Agricultural decision reasoning large model based on large language model and construction method thereof

Abstract

The invention discloses an agricultural decision reasoning large model based on a large language model and a construction method thereof. The method comprises the following steps: constructing a space-ground-air integrated multi-modal agricultural data acquisition and fusion module, integrating data of a ground sensor, unmanned aerial vehicle remote sensing, meteorological sensing and Internet of Things, solving the problem of data isomerism through reinforcement learning and a Transform architecture, and ensuring real-time integrity of the data. In combination with crop simulation model mechanisms such as DSSAT and the like, the inference rule base is dynamically updated to realize rapid iteration of agricultural knowledge; an agricultural feature extraction module and an adaptation layer are integrated, so that professional data understanding is enhanced; a multi-stage training strategy is adopted, and large-scale agricultural text fine tuning is carried out; reinforcement learning is introduced and expert feedback is combined to optimize decision generation, and a causal inference module is innovatively added to improve the inference precision. The model fuses the language ability of a large language model and agricultural knowledge, can accurately predict and diagnose agricultural problems and provide decision suggestions, and significantly improves the intelligence, precision and high efficiency level of agricultural production.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of agricultural information technology and computer science and technology, in particular, to the field of natural language processing (NLP), database technology and large language model, and particularly relates to an agricultural decision reasoning large model based on a large language model and a construction method thereof. BACKGROUND

[0002] Current agricultural production is facing the contradiction between large-scale planting and fine management, and the traditional experience-driven decision-making mode has been unable to meet the demand of modern agriculture for yield improvement, quality guarantee and risk prevention. With the penetration of Internet of Things and big data technology in the field of agriculture, a large amount of agricultural data has been generated, but the data island phenomenon is serious, and the fusion analysis of multi-source heterogeneous data has become a problem.

[0003] The existing agricultural decision system mainly depends on rule engine or traditional machine learning model, and has the problems of weak reasoning ability, poor generalization, slow knowledge updating, etc. Although the general large language model such as LLama has strong text understanding and generation ability, it has the defect of insufficient knowledge adaptability in the agricultural professional field: first, it lacks the processing ability of agricultural spatio-temporal dynamic data, second, it is difficult to realize the causal relationship reasoning, and third, the decision result lacks explainability. Therefore, constructing a large model in the field of agriculture with professional reasoning ability has become the key to breaking through the bottleneck of agricultural intelligence. SUMMARY

[0004] The present application aims to provide an agricultural decision reasoning large model based on a large language model and a construction method thereof, which realizes intelligent decision support for the whole process of agricultural production by constructing an agricultural exclusive knowledge system, optimizing the model reasoning mechanism and strengthening the multi-source data fusion capability, and solves the problems of low reasoning precision and poor scene adaptability in the prior art.

[0005] The present application discloses an agricultural decision reasoning large model based on a large language model and a construction method thereof, mainly comprising the following steps:

[0006] 1. Multi-modal agricultural data acquisition and fusion module.

[0007] 1) Agricultural data acquisition network: including ground sensor array (acquiring soil moisture condition and crop physiological parameters), unmanned aerial vehicle remote sensing system (acquiring multi-spectral images of crop growth), meteorological satellite data interface (acquiring regional climate data) and Internet of Things equipment (acquiring agricultural machinery operation data).

[0008] 2) Construction of data fusion framework: adopting federated learning technology to realize distributed data desensitization processing, and through the multi-head attention mechanism of the Transformer architecture, converting structured data (meteorological index, soil index) and unstructured data (agricultural text) into a unified feature space to solve the problem of data heterogeneity.

[0009] 3) Establish a dynamic data update mechanism: set data quality evaluation indicators, realize real-time data cleaning and feature extraction through edge computing nodes, and ensure the effectiveness of the input model data.

[0010] 2. Agricultural knowledge enhancement system construction module.

[0011] 1) Develop a knowledge dynamic update engine: extract new knowledge from agricultural journal papers and agricultural technology promotion files through the BERT model, and use knowledge distillation technology to convert expert experience into computable rules, and automatically update agricultural knowledge content every month.

[0012] 2) Build a reasoning rule base: organize 1000+ typical production scenarios decision trees, such as "drought stress-irrigation strategy" reasoning chain, to provide explicit reasoning basis for the model.

[0013] 3. Large language model agricultural adaptation module.

[0014] 1) Model structure optimization: add an agricultural feature adaptation layer to the large language base architecture, extract crop features through a 12-layer fully connected neural network, and expand the model's professional knowledge capacity.

[0015] 2) Staged training strategy: use more than 3 million agricultural text corpus for domain fine-tuning in the first stage, use 500,000 annotated production cases for full-parameter fine-tuning in the second stage, and use GRPO algorithm to integrate expert feedback to optimize decision preferences in the third stage.

[0016] 3) Reasoning mechanism innovation: introduce causal inference module, identify key influencing factors (such as the causal weight of precipitation on wheat lodging) through Do-Calculus framework; design multi-step reasoning chain, decompose complex decision-making problems into "problem diagnosis-solution generation-effect prediction" three-step reasoning process.

[0017] 4. Intelligent decision generation and interaction module.

[0018] 1) Decision explanation generation system: use comparative explanation method to generate natural language explanation, show key data and reasoning path visualization of decision basis.

[0019] 2) Adaptive decision optimization: according to user feedback, including decision adoption rate, actual effect score, etc., real-time adjustment of model parameters, through transfer learning to adapt to crop production characteristics in different regions.

[0020] 5. Model evaluation and operation module.

[0021] 1) Establish a multi-dimensional evaluation system: including accuracy, recall rate, robustness, timeliness and other core indicators.

[0022] 2) Develop a model health monitoring system: use PSI indicators to detect distribution changes and monitor parameter abnormalities and performance degradation in real time.

[0023] 3) Build a security protection mechanism: prevent malicious queries through input filtering, protect user data using differential privacy technology, and set decision risk level indicators to provide early warnings for users.

[0024] The beneficial effects of the present application are: 1) very high pattern linking accuracy: through the causal reasoning module and the knowledge enhancement mechanism, the decision accuracy is improved compared with traditional machine learning models, realizing the transition from experience-based decision to data-driven decision.

[0025] 2) Data fusion technology breaks down data barriers, improves the utilization of scattered agricultural data, and provides data support for precision agriculture. 3) Natural language interaction and visual explanation function enable ordinary farmers to obtain decision support without professional knowledge. 4) In the pilot area, the use of water, fertilizer and pesticide is reduced, crop yield is increased, and the economic and ecological benefits of agricultural production are significantly improved. BRIEF DESCRIPTION OF DRAWINGS

[0026] Figure 1 is the framework diagram of the present application. DETAILED DESCRIPTION

[0027] In order to make the technical solutions, advantages and purposes of the present application clearer and more clear, the method flow is illustrated combined with specific examples, and the technical solutions of the present application are further illustrated with reference to the drawings.

[0028] Embodiment 1 of the present application is a large language model-based agricultural decision reasoning large model and a construction method thereof, which is performed according to the following steps:

[0029] 1. Data acquisition, specifically including:

[0030] Deploy a monitoring network in the cotton planting area of Changji Huaxing Farm in Xinjiang: set 1 group of soil moisture sensors on average every 300 mu of farmland, equipped with multi-spectral unmanned aerial vehicles, and access to small weather stations in the entire range of the farm to obtain hourly weather data.

[0031] 2. Data fusion, specifically including:

[0032] The original data is preprocessed by the 3σ principle to eliminate outliers, denoised by the smoothing filter method, and missing values are filled by the random forest model. Then, the soil data, weather data and crop data are mapped to 1024-dimensional agricultural feature vectors using the multi-head attention mechanism.

[0033] 3. Model fine-tuning, specifically including:

[0034] 1) Model structure optimization. For text data, semantic features are first extracted by the BERT model, and then a 256-dimensional feature vector is obtained through linear transformation. A multi-head latent attention mechanism is used to construct an adaptation layer, with the input being the agricultural data features output by the agricultural feature extraction module and the basic semantic features output by the large language model encoder. By calculating the attention weights between different modal features, dynamic fusion of features is achieved, generating a fused agricultural enhanced feature vector as input for the large language model decoder, allowing the model to understand both general semantics and agricultural professional features.

[0035] 2) Training in stages. Collect 3 million agricultural field text data, build a field fine-tuning corpus, and after de-duplication and filling, convert it to a text format supported by the large language model. Use LoRA technology to freeze most parameters of the LLama model, and only fine-tune 10% of the key parameters. The training goal is the language modeling task, i.e., predicting the next token, and optimizing the model parameters through the cross-entropy loss function, so that the model learns the professional terminology, grammar rules, and knowledge expression patterns in the agricultural field. Then, build a 500,000 expert-level agricultural production case dataset, each case containing "input data - decision problem - expert solution" in three parts. Next, unfreeze all parameters of the agricultural feature adaptation layer and extraction module, and optimize the model through the contrast loss function, so that the model can generate decision-making suggestions that conform to expert logic based on agricultural scene data.

[0036] 3) Reinforce inference mechanism. Select more than 10,000 representative agricultural decision-making problems, and score the decision-making suggestions generated by the model from more than 10 agricultural experts, and build a reinforcement learning feedback dataset with "question - suggestion - score" triplets. Use the GRPO algorithm for reinforcement learning training, with the large language model as the policy network, the input being the agricultural decision-making problem and the context data, and the output being the probability distribution of the decision-making suggestion. Calculate the reward value through expert scoring, and maximize the cumulative reward by adjusting the policy network parameters. During model inference, when a decision-making problem is received, the causal inference module first identifies key influencing factors through causal graphs, and then uses the Do-Calculus algorithm to calculate the causal effect of the intervention variable on the result variable, excluding confounding factor interference.

[0037] 4. Decision output, specifically including:

[0038] The model receives agricultural production problem input (such as crop abnormal state, planting scheme, etc.), calls the fused data and updated knowledge system, analyzes through the optimized inference mechanism, generates accurate prediction, intelligent diagnosis result and scientific decision-making suggestion, and outputs to the user end.

[0039] The contents not described in detail in the specification of the present application are the existing technology known to those skilled in the art.

Claims

1. A large language model-based agricultural decision reasoning large model and a construction method thereof, characterized in that, The application relates to an agricultural knowledge dynamic updating module, a large language model agriculturalization module and a decision reasoning output module. The agricultural data acquisition and fusion module specifically comprises: 1) a data acquisition unit, which is used for collecting soil data such as soil temperature and humidity, nitrogen, phosphorus and potassium and pH value through ground sensors, acquiring visual data such as crop growth multi-spectral images through unmanned aerial vehicle remote sensing equipment, collecting meteorological data such as temperature, precipitation, illumination and wind speed through meteorological sensing equipment and collecting agricultural machinery data such as agricultural machinery operation position and operation speed through Internet of Things equipment; 2) a data fusion unit, which is used for performing feature mapping on structured data and unstructured data through a multi-head attention mechanism of a Transformer architecture, performing cleaning, denoising and filling processing on the data through reinforcement learning technology and converting multi-source heterogeneous data to a unified feature space, so as to ensure data quality. The implementation mode of the agricultural knowledge dynamic updating module is as follows: based on the correlation mechanism of crop growth and environmental factors in a DSSAT crop simulation model, an initial reasoning rule base is constructed, the rule base comprises crop growth period management rules, pest control rules and water and fertilizer regulation rules; new agricultural knowledge is extracted from agricultural scientific research literature, agricultural technology popularization files and expert experience through a knowledge extraction technology, the rule contents in the reasoning rule base are updated in real time, and dynamic iteration of agricultural knowledge is realized.

2. The large language model-based agricultural decision reasoning large model and its construction method according to claim 1, characterized in that, The architecture optimization in the large language model agriculturalization module specifically comprises: 1) an agricultural feature extraction module, which is used for extracting semantic features of text data through a BERT model and obtaining a 256-dimensional feature vector through linear transformation; 2) an agricultural feature adaptation layer, which is used for inputting features output by the agricultural feature extraction module and basic semantic features output by a large language model encoder through a multi-head latent attention mechanism, realizing dynamic fusion through calculation of attention weights between the features and generating a 1024-dimensional agricultural enhanced feature vector as input of a large language model decoder.

3. The large language model-based agricultural decision reasoning large model and its construction method according to claim 1, characterized in that, The training optimization in the large language model agriculturalization module specifically comprises:

4. The large language model-based agricultural decision reasoning large model and its construction method according to claim 1, characterized in that, ​ ​ ​ 5. The LLama model-based agricultural decision reasoning large model and its construction method according to claim 1, characterized in that, ​ 1) Stage training strategy, the first stage uses 3 million agricultural field text data for field fine-tuning, uses LoRA technology to freeze most parameters of the large language model, and optimizes the model through the cross-entropy loss function for language modeling tasks; The second stage is based on 50 million expert-labeled agricultural production cases for targeted training, unfreezing the agricultural feature adaptation layer and extraction module parameters, and optimizing the model through the contrast loss function for question and answer generation tasks; 2) Reinforcement learning mechanism, build a feedback dataset containing 10,000 "question-suggestion-expert score" triplets, use the GRPO algorithm to use the large language model as the strategy network, calculate the reward value according to the expert score, maximize the cumulative reward by adjusting the network parameters, and optimize the decision generation mechanism.

6. The large language model-based agricultural decision reasoning large model and its construction method according to claim 1, characterized in that, The implementation of the causal inference module in the large language model agriculturalization module is: based on agricultural knowledge and expert experience, build a triple including "meteorological factors-soil state-crop growth", use Do-Calculus algorithm to calculate the causal effect of intervention variable on result variable, and use the causal inference result as attention weight to guide model to generate decision suggestion.

Images