims 1-5.

Images