Method and system for realizing intelligent decision-making of sewage plant through cooperation of large and small models

Abstract

The invention discloses a method and system for realizing intelligent decision-making of a sewage plant through cooperation of large and small models, and relates to the field of sewage plant operation management. Constructing a large model collaboration module by fine tuning an open source large model, collecting and preprocessing operation data to form real-time feature vectors, and storing the real-time feature vectors in a constructed transfer database; the small model prediction module predicts according to the real-time feature vector and transmits a result to the large model collaboration module; the large model collaboration module integrates the prediction result, outputs an analysis result, displays the analysis result to a user interface, writes the adoption content into a transfer database and calculates the adoption rate; the user can correct and feed back the result, and feedback information is stored in a transfer database; the model effect evaluation module realizes adaptive learning according to errors of output and correction results, and a progressive permission security mechanism is followed when a large model generation strategy is executed; according to the method and the system for cooperatively realizing the intelligent decision of the sewage plant by using the large and small models, the interpretability, the evolutionary property and the landing property of the intelligent decision are realized.

Description

Technical Field

[0001] This invention relates to the field of wastewater treatment plant operation and management, and in particular to a method and system for achieving intelligent decision-making in wastewater treatment plants through the collaboration of large and small models. Background Technology

[0002] Digital and intelligent transformation and upgrading are important trends in promoting high-quality development in the wastewater sector. However, their practical application in engineering projects still faces many challenges. The technical barriers to model application and adjustment are high, making it difficult for wastewater treatment plant operators to master them. Models lack two-way interactive capabilities; most models can only provide one-way predictive outputs, unable to understand or accept operator prompts and feedback, and are difficult to update and evolve. Model trust is low, making engineering implementation difficult; there is a significant gap between laboratory results and engineering applications. Model stability and reliability need to be verified and improved in engineering practice, but wastewater treatment plants cannot afford the verification risks. Furthermore, existing models have relatively singular functions, often focusing on specific tasks and lacking comprehensive decision support capabilities. Summary of the Invention

[0003] The purpose of this invention is to provide a method and system for intelligent decision-making in wastewater treatment plants through the collaborative use of small and large models. By simultaneously deploying specialized small models and generative AI large models in the intelligent decision-making system of wastewater treatment plants, and using a transit database and control plane to realize information flow and continuous optimization between models, the small model is responsible for high-precision numerical prediction, while the large model is responsible for knowledge reasoning, explanation, and optimization suggestions. The system uses a feedback mechanism to realize a closed loop of "prediction-explanation-feedback-relearning", thereby achieving interpretable, evolvable, and implementable intelligent decision-making.

[0004] To achieve the above objectives, this invention provides a method for intelligent decision-making in wastewater treatment plants through the collaborative implementation of large and small models, comprising the following steps: S1. Utilizing general knowledge of wastewater treatment and specific information about wastewater treatment plants, a large-scale model of the wastewater treatment plant is obtained through fine-tuning and training based on an open-source large model; real-time operational data of the wastewater treatment plant is collected and preprocessed using feature engineering to form real-time feature vectors. And build a transit database about wastewater treatment plants; S2, the real-time feature vector of S1 The data is input into the small model prediction module for prediction, and the prediction results are transmitted to the large model collaboration module for building the large model of the S1 wastewater treatment plant. S3, the large model collaboration module integrates the prediction results of S2, outputs analysis results and displays them on the user interface for user adoption, writes the user's adopted content into the transit database and calculates the adoption rate; S4. Users can correct and provide feedback on the analysis results output by S3 according to their needs and write the corrections into the transit database. S5. Based on the model output results of S3 and the corrected output results of S4, the model effect evaluation module detects errors and performs adaptive learning optimization and verification based on the errors. The results of the S6 and S4 modifications are executed in accordance with the security mechanism of progressive permissions.

[0005] Preferably, the real-time feature vector in S1 The collected operational data includes influent water quality parameters, effluent water quality parameters, process control variables, equipment operating status, and alarm information; The S1 transit database stores information including process principles, standards and specifications, equipment descriptions, common abnormal operating conditions and handling strategies, historical operating data of the target plant area, equipment records, emergency events and experience knowledge.

[0006] Preferably, the small model prediction module in S2 selects a learning model that includes Multilayer Perceptron (MLP), Long Short-Term Memory (LSTM) network, Random Forest, and several integrated models.

[0007] Preferably, the specific process of the large model collaboration module in S3 is as follows: S31. The large model collaboration module reads and integrates the prediction results of the current small model prediction module, the recent period's operating conditions data, and relevant data from historical trends and transit databases. S32. After comprehensive analysis by the large model collaborative module, three types of outputs are generated, including explanatory text, operation suggestions, parameter corrections or residual estimates, and displayed on the user interface. S33. Users select and adopt content based on the analysis presented in the user interface in S32. The user's adoption status is recorded in the transit database to calculate the adoption rate.

[0008] Preferably, the calculation process for the adoption rate in S33 is as follows: The large model collaboration module generates several decision options based on the prediction results of the small model prediction module, and assigns different scores to each decision option according to the recommendation order. Then, it records each choice made by the user. When the number of questions and answers between the user and the large model collaboration module reaches a set number, the score of each option is calculated by multiplying the score percentage of each selected option by the cumulative number of times that option is selected. Finally, the total score of several decision options is divided by the total number of questions and answers to obtain the adoption rate.

[0009] Preferably, the specific process of S4 is as follows: S41. The analysis results presented by the large model collaboration module can be modified or partially modified by the user according to actual needs. The corrections made in S42 and S41 are written into the transfer database in the form of feedback samples for subsequent learning and optimization.

[0010] Preferably, the specific process of the S5 model performance evaluation module is as follows: S51. Periodically read the deviation between the output results of the large model collaboration module in the transit database and the output results after user correction, and automatically identify and mark high error samples; S52. Send the error samples to the large model collaboration module to generate the error causes and specific optimization directions; S53. Based on the optimization direction of the large model collaborative module, provide an overall corrected bias and perform automatic optimization; S54. Verify the optimization effect of S53. If the error decreases, update the version of the small model prediction module. If the error does not decrease, retrain the small model.

[0011] Preferably, the specific process of S6 is as follows: S61. The large model collaboration module only outputs results and does not automatically execute operations; all decisions require manual confirmation. S62. For low-risk tasks, a suggestion-confirmation-execution mode is adopted. When the user adoption rate rises to the set custom threshold, it starts to enter the semi-automatic execution mode. The semi-automatic execution mode guides the operation of some devices, but will ensure that the actual execution strategy is implemented within the safe range according to the set safety value. S63. Once the large model collaboration module is stable and authorized, it allows automatic execution of regular high-frequency tasks and manual intervention for abnormal operating conditions.

[0012] To achieve the above objectives, the present invention also provides a system for intelligent decision-making in wastewater treatment plants through the collaboration of large and small models, including a small model prediction module, a large model collaboration module, a transfer database, a workflow control module, and a model performance evaluation module. The core of the small model prediction module is a data-driven model, which is trained using influent, effluent and process data from the wastewater treatment plant. It is responsible for performing high-precision numerical prediction of key wastewater treatment indicators, receiving instructions from the workflow control module, reading specific data from the transfer database, training model parameters, and outputting the calculation results to specific columns in the transfer database. The large model collaboration module is deployed locally based on the open-source large model collaboration module. It performs secondary training using general wastewater treatment data and wastewater plant-specific data, calls the output results of the small model prediction module and parses and generates various forms of output results. It is responsible for interacting with users in natural language, providing indicator prediction and operating parameter calculation based on the content written to the transit database, parsing user commands, and updating the database based on the parsing results. The transit database is a database designed according to the needs of wastewater treatment management. It has standardized data types and names and other attributes that are easy for both large and small models to read and write. It is mainly used to store historical data of wastewater treatment plants, content generated by the large model collaborative module, user feedback results, actual monitoring data, and output results of the small model prediction module, providing a transit platform for interaction between large and small models. The workflow control module is a core control program containing a specific workflow. It is responsible for executing the workflow, controlling the large model collaboration module to write the parsing results into the transit database, calling the small model prediction module to calculate and store the results, and controlling the large model collaboration module to read specific data from the transit database, thereby realizing asynchronous interaction and continuous learning between the large and small models. The model performance evaluation module is responsible for statistics and correction. Based on the error and user feedback, it dynamically calculates the model performance and provides an overall correction bias to perform automatic optimization according to the optimization direction of the large model collaboration module.

[0013] Therefore, the present invention provides a method and system for intelligent decision-making in wastewater treatment plants that utilizes the above-mentioned large and small model collaboration, which has the following advantages compared with the prior art: This application features a clear division of labor: the small model prediction module manages accuracy, the large model collaboration module manages solutions and optimizations, data flow is decoupled, a transit database enables asynchronous communication and log management between models, feedback-driven optimization, a reinforcement learning mechanism based on user ratings and error calculation, a secure and gradual implementation, a hierarchical permission strategy from manual assistance to semi-autonomous operation, a platform-based interface, and the large model collaboration module serving as a unified interaction entry point to schedule underlying models and databases, thus constructing an intelligent management hub.

[0014] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0015] Figure 1 This is an overall flowchart of a method for intelligent decision-making in wastewater treatment plants through the collaborative use of large and small models, as described in this invention. Figure 2 This is an overall structural diagram of a system for intelligent decision-making in wastewater treatment plants, based on the collaborative implementation of large and small models according to the present invention. Detailed Implementation

[0016] In the description of this invention, it should be noted that the terms "upper," "lower," "inner," "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship in which the product of this invention is usually placed when in use. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limiting this invention.

[0017] Example like Figure 1 As shown, the present invention provides a method for intelligent decision-making in wastewater treatment plants through the collaborative implementation of large and small models, comprising the following steps: S1. Utilizing general knowledge of wastewater treatment and specific information about wastewater treatment plants, a large-scale model of the wastewater treatment plant is obtained through fine-tuning and training based on an open-source large model; real-time operational data of the wastewater treatment plant is collected and preprocessed using feature engineering to form real-time feature vectors. And build a transit database about wastewater treatment plants; Among them, real-time feature vectors The collected operational data includes influent water quality parameters, effluent water quality parameters, process control variables, equipment operating status, and alarm information; The transit database stores information including process principles, standards and specifications, equipment descriptions, common abnormal operating conditions and handling strategies, historical operating data of the target plant, equipment records, emergency events and experience; as shown in Table 1: Table 1. Transit database records various interaction and operational information.

[0018] S2, the real-time feature vector of S1 The input is fed into the small model prediction module for prediction, and the prediction results are transmitted to the large model collaboration module for building the large model of the S1 wastewater treatment plant. The small model prediction module selects models including Multilayer Perceptron (MLP), Long Short-Term Memory Network (LSTM), Random Forest, and a learning model ensembled from several models. Real-time feature vectors The specific calculation process input into the small model prediction module is as follows: ,in This is the initial prediction result. This is the running function of the small model prediction module. The small model prediction module predicts the specified data and stores the prediction results in the intermediate database. When the large model collaboration module and the user interact through the interaction module, the large model collaboration module calls the small model prediction module in the intermediate database to analyze and answer the questions. Taking an ensemble learning model as an example to obtain the predicted output, compared with a single model, it effectively reduces the overall variance of the model statistically by introducing model diversity, thereby smoothing out the instability caused by the overfitting of a single model to specific data noise and significantly improving the generalization ability. Computationally, by integrating models that converge to different local optima from different starting points, it essentially constructs a solution that is closer to the global optimum. In terms of representation ability, the ensemble model combines multiple hypothesis spaces to form a more expressive and complex composite hypothesis space, enabling it to capture deeper and more complex patterns in the data. Finally, this structure endows the ensemble model with inherent robustness far exceeding that of a single model. Even if individual component models have prediction biases, the overall accuracy and stability of the output can be maintained through the collective decision-making mechanism.

[0019] S3, the large model collaboration module integrates the prediction results of S2, outputs analysis results and displays them on the user interface for user adoption, writes the user's adopted content into the transit database and calculates the adoption rate; The selection of large models includes generative AI models such as Gemini, GPT-4o, Qwen, and GLM. The specific process of the large model collaboration module is as follows: S31. The large model collaboration module reads and integrates the prediction results of the current small model prediction module, the recent period's operating conditions data, and relevant data from historical trends and transit databases. S32. After comprehensive analysis by the large model collaborative module, three types of outputs are generated, including explanatory text, operation suggestions, parameter corrections or residual estimates, and displayed on the user interface. S33. Users select and adopt based on the analysis content presented in the user interface in S32. The user's adoption status is recorded in the transit database to calculate the adoption rate. Explanatory text: Analysis of the reasons for the prediction results (e.g., fluctuations in influent load leading to higher COD); Operational recommendations: Implementable process optimization measures (e.g., increase aeration rate by 5% to increase organic matter removal efficiency). Parameter correction or residual estimation: Numerical compensation terms used to correct the results of the small model prediction module. The compensation formula for the prediction results of the small model prediction module is as follows: ; in, The result is the compensated prediction. The adoption rate is calculated as follows: The large model collaboration module generates several decision options based on the prediction results of the small model prediction module, and assigns different scores to each decision option according to the recommendation order. Then, it records each choice made by the user. When the number of questions and answers between the user and the large model collaboration module reaches a set number, the score of each option is calculated by multiplying the score percentage of each option chosen by the cumulative number of times that option is chosen. Then, the total score of several decision options is divided by the total number of questions and answers to obtain the adoption rate. For example, the large model generates five decision options. Since each selected option has a different score percentage [1, 0.75, 5, 0.25, 0], and the user's choice is recorded each time, when the number of questions and answers reaches a certain number, the score for each option is calculated by multiplying the score percentage of each selected option by the cumulative number of times that option was chosen. Then, the total score of the five options is divided by the total number of questions and answers to obtain the adoption rate. The results of the adoption rate calculation are shown in Table 2. Table 2 Adoption Rate Calculation Table

[0020] S4. Users can correct and provide feedback on the analysis results output by S3 according to their needs and write the corrections into the transit database. S41. The analysis results presented by the large model collaboration module can be modified or partially modified by the user according to actual needs. The corrections made in S42 and S41 are written into the transfer database in the form of feedback samples for subsequent learning and optimization. S5. Based on the model output results of S3 and the corrected output results of S4, the model effect evaluation module detects errors and performs adaptive learning optimization and verification based on the errors. The specific process of the model performance evaluation module is as follows: S51. Periodically read the deviation between the output results of the large model collaboration module in the transit database and the output results after user correction, and automatically identify and mark high error samples; S52. Send the error samples to the large model collaboration module to generate the error causes and specific optimization directions; S53. Based on the optimization direction of the large model collaborative module, provide an overall corrected bias and perform automatic optimization; S54. Verify the optimization effect of S53. If the error decreases, update the version of the small model prediction module. If the error does not decrease, retrain the small model. The results of the S6 and S4 modifications are executed in accordance with the security mechanism of progressive permissions.

[0021] S61. The large model collaboration module only outputs results and does not automatically execute operations; all decisions require manual confirmation. S62. For low-risk tasks, a suggestion-confirmation-execution mode is adopted. When the user adoption rate rises to the set custom threshold, it starts to enter the semi-automatic execution mode. The semi-automatic execution mode guides the operation of some devices, but will ensure that the actual execution strategy is implemented within the safe range according to the set safety value. S63. Once the large model collaboration module is stable and authorized, it allows automatic execution of regular high-frequency tasks and manual intervention for abnormal operating conditions.

[0022] The overall input and output flow processing table is shown in Table 3: Table 3 Input / output flow processing table of the overall method

[0023] like Figure 2 As shown, the present invention provides a system for intelligent decision-making in wastewater treatment plants through the collaboration of large and small models, comprising a small model prediction module, a large model collaboration module, a transfer database, a workflow control module, and a model performance evaluation module; The core of the small model prediction module is a data-driven model (such as a deep neural network model or a hybrid model), trained using influent, effluent, and process data from the wastewater treatment plant. This model is responsible for executing key wastewater treatment indicators (including effluent chemical oxygen demand (COD) and ammonia nitrogen (NH4)). + High-precision numerical prediction of N, total nitrogen TN, total phosphorus TP, dissolved oxygen DO, etc., accepts instructions from the workflow control module, reads specific data from the transfer database, trains model parameters, and outputs the calculation results to a specific column in the transfer database; The large model collaboration module is deployed locally based on the open-source large model collaboration module. It performs secondary training using general wastewater treatment data and wastewater plant-specific data, calls the output results of the small model prediction module and parses and generates various forms of output results. It is responsible for interacting with users in natural language, providing indicator prediction and operating parameter calculation based on the content written to the database, parsing user commands (including feedback content), and updating the database based on the parsing results. The transit database is a database designed according to the needs of wastewater treatment management. It has standardized data types and names and other attributes that are easy for both large and small models to read and write. It is mainly used to store historical data of wastewater treatment plants, content generated by the large model collaborative module, user feedback results, actual monitoring data, and output results of the small model prediction module, providing a transit platform for interaction between large and small models. The workflow control module is a core control program containing a specific workflow. It is responsible for executing the workflow, controlling the large model collaboration module to write the parsing results into the transit database, calling the small model prediction module to calculate and store the results, and controlling the large model collaboration module to read specific data from the transit database, thereby realizing asynchronous interaction and continuous learning between the large and small models. The model performance evaluation module is responsible for statistics and correction. Based on the error and user feedback, it dynamically calculates the model performance, including the adoption rate of the model's generated results, the magnitude of the error, etc., and provides an overall correction bias based on the optimization direction of the large model collaboration module to perform automatic optimization.

[0024] Therefore, this invention employs a method and system for intelligent decision-making in wastewater treatment plants that utilizes both small and large models in synergy. By simultaneously deploying specialized small models and generative AI large models within the intelligent decision-making system for wastewater treatment plants, and leveraging a transit database and control plane to achieve information flow and continuous optimization between models, the small model is responsible for high-precision numerical prediction, while the large model is responsible for knowledge reasoning, explanation, and optimization suggestions. The system utilizes a feedback mechanism to achieve a closed loop of "prediction-explanation-feedback-relearning," thereby realizing interpretable, evolvable, and implementable intelligent decision-making.

[0025] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for intelligent decision-making in wastewater treatment plants through collaborative use of large and small models, characterized in that: Includes the following steps: S1. Utilizing general knowledge of wastewater treatment and specific information about wastewater treatment plants, a large-scale model of the wastewater treatment plant is obtained through fine-tuning and training based on an open-source large model. Real-time operational data of the wastewater treatment plant is collected and preprocessed using feature engineering to form real-time feature vectors. And build a transit database about wastewater treatment plants; S2, the real-time feature vector of S1 The data is input into the small model prediction module for prediction, and the prediction results are transmitted to the large model collaboration module for building the large model of the S1 wastewater treatment plant. S3, the large model collaboration module integrates the prediction results of S2, outputs analysis results and displays them on the user interface for user adoption, writes the content adopted by the user into the transit database and calculates the adoption rate; S4. Users can correct and provide feedback on the analysis results output by S3 according to their needs, and write the corrections into the transit database. S5. Based on the model output results of S3 and the corrected output results of S4, the error is detected through the model performance evaluation module, and adaptive learning optimization and verification are performed through the error. The results of the S6 and S4 modifications are executed in accordance with the security mechanism of progressive permissions.

2. The method for intelligent decision-making in wastewater treatment plants through collaborative large and small model implementation according to claim 1, characterized in that: Real-time feature vector in S1 The collected operational data includes influent water quality parameters, effluent water quality parameters, process control variables, equipment operating status, and alarm information; The S1 transit database stores information including process principles, standards and specifications, equipment descriptions, common abnormal operating conditions and handling strategies, historical operating data of the target plant area, equipment records, emergency events and experience knowledge.

3. The method for intelligent decision-making in wastewater treatment plants through collaborative large and small model implementation according to claim 2, characterized in that: The small model prediction module in S2 selects models including Multilayer Perceptron (MLP), Long Short-Term Memory (LSTM) network, Random Forest, and several integrated learning models.

4. The method for intelligent decision-making in wastewater treatment plants through collaborative large and small model implementation according to claim 3, characterized in that: The specific process of the large model collaboration module in S3 is as follows: S31. The large model collaboration module reads and integrates the prediction results of the current small model prediction module, the recent period's operating conditions data, and relevant data from historical trends and transit databases. S32. After comprehensive analysis by the large model collaborative module, three types of outputs are generated, including explanatory text, operation suggestions, parameter corrections or residual estimates, and displayed on the user interface. S33. Users select and adopt the analysis content presented on the user interface in S32. The large model collaboration module records the user's adoption status to the transit database and calculates the adoption rate.

5. The method for intelligent decision-making in wastewater treatment plants through collaborative large and small model implementation according to claim 4, characterized in that: The calculation process for the adoption rate in S33 is as follows: The large model collaboration module generates several decision options based on the prediction results of the small model prediction module, and assigns different scores to each decision option according to the recommendation order. Then, it records each choice made by the user. When the number of questions and answers between the user and the large model collaboration module reaches a set number, the score of each option is calculated by multiplying the score percentage of each selected option by the cumulative number of times that option is selected. Finally, the total score of several decision options is divided by the total number of questions and answers to obtain the adoption rate.

6. The method for intelligent decision-making in wastewater treatment plants through collaborative large and small model implementation according to claim 5, characterized in that: The specific process of S4 is as follows: S41. The analysis results presented by the large model collaboration module can be modified or partially modified by the user according to actual needs. The corrections made in S42 and S41 are written into the intermediate database in the form of feedback samples for subsequent learning and optimization.

7. The method for intelligent decision-making in wastewater treatment plants through collaborative large and small model implementation according to claim 6, characterized in that: The specific process of the S5 model performance evaluation module is as follows: S51. Periodically read the deviation between the output results of the large model collaboration module in the transit database and the output results after user correction, and automatically identify and mark high error samples; S52. Send the error samples to the large model collaboration module to generate the error causes and specific optimization directions; S53. Based on the optimization direction of the large model collaborative module, provide an overall corrected bias and perform automatic optimization; S54. Verify the optimization effect of S53. If the error decreases, update the version of the small model prediction module. If the error does not decrease, retrain the small model.

8. The method for intelligent decision-making in wastewater treatment plants through collaborative large and small model implementation according to claim 7, characterized in that: The specific process of S6 is as follows: S61. The large model collaboration module only outputs results and does not automatically execute operations; all decisions require manual confirmation. S62. For low-risk tasks, adopt the suggestion-confirmation-execution mode. When the user adoption rate rises to the set custom threshold, start to enter the semi-automatic execution mode. S63. Once the large model collaboration module is stable and authorized, it allows automatic execution of regular high-frequency tasks and manual intervention for abnormal operating conditions.

9. A system for intelligent decision-making in wastewater treatment plants through collaborative use of large and small models, characterized in that: The method for intelligent decision-making in wastewater treatment plants using a combination of large and small models as described in any one of claims 1-8, the system comprising a small model prediction module, a large model collaboration module, a transfer database, a workflow control module, and a model performance evaluation module; The core of the small model prediction module is a data-driven model, which is trained using influent, effluent and process data from the wastewater treatment plant. It is responsible for performing high-precision numerical prediction of key wastewater treatment indicators, receiving instructions from the workflow control module, reading specific data from the transfer database, training model parameters, and outputting the calculation results to specific columns in the transfer database. The large model collaboration module is deployed locally based on the open-source large model collaboration module. It performs secondary training using general wastewater treatment data and wastewater plant-specific data, calls the output results of the small model prediction module and parses and generates various forms of output results. It is responsible for interacting with users in natural language, providing indicator prediction and operating parameter calculation based on the content written to the transit database, parsing user commands, and updating the database based on the parsing results. The transit database is a database designed according to the needs of wastewater treatment management. It has standardized data types and names that are easy for both large and small models to read and write. It is mainly used to store historical data of wastewater treatment plants, content generated by the large model collaborative module, user feedback results, actual monitoring data, and output results of the small model prediction module, providing a transit platform for interaction between large and small models. The workflow control module is a core control program containing a specific workflow. It is responsible for executing the workflow, controlling the large model collaboration module to write the parsing results into the transit database, calling the small model prediction module to calculate and store the results, and controlling the large model collaboration module to read specific data from the transit database, thereby realizing asynchronous interaction and continuous learning between the large and small models. The model performance evaluation module is responsible for statistics and correction. Based on the error and user feedback, it dynamically calculates the model performance and provides an overall correction bias to perform automatic optimization according to the optimization direction of the large model collaboration module.

Images