Method and system for automatically repairing control system based on large language model

Abstract

The invention belongs to the field of automatic control, and relates to a control system automatic repair method and system based on a large language model. The method comprises the following steps: initializing and pre-training a large language model to obtain an adjusted large language model; an interactive target conflict resolution framework is established, and three roles of an analyst, a repairer and an inspector are set; and iteratively generating and evaluating a conflict solution based on the adjusted large-scale language model according to the target, the domain attributes and the boundary condition data, generating a corrected target by a repairer by using an analysis result obtained from an analyst, checking whether the corrected target meets a conflict solution requirement or not by a checker, and finally obtaining the repaired target. Obtaining a solution set; needed information is obtained from the solution set, and repair of the control system is achieved. According to the method, conflicts are analyzed in an iterative interaction mode, the solution is generated, target conflicts can be rapidly and accurately analyzed and solved, and the quality and efficiency of the solution are remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the field of automation control, and particularly relates to a method and system for automatically repairing a control system based on a large language model. Background Art

[0002] The field of automation control is an important part of industrial production, and its task is to control different software and hardware in the production process to obtain a stable and safe production environment. In reality, such as the water pump control system inside a mine, the production control system of an assembly line, the control of traffic lights, etc., there are dedicated software and hardware for control. In these control software, a series of goals (purposes to be achieved), domain attributes (working effects of different hardware), and boundary conditions (situations where goals and domain attributes conflict) can usually be abstracted. Among them, goals, domain attributes, and boundary conditions can be represented by linear temporal logic formulas because linear temporal logic formulas can formally abstract the specifications of a large class of goals, domain attributes, and boundary conditions.

[0003] To solve the possible conflicts between goals and domain attributes, requirements engineers need to formally describe complex requirement conflicts through linear temporal logic to better understand and solve these conflicts. Since requirements come from different aspects, requirements engineers need to analyze and understand the specific situations and backgrounds (i.e., boundary conditions) of these requirements to clarify the interaction between requirements and the nature of conflicts, so as to achieve the stability and security of the system by modifying the goals. This process requires a high degree of professional knowledge and systematic analysis methods to ensure that various control systems can effectively meet various complex and changing business requirements.

[0004] In this process, traditional goal conflict resolution methods usually rely on search algorithms to explore different goal configurations and solutions, rather than deeply analyzing requirement specifications and boundary conditions. Existing traditional methods may ignore the specific information of the requirements themselves, similar to understanding a sentence only by looking up keywords without deeply understanding the meaning of the text. In short, traditional conflict resolution methods require the participation of a large number of requirements engineers and ignore the information provided by the analysis of requirement specifications and boundary conditions, resulting in the inability to fully utilize the detailed information in requirement analysis to solve conflicts. Therefore, these traditional methods are not efficient and accurate enough in dealing with complex requirement conflicts. Summary of the Invention

[0005] To solve the problems existing in the prior art, the present invention proposes a method and system for automatically repairing a control system based on a large language model, which adopts an automated interactive goal conflict resolution framework based on a large language model, analyzes conflicts and generates solutions through an iterative interaction method, can quickly and accurately analyze and solve goal conflicts, and significantly improves the quality and efficiency of the solutions.

[0006] On the one hand, an embodiment of the present invention provides a method for automatically repairing a control system based on a large language model, including the following steps: S1. Initialize and pre-train a large language model to obtain an adjusted large language model; S2. Establish an interactive target conflict resolution framework with three roles: an analyst, a fixer, and an inspector; based on the target, domain attributes, and boundary condition data, iteratively generate and evaluate conflict resolution solutions based on the adjusted large language model. The fixer uses the analysis results obtained from the analyst to generate a revised target, and the inspector is responsible for verifying whether the revised target meets the requirements of conflict resolution, and finally obtains a repaired target to obtain a solution set; S3. Obtain the required information from the solution set to implement the repair of the control system; The boundary conditions are used in requirements engineering to capture specific situations that lead to target conflicts.

[0007] On the other hand, an embodiment of the present invention also provides a control system automatic repair system based on a large language model for implementing the above automatic repair method, including the following modules: A model adjustment module for initializing and pre-training a large language model to obtain an adjusted large language model; A solution set generation module that establishes an interactive target conflict resolution framework with three roles: an analyst, a fixer, and an inspector; based on the target, domain attributes, and boundary condition data, iteratively generate and evaluate conflict resolution solutions based on the adjusted large language model. The fixer uses the analysis results obtained from the analyst to generate a revised target, and the inspector is responsible for verifying whether the revised target meets the requirements of conflict resolution, and finally obtains a repaired target to obtain a solution set; A repair module for obtaining the required information from the solution set to implement the repair of the control system; The boundary conditions are used in requirements engineering to capture specific situations that lead to target conflicts.

[0008] Compared with the prior art, the beneficial effects obtained by the present invention include: 1. Improve the efficiency and quality of solutions: By utilizing the extensive knowledge base and reasoning ability of the large language model, the present invention can quickly and accurately analyze and resolve target conflicts, significantly improving the quality and efficiency of solutions. Compared with traditional methods that rely on manual or basic algorithms, it can handle more complex problems more effectively.

[0009] 2. Reduce human resource requirements and costs: The present invention reduces the dependence on professional requirements engineers through automated and iterative methods, which not only alleviates the burden on the engineering team but also significantly reduces the labor costs required to resolve target conflicts.

[0010] 3. Improve the adaptability and reliability of the solution: The interactive and iterative method allows the present invention to be adjusted according to the actual situation during the resolution process, which improves the adaptability and reliability of the solution. In the process of continuous feedback and optimization, specific requirements and conditions can be more accurately met.

[0011] 4. Simple operation and easy to implement: The role-based operation and chain-of-thought prompt engineering make the operation process of the present invention more intuitive and simple, and even non-professionals can start and control the solution generation process through simple interactions.

[0012] 5. Improve the innovation and customization of the solution: The present invention supports the generation of highly customized solutions and can generate innovative and effective solutions for specific conflict scenarios, which is of great value for dealing with specific or unconventional target conflicts. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 is the overall process of the control system automation repair method based on the large language model in the embodiment of the present invention; Figure 2 is the flowchart of obtaining the solution by analyzing information through role interaction in the embodiment of the present invention; Figure 3 is the flowchart of updating the solution in the embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0014] To make the invention purpose, technical solutions and advantages of the present invention clearer, the present invention will be further described below in conjunction with the drawings and embodiments, but the embodiments of the present invention are not limited thereto.

[0015] In addition, it should be noted that only parts related to the present application are shown in the drawings for the sake of convenience of description. It is understandable that some well-known structures and their descriptions in the drawings may be omitted for those skilled in the art.

[0016] The present invention provides a control system automation repair method and system based on a large language model, which iteratively analyzes, generates and validates solutions through role-based (analyzer, repairer and checker) operations and chain-of-thought prompt engineering. To better describe the technical solutions of the present invention, the following basic concepts are first described: 1. Goal-Conflict Analysis: Goal-conflict analysis is usually driven by an identification-evaluation-control loop, aiming to identify, evaluate, and resolve inconsistencies that may impede the achievement of expected goals. In this work, the present invention focuses on the resolution phase. Based on competing hypotheses techniques and argumentation pattern analysis, and using genetic algorithms, existing methods cannot be directly used to resolve the divergences captured by boundary conditions. Modern algorithms based on search explore syntactic variants of goals and implement candidate solutions through multi-objective adaptation functions; however, this algorithm ignores the information provided by the requirements specification and also ignores the analysis of the reasons for goal conflicts caused by given boundary conditions.

[0017] 2. Linear Temporal Logic: This is a logical system used to describe the behavior and relationships of events in a time sequence. In software engineering, linear temporal logic is used to formalize requirement specifications, especially those involving time order or constraints. By using linear temporal logic, developers can precisely describe the conditions that software should meet at different time points or time periods, as well as the dependencies and interactions between various functions. In addition to basic logical rules (AND, OR, NOT), a series of time-related logical rules are added, such as: eventually, always, until, next, etc., representing the change relationships of different atomic propositions on the time axis.

[0018] 3. Boundary Conditions: In requirements engineering, boundary conditions are used to capture specific situations that lead to goal conflicts, manifested as situations where the satisfaction of some goals inhibits the satisfaction of other goals.

[0019] 4. Large Language Model: A pre-trained language model based on transformers is one of the major advancements in the field of artificial intelligence. A large language model is a large neural network consisting of up to 176 billion parameters, such as the pre-trained generative transformer series. Large language models perform well in verification, reasoning, and the automatic formalization of mathematics and formal specifications.

[0020] 5. Interactive Goal Conflict Resolution Framework: This is a collaborative framework proposed by the present invention, with three roles: analyzer, fixer, and checker, using a large language model to iteratively generate and evaluate conflict solutions. The fixer uses the analysis results obtained from the analyzer to generate revised goals, and the checker is responsible for verifying whether these revised goals meet the requirements of conflict resolution.

[0021] The present invention will be described in detail below with reference to the accompanying drawings and in conjunction with embodiments. Embodiment

[0022] See Figure 1 , a control system automation repair method based on a large language model provided by this embodiment includes the following steps: S1. Initialize and pre-train a large language model to obtain an adjusted large language model.

[0023] In an automatic control system and / or software, the objectives, domain attributes, and boundary conditions corresponding to the control system are manually or automatically extracted as data. The objectives can be safety indicators for production or quantity and quality requirements, the domain attributes can be the utilities of different hardware, and the boundary conditions are the states where the objectives and domain attributes conflict. Using the above data as a training set, an existing large language model is pre-trained to obtain the large language model required by the present invention.

[0024] The specific process of pre-training the large language model includes the following steps: S11. Data collection and processing: Obtain relevant text data from the automatic control system and / or automatic control software. The text data includes operation logs, fault reports, operation manuals, etc. of the automatic control system.

[0025] S12. Preprocessing: Perform operations such as cleaning, word segmentation, and removing irrelevant text on these text data for subsequent training and fine-tuning.

[0026] S13. Model selection: Select a text-based large language model, such as a bidirectional encoder based on transformers, as the base model.

[0027] S14. Pre-training: On the selected base model, use the text data of the control system collected for pre-training. The training process uses methods such as masked language model and next sentence prediction, that is, predicting the tokens in the mask and the content of the next sentence.

[0028] S15. Adjustment: After the pre-training is completed, use the data of the manually annotated target conflict resolution task to adjust the large language model. That is, given the input objectives, domain attributes, and boundary conditions, manually answer the repaired objectives, input each set of corresponding data into the large language model, freeze the lower-layer parameters of the large language model, and only adjust the upper-layer parameters of the large language model, so as to improve the performance of the model for subsequent target conflict resolution tasks.

[0029] S16. Evaluation and optimization: After the pre-training and fine-tuning are completed, evaluate the large language model so that the accuracy of the large language model in the target conflict resolution task is significantly improved compared with the existing model and the manual resolution framework.

[0030] S2. Establish an interactive target conflict resolution framework with three roles: analyzer, fixer, and checker; based on the data of objectives, domain attributes, and boundary conditions, iteratively generate and evaluate conflict resolution solutions based on the adjusted large language model. The fixer uses the analysis results obtained from the analyzer to generate a corrected objective, and the checker is responsible for verifying whether the corrected objective meets the requirements of conflict resolution, and finally obtains the repaired objective and obtains a solution set.

[0031] As shown in Figure 2 , 3 , the specific steps are as follows: S21. Initialize the large language models of the analyzer and the fixer with role instructions respectively, and start the background service of the checker.

[0032] In this embodiment, the role instructions for step S21 are as follows: The role instruction for the analyzer is: Your task is to help me analyze the reason why a certain formula does not hold under certain conditions.

[0033] The role instruction for the fixer is: Your task is to generate a repair target that satisfies the boundary conditions in the domain.

[0034] The domain, the target, and the boundary conditions are all written as linear temporal logic formulas. Both the domain and the repaired target can be satisfied. The repaired target should be syntactically and semantically similar to the target. The repaired target can only be modified on the given target. The returned format is the repaired target.

[0035] S22. Input the linear temporal logic formulas serving as the target, the domain attribute, and the boundary condition into the analyzer respectively.

[0036] The domain attribute in step S22 is a descriptive statement of the problem domain, used to capture the scope of requirements; the target is a prescriptive statement, which is the goal that the system must achieve; the boundary condition describes the combination under which the target cannot be fully satisfied in a specific situation, used to capture the occurrence conditions of target divergence. The above three inputs can all be expressed as legal linear temporal logic formulas and can be verified by a logic checker.

[0037] In this embodiment, the control system is the water pump control system inside the mine, the target can be "no flood, no explosion", the domain attribute is "water pump effect", and the boundary condition can be that "high water level and methane exist simultaneously" is true.

[0038] S23. Obtain analysis information through the analyzer, and add the analysis information, the input target, the domain attribute, and the boundary condition to all information. All information will be input into the trained large language model.

[0039] S24. Iteratively interact, allowing the analyzer, the fixer, and the checker to continuously interact until the condition for the end of the interaction is met; when the condition for the end of the interaction is met, the loop of continuous interaction ends, and finally a solution set is extracted from all the information in step S23.

[0040] During the role interaction process, the analyst provides analysis information, the fixer gives the repaired target based on the analyst's information, and the checker checks whether the fixer's answer meets the conditions until all conditions are met to gradually generate a solution.

[0041] In this embodiment, the condition for the interaction to end is that the repaired target can pass the syntax check, semantic check, and satisfiability check. Among them, the syntax check is to check whether the repaired target is a legal linear temporal logic formula; the semantic check is to check whether the repaired target can contain the semantics of the original target; the satisfiability check is to check whether the repaired target meets the domain attributes and boundary conditions. Using regular expressions, locate the matching information in the text, such as "repaired target:", and extract the relevant content after the colon. In a similar way, extract the solution from all the information obtained in step S23 to form the final result.

[0042] See Figure 3 , this step further includes the following steps: S241. Initialize the solution set as an empty set.

[0043] S242. If the solution set is not empty, input the check information obtained by the logic checker to the fixer. The fixer gets the solution information, then adds the solution information to all the information, and at the same time extracts the solution and adds it to the solution set.

[0044] S243. If the solution set is empty, input all the information to the fixer. The fixer gets the solution information, extracts the solution, and adds it to the solution set.

[0045] Since the solution set is not empty at this time, the check information obtained by the logic checker is: there is no methane at the next time of the high water level, and the pump is on and the low water level and the pump is off at the next time of the methane presence, which cannot meet the boundary condition "the high water level and methane exist simultaneously", and a repaired target needs to be given again.

[0046] Input the above check information to the fixer, and the fixer gets the solution information: the repaired target is "when the water level is high and there is no methane, turn on the pump at the next time; when the water level is low and there is methane, turn off the pump at the next time".

[0047] Then add the solution information to all the information, and at the same time extract the solution and add it to the solution set.

[0048] S244. Repeat steps S242 and S243 until the end condition is met, and return the final solution set R.

[0049] The logic checker includes syntax checking, semantic checking, and satisfiability checking, and is implemented based on the code library. After being verified by the logic checker, the solution can pass the syntax checking, semantic checking, and satisfiability checking, so the loop ends.

[0050] Finally, the returned set of final solutions can be interpreted as follows: 1. When the water level is high and there is no methane, turn on the water pump at the next time; 2. When the water level is low and there is methane, turn off the water pump at the next time.

[0051] S3. The engineer obtains the required information from the solution set and repairs the control system by fixing the corresponding content in the control system or software.

[0052] This embodiment also provides a control system automatic repair system based on a large language model for implementing the above automatic repair method, including the following modules: The model adjustment module is used to initialize and pre-train a large language model to obtain an adjusted large language model; The solution set generation module establishes an interactive target conflict resolution framework with three roles: an analyst, a fixer, and an inspector; based on the target, domain attributes, and boundary condition data, iteratively generates and evaluates conflict solutions based on the adjusted large language model. The fixer uses the analysis results obtained from the analyst to generate a revised target, and the inspector is responsible for verifying whether the revised target meets the requirements of conflict resolution, and finally obtains a repaired target to obtain a solution set; The repair module obtains the required information from the solution set and realizes the repair of the control system; The boundary conditions are used in requirements engineering to capture specific situations that lead to target conflicts.

[0053] For the various modules of the above automatic repair system, the detailed implementation process can refer to the foregoing steps.

[0054] The present invention not only improves the overall process of target conflict resolution, but also improves the efficiency and quality of the overall solution through technological innovation. It shows its high efficiency and innovation in dealing with complex problems. Especially in comparison with manual solutions, the high consistency shown further verifies its practicality and reliability.

[0055] Table 1 shows the conflict resolution results of the present invention under five different requirement specifications, which come from control systems in different fields. The present invention can handle specific boundary conditions in each requirement specification and generate solutions. These solutions are verified by the inspection module to confirm that they can successfully resolve the conflicts in the boundary conditions.

[0056]

[0057] Table 1: Effectiveness of the present invention in generating solutions

[0058] Table 2: Comparison of experimental results of two methods in solving the problem of target conflict As can be seen from Table 2, compared with the solutions written manually, the present invention can generate more precise or more general solutions more effectively.

[0059] From the above, in the present invention, first, a large language model is used as the core technology to automatically process target conflicts with its rich knowledge base and powerful reasoning ability. Second, through an interactive method, the large language model is allowed to continuously interact as user roles (analyzer, fixer, checker) to gradually generate and optimize solutions. The design of role-based operations assigns tasks to different roles. For example, the analyzer is responsible for inputting requirement specifications and boundary conditions, the fixer uses the large language model to generate corrected targets, and the checker verifies the generated solutions. In addition, the chain-of-thought prompt engineering provides guiding information to help the large language model more accurately understand and solve target conflicts. Finally, an iterative method is adopted to continuously optimize the solutions to ensure their adaptation to different requirements and boundary conditions. In summary, the present invention combines the large language model with interactive role-based operations and chain-of-thought prompt engineering to achieve the purpose of automatically analyzing, generating, and verifying target conflict solutions.

[0060] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications made without departing from the spirit and principle of the present invention shall be equivalent replacement methods and are all included in the protection scope of the present invention.

Claims

1. A control system automatic repair method based on a large language model, characterized in that: The following steps are involved: S1, initializing and pre-training a large language model to obtain an adjusted large language model; S2. Establish an interactive goal conflict resolution framework with three roles: analyst, fixer, and checker. According to the goals, domain attributes, and boundary condition data, conflict resolution solutions are iteratively generated and evaluated based on the adjusted large language model. The fixer uses the analysis results obtained from the analyst to generate the revised goals, while the checker is responsible for checking whether the revised goals meet the requirements of conflict resolution. Finally, the repaired goals are obtained and the solution set is obtained. S3, obtain the required information from the solution set to repair the control system; The boundary conditions are used in requirements engineering to capture specific situations that lead to conflicting goals.

2. The automated repair method according to claim 1, characterized in that: Step S2 includes: S21, using role instructions to initialize the large language models of the analyzer and the repairer respectively, and start the background service of the checker; S22, inputting the linear temporal logic formulas as the target, domain attributes and boundary conditions to the analyst respectively; S23, obtaining analysis information through the analyst, and adding the analysis information, input target, domain attributes and boundary conditions to the entire information; S24, iterative interaction, allowing the analyst, repairer, and inspector to continue interacting until the conditions for ending the interaction are met; when the conditions for ending the interaction are met, the cycle of continuous interaction ends, and finally a solution set is extracted from all the information in step S23.

3. The automated repair method according to claim 2, characterized in that: Step S24 includes: S241, initializing the solution set to an empty set; S242, if the solution set is not empty, the inspection information obtained by the logic checker is input to the repairer, and the repairer obtains the solution information, and then adds the solution information to all the information, and extracts the solution at the same time and adds it to the solution set; S243. If the solution set is empty, all information is input to the repairer, and the repairer obtains the solution information, extracts the solution, and adds it to the solution set; S244. Repeat steps S242 and S243 until the end condition is met, and return the final solution set.

4. The automated repair method according to claim 2, characterized in that: In step S22, the domain attributes are descriptive statements of the problem domain, which are used to capture the scope of the requirements; the goals are prescriptive statements, which require the system to achieve the goals; the boundary conditions describe the combinations that cannot be met as a whole in specific situations, which are used to capture the conditions for the occurrence of goal divergence.

5. The automated repair method according to claim 2, characterized in that: The condition for the termination of the interaction in step S24 is that the repaired target can pass the syntax check, semantic check and satisfiability check; wherein the syntax check is to check whether the repaired target is a legal linear temporal logic formula; Semantic checking is to check whether the repaired target can contain the semantics of the original target; Satisfiability checking is to check whether the repaired target satisfies the domain properties and boundary conditions.

6. The automated repair method according to claim 2, characterized in that: The role instructions of step S21 include: The role instruction of the analyst is: Please help analyze the reason why a certain formula does not satisfy a certain condition; The role instruction of the repairer is: please generate a repair target in the domain that satisfies the boundary conditions.

7. The automated repair method according to claim 1, characterized in that: Step S1 manually or automatically extracts the objectives, domain attributes and boundary conditions corresponding to the control system from the automatic control system and / or software as data for pre-training a large language model; the objectives are safety indicators or quantity and quality requirements of production, the domain attributes are the utility of different hardware, and the boundary conditions are the state where the objectives and domain attributes conflict.

8. The automated repair method according to claim 7, characterized in that: After pre-training, step S1 uses manually annotated data about the target conflict resolution task to adjust the large language model, that is, given the input target, domain attributes and boundary conditions, manually answer the repaired target, input each set of corresponding data into the large language model, freeze the lower-level parameters of the large language model, and only adjust the upper-level parameters of the large language model.

9. The automated repair method according to claim 3, characterized in that: In step S22, the control system is a water pump control system inside the mine, the goal is "no flooding, no explosion", the domain attribute is "water pump effect", and the boundary condition is "high water level and methane exist at the same time" is true; In step S24, when the solution set is not empty, the check information obtained by the logic checker is: the next time of high water level, there is no methane and the pump is turned on, and the next time of methane existence, there is low water level and the pump is turned off, which cannot meet the boundary condition "high water level and methane exist at the same time", and it is necessary to give a repaired target again; The inspection information is input to the repairer, who obtains solution information: the goal after repair is "when the water level is high and there is no methane, turn on the water pump at the next time; When the water level is low and methane is present, the pump is turned off next.” 10. A control system automatic repair system based on a large language model, used to implement the automatic repair method according to any one of claims 1 to 9, characterized in that: Includes the following modules: A model adjustment module is used to initialize and pre-train a large language model to obtain an adjusted large language model; The solution set generation module establishes an interactive target conflict resolution framework with three roles: analyst, fixer, and checker. According to the target, domain attribute, and boundary condition data, the conflict resolution scheme is iteratively generated and evaluated based on the adjusted large language model. The fixer uses the analysis results obtained from the analyst to generate the revised target, while the checker is responsible for checking whether the revised target meets the requirements of conflict resolution. Finally, the repaired target is obtained, and the solution set is obtained. The repair module obtains the required information from the solution set to repair the control system; The boundary conditions are used in requirements engineering to capture specific situations that lead to conflicting goals.

Images