Method for generating animal epidemic prevention and control scheme by using large model

Abstract

The invention provides a method for generating an animal epidemic prevention and control scheme by using a large model, and belongs to the technical field of animal epidemic prevention and control. Inputting the fusion parameter vector and the epidemiological parameter set into a dual-LoRA fine-tuning domestic open-source large model to output a preliminary prevention and control scheme text set, and performing one-by-one cross check on prevention and control scheme texts through a physical feasibility verification rule base to generate a feasibility score; and calling a correction cue word for the prevention and control scheme text of which the feasibility score does not meet the requirement to trigger regeneration of the large model, performing similarity matching on the prevention and control scheme text passing verification and a historical prevention and control effect database, and performing sorting to output an optimal prevention and control scheme text sequence. The technical problem that the reliability of generation of an animal epidemic prevention and control scheme is insufficient due to low data fusion precision of a multi-source sensor is solved.

Description

Technical Field

[0001] This invention belongs to the field of animal disease prevention and control technology, and specifically relates to a method for generating animal disease prevention and control plans using a large model. Background Technology

[0002] In the field of animal disease prevention and control, traditional technologies collect environmental data by deploying multiple sensors such as temperature sensors, humidity sensors, gas concentration sensors, and pathogen concentration sensors, and generate prevention and control plans based on single sensor data or simple weighted averaging methods. However, existing technologies suffer from measurement errors, data heterogeneity, and time asynchrony issues among different sensors. Traditional data fusion methods, using linear fusion algorithms such as fixed-weighted averaging or Kalman filtering, fail to dynamically adjust fusion weights based on the real-time reliability of the sensors, resulting in significant errors in the fusion parameter vector. Furthermore, existing technologies lack mechanisms for evaluating the quality of fused data and verifying physical feasibility when generating prevention and control plans from large models. When the fused data contains significant errors, the generated prevention and control plan may include unreasonable disinfectant concentrations or drug dosages. In other words, existing technologies suffer from low accuracy in multi-source sensor data fusion, leading to insufficient reliability in the generation of animal disease prevention and control plans. Summary of the Invention

[0003] In view of this, the present invention provides a method for generating animal epidemic prevention and control plans using large models, which can solve the technical problem of insufficient reliability of animal epidemic prevention and control plan generation due to low accuracy of multi-source sensor data fusion in the prior art.

[0004] This invention is implemented as follows: It provides a method for generating animal disease prevention and control plans using a large model. This method involves collecting multi-source sensor data from a farm environment and performing time-calibrated processing to obtain a time-calibrated multi-source sensor dataset. The time-calibrated multi-source sensor dataset and the historical reliability scores of each sensor are input into a sparse coding-deep unrolling fusion network model, which outputs a fusion parameter vector. Based on the fusion parameter vector, an animal epidemiological parameter coupling propagation algorithm is used to calculate an epidemiological parameter set. The fusion parameter vector and the epidemiological parameter set are then input into an open-source large model fine-tuned using dual LoRA, which outputs a preliminary prevention and control plan text set. A physical feasibility verification rule base is used to cross-check the preliminary prevention and control plan text set to generate a feasibility score. For prevention and control plan texts that do not meet the feasibility score requirements, correction prompts are invoked to trigger the large model to regenerate until feasibility verification is passed. Finally, the verified prevention and control plan texts are matched for similarity with a historical prevention and control effect database, and the optimal prevention and control plan text sequence is output based on the goal of minimizing the epidemic transmission rate.

[0005] The multi-source sensor data includes temperature parameters, relative humidity parameters, gas concentration parameters, and pathogen concentration parameters.

[0006] The sparse coding-deep unrolling fusion network model includes a dynamic weight adjustment layer, a sparse coding dictionary learning layer, a deep unrolling iterative layer, and a fusion output layer.

[0007] The dynamic weight adjustment layer adaptively adjusts the weight coefficients of each sensor data in the time-calibrated multi-source sensor dataset based on the historical reliability scores of each sensor.

[0008] The sparse coding dictionary learning layer uses the K-SVD algorithm to learn from the dynamically weighted multi-source sensor data to generate a sparse dictionary atom set.

[0009] The deep unfolding iterative layer unfolds the iterative optimization algorithm for solving sparse representations into a multi-layer neural network. The number of layers is equal to the number of sensor types plus 2 layers. Each layer learns the soft threshold function parameters and intermediate sparse representation coefficients.

[0010] The fusion output layer generates a fusion parameter vector by linearly combining the sparse representation coefficients output by the depth expansion iteration layer with sparse dictionary atoms.

[0011] The fusion parameter vector includes four components: fusion temperature parameter, fusion relative humidity parameter, fusion gas concentration parameter, and fusion pathogen concentration parameter.

[0012] The animal epidemiological parameter coupling propagation algorithm is based on the SEIR infection dynamics model to establish differential equations that integrate the influence of environmental parameters on pathogen survival.

[0013] The set of epidemiological parameters calculated by the animal epidemiological parameter coupling propagation algorithm includes pathogen generation time, pathogen transmission rate, and disinfectant inactivation biomechanical parameters.

[0014] Specifically, the first LoRA fine-tuning part of the dual LoRA fine-tuning mechanism fine-tunes the first 12 layers of the base model to receive the fused parameter vector and the epidemiological parameter set. The second LoRA fine-tuning part fine-tunes the last 12 layers of the base model to generate a preliminary prevention and control plan text set that conforms to the prevention and control specifications.

[0015] The physical feasibility verification rule base includes tables of safe concentration ranges, contraindications, and interactions for disinfectants, veterinary drugs, and nutritional additives. This rule base extracts disinfectant concentration values, contact time values, and drug concentration values ​​from the prevention and control plan text and compares them with the safe concentration ranges to generate a feasibility score.

[0016] The correction prompt includes the name of the parameter that was exceeded, the actual value, the lower limit of the safe range, and the upper limit of the safe range.

[0017] When the uncertainty score of any parameter in the fusion parameter vector exceeds the uncertainty threshold, the system automatically adjusts to a conservative prevention and control state.

[0018] This invention constructs a sparse coding-deep unrolling fusion network model. A dynamic weight adjustment layer adaptively adjusts the weights of each sensor data based on historical sensor reliability scores. A sparse coding dictionary learning layer and a deep unrolling iterative layer convert high-dimensional heterogeneous sensor data into a unified-dimensional fusion parameter vector, overcoming the limitation of traditional fixed-weight fusion methods that cannot adapt to dynamic changes in sensor reliability. This invention combines sparse coding dictionary learning with deep unrolling networks, maintaining the interpretability of iterative optimization during the fusion process while leveraging the adaptive learning capabilities of deep learning. End-to-end training stabilizes the fusion parameter vector error within ±5%, significantly improving the accuracy of multi-source sensor data fusion. Furthermore, this invention uses a physical feasibility verification rule base to check the parameter range of the prevention and control plan text generated by the large model. Correction prompts trigger plan regeneration, ensuring that the output prevention and control plan meets safety standards. In summary, this invention solves the technical problem mentioned in the background art where low accuracy of multi-source sensor data fusion leads to insufficient reliability in animal epidemic prevention and control plan generation. Attached Figure Description

[0019] Figure 1 This is a flowchart of the method of the present invention.

[0020] Figure 2 A comparison chart of fusion errors for sparse coding-deep unfolding fusion network models.

[0021] Figure 3 A flowchart is generated for the dual LoRA fine-tuning large model scheme.

[0022] Figure 4 This is a graph showing the similarity matching results between the prevention and control plan and historical plans.

[0023] Figure 5 This is a graph showing the change in pathogen concentration after the implementation of the optimal prevention and control plan. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0025] like Figure 1 The diagram shown is a flowchart of a method for generating animal disease prevention and control plans using a large model, provided by this invention. This method includes the following steps:

[0026] S01. Collect multi-source sensor data in the breeding farm environment, including temperature parameters, relative humidity parameters, gas concentration parameters, and pathogen concentration parameters. Perform time calibration processing on the collected raw multi-source sensor data to obtain a time-calibrated multi-source sensor dataset.

[0027] S02. Input the time-calibrated multi-source sensor dataset and the historical reliability scores of each sensor into the sparse coding-deep unrolling fusion network model, and the sparse coding-deep unrolling fusion network model outputs the fusion parameter vector.

[0028] S03. Based on the temperature parameter, relative humidity parameter H, and pathogen concentration parameter in the fusion parameter vector F, the animal epidemiological parameter coupling propagation algorithm is used to calculate the pathogen generation time, pathogen transmission rate, and disinfectant inactivation dynamic parameters, and generate an epidemiological parameter set.

[0029] S04. Input the fusion parameter vector F and the epidemiological parameter set into the domestic open-source large model with dual LoRA fine-tuning. Through the cascaded processing of the front LoRA layer and the back LoRA layer, output the preliminary prevention and control plan text set.

[0030] S05. Each prevention and control plan text in the preliminary prevention and control plan text set is cross-checked one by one through the physical feasibility verification rule base. The disinfectant concentration value, contact time value, and drug concentration value in the prevention and control plan text are extracted and compared with the corresponding safe concentration range in the rule base to generate a feasibility score.

[0031] S06. For prevention and control plan texts that do not meet the feasibility score requirements, call the correction prompt words to trigger the large model to regenerate the plan text until a prevention and control plan text that has passed the feasibility verification is generated. Then, the prevention and control plan text that has passed the feasibility verification is matched with the historical prevention and control effect database for similarity. The candidate prevention and control plan texts are sorted according to the goal of minimizing the epidemic transmission rate, and the optimal prevention and control plan text sequence after sorting is output.

[0032] S07. When the uncertainty score U of any parameter in the fusion parameter vector F exceeds the uncertainty threshold, the system automatically adjusts to a conservative prevention and control working state, increases the execution frequency of key disinfection measures, increases the sampling frequency of environmental monitoring, and strengthens the level of personnel protection.

[0033] S08. When the feasibility score W in the physical feasibility verification rule base is lower than the feasibility threshold, the system automatically marks the corresponding prevention and control plan text as a candidate prevention and control plan text, generates targeted correction prompts based on the parameter types that exceed the safety threshold, calls the large model to regenerate the prevention and control plan text, and the newly generated prevention and control plan text re-enters the feasibility verification process.

[0034] The sparse coding-deep unrolling fusion network model has the following structure: a dynamic weight adjustment layer, a sparse coding dictionary learning layer, a deep unrolling iterative layer, and a fusion output layer. The dynamic weight adjustment layer adaptively adjusts the weight coefficients of each sensor in the time-calibrated multi-source sensor dataset based on the historical reliability scores of each sensor, ensuring that sensor data with high quality has higher weights and sensor data with lower quality has lower weights. The sparse coding dictionary learning layer learns from the dynamically weighted multi-source sensor data using the K-SVD algorithm, generating a sparse dictionary atom set, enabling high-dimensional heterogeneous sensor data to be represented by a linear combination of a small number of sparse dictionary atoms. The deep unrolling iterative layer expands the iterative optimization algorithm for solving sparse representations into a multi-layer neural network, with the number of layers equal to the number of sensor types plus two layers. Each layer corresponds to one iteration step of the iterative optimization algorithm, and each layer learns the soft thresholding function parameter θ and the intermediate sparse representation coefficient β. Backpropagation makes the soft thresholding function parameter θ and the intermediate sparse representation coefficient β learnable. The fusion output layer linearly combines the sparse representation coefficients output by the depth expansion iteration layer with the sparse dictionary atoms to generate a fusion parameter vector F with uniform dimension and consistent scale. The fusion parameter vector F contains four components: fusion temperature parameter T, fusion relative humidity parameter H, fusion gas concentration parameter G, and fusion pathogen concentration parameter P.

[0035] The steps for establishing the training dataset for the sparse coding-deep unfolding fusion network model specifically include: collecting multi-source sensor data records from the farm that has been operating normally for more than 6 months, and simultaneously recording environmental indicator labels and reliability scores for each sensor; establishing a time-aligned set of data pairs, each data pair containing an unfused time-calibrated multi-source sensor dataset and a manually labeled optimal fusion parameter vector; dividing the dataset into training, validation, and test datasets in an 8:1:1 ratio; performing dimensionless processing on the training, validation, and test datasets, and normalizing the historical data of each sensor according to its measurement range so that the normalized data are all within the range of 0 to 1.

[0036] The training steps of the sparse coding-deep unrolling fusion network model specifically include: learning the training dataset using the K-SVD algorithm to initialize the sparse dictionary atom set; setting the number of layers in the deep unrolling iteration layer to be equal to the number of sensor types plus 2; and training in an end-to-end manner using the fusion parameter vector reconstruction error and fusion accuracy as the joint loss function, the joint loss function being expressed as follows: ,in This is the predicted fusion parameter vector of the sparse coding-deep unrolling fusion network model. The optimal fusion parameter vector is manually labeled, MSE is the mean squared error, Accuracy is the fusion accuracy, and α and β are weight coefficients. The fusion error is monitored on the validation dataset to keep it within ±5%. Training is stopped early when the loss function value on the validation dataset does not decrease for 5 consecutive epochs. The weight parameters of the trained sparse coding-deep unrolling fusion network model are saved as the initial state for processing real-time multi-source sensor data.

[0037] The physical feasibility verification rule base includes tables of safe concentration ranges, contraindications, and interactions for commonly used disinfectants, veterinary drugs, and nutritional additives. For disinfection protocols, it is recommended to extract the disinfectant name, disinfectant concentration value (V), and contact time value from the prevention and control protocol text. The disinfectant name and concentration value V are compared with the corresponding safe concentration range in the rule base; for medication recommendations, the drug name and concentration value are extracted from the prevention and control plan text. The usage route and course of treatment are compared and verified with the rule base, and the interaction with other medication recommendations during the same period is checked. When any parameter exceeds the safety threshold, the feasibility score W is marked as unsatisfactory. The correction prompt includes information such as the name of the parameter that exceeds the threshold, the range of the safe parameter, and the degree of exceedance. The large model is called to regenerate the prevention and control plan text.

[0038] The specific principle of the animal epidemiological parameter coupled transmission algorithm is as follows: Based on the SEIR infection dynamics model, differential equations are established for four states: susceptible individuals (S), latent infected individuals (E), clinically infected individuals (I), and recovered individuals (R), incorporating the influence of environmental parameters on pathogen survival. The relationship between pathogen generation time τ and environmental temperature parameter T and relative humidity parameter H is obtained by curve fitting of infection experimental data. The disinfectant inactivation kinetics follow first-order reaction kinetics, and the disinfectant concentration value V is related to the contact time value... The inactivation effect on the pathogen concentration parameter P is described by the inactivation kinetic equation, which is expressed as follows: Where P is the concentration of pathogens after treatment. K represents the pathogen concentration before treatment, K represents the biomechanical parameter of the disinfectant, and V represents the disinfectant concentration. The contact time value is used. Environmental parameters, pathogen characteristics, and disinfection measures are dynamically simulated within the same coupled equation system. The epidemic transmission trajectory under different prevention and control schemes is obtained by numerically solving the equations. The inverse solution method is used to minimize the pathogen transmission rate R. By solving the partial derivative of the pathogen transmission rate R with respect to the prevention and control parameters, the extreme point is found. This enables the generation of more than 1000 ranked candidate prevention and control scheme texts within seconds. The candidate prevention and control scheme texts are arranged from low to high according to the pathogen transmission rate R.

[0039] The technical benefits of the animal epidemiology parameter coupling propagation algorithm are as follows: By establishing a multi-parameter coupled differential equation for pathogen transmission, it unifies environmental parameter changes, pathogen biological characteristics, and the chemical kinetics of disinfection measures within a single mathematical framework. This transforms the generation of prevention and control plan texts from experience-based recommendations to rational design based on infectious dynamics models. By inversely solving for the parameter combination that minimizes the pathogen transmission rate, it shifts the focus from subjective assessment to objective prediction of prevention and control effectiveness, improving the scientific rigor and timeliness of prevention and control plan texts, and rapidly traversing and filtering the optimal prevention and control plan text within the parameter space.

[0040] The dual LoRA fine-tuning mechanism uses a domestically developed open-source large model as its base model. The first LoRA fine-tuning part fine-tunes the first 12 layers of the base model, receiving the fusion parameter vector F and the epidemiological parameter set D output by the sparse coding-deep unrolling fusion network model. It learns the mapping relationship between aquaculture environment parameters and the semantics of the prevention and control plan text through low-rank matrix factorization. The second LoRA fine-tuning part fine-tunes the last 12 layers of the base model, receiving the output of the first LoRA fine-tuning part and generating a preliminary prevention and control plan text set S that conforms to the prevention and control specifications. The number of heads in the multi-head attention mechanism in the second LoRA fine-tuning part... Based on the dimension of the fusion parameter vector F The number of parameter types in the epidemiological parameter set D Dynamic settings, number of heads The calculation method is as follows: ,in This represents the function for rounding up.

[0041] The specific implementation of combining sparse coding dictionary learning with deep unfolding networks in the sparse coding-deep unfolding fusion network model is as follows: After weight adjustment is completed in the dynamic weight adjustment layer, the time-calibrated multi-source sensor dataset undergoes atomic-level sparse decomposition through the sparse coding dictionary learning layer, and the sparse dictionary atom set is learned through the K-SVD algorithm. This allows high-dimensional heterogeneous sensor data to be represented by a linear combination of a small number of sparse dictionary atoms, the mathematical form of which is: Where X is the time-calibrated multi-source sensor dataset, Let be the set of sparse dictionary atoms, and β be the sparse representation coefficients. The deep expansion iterative layer expands the iterative optimization algorithm for solving the sparse representation coefficients β into a multi-layer neural network. Each layer corresponds to one iteration step of the algorithm. In the j-th layer iteration, the soft threshold function parameter θ and the intermediate sparse representation coefficients β are adaptively updated by learning adjustable parameters. The mathematical form of the iterative process is: ,in Let θ be the soft thresholding function and λ be the step size parameter. Backpropagation enables the soft thresholding function parameters θ and the intermediate sparse representation coefficients β to be learnable in the deep unfolding iterative layers. The combination of dictionary learning for sparse coding and the deep unfolding network allows the sparse fusion process to maintain the interpretability and mathematical intelligibility of iterative optimization while also gaining the adaptive learning and end-to-end optimization capabilities of deep learning. Compared to traditional data fusion methods, this significantly improves the accuracy of multi-source sensor data fusion, enhances the robustness of the fusion results to sensor faults and data anomalies, and keeps the error of the fusion parameter vector F stable within ±5%.

[0042] The correction prompt is constructed as follows: when a parameter in the prevention and control plan text exceeds a safety threshold, the correction prompt includes the name of the parameter that exceeds the threshold, the actual value of the parameter, the lower limit of the safety range, and the upper limit of the safety range. The text template for the correction prompt is: "According to the prevention and control specifications, the safety range for parameter [parameter name] is from [lower limit] to [upper limit]. The current recommended value is [actual value]. Please regenerate the prevention and control plan within the range of [lower limit] to [upper limit]."

[0043] The specific process of the failure backtracking mechanism is as follows: when the prevention and control plan texts in the preliminary prevention and control plan text set S generated by the large model are checked by the physical feasibility verification rule base, and the feasibility score W is lower than the feasibility threshold... When the system fails to meet the requirements, it automatically records the parameter information that does not meet the requirements. Based on the type and degree of the parameter that exceeds the safety threshold, it generates targeted correction prompts. The correction prompts are then input into the large model along with the fusion parameter vector F and the epidemiological parameter set D to regenerate the prevention and control plan text. The newly generated prevention and control plan text is then entered into the physical feasibility verification rule base for verification until a prevention and control plan text with a feasibility score W that meets the requirements is generated.

[0044] The historical prevention and control effectiveness database contains historical prevention and control plan texts, corresponding prevention and control effectiveness indicators, and evaluations of prevention and control results. Prevention and control effectiveness indicators include the rate of decrease in disease incidence, the degree of improvement in animal welfare, and prevention and control costs. For prevention and control plan texts that pass the physical feasibility verification rule base, the system extracts their key prevention and control measures and parameter values, performs text similarity matching with the prevention and control plan texts in the historical prevention and control effectiveness database, identifies historical prevention and control plan texts with high similarity, and their prevention and control effects. Based on the objectives of minimizing the epidemic transmission rate and maximizing the disease prevention and control effect, the system comprehensively ranks the candidate prevention and control plan texts and outputs the optimal sequence of ranked prevention and control plan texts.

[0045] Optionally, the present invention also provides a method for generating an animal epidemic prevention and control plan by means of a computer, wherein the computer is provided with a readable storage medium, the readable storage medium stores program instructions, and the program instructions are used to execute the above-mentioned method when the computer is run.

[0046] The specific implementation methods of the above steps are described in detail below.

[0047] The specific implementation of step S01 involves deploying various sensor devices within the farm, including temperature sensors, relative humidity sensors, gas concentration sensors, and pathogen concentration detection devices. Each sensor collects data in real time according to its factory-set sampling frequency. Since the sampling frequencies of the various sensors differ, time synchronization processing is required for the original multi-source sensor data. The time calibration process employs the Kalman filter principle. Kalman filtering establishes a recursive relationship between state prediction and measurement updates, mapping data from different sampling frequencies to a unified time coordinate system. Its core principle is to calculate the optimal state estimate for the current moment based on the state estimate from the previous moment and the measurement value at the current moment, thus achieving precise alignment of the time-misaligned multi-source sensor data. After time calibration, the timestamp error of the multi-source sensor dataset is reduced to within 100ms, ensuring time consistency in subsequent fusion processing.

[0048] The specific implementation of step S02 involves inputting the time-calibrated multi-source sensor dataset and the historical reliability scores of each sensor into a sparse coding-deep unfolding fusion network model for data fusion processing. The sparse coding-deep unfolding fusion network model first assigns weights to each sensor data in a dynamic weight adjustment layer based on the historical reliability scores of each sensor. Sensor data with high reliability scores are assigned greater weights, and sensor data with low reliability scores are assigned less weight. The goal of weight allocation is to reduce the interference of low-quality data on the fusion result. The multi-source sensor data after weight adjustment enters the sparse coding dictionary learning layer. Dictionary learning employs the K-SVD iterative optimization algorithm, which achieves dimensionality compression and feature extraction through atomic decomposition of the data, enabling high-dimensional heterogeneous sensor data to be sparsely represented through a linear combination of a small number of basic dictionary atoms. The deep unfolding iterative layer unfolds the iterative optimization steps in dictionary learning into learnable neural network layers. Each layer corresponds to one iteration. Through backpropagation, the soft threshold function parameters and sparse representation coefficients are automatically optimized during gradient descent. Compared to the fixed iterative process of traditional algorithms, the deep unfolding network achieves adaptive learning of iterative parameters. The fusion output layer reconstructs the sparse representation coefficients and dictionary atoms through linear combination, and outputs the fusion parameter vector F. The fusion errors of the four components—fusion temperature parameter T, fusion relative humidity parameter H, fusion gas concentration parameter G, and fusion pathogen concentration parameter P—are all controlled within ±5%.

[0049] The specific implementation of step S03 is as follows: based on the temperature parameter T, relative humidity parameter H, and pathogen concentration parameter P extracted from the fusion parameter vector F, the animal epidemiological parameter coupling propagation algorithm is used to calculate the epidemiological parameter set D. The core principle of the animal epidemiological parameter coupling propagation algorithm is to establish a differential equation system based on the SEIR infection dynamics model. The SEIR model divides the population into four states: susceptible individuals, latent infected individuals, clinically infected individuals, and recovered individuals, and describes the transition between states through the transfer rate coefficient. The quantitative relationship between pathogen generation time τ and environmental temperature T and relative humidity H is obtained by nonlinear regression of historical infection experimental data, establishing a functional relationship between τ and T and H. The disinfectant inactivation kinetics follows the first-order reaction kinetics principle, and the disinfectant concentration V and contact time... The inactivation effect on pathogen concentration P is described by the classical microbial inactivation equation, which states that the logarithm of the pathogen concentration is proportional to the product of the disinfectant concentration and the contact time. Environmental parameters affect the transmission rate by influencing the pathogen's generation time. Disinfection measures directly affect the pathogen concentration through the inactivation equation. By coupling these factors within a unified differential equation framework, a numerical solution method is used to calculate the pathogen transmission rate R and the disinfectant inactivation kinetic parameter K, resulting in the epidemiological parameter set D.

[0050] The specific implementation of step S04 involves simultaneously inputting the fusion parameter vector F and the epidemiological parameter set D into a domestically developed open-source large-scale model employing dual LoRA fine-tuning. Semantic mapping and text generation are then performed through the pre-LoRA fine-tuning layer and the post-LoRA fine-tuning layer of the large-scale model. The basic architecture of the large-scale model is based on a domestically developed open-source model. The pre-LoRA fine-tuning layer efficiently fine-tunes the parameters of the first 12 layers of the basic model, receiving the fusion parameter vector F and the epidemiological parameter set D as input. It learns the mapping relationship between the aquaculture environment parameter vector and the prevention and control semantic space through low-rank matrix factorization. Its core principle is to utilize the mathematical properties of low-rank matrix factorization, approximating the complete weight matrix with the product of two small-dimensional matrices, significantly reducing the number of fine-tuning parameters. The post-LoRA fine-tuning layer fine-tunes the last 12 layers of the basic model, receiving the output of the pre-LoRA layer as input. In the multi-head attention mechanism within the post-LoRA layer, the number of attention heads... Based on the dimension of the fusion parameter vector F In addition to the number of parameter types in the epidemiological parameter set D The attention decomposition granularity is dynamically set according to the rounding-up division by 32, which allows for adaptive adjustment based on the richness of the input parameters. The large model outputs a preliminary prevention and control plan text set S through cascaded processing of consecutive LoRA layers. This set contains several candidate prevention and control plan texts, each describing different prevention and control measures and parameter suggestions.

[0051] The specific implementation of step S05 involves performing physical feasibility verification on each prevention and control plan text in the preliminary prevention and control plan text set S. The purpose is to eliminate plans generated by the large model that do not conform to actual prevention and control standards. The physical feasibility verification rule base includes tables of safe concentration ranges, contraindications, and drug interactions for commonly used disinfectants, veterinary drugs, and nutritional additives. These tables are derived from veterinary pharmacy standards and prevention and control practices. For disinfection recommendations in the prevention and control plan texts, the disinfectant name, disinfectant concentration value V, and contact time value are extracted. The system matches the disinfectant name against the disinfectant list in the rule base. If a match is found, the disinfectant concentration value V is further compared with the corresponding safe concentration range. If the concentration value exceeds the upper limit of the safe range, it is judged as an excessively high concentration risk. For medication recommendations, the drug name and medication concentration value are extracted. The system retrieves information on usage routes and treatment durations, and extracts other medication recommendations within the same time period, checking for any incompatibilities in the drug interaction table. If any parameter exceeds the safety threshold or an interaction incompatibility exists, the resulting feasibility score W is deemed unsatisfactory. The reference threshold is that the feasibility score W must be greater than or equal to 0.8, meaning that at least 80% of the parameter recommendations in the prevention and control plan text must comply with safety standards.

[0052] The specific implementation of step S06 involves correcting and regenerating prevention and control plan texts that do not meet the feasibility score W, followed by sorting the verified prevention and control plan texts. The correction process first identifies the specific parameters in the prevention and control plan texts that do not meet the requirements. The names, actual values, lower and upper limits of the safe range of parameters exceeding the safety threshold are organized into correction prompts. These prompts guide the large model to regenerate the plan within the safe parameter range, thus preventing the direct adoption of infeasible plans. After receiving the correction prompts, the large model regenerates the prevention and control plan text. The newly generated text is then checked again in the physical feasibility verification rule base, iterating until a valid prevention and control plan text is generated. For verified prevention and control plan texts, key prevention and control measures and parameter values ​​are extracted and matched with a historical prevention and control effect database. This database contains previous prevention and control plan texts, corresponding prevention and control effect indicators, and evaluation results. Candidate prevention and control plan texts are comprehensively sorted based on the optimization objectives of minimizing the epidemic transmission rate R and maximizing the disease prevention and control effect, outputting the optimal sequence of sorted prevention and control plan texts.

[0053] The specific implementation of step S07 is to continuously monitor the uncertainty score U of each parameter in the fusion parameter vector F, and when the uncertainty score U of any parameter exceeds the uncertainty threshold... When the system automatically adjusts to a conservative prevention and control mode, the uncertainty score is calculated using the prediction variance of the fusion model or a Bayesian probability framework, and the uncertainty threshold is... The reference value is 0.3, meaning that conservative prevention and control measures are triggered when the uncertainty score of a certain parameter exceeds 0.3. Under conservative prevention and control measures, the system increases the execution frequency of key disinfection measures to 1.5 times the original plan, increases the environmental monitoring sampling frequency to once per hour, and strengthens the personnel protection level to Level 2 protection. The purpose of these measures is to strengthen prevention and control efforts when environmental parameters are uncertain, and to prevent prevention and control failures due to parameter estimation errors.

[0054] The specific implementation of step S08 is as follows: when the feasibility score W generated by the physical feasibility verification rule base is lower than the feasibility threshold... At that time, the system automatically marks the corresponding prevention and control plan text as a candidate prevention and control plan text. Feasibility threshold. The reference value is 0.8, meaning this step is triggered when the feasibility score W is less than 0.8. The system generates targeted correction prompts based on the type and degree of the parameter exceeding the safety threshold and relevant regulatory requirements. These prompts include the actual violation of the parameter and the direction of correction. The system then calls the large model to regenerate the prevention and control plan text. The newly generated prevention and control plan text re-enters the physical feasibility verification rule base for item-by-item checking until a prevention and control plan text with a feasibility score W that meets the threshold requirement is generated. This iterative mechanism ensures that the final output prevention and control plan text is physically feasible.

[0055] It should be noted that the key technical ideas of this invention include the following aspects: The first key idea is the application of a sparse coding-deep unfolding fusion network model. This model unfolds the iterative optimization algorithm process into a learnable neural network layer, enabling end-to-end adaptive learning while maintaining mathematical interpretability. Compared with the fixed fusion weights or fusion rules in traditional data fusion methods, the sparse coding-deep unfolding fusion network model can dynamically adapt to the quality changes of each sensor, significantly improving fusion accuracy in aquaculture environments with large differences in the quality of multi-source heterogeneous sensor data. The second key idea is the combination of an animal epidemiological parameter coupling propagation algorithm and a large model. This algorithm establishes the coupling relationship between environmental parameters, pathogen characteristics, and disinfection measures based on the SEIR infectious disease dynamics model. By inversely solving for the parameter combination that minimizes the propagation rate, compared with traditional empirical decision-making methods, the algorithm provides a rational prevention and control scheme generation mechanism based on a physical model, transforming the generation of prevention and control schemes from subjective judgment to objective mathematical optimization. The third key approach is the combined application of a physical feasibility verification rule base and a failure backtracking mechanism. This mechanism systematically checks the physical constraints of the large model's output and guides the large model to regenerate a solution within a safe range through corrective prompts, effectively preventing the harm caused by the large model illusion problem in prevention and control practice. The synergistic effect of these three key approaches is that the sparse coding-deep unfolding fusion network model provides high-precision environmental parameter inputs, the animal epidemiology parameter coupling propagation algorithm performs scientific prevention and control optimization based on these precise parameters, and the physical feasibility verification rule base ensures the actual executability of the final prevention and control plan. Together, they form a complete closed loop from data collection and scientific optimization to feasibility assurance, transforming the application of large models in animal epidemic prevention and control from theoretical research into practical production tools.

[0056] It should be noted that this invention also solves the following technical problem: Existing large-scale animal disease prevention and control plans lack epidemiological transmission dynamics support, leading to inaccurate predictions of control effects. This invention, through an animal epidemiological parameter-coupled transmission algorithm, establishes a multi-parameter coupled differential equation for pathogen transmission based on the SEIR infectious dynamics model. It unifies the influence of environmental parameters on pathogen survival, the relationship between pathogen generation time and temperature and humidity, and the dynamics of disinfectant deactivation within a mathematical framework. Using a reverse solution method, it minimizes the pathogen transmission rate, generating over 1000 ranked candidate control plans within seconds. This transforms control plan generation from experience-based recommendations to rational design based on infectious dynamics models, shifting from subjective assessment to objective prediction of control effects. This solves the problem of insufficient scientific rigor in control plans due to the lack of epidemiological theoretical support in traditional methods.

[0057] Specifically, the principle of this invention is as follows: The sparse coding-deep unfolding fusion network model solves the aforementioned technical problems by using a dynamic weight adjustment layer to adaptively allocate sensor data weights. This ensures that sensors with high-quality data receive higher weights, while those with lower-quality data receive lower weights, reducing the impact of low-quality data on the fusion result from the source. The sparse coding dictionary learning layer learns the sparse dictionary atom set through the K-SVD algorithm, representing high-dimensional heterogeneous sensor data as a linear combination of a small number of sparse dictionary atoms, achieving data dimensionality reduction and feature extraction. The deep unfolding iterative layer unfolds the iterative optimization algorithm for solving sparse representations into a multi-layer neural network, allowing the soft threshold function parameters and intermediate sparse representation coefficients to be learned through backpropagation. This maintains the mathematical interpretability of iterative optimization while achieving end-to-end optimization capabilities of deep learning. The fusion output layer generates a fusion parameter vector with uniform dimension and consistent scale, eliminating the impact of sensor data heterogeneity on subsequent processing. The physical feasibility verification rule base extracts key parameters from the prevention and control plan text and compares them with the safety range. When a parameter exceeds the safety threshold, it generates a correction prompt word to trigger the large model to regenerate the plan, forming a closed-loop error correction mechanism to ensure the physical feasibility and safety of the output plan.

[0058] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0059] The specific implementation of step S01 is as follows: Real-time data from temperature sensors, humidity sensors, gas sensors, and pathogen detection sensors in the breeding farm environment are collected. The temperature parameter is in °C, the relative humidity parameter is in %, and the gas concentration parameter is in [missing information]. The pathogen concentration parameter is in units of The raw multi-source sensor data collected is time-calibrated by using a network time protocol to unify the timestamps of each sensor to the same time base, eliminating time deviations in sensor sampling times, and obtaining a time-calibrated multi-source sensor dataset.

[0060] The specific implementation of step S02 is the same as described above, and will not be repeated in detail here.

[0061] The specific implementation of step S03 is as follows: based on the fusion parameter vector Fusion temperature parameters , Integration of relative humidity parameters and fusion pathogen concentration parameters The generation time of pathogens was calculated using an animal epidemiological parameter coupled propagation algorithm. Pathogen transmission rate Disinfectant activity dynamic parameters Pathogen generation time With ambient temperature parameters and relative humidity parameters The relationship is obtained through curve fitting, and it is expressed as follows:

[0062] ;

[0063] In the formula, The generation time of pathogens, in units of ; The generation time of pathogens under standard conditions, in units of The value is usually 24; The fusion temperature parameter is in °C. Standard temperature, in °C, default is 25; To incorporate relative humidity parameters, the unit is % This is the standard relative humidity, expressed in %, with a default value of 60. and The fitting coefficients are dimensionless and obtained by nonlinear least squares fitting of the infection experimental data. The fitting experimental steps include: Step 1, setting up 5 combinations of different temperature gradients and 5 combinations of different humidity gradients in a constant temperature and humidity incubator; Step 2, inoculating each condition with a standard concentration of pathogen and monitoring its reproduction cycle; Step 3, fitting the experimental data using the nonlinear least squares method. and The values ​​are empirically 0.05 and 0.03, respectively. The kinetic equation for the disinfectant's failure is expressed as follows:

[0064] ;

[0065] In the formula, The concentration of pathogens after treatment, in units of ; The concentration of pathogens before treatment, in units of ; These are dimensionless, mechanical parameters related to the loss of activity of disinfectants. They were determined experimentally for different disinfectants. The value range is 0.5 to 2.5; This is the disinfectant concentration value, in units of... ; This refers to the standard concentration of disinfectant, in units of... The default value is 1000; This is the contact time value, in units of ; Standard contact time, in units of The default value is 30. A system of differential equations is established based on the SEIR infection dynamics model, and its expression is as follows:

[0066] ;

[0067] ;

[0068] ;

[0069] ;

[0070] In the formula, Number of susceptible animals, in units of individuals; This represents the number of animals infected during the incubation period, expressed in individual animals. This refers to the number of clinically infected animals, expressed in units of animals. The number of animals that have recovered is expressed in units of animals. This represents the total number of animals, expressed in units of individuals. ; The rate of pathogen transmission, in units of ; The conversion rate from latent to clinical stage is expressed in units of... The calculation method is as follows 24 is a unit conversion factor, dimensionless, used to convert pathogen generation time. The unit from Convert to ; The recovery rate is expressed in units of... It is usually set to a value of 0.1; For time, the unit is The transmission trajectory of the epidemic under different prevention and control schemes was obtained by numerically solving a system of differential equations. The pathogen transmission rate was then determined using a reverse solution method. By solving for the partial derivatives of the prevention and control parameters, we can find the extreme points and generate a set of epidemiological parameters. Epidemiological parameter set Includes pathogen generation time Pathogen transmission rate Disinfectant activity dynamic parameters and recovery rate .

[0071] The specific implementation of step S04 is the same as described above, and will not be repeated in detail here.

[0072] The specific implementation method of step S05 is: to compile the preliminary prevention and control plan text set. Each prevention and control plan text in the document undergoes a line-by-line cross-check using a physical feasibility verification rule base to extract disinfectant concentration values ​​from the prevention and control plan text. Contact time value Drug concentration value Compare with the corresponding safe concentration range in the rule base. Feasibility score. The calculation is expressed as follows:

[0073] ;

[0074] In the formula, The feasibility score is dimensionless and ranges from 0 to 1. This is a dimensionless indicator of the compliance of disinfectant concentration values. Within a safe range ,otherwise ; The contact time value conformity index is dimensionless. Within a safe range ,otherwise ; This is a dimensionless index for the compliance of drug concentration values. Within a safe range ,otherwise ; , , The weighting coefficients are dimensionless and satisfy the following conditions: The empirical values ​​are 0.4, 0.3, and 0.3 respectively; This represents the drug concentration value, in units of... .

[0075] The specific implementation method of step S06 is: calculating the feasibility score. The feasibility threshold is not met. The prevention and control plan text is invoked with correction prompts to trigger the large model to regenerate the plan text until a prevention and control plan text that has passed feasibility verification is generated, and the feasibility threshold is reached. The empirical value is 0.8, dimensionless. The text of the feasibility-verified prevention and control plan is matched with a database of historical prevention and control effectiveness based on text similarity. The calculation uses the cosine similarity algorithm, which is described as follows:

[0076] ;

[0077] In the formula, This represents text similarity, is dimensionless, and ranges from 0 to 1. This is the first section of the current prevention and control plan text. Each eigenvector component is dimensionless. The first section of the historical prevention and control plan text Each eigenvector component is dimensionless. The dimension of the feature vector is dimensionless and defaults to 512. The Euclidean norm of the feature vector of the current prevention and control plan text is dimensionless and calculated as follows: ; The Euclidean norm of the feature vector of the historical prevention and control plan text is dimensionless and calculated as follows: Based on minimizing the rate of epidemic transmission. The objective is to rank the candidate prevention and control plan texts, using ranking indicators. The statement is as follows:

[0078] ;

[0079] In the formula, This is a ranking index, dimensionless, with a value range of 0 to 1; The pathogen transmission rate corresponding to the current prevention and control plan, in units of... ; The maximum pathogen transmission rate without any control measures, in units of ; and The weighting coefficients are dimensionless and satisfy the following conditions: The default values ​​are 0.7 and 0.3 respectively. Output the sorted optimal prevention and control plan text sequence.

[0080] The specific implementation of step S07 is as follows: when the fusion parameter vector Uncertainty score of any parameter Exceeding the uncertainty threshold When this happens, the system automatically adjusts to a conservative prevention and control mode, increasing the frequency of key disinfection measures, increasing the frequency of environmental monitoring and sampling, strengthening personnel protection levels, and adjusting uncertainty thresholds. The empirical value is 0.15, dimensionless; uncertainty score. The range is 0 to 1, and it is dimensionless.

[0081] The specific implementation method of step S08 is the same as described above, and will not be repeated in detail here.

[0082] The sparse coding dictionary learning layer learns from the dynamically weighted multi-source sensor data using the K-SVD algorithm. The mathematical form of its linear representation is: In the formula This is a time-calibrated multi-source sensor dataset with matrix dimensions of [missing information]. , The number of sensors is dimensionless. The number of sampling times is dimensionless. It is a sparse dictionary set of atoms, with matrix dimension O(n). , The number of atoms in the dictionary is dimensionless and is typically twice the number of sensors. The coefficients are sparse representations, and the matrix dimension is . The deep unfolding iterative layer will solve for the sparse representation coefficients. The iterative optimization algorithm is expanded into a multi-layer neural network, and in the th... The mathematical form of the iterative process in layer iteration is:

[0083] ;

[0084] In the formula, For the first The sparse representation coefficients after layer iteration have a matrix dimension of ; For the first The sparse representation coefficients of the layer have a matrix dimension of ; This is the iteration layer number, dimensionless, with a value range from 0 to the number of sensor types plus 1; This is the soft threshold function, and its functional form is: , For input variables, For symbolic functions, To find the maximum value function, for The absolute value; For the first The soft threshold function parameter for the layer is dimensionless and has an initial value of 0.1. This is the step size parameter, which is dimensionless and typically takes a value of 0.01. Let be the transpose of the sparse dictionary set of atoms, with dimension O(n). The joint loss function is expressed as follows:

[0085] ;

[0086] In the formula, The value of the joint loss function is dimensionless. The predicted fusion parameter vector for the sparse coding depth unfolded fusion network model; The optimal fusion parameter vector is manually annotated; The mean square error is dimensionless and is calculated as follows: , and The predicted fusion parameter vector and the optimal fusion parameter vector are respectively the first two. One portion, For the first Reference values ​​for each component are used for dimensionless processing. ℃, %, mg / m^3 F_{ref,4} = 10^6 ; For fusion accuracy, dimensionless, its calculation form is as follows: ; and These are weighting coefficients, dimensionless, with empirical values ​​of 0.6 and 0.4, respectively.

[0087] Number of heads in the multi-head attention mechanism in the dual LoRA fine-tuning mechanism The calculation method is as follows:

[0088] ;

[0089] In the formula, The number of heads in the multi-head attention mechanism is dimensionless. For the fusion parameter vector The dimension is dimensionless and defaults to 4. For epidemiological parameter set The number of parameter types, dimensionless, defaults to 4; This represents the function for rounding up.

[0090] It should be noted that the variables involved in this embodiment are explained in detail in Tables 1 and 2.

[0091] Table 1. Variable Explanation Table (Part 1)

[0092]

[0093] Table 2. Variable Explanation Table (Part Two)

[0094]

[0095] To better understand and implement this invention, a specific application scenario of the invention is provided below as Example 2: To verify the effectiveness of the invention, technicians built a test environment and deployed a multi-source sensor network in a large broiler farm, running the animal disease prevention and control plan generation system of this invention. The farm has 8 chicken houses, each housing 15,000 chickens, with a breeding cycle of 42 days. Technicians deployed temperature sensors, relative humidity sensors, and other sensors in each chicken house. Gas concentration sensor The gas concentration sensor and pathogen concentration sensor collect data every 5 minutes.

[0096] On the 15th day of system operation, technicians detected suspected early signs of avian influenza in chicken house No. 2. The raw data collected by the multi-source sensors at this time is shown in Table 3.

[0097] Table 3 Raw data from multi-source sensors

[0098]

[0099] The system first performs time calibration on the collected raw multi-source sensor data, unifying the timestamp differences caused by sampling clock deviations between different sensors to the same time base. The time-calibrated multi-source sensor dataset and the historical reliability scores of each sensor are input into the sparse coding-deep unrolling fusion network model. Based on the historical reliability scores in Table 1, the model's dynamic weight adjustment layer assigns a weight coefficient of 0.511 to temperature sensor A, 0.489 to temperature sensor B, 0.486 to relative humidity sensor A, and 0.514 to relative humidity sensor B, ensuring that the data from the more reliable sensors have a greater influence during the fusion process.

[0100] The sparse coding dictionary learning layer learns from the dynamically weighted multi-source sensor data using the K-SVD algorithm, generating a sparse dictionary atom set containing 64 atoms. The deep unrolling iterative layer is configured as a 6-layer neural network, corresponding to a configuration of 4 sensor types plus 2 layers. After 6 iterations of the deep unrolling iterative layer, the system outputs 64 sparse representation coefficients at the 6th layer, with values ​​of 0.73, 0.18, 0.06, and 0.03. The fusion output layer linearly combines these sparse representation coefficients with the sparse dictionary atoms to generate a fusion parameter vector, where the fused temperature parameter is 28.85℃, the fused relative humidity parameter is 66.62%, and the fusion... The gas concentration parameter is 18.5 mg / L. fusion The gas concentration parameter is 3.2 mg / L. The fusion pathogen concentration parameter is CFU / m .like Figure 2 As shown, the fusion parameter vector error of the sparse coding-deep unfolding fusion network model is stable within ±4.8%, which is significantly better than the ±12.3% error range of the traditional weighted average method.

[0101] The system calculates the epidemiological parameter set using an animal epidemiological parameter coupling propagation algorithm based on the fused parameter vector. Based on the SEIR transmission dynamics model, the system establishes a set of differential equations for four states: susceptible individuals, latent infected individuals, clinically infected individuals, and recovered individuals. Based on the fused temperature parameter of 28.85℃ and the fused relative humidity parameter of 66.62%, the system calculates the pathogen generation time to be 3.2 days through curve fitting of infection experimental data. The system calculates the pathogen transmission rate to be 2.8, indicating that under the current environmental conditions, one infected individual can infect an average of 2.8 susceptible individuals. For the commonly used disinfectant peracetic acid, the system calculates the dissociation kinetic parameter to be 0.45 based on the disinfectant dissociation kinetic equation, indicating that under the conditions of a disinfectant concentration of 0.3% and a contact time of 30 minutes, the pathogen concentration can be reduced to [a certain percentage] of its original concentration. times.

[0102] like Figure 3 As shown, the system inputs the fused parameter vector and the epidemiological parameter set into a domestically developed open-source large-scale model employing dual LoRA fine-tuning. The first LoRA fine-tuning section fine-tunes the first 12 layers of the base model, learning the mapping relationship between aquaculture environment parameters and the semantics of the prevention and control plan text through low-rank matrix factorization. The second LoRA fine-tuning section fine-tunes the last 12 layers of the base model, dynamically setting the number of heads in the multi-head attention mechanism to 1 based on the fused parameter vector dimension of 4 and the number of parameter types in the epidemiological parameter set of 3. The system outputs a preliminary prevention and control plan text set containing 5 candidate prevention and control plans, each including disinfection measures, isolation measures, medication recommendations, and environmental control recommendations.

[0103] Among them, domestically developed open-source models can use ChatGLM6B or Qwen7B, etc.

[0104] The system performs a line-by-line cross-check of each prevention and control plan text in the preliminary prevention and control plan text set using a physical feasibility verification rule base. For the disinfection recommendation in Plan 1, the system extracts the peracetic acid concentration value of 0.5% and the contact time of 45 minutes from the prevention and control plan text. This is compared with the safe concentration range of peracetic acid (0.2% to 0.4%) in the rule base. The system finds that the concentration value of 0.5% exceeds the upper limit of the safe range (0.4%). The system generates a feasibility score of 0.62, which is lower than the feasibility threshold of 0.75, and automatically marks Plan 1 as unqualified. Based on the parameter type exceeding the safe threshold, the system generates a correction prompt: "According to the prevention and control specifications, the safe range for the peracetic acid concentration parameter is 0.2% to 0.4%, and the current recommended value is 0.5%. Please regenerate the prevention and control plan within the range of 0.2% to 0.4%." The correction prompts, along with the fusion parameter vector and the epidemiological parameter set, were input into the large model to regenerate the prevention and control plan text. The newly generated Plan 1 (revised version) adjusted the peracetic acid concentration to 0.35% and the contact time to 50 minutes. It was then re-entered into the physical feasibility verification rule base for verification. This time, the feasibility score was 0.86, meeting the feasibility threshold requirements.

[0105] The system extracts key prevention and control measures and parameter values ​​from the texts of five prevention and control plans that have passed physical feasibility verification, and performs text similarity matching with 327 historical prevention and control plan texts in the historical prevention and control effect database. For example... Figure 4As shown, the similarity between the revised version of Scheme 1 and the historical scheme HC0158 is 0.83, and the historical scheme reduced the disease incidence rate by 78% under similar environmental conditions. Scheme 2 has a similarity of 0.79 with the historical scheme HC0203, and the historical scheme reduced the disease incidence rate by 72%. Scheme 3 has a similarity of 0.81 with the historical scheme HC0176, and the disease incidence rate reduced by 75%. Scheme 4 has a similarity of 0.77 with the historical scheme HC0192, and the disease incidence rate reduced by 69%. Scheme 5 has a similarity of 0.75 with the historical scheme HC0215, and the disease incidence rate reduced by 66%.

[0106] The system comprehensively ranks candidate prevention and control plan texts based on the objectives of minimizing the epidemic transmission rate and maximizing disease control effectiveness. Using an animal epidemiological parameter-coupled propagation algorithm, the system simulates the epidemic transmission trajectory for each plan. Plan 1 (modified version) can reduce the pathogen transmission rate from 2.8 to 0.6 after implementation; Plan 2 can reduce it to 0.8; Plan 3 to 0.7; Plan 4 to 0.9; and Plan 5 to 1.1. Combining this with the disease incidence rate reduction data from the historical prevention and control effectiveness database, the system outputs the optimal prevention and control plan text sequence as follows: Plan 1 (modified version), Plan 3, Plan 2, Plan 4, and Plan 5. Figure 5 As shown, the revised version of the optimal prevention and control plan 1 includes the following main measures: Immediately initiate a comprehensive disinfection program for the chicken house, using a 0.35% peracetic acid solution for spray disinfection, with a contact time of 50 minutes, and disinfection frequency twice daily for 7 consecutive days; strictly isolate chicken house No. 2, suspending the flow of personnel and materials to other chicken houses; control the ambient temperature to the range of 26℃ to 28℃, and the relative humidity to the range of 55% to 65%; strengthen ventilation and air exchange. Concentration reduced to 10 mg / The following is a preventative administration of antiviral drugs to suspected infected chicken flocks, using Astragalus polysaccharide injection at a dose of 2 ml per kilogram of body weight, administered intramuscularly once daily for 5 consecutive days.

[0107] On the third day after implementing the revised version of the optimal prevention and control plan, technicians monitored the pathogen concentration in chicken house No. 2 using a pathogen concentration sensor and found that the pathogen concentration had decreased from [previous level]. CFU / m Reduce to CFU / m The decrease reached 82.9%. On day 7, the pathogen concentration further decreased to... CFU / m The spread of the epidemic was successfully controlled. Only house number 2 out of the eight chicken houses showed clinical symptoms, with a morbidity rate of 0.13%. The remaining seven houses did not experience any further spread. Technical personnel found through comparison that traditional methods rely on manual experience to formulate prevention and control plans, often using fixed disinfectant concentrations and dosages. This failed to dynamically adjust prevention and control measures based on real-time environmental parameters and pathogen concentrations, leading to significant uncertainty in the effectiveness of the control measures.

[0108] This invention achieves high-precision fusion of multi-source sensor data through a sparse coding-deep unrolling fusion network model, overcoming the limitations of traditional fixed-weight fusion methods that cannot adapt to dynamic changes in sensor reliability, and significantly reducing the error of the fusion parameter vector. The dynamic weight adjustment layer adaptively adjusts the weights of each sensor data based on the historical reliability scores of the sensors, ensuring that sensors with high-quality data play a leading role in the fusion process and effectively suppressing the interference of low-quality sensor data on the fusion results. The sparse coding dictionary learning layer uses the K-SVD algorithm to represent high-dimensional heterogeneous sensor data as a linear combination of a small number of sparse dictionary atoms, achieving data dimensionality reduction and feature extraction, and eliminating dimensional differences between different sensor data. The deep unrolling iterative layer unfolds the iterative optimization algorithm into a multi-layer neural network, allowing the soft threshold function parameters and intermediate sparse representation coefficients to be learned through backpropagation, maintaining the mathematical interpretability of iterative optimization while achieving the end-to-end optimization capabilities of deep learning. The physical feasibility verification rule base extracts key parameters from the prevention and control plan text and compares them with safe ranges. When a parameter exceeds the safe threshold, it generates a correction prompt, triggering the large model to regenerate the plan, forming a closed-loop error correction mechanism. This ensures the physical feasibility and safety of the output plan, avoiding unreasonable disinfectant concentrations or drug dosages that may occur due to the lack of verification mechanisms in traditional methods. The animal epidemiology parameter coupling propagation algorithm establishes a multi-parameter coupled differential equation for pathogen transmission based on the SEIR infectious dynamics model. It unifies environmental parameters, pathogen biological characteristics, and disinfection measures within a mathematical framework, transforming the generation of prevention and control plans from experience-based recommendations to rational design based on infectious dynamics models. This achieves a shift from subjective assessment to objective prediction of prevention and control effects, improving the scientific rigor and reliability of prevention and control plans.

[0109] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for generating animal disease prevention and control plans using large models, characterized in that, Multi-source sensor data from the farm environment is collected and time-calibrated to obtain a time-calibrated multi-source sensor dataset. The time-calibrated multi-source sensor dataset and the historical reliability scores of each sensor are input into a sparse coding-deep unrolling fusion network model to output a fusion parameter vector. Based on the fusion parameter vector, an animal epidemiological parameter coupling propagation algorithm is used to calculate the epidemiological parameter set. The fusion parameter vector and the epidemiological parameter set are input into an open-source large model with dual LoRA fine-tuning to output a preliminary prevention and control plan text set. The preliminary prevention and control plan text set is cross-checked using a physical feasibility verification rule base to generate a feasibility score. For prevention and control plan texts that do not meet the feasibility score requirements, correction prompts are invoked to trigger the large model to regenerate until the feasibility verification is passed. The verified prevention and control plan texts are matched with the historical prevention and control effect database for similarity and sorted according to the objective of minimizing the epidemic transmission rate to output the optimal prevention and control plan text sequence.

2. The method for generating animal disease prevention and control plans using a large model according to claim 1, characterized in that, The multi-source sensor data includes temperature parameters, relative humidity parameters, gas concentration parameters, and pathogen concentration parameters.

3. The method for generating animal disease prevention and control plans using a large model according to claim 2, characterized in that, The sparse coding-deep unrolling fusion network model includes a dynamic weight adjustment layer, a sparse coding dictionary learning layer, a deep unrolling iterative layer, and a fusion output layer.

4. The method for generating animal disease prevention and control plans using a large model according to claim 3, characterized in that, The dynamic weight adjustment layer adaptively adjusts the weight coefficients of each sensor data in the time-calibrated multi-source sensor dataset based on the historical reliability scores of each sensor.

5. The method for generating animal disease prevention and control plans using a large model according to claim 4, characterized in that, The sparse coding dictionary learning layer uses the K-SVD algorithm to learn from the dynamically weighted multi-source sensor data to generate a sparse dictionary atom set.

6. The method for generating animal disease prevention and control plans using a large model according to claim 5, characterized in that, The deep unfolding iterative layer unfolds the iterative optimization algorithm for solving sparse representations into a multi-layer neural network with the number of layers equal to the number of sensor types plus 2 layers. Each layer learns the soft threshold function parameters and intermediate sparse representation coefficients.

7. The method for generating animal disease prevention and control plans using a large model according to claim 6, characterized in that, The fusion output layer linearly combines the sparse representation coefficients output by the depth expansion iteration layer with the sparse dictionary atoms to generate a fusion parameter vector.

8. The method for generating animal disease prevention and control plans using a large model according to claim 7, characterized in that, The fusion parameter vector includes four components: fusion temperature parameter, fusion relative humidity parameter, fusion gas concentration parameter, and fusion pathogen concentration parameter.

9. The method for generating animal disease prevention and control plans using a large model according to claim 8, characterized in that, The animal epidemiological parameter coupling propagation algorithm is based on the SEIR infection dynamics model to establish differential equations that integrate the influence of environmental parameters on pathogen survival.

10. The method for generating animal disease prevention and control plans using a large model according to claim 9, characterized in that, The set of epidemiological parameters calculated by the animal epidemiological parameter coupling propagation algorithm includes pathogen generation time, pathogen transmission rate, and disinfectant inactivation biomechanical parameters.

Images