Poultry breeding record analysis method and system

Abstract

The invention discloses a poultry breeding record analysis method and system, and relates to the technical field of poultry farming, and the system comprises a multi-modal data dynamic fusion module, a self-adaptive dynamic modeling module, a multi-mechanism verification and screening module, and a system cooperation interface module. And the multi-modal data dynamic fusion module is used for carrying out dynamic knowledge graph construction and conflict cleaning on genotype data, environment time sequence data and phenotype data. According to the poultry breeding record analysis method and system, deep association and real-time fusion of genotype, environment and phenotype multi-source data are realized through a dynamic knowledge graph and a federated space-time embedding model, a data island barrier of a traditional system is effectively broken through, autonomous balance between a long-term breeding target and short-term dynamic response can be realized, and meanwhile, the method and the system have good application prospects. And a double verification mechanism of causal reasoning and extreme confrontation testing is adopted, so that the decision confidence is improved, and resource waste and risk lag caused by empirical decision are remarkably reduced.

Description

Technical Field

[0001] The invention relates to the technical field of poultry breeding, and in particular to a poultry breeding record analysis method and system. Background Art

[0002] As the poultry breeding industry moves towards intensive and precise production, the technical bottlenecks of traditional breeding record analysis systems are becoming increasingly prominent. Although the current mainstream system can realize the electronic storage and statistics of basic data, the underlying architecture based on the design of a single-dimensional structured database makes it difficult to integrate multi-source heterogeneous data. Data such as the genetic pedigree, dynamic phenotype and environmental monitoring of poultry are scattered in different subsystems or manual ledgers, lacking unified standards and association models, resulting in a serious "data island" phenomenon. Breeding analysis only stays at single-point statistics and cannot explore potential laws across dimensions. Moreover, the existing systems mostly set breeding target parameters based on fixed thresholds or manual experience. In the face of sudden epidemics, market fluctuations or genotype iterations, static models cannot be adjusted and optimized in real time, resulting in lagging breeding strategies. The algorithm module lacks dynamic learning capabilities, and traditional statistical analysis methods are insufficient for nonlinear association mining. It is difficult to build an iterative prediction model. When selecting disease-resistant strains, it is impossible to dynamically adjust the weight coefficient, and the selection decision relies on post-tracing. In addition, the current breeding analysis system has poor synergy with the downstream of the industrial chain. The changes in market demand for traits such as poultry meat quality cannot be quickly fed back to the breeding model; the production data on the breeding side cannot be optimized with the closed loop of genetic evaluation. The "analysis-decision-application" chain is broken, resulting in a waste of resources. Therefore, new analysis systems are urgently needed to break through data barriers and achieve intelligent decision-making to meet the complex challenges of modern poultry breeding. Summary of the invention

[0003] 1. Technical issues to be resolved

[0004] In view of the shortcomings of the prior art, the present invention provides a poultry breeding record analysis method and system, which solves the problem of how to build an environment-gene-phenotype multidimensional association analysis system through real-time fusion and dynamic modeling of multi-source heterogeneous data, and realize adaptive breeding strategy recommendation based on machine learning, so as to break through the technical bottlenecks of traditional system data siloization, static analysis, and empirical decision-making.

[0005] (II) Technical solution

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: a poultry breeding record analysis method and system, including the following functional modules: a multimodal data dynamic fusion module, an adaptive dynamic modeling module, a multi-mechanism verification and screening module, and a system collaborative interface module;

[0007] The multi-modal data dynamic fusion module is used to construct a dynamic knowledge graph and clean conflicts for genotype data, environmental time-series data, and phenotypic data. It should be further noted that environmental time-series data, including temperature, humidity, and light, SNP sequences of genotype data, and egg production rate and body weight in phenotypic data, are collected in real time through temperature and humidity sensors and gene sequencers in Internet of Things devices. The data is standardized into JSON or Protobuf in a unified format and transmitted to the central server.

[0008] The adaptive dynamic modeling module realizes multi-modal spatio-temporal embedding modeling based on the federated learning framework and generates dynamic breeding strategies through a two-layer optimization engine. It should be further noted that under the federated learning framework, local data from multiple farms are jointly trained in a model through gradient aggregation to avoid the leakage of raw data. Among them, the local data is encrypted and stored on the local server.

[0009] The multi-mechanism verification and screening module uses causal inference and adversarial stress testing to double-verify the strategy. It should be further noted that after the strategy is generated, a causal inference engine, such as the DoWhy library, is called to construct counterfactual scenarios, and at the same time, an adversarial test script is started to simulate extreme events.

[0010] The system collaboration interface module realizes protocol adaptation and visual decision-making interaction for multi-source heterogeneous devices. It should be further noted that in the specific implementation process, an adapter middleware is developed to parse heterogeneous interfaces such as the Modbus protocol of temperature control devices and the VCF files of gene sequencers, and push them to the visual interface in real time through WebSocket.

[0011] Preferably, the multi-modal data dynamic fusion module includes:

[0012] A dynamic ontology modeling unit that defines three-dimensional association rules for genetic lineages, environmental parameters, and phenotypic indicators based on a dedicated ontology library in the field of poultry breeding;

[0013] A knowledge graph evolution unit that uses the spatio-temporal sequence clustering algorithm ST-DBSCAN to identify the temporal coupling patterns between environmental parameters and phenotypic data and encodes them as dynamic weighted edges;

[0014] A dual-channel data cleaning unit that sets hard filtering thresholds for genotype and phenotypic data through a rule engine and soft-corrects sensor time-series data based on the generative adversarial network GAN.

[0015] Preferably, the adaptive dynamic modeling module includes:

[0016] The Heterogeneous Federated Spatiotemporal Embedding Model (Hetero-FSTE) maps genotype SNP sequences, phenotypic time series, and environmental tensors to a unified vector space and uses the Gated Spatiotemporal Attention Mechanism (GSTA) to capture gene-environment interaction effects. It should be further noted that in the specific implementation process:

[0017] Genotype embedding includes encoding the SNP sequence into a 128-dimensional vector using One-hot + self-attention mechanism, where the SNP sequence includes AA, AT, and TT.

[0018] Phenotype embedding includes performing wavelet transform on the daily recorded egg production rate time series to extract frequency domain features, and then generating a 64-dimensional vector through a GRU network.

[0019] Environmental embedding includes converting the temperature and humidity heat map divided by chicken coop areas into a three-dimensional tensor and outputting a 32-dimensional vector through a 3D convolutional layer.

[0020] The Gated Spatiotemporal Attention (GSTA) includes calculating the cross-attention scores between gene vectors and environmental vectors and screening significant interaction terms, such as the weight of SNP_A on egg production rate increasing in a high-temperature environment.

[0021] The longitudinal federated learning architecture allows local servers to retain the original data and only upload encrypted model gradients to the central server for aggregation to achieve cross-subject data privacy protection and format heterogeneity compatibility. It should be further noted that in the specific implementation process, it includes:

[0022] Data alignment: Based on differential privacy technology, align the individual IDs of multiple farms in the encrypted state, such as SHA-256 hash encryption.

[0023] Gradient aggregation: The local model, i.e., Farm A, calculates the gradient uploads it to the central server after encryption, and performs weighted averaging to update the global model.

[0024] Preferably, the adaptive dynamic modeling module further includes a two-layer optimization decision engine, and the two-layer optimization decision engine includes a global Deep Q-Network (DQN) and a Real-time Proximal Policy Optimization (PPO) algorithm; the global Deep Q-Network generates a long-term breeding target genotype combination based on historical data; the real-time proximal policy optimization algorithm receives environmental monitoring data to dynamically adjust the weights of the objective function, and verifies the policy compatibility through Monte Carlo Tree Search (MCTS).

[0025] Preferably, the multi-mechanism verification and screening module includes:

[0026] The counterfactual causality verification unit constructs a counterfactual scenario for the strategy based on the potential outcome model (PO Model), compares the actual data with the counterfactual prediction results, and quantifies the confidence level of the causal effect of the strategy. It should be further noted that in the specific implementation process:

[0027] Construct a counterfactual scenario: Assume that the "breeding SNP_A" strategy recommended by the system is not adopted, and predict the egg production rate of the non-breeding population based on the potential outcome model (PO Model).

[0028] Calculate the causal effect: The difference between the actual egg production rate Y1 and the counterfactual egg production rate Y0, and its calculation formula is: Δ = Y1 - Y0. If Δ > 5% and the p-value < 0.05, it is determined that the strategy is effective.

[0029] The Extreme-AGAN (Extreme Scenario Adversarial Generator) simulates genotype mutations, drastic environmental changes, and sudden market reversals, locates the vulnerable links of the strategy through Shapley value analysis, and triggers model retraining.

[0030] It should be further noted that in the specific implementation process:

[0031] Generate genotype mutations: Randomly mask 10% of the SNP sites to simulate sequencing errors.

[0032] Drastic environmental changes: Inject a continuous 7-day high-temperature (+5°C) event into the temperature and humidity data.

[0033] Sudden market reversal: Assume a 30% decrease in poultry meat demand and recalculate the breeding target benefits.

[0034] Shapley value analysis: Locate the features that have the greatest impact on the stability of the strategy. For example, if the weight shift caused by temperature control equipment failure > 20%, trigger model retraining, thereby ensuring the scientific nature and risk resistance ability of the strategy.

[0035] Preferably, the counterfactual causality verification unit performs the following operations:

[0036] Construct a counterfactual hypothesis scenario for the system-recommended strategy and calculate the differential contribution before and after the implementation of the strategy.

[0037] When the confidence level of the causal effect is lower than the preset threshold, automatically trigger strategy rollback and model iterative update.

[0038] The differential contribution is calculated jointly through Bayesian network and intervention factor decomposition.

[0039] Preferably, the system collaboration interface module includes:

[0040] Intelligent Protocol Adapter IPA, which supports second-level protocol parsing for the VCF file interface of the gene sequencing platform, the Modbus protocol of intelligent environmental control devices, and the market data API; it should be further noted that in the specific implementation process, it includes:

[0041] Gene sequencing platform interface: Parse SNP site information (such as chr1: 123456A / T) in the VCF file and convert it into a unified genotype code (0 / 1 / 2 corresponding to AA / AT / TT);

[0042] Modbus protocol parsing: Read the register address of the environmental control device (such as 0x0001 for the temperature value), convert it into a floating-point number and add a timestamp;

[0043] Market data API: Obtain the poultry futures price through OAuth2.0 authorization and integrate it into the breeding income model on a daily granularity.

[0044] Three-dimensional breeding situation map, dynamically visualizing the genotype diffusion trend, environmental hot zone distribution, and strategy income heat map, and providing an interactive tuning interface for manual expert annotation and AI strategies. It should be further noted that in the specific implementation process, use WebGL to render the three-dimensional model of the chicken coop, mark the genotype diffusion area with different colors, and red is the high-incidence area; experts can manually adjust the strategy parameters, such as the laying rate weight from 0.6 → 0.8, and the system will simulate the adjusted income heat map in real time. Through the setting of the system collaboration interface module, seamless access of multi-source devices and visual decision support are realized.

[0045] Preferably, the dual-channel data cleaning unit includes:

[0046] Adversarial neural network discriminator, which performs secondary noise detection on the data preliminarily cleaned by the rule engine;

[0047] Dynamic weight allocator, which adjusts the cleaning intensity according to the data source credibility. Among them, in the data source credibility: gene sequencing platform > manual entry > edge sensor;

[0048] The data repair result is finally arbitrated by comparing the prediction with the true value through the bidirectional long short-term memory network Bi-LSTM.

[0049] Preferably, the Monte Carlo Tree Search (MCTS) verification process includes: simulating the environmental fluctuations and market change paths in the next 30 days before policy deployment, generating a Markov chain model based on historical data, predicting the temperature and humidity transfer probabilities in the next 30 days. For example, when the temperature on the t-th day > 28°C, the probability of temperature drop on the (t + 1)-th day is 60%; evaluating the stability of the target traits and the growth rate of resource consumption of the policy under the simulated path, calculating the feed increment, immunization cost, etc. required for policy execution. If the growth rate > 20%, i.e., the preset threshold, an alarm is triggered; when the growth rate of resource consumption exceeds the preset threshold, the real-time layer policy weight reallocation is automatically triggered, reducing the Q-value weight of the high-resource consumption policy, such as from 0.8 → 0.6, and preferentially selecting the policy branch with a cost growth rate < 15%. Through the implementation of the Monte Carlo Tree Search (MCTS) verification process, the stability of the policy in the simulated environment is verified.

[0050] Preferably, the Monte Carlo Tree Search (MCTS) verification process further includes:

[0051] A dynamic environment path generator that constructs a probability distribution model of temperature, humidity, and light intensity within the next 30 days based on historical environmental data and real-time meteorological prediction APIs;

[0052] A policy robustness evaluation unit that quantifies the policy stability through the following steps:

[0053] 1) Inject random perturbation events into the simulated path, such as temperature control failure caused by equipment failure and feed supply interruption;

[0054] 2) Calculate the coefficient of variation of the target traits before and after the perturbation. If the coefficient of variation exceeds the preset threshold, it is determined that the policy robustness is insufficient. The target traits include egg production rate and weight growth rate;

[0055] 3) According to the evaluation results, automatically adjust the exploration-exploitation balance parameter in the real-time proximal policy optimization (PPO) algorithm, and preferentially select the policy branch that is insensitive to perturbations;

[0056] The dynamic adjustment process is achieved through the joint optimization of gradient backpropagation and policy entropy value, ensuring the self-adaptability of the model in complex scenarios.

[0057] A method for analyzing poultry breeding records includes the following steps:

[0058] S1: Conduct dynamic knowledge graph construction and conflict cleaning on genotype data, environmental time series data, and phenotype data;

[0059] S2: Implement multi-modal spatio-temporal embedding modeling based on the federated learning framework, and generate dynamic breeding strategies through a two-layer optimization engine;

[0060] S3: Conduct double verification on the policy using causal reasoning and adversarial stress testing;

[0061] S4: Perform protocol adaptation and visualized decision interaction for multi-source heterogeneous devices.

[0062] (III) Beneficial effects

[0063] The present invention provides a poultry breeding record analysis method and system, which has the following beneficial effects:

[0064] (I) The poultry breeding record analysis method and system, through dynamic knowledge graph and federated spatiotemporal embedding model, realizes the deep association and real-time integration of multi-source data of genotype, environment and phenotype, effectively breaking the data island barrier of traditional systems. Based on the collaborative optimization of two-layer reinforcement learning and Monte Carlo tree search verification, the system can autonomously balance between long-term breeding goals and short-term dynamic responses, improve the efficiency of identifying individuals with high genetic value, and shorten the response time of strategy adjustment. At the same time, the dual verification mechanism of causal reasoning and extreme adversarial testing will improve the confidence of decision-making and significantly reduce the waste of resources and risk lag caused by empirical decision-making.

[0065] (ii) The poultry breeding record analysis method and system can simulate complex scenarios such as epidemics, extreme climates, and market fluctuations through a dynamic environmental path generator and quantitative assessment of strategy robustness, and automatically optimize strategy parameters based on the coefficient of variation threshold to ensure improved trait stability under disturbance events. In addition, the intelligent protocol adapter and three-dimensional visualization interactive design have opened up the closed loop of the entire industry chain of gene sequencing, environmental monitoring, and market data, thereby improving the matching degree between breeding strategies and industry needs and optimizing resource utilization. This solution combines technical foresight with feasibility, providing core support for the intelligent upgrade of the poultry breeding industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 It is a schematic diagram of the overall framework of the present invention. DETAILED DESCRIPTION

[0067] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0068] See also Figure 1 ,The present invention provides a technical solution: a poultry breeding record analysis method and system, including the following functional modules: a multimodal data dynamic fusion module, an adaptive dynamic modeling module, a multi-mechanism verification and screening module, and a system collaborative interface module;

[0069] The multi-modal data dynamic fusion module is used to construct a dynamic knowledge graph and clean conflicts for genotype data, environmental time-series data, and phenotypic data;

[0070] The adaptive dynamic modeling module realizes multi-modal spatio-temporal embedding modeling based on the federated learning framework and generates dynamic breeding strategies through a two-layer optimization engine;

[0071] The multi-mechanism verification and screening module uses causal reasoning and adversarial stress testing to double-verify the strategies;

[0072] The system collaboration interface module realizes protocol adaptation of multi-source heterogeneous devices and visual decision-making interaction.

[0073] The multi-modal data dynamic fusion module includes:

[0074] The dynamic ontology modeling unit defines three-dimensional association rules for genetic pedigrees, environmental parameters, and phenotypic indicators based on a dedicated ontology library in the field of poultry breeding; it should be further noted that in the specific implementation process, defining the three-dimensional association rules of the ontology library includes the genetic dimension, environmental dimension, and phenotypic dimension; among them, in the genetic dimension, the association rule between SNP loci and dominant traits, such as the rs123 locus A → egg production rate +5%; in the environmental dimension, the influence function of temperature and humidity thresholds on growth rate, such as the body weight growth rate decreases when the temperature > 28°C; in the phenotypic dimension, the non-linear relationship model between egg production rate and feed intake.

[0075] The knowledge graph evolution unit uses the spatio-temporal sequence clustering algorithm ST-DBSCAN to identify the temporal coupling patterns of environmental parameters and phenotypic data and encodes them as dynamic weight edges; it should be further noted that in the specific implementation process, the ST-DBSCAN algorithm is used to process environmental time-series data, including inputting the temperature sequence T = {t1, t2,..., t n}, humidity sequence H = {h1, h2,..., h n}; clustering to identify the period with a temperature fluctuation > 5°C and humidity > 80% for 3 consecutive days as the "high temperature and high humidity cluster"; it also includes association, that is, associating this cluster with individuals whose egg production rate has decreased by more than 10% during the same period to generate a dynamic edge weight, weight = correlation coefficient × 0.8.

[0076] Dual-channel data cleaning unit, which sets the hard filtering threshold for genotype and phenotype data through a rule engine and soft-corrects the sensor time-series data based on the Generative Adversarial Network (GAN). It should be further noted that in the specific implementation process, it includes rule engine cleaning and GAN soft correction. Rule engine cleaning includes hard filtering, that is, deleting SNP sites with sequencing depth < 20X and abnormal records with daily weight gain > 15%; GAN soft correction includes a generator and a discriminator. The LSTM network of the generator simulates the normal temperature and humidity fluctuation pattern; the Convolutional Neural Network (CNN) of the discriminator distinguishes real data from generated data; after adversarial training, smooth repair is performed on sensor jump data, such as the temperature instantaneously changing from 25°C → 40°C → 25°C.

[0077] The adaptive dynamic modeling module includes:

[0078] Heterogeneous Federated Spatio-Temporal Embedding Model (Hetero-FSTE), which maps genotype SNP sequences, phenotype time series, and environmental tensors to a unified vector space and uses the Gated Spatio-Temporal Attention Mechanism (GSTA) to capture gene-environment interaction effects;

[0079] Longitudinal federated learning architecture, which allows local servers to retain the original data and only upload encrypted model gradients to the central server for aggregation to achieve cross-subject data privacy protection and format heterogeneity compatibility.

[0080] The adaptive dynamic modeling module also includes a two-layer optimization decision engine, which contains a global Deep Q-Network (DQN) and a Real-Time Proximal Policy Optimization (PPO) algorithm; the global Deep Q-Network generates long-term breeding target genotype combinations based on historical data; the Real-Time Proximal Policy Optimization algorithm receives environmental monitoring data to dynamically adjust the weights of the objective function and verifies the policy compatibility through Monte Carlo Tree Search (MCTS). It should be further noted that in the specific implementation process, it includes:

[0081] Global Deep Q-Network (DQN):

[0082] Input: Breeding data (genotype, environment, phenotype) for the past 3 years;

[0083] Output: Q-value matrix, representing the long-term benefits of different genotype combinations, such as the total number of eggs produced within 5 years, where different genotype combinations include SNP_A + SNP_B;

[0084] Training: Explore the optimal combination through the ε-greedy strategy, and the objective function is to maximize the discounted cumulative reward.

[0085] Real-Time Proximal Policy Optimization (PPO):

[0086] Input: Real-time environmental data, such as the latest epidemic report, market price fluctuations;

[0087] Policy Update: Sample the latest data every 2 hours, calculate the policy gradient and limit the update amplitude. The Clip threshold ε = 0.2 to prevent model oscillation;

[0088] Monte Carlo Tree Search MCTS Verification:

[0089] Simulation: Generate 1000 environmental paths for the next 30 days, with temperature and humidity normally distributed and perturbed by ±10%;

[0090] Evaluation: Calculate the variance of the target trait under each path. If the variance > 15%, the policy is determined to be unstable. Among them, the target trait is the egg production rate;

[0091] Backtracking: Select the top 10% of the paths with the smallest variance, backpropagate to the advantage function of the PPO algorithm, and adjust the policy weights.

[0092] The multi-mechanism verification and screening module includes:

[0093] Counterfactual Causal Verification Unit: Based on the Potential Outcome Model PO Model, construct a counterfactual scenario for the policy, compare the actual data with the counterfactual prediction results, and quantify the confidence of the policy causal effect;

[0094] Extreme Scenario Adversarial Generator Extreme-AGAN: Simulate genotype mutations, environmental upheavals, and market reversals, and locate the vulnerable links of the policy through Shapley value analysis and trigger model retraining.

[0095] The counterfactual causal verification unit performs the following operations:

[0096] Construct a counterfactual hypothesis scenario for the system-recommended policy, and calculate the differential contribution before and after the policy implementation;

[0097] When the confidence of the causal effect is lower than the preset threshold, automatically trigger policy rollback and model iterative update;

[0098] The differential contribution is calculated jointly through Bayesian network and intervention factor decomposition.

[0099] It should be further noted that in the specific implementation process, it includes:

[0100] Calculation of differential contribution:

[0101] Bayesian network: Construct a causal graph of SNP loci, environmental parameters, and phenotypic indicators, calculate the posterior probability P, egg production rate ↑|SNP_A = 1, temperature < 25°C;

[0102] Intervention factor decomposition: Separate the environmental interference term through do-calculus, such as the expected change in egg production rate when do(temperature = 25°C);

[0103] Policy rollback mechanism:

[0104] If the causal effect confidence level < 90%, the system automatically rolls back to the previous stable version policy and marks the current policy as "high risk".

[0105] Trigger model iteration: Increase the weight of adversarial training samples and preferentially learn historical successful policy patterns. Through the above operations performed by the counterfactual causal verification unit, quantify the causal effect and achieve dynamic policy iteration.

[0106] The system cooperation interface module includes:

[0107] Intelligent Protocol Adapter IPA, which supports second-level protocol parsing for the VCF file interface of the gene sequencing platform, the Modbus protocol of intelligent environmental control devices, and the market data API.

[0108] Three-dimensional breeding situation map, which dynamically visualizes the genotype diffusion trend, environmental hot zone distribution, and policy benefit heat map, and provides an interactive tuning interface for manual expert annotation and AI strategies.

[0109] The dual-channel data cleaning unit includes:

[0110] Adversarial neural network discriminator, which performs secondary noise detection on the data preliminarily cleaned by the rule engine; including: Input: temperature and humidity data after rule cleaning. During training, the generator generates synthetic data, and the discriminator distinguishes between real data and synthetic data; Output: Probability marking is performed on suspected noise points, such as a temperature drop of 3°C for 10 minutes. If the probability > 70%, repair is triggered.

[0111] Dynamic weight allocator, which adjusts the cleaning intensity according to the data source credibility. Among them, in the data source credibility: gene sequencing platform > manual entry > edge sensor; the weight of gene sequencing platform data in the credibility rule = 0.9, manual entry = 0.7, edge sensor = 0.5. Among them, the cleaning intensity includes GAN correction with an allowable correction range of ±5% for sensor data, while gene data only allows ±1% correction.

[0112] The data repair result is finally arbitrated by comparing the prediction with the true value through the bidirectional long short-term memory network Bi-LSTM. It includes the difference sequence of the input rule-cleaned data and the GAN-repaired data; the reasonable value range at the next time point is predicted through Bi-LSTM. If both sets of data exceed the range, it is marked as "pending manual review". Through the application of the dual-channel data cleaning unit, the data cleaning accuracy and efficiency are improved.

[0113] The Monte Carlo Tree Search MCTS verification process includes:

[0114] Simulate the environmental fluctuations and market change paths in the next 30 days before policy deployment.

[0115] Evaluate the stability of the target traits and the growth rate of resource consumption under the simulation path;

[0116] When the growth rate of resource consumption exceeds the preset threshold, automatically trigger the reallocation of the real-time layer policy weights.

[0117] The Monte Carlo Tree Search (MCTS) verification process further includes:

[0118] A dynamic environment path generator, based on historical environmental data and real-time meteorological prediction APIs, constructs a probability distribution model of temperature, humidity, and light intensity within the next 30 days; accesses the meteorological API to obtain 7-day forecast data and constructs an ARIMA model to predict the probability distribution of temperature and humidity for the subsequent 23 days with a 95% confidence interval;

[0119] A policy robustness evaluation unit quantifies the policy stability through the following steps:

[0120] 1) Inject random perturbation events into the simulation path, including temperature control failures caused by equipment failures and feed supply interruptions. Among them, for equipment failures: randomly select 5% of the temperature control nodes and simulate a 48-hour failure with random temperature fluctuations of ±8°C; for feed interruptions: randomly insert an event of a 50% decrease in feed supply for 3 days in the simulation path;

[0121] 2) Calculate the coefficient of variation of the target traits (egg production rate, weight growth rate) before and after the perturbation. If the coefficient of variation exceeds the preset threshold, it is determined that the policy robustness is insufficient. The formula for calculating the coefficient of variation is: the coefficient of variation of egg production rate CV = standard deviation / mean. If CV > 0.25 threshold, it is determined that the policy fails;

[0122] 3) According to the evaluation results, automatically adjust the exploration-exploitation balance parameter in the Proximal Policy Optimization (PPO) algorithm for real-time proximal policy optimization, and preferentially select the policy branches that are insensitive to perturbations; the exploration rate ε of the original PPO algorithm is 0.3. When the policy robustness is insufficient, it is reduced to ε = 0.1, and the historical stable policies are preferentially utilized;

[0123] The dynamic adjustment process is achieved through the joint optimization of gradient backpropagation and policy entropy value, ensuring the self-adaptability of the model in complex scenarios, that is, increasing the entropy penalty term of the policy distribution to prevent over-reliance on a single policy branch. It should be further noted that in the specific implementation process, the Monte Carlo Tree Search (MCTS) verification process enhances the robustness of the policy under extreme perturbations.

[0124] It should be further explained that in the specific implementation process, this solution realizes the deep association and real-time integration of multi-source data such as genotype, environment, and phenotype through dynamic knowledge graph and federated spatiotemporal embedding model, effectively breaking the data island barriers of traditional systems. Based on the collaborative optimization of two-layer reinforcement learning and Monte Carlo tree search verification, the system can autonomously balance between long-term breeding goals and short-term dynamic responses, improve the efficiency of identifying individuals with high genetic value, and shorten the response time of strategy adjustment. At the same time, the dual verification mechanism of causal reasoning and extreme adversarial testing will improve the confidence of decision-making and significantly reduce the waste of resources and risk lag caused by empirical decision-making.

[0125] Through the dynamic environmental path generator and quantitative evaluation of strategy robustness, the system can simulate complex scenarios such as epidemics, extreme climates, market fluctuations, and automatically optimize strategy parameters based on coefficient of variation thresholds, such as egg production rate CV <0.25, to ensure that trait stability is improved under disturbance events. In addition, the intelligent protocol adapter and three-dimensional visualization interactive design have opened up the closed loop of the entire industry chain of gene sequencing, environmental monitoring, and market data, improving the matching degree between breeding strategies and industry needs and optimizing resource utilization. This solution combines technical foresight with feasibility, providing core support for the intelligent upgrade of the poultry breeding industry.

[0126] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0127] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A poultry breeding record analysis system, characterized in that, Including: Multimodal data dynamic fusion module, adaptive dynamic modeling module, multi-mechanism verification and screening module, system collaborative interface module; The multimodal data dynamic fusion module is used to construct a dynamic knowledge graph and clean conflicts for genotype data, environmental time-series data, and phenotypic data; The adaptive dynamic modeling module realizes multimodal spatio-temporal embedding modeling based on the federated learning framework and generates dynamic breeding strategies through a two-layer optimization engine; The multi-mechanism verification and screening module uses causal reasoning and adversarial stress testing to double-verify the strategy; The system collaborative interface module realizes protocol adaptation of multi-source heterogeneous devices and visual decision-making interaction.

2. The poultry breeding record analysis system according to claim 1, wherein: The multimodal data dynamic fusion module includes: Dynamic ontology modeling unit, which defines three-dimensional association rules for genetic pedigree, environmental parameters, and phenotypic indicators based on a dedicated ontology library in the poultry breeding field; Knowledge graph evolution unit, which uses the spatio-temporal sequence clustering algorithm ST-DBSCAN to identify the temporal coupling patterns of environmental parameters and phenotypic data and encodes them as dynamic weight edges; Dual-channel data cleaning unit, which sets hard filtering thresholds for genotype and phenotypic data through a rule engine and soft-corrects sensor time-series data based on the generative adversarial network GAN.

3. The poultry breeding record analysis system according to claim 2, characterized in that: The adaptive dynamic modeling module includes: Heterogeneous federated spatio-temporal embedding model Hetero-FSTE, which maps genotype SNP sequences, phenotypic time series, and environmental tensors to a unified vector space and uses the gated spatio-temporal attention mechanism GSTA to capture gene-environment interaction effects; Longitudinal federated learning architecture, which allows local servers to retain the original data and only upload encrypted model gradients to the central server for aggregation, for cross-subject data privacy protection and format heterogeneity compatibility.

4. The poultry breeding record analysis system according to claim 3, wherein: The adaptive dynamic modeling module also includes a two-layer optimization decision engine, which includes a global deep Q-network DQN and a real-time proximal policy optimization PPO algorithm; the global deep Q-network generates long-term breeding target genotype combinations based on historical data; the real-time proximal policy optimization algorithm receives environmental monitoring data to dynamically adjust the weights of the objective function and verifies the strategy compatibility through Monte Carlo tree search MCTS.

5. The poultry breeding record analysis system according to claim 4, characterized in that: The Monte Carlo tree search MCTS verification process includes: Simulating future 30-day environmental fluctuations and market change paths before strategy deployment; Evaluating the target trait stability and resource consumption growth rate of the strategy under the simulated path; When the resource consumption growth rate exceeds the preset threshold, automatically trigger real-time layer strategy weight reallocation; Among them, the Monte Carlo tree search MCTS verification process further includes: Dynamic environmental path generator, which constructs a probability distribution model of temperature, humidity, and light intensity within the next 30 days based on historical environmental data and real-time meteorological prediction API; Strategy robustness evaluation unit, which quantifies strategy stability through the following steps: 1) Inject random perturbation events into the simulated path; 2) Calculate the coefficient of variation of the target trait before and after the perturbation. If the coefficient of variation exceeds the preset threshold, it is determined that the strategy robustness is insufficient. 3) According to the evaluation results, the exploration-exploitation balance parameters in the real-time proximal policy optimization (PPO) algorithm are automatically adjusted to give priority to the policy branches that are insensitive to disturbances.

6. The poultry breeding record analysis system according to claim 5, characterized in that: The multi-mechanism verification and screening module includes: The counterfactual causal verification unit builds a strategy counterfactual scenario based on the potential outcome model PO Model, compares the actual data with the counterfactual prediction results, and quantifies the confidence of the strategy causal effect; The extreme scenario adversarial generator Extreme-AGAN simulates genotype mutations, environmental upheavals, and market changes, locates strategy vulnerabilities through Shapley value analysis, and triggers model retraining.

7. The poultry breeding record analysis system according to claim 6, characterized in that: The counterfactual causal verification unit performs the following operations: Construct counterfactual scenarios for the system's recommended strategies and calculate the difference in contribution before and after the strategy is implemented; When the confidence level of the causal effect is lower than the preset threshold, the strategy rollback and model iterative update are automatically triggered; The difference contribution is calculated jointly by Bayesian network and intervention factor decomposition.

8. The poultry breeding record analysis system according to claim 7, characterized in that: The system coordination interface module includes: Intelligent protocol adapter IPA supports second-level protocol parsing of VCF file interface of gene sequencing platform, Modbus protocol of intelligent environmental control equipment and market data API; The three-dimensional breeding situation map dynamically visualizes the genotype diffusion trend, environmental hot zone distribution and strategy benefit heat map, and provides an interactive tuning interface for manual expert annotation and AI strategy.

9. The poultry breeding record analysis system according to claim 8, characterized in that: The dual-channel data cleaning unit comprises: Adversarial neural network discriminator, performs secondary noise detection on the data that has been initially cleaned by the rule engine; Dynamic weight allocator adjusts the cleaning intensity according to the credibility of the data source, where the credibility of the data source is: gene sequencing platform > manual input > edge sensor; The data restoration results are finally arbitrated by comparing the predictions of the bidirectional long short-term memory network Bi-LSTM with the actual values.

10. A method for analyzing poultry breeding records, characterized in that, The steps include: S1: Dynamic knowledge graph construction and conflict cleaning for genotype data, environmental time series data and phenotypic data; S2: Implement multimodal spatiotemporal embedding modeling based on a federated learning framework and generate dynamic breeding strategies through a two-layer optimization engine; S3: Use causal reasoning and adversarial stress testing to double-validate the strategy; S4: Perform protocol adaptation and visualized decision interaction for multi-source heterogeneous devices.

Images