Method for setting PID (Proportion Integration Differentiation) parameters by using large language model

Abstract

The invention provides a PID parameter setting method based on a large language model, and belongs to the technical field of industrial automation control. Aiming at the problems that a traditional PID setting method depends on expert experience, is long in debugging period and lacks theoretical interpretation, adaptive optimization of PID parameters is realized by fusing dynamic response characteristics of a control system and inference capability of a large language model. The method comprises the following steps: collecting time sequence data of process variables, and extracting dynamic characteristics such as overshoot, oscillation period, attenuation ratio and integral absolute error (IAE); comparing the feature vector with a preset industrial standard threshold value, and generating a natural language cue word containing quantitative analysis; inputting the large language model, and outputting adjustment suggestions including proportion, integral and differential parameters; evaluating a parameter effect through a closed-loop verification experiment, and if the parameter effect does not reach the standard, combining historical debugging data to carry out iterative optimization; and finally, an optimal PID parameter is generated. Experiments prove that the method has a good effect in the industrial process containing pure delay and large inertia.

Description

Technical Field

[0001] The present invention relates to the field of process industrial control, and particularly to a method for tuning PID parameters using a large language model. Background Art

[0002] The proportional-integral-derivative (PID) controller, as the most widely used regulating device in the field of industrial control, the quality of its parameter tuning directly determines the stability and dynamic performance of the control system. According to the statistics of the International Society of Automation (ISA), more than 90% of industrial control loops globally adopt the PID control strategy. However, the core contradiction faced by this classical control method lies in the mismatch between the simple and understandable control structure (three-parameter regulation) and the complex and variable parameter tuning requirements. Traditional PID tuning methods mainly rely on engineers' empirical trial and error, resulting in a long debugging cycle and high costs, especially when dealing with large inertia, nonlinear, or time-varying systems.

[0003] Based on a comprehensive analysis of existing patented technologies, the traditional PID parameter tuning method has the following significant limitations in industrial applications:

[0004] Step response methods represented by the Ziegler-Nichols method (such as the critical ratio method mentioned in patent CN101900991A) determine parameters by measuring the critical gain, but there are serious operation risks: critical oscillations in the actual system may cause equipment safety hazards. The Cohen-Coon method (the multi-loop system analysis involved in patent CN115562221A) although introduces delay parameters to improve the regulation rule, is still limited by the strong assumption conditions of the first-order plus pure delay model and cannot adapt to multivariable coupling systems (such as the heat exchange system described in patent CN115562221A) and nonlinear time-varying conditions. Patent CN109696827B points out that it is difficult for such pre-set rule parameter combinations to maintain optimal control performance in a complex dynamic environment.

[0005] Model-driven methods such as internal model control (IMC) (such as the multi-loop decoupling strategy proposed in patent CN115562221A) rely on accurate transfer function modeling, but actual industrial objects (such as the time-varying time-delay process mentioned in patent CN115657453A) often have unmodeled dynamic characteristics. Patent CN115562221A reveals that traditional IMC needs to frequently reconstruct the decoupling matrix in a multi-loop system. For a chemical process with strong coupling characteristics (such as the temperature-pressure control of a reactor), model mismatch will lead to deterioration of control quality. In addition, patent CN116165880B emphasizes that traditional modeling methods are difficult to handle complex conditions with the superposition of sensor noise and load disturbance.

[0006] There are contradictions in the convergence speed and optimization accuracy between the traditional Genetic Algorithm (GA) and the Particle Swarm Optimization (PSO): Patent CN115657453A points out that the lack of population diversity in the standard GA is prone to falling into local optima, while Patent CN109696827B confirms that the fixed inertia weight mechanism of the classic PSO loses its global search ability in the later stage of iteration. Patent CN116165880B further reveals that such algorithms require thousands of simulation iterations (such as the reinforcement learning trial-and-error mechanism in Patent CN115169520A), and the calculation time-consuming is difficult to meet the real-time control requirements, and the algorithm parameters (such as the cellular space structure in Patent CN115657453A) need to be adjusted empirically, increasing the deployment complexity.

[0007] The parameter self-tuning method based on Reinforcement Learning (RL) (as described in Patents CN118244618A and CN116165880B) avoids precise modeling, but there are three obstacles: First, as described in Patent CN115562221A, the RL policy training depends on a high-fidelity simulation environment, and the sensor noise and actuator nonlinearity in the actual system will cause the policy migration to fail; Second, Patent CN118244618A points out that traditional RL requires more than ten thousand interactive trainings, which cannot meet the online learning requirements of devices such as AGVs; Finally, as emphasized in Patent CN116165880B, the black-box neural network decision logic lacks interpretability and is difficult to verify the control stability through traditional tools such as Bode diagrams.

[0008] The existing patent technologies show an obvious improvement path: Patent CN109696827B enhances the global search ability of PSO through cosine inertia weight adjustment; Patent CN115657453A adopts a cellular space structure and a diversity preservation mechanism to optimize the convergence characteristics of GA; Patent CN115562221A innovatively combines inverse decoupling and all-pole approximation to improve the robustness of IMC in multi-loop systems; Patent CN116165880B proposes a hybrid architecture of genetic algorithm and convolutional neural network, reducing the computational complexity through data preprocessing. These progress provide new ideas for breaking through the limitations of traditional methods, but have not fully solved the collaborative optimization problems of multi-objective optimization, real-time response and model interpretability. Summary of the Invention

[0009] In view of the technical deficiencies of existing PID parameter tuning methods in terms of model dependence, computational efficiency, and human-machine collaboration, the present invention proposes a model-free adaptive tuning method based on large language models. The present invention is expected to specifically solve the following technical problems: (1) Breaking the model dependence bottleneck: Traditional methods require obtaining the accurate mathematical model of the controlled object (such as transfer function, state equation), while actual industrial systems often have problems of unmodeled dynamics and parameter time-variation; (2) Reducing computational complexity: Intelligent optimization algorithms (genetic algorithms, reinforcement learning) require thousands of simulation iterations and are difficult to meet the requirements of real-time control; (3) Enhancing human-machine collaboration ability: Existing automation methods output black-box parameters and lack natural language explanations for adjustment strategies, hindering engineers' decision-making; (4) Achieving multi-objective dynamic balance: Optimizing a single performance index results in mutual constraints among parameters such as overshoot and adjustment time, and cannot meet the requirements of complex working conditions.

[0010] To achieve the above object, the present invention provides the following technical solutions:

[0011] The present invention first proposes a method for tuning PID parameters using a large language model, which realizes the tuning of PID parameters by integrating the dynamic response characteristics of the control system and the reasoning ability of the large language model. The method includes the following steps:

[0012] S1. Collect the process variable time series data of the industrial system controlled by the PID controller, and extract the control system quality indexes of the process variables;

[0013] S2. Compare and analyze the control system quality indexes with the preset thresholds. If they meet the preset threshold range, no adjustment is made; if they exceed the preset threshold range, a prompt word including the control system quality indexes and their natural language descriptions is generated;

[0014] S3. Input the prompt word into the large language model to generate PID parameter adjustment suggestions;

[0015] S4. Optimize the PID parameters according to the adjustment suggestions, perform a closed-loop control verification experiment, and collect the optimized process variable time series data;

[0016] S5. Verify whether the optimized control system quality indexes meet the preset threshold requirements. If not, regenerate the prompt word based on the current control system quality indexes and return to step S3 until the control system quality indexes meet the preset threshold requirements.

[0017] According to the preferred embodiment of the present invention, the control system quality indexes include at least three characteristic quantities representing the dynamic response of the system. Further, the characteristic quantities representing the dynamic response of the system are selected from indexes such as overshoot, oscillation amplitude, decay ratio, oscillation period, time constant, settling time, steady-state error, and / or integral absolute error (IAE), etc.

[0018] Preferably, the large language model is selected from pre-trained language models including but not limited to at least one of the following architectures: ChatGPT-4o, Llama3.1-8b, or DeepSeek-R1 based on the Transformer architecture. The model establishes the mapping relationship between the PID parameter adjustment strategy and the control system quality index through the transfer learning mechanism, and realizes the multi-dimensional PID parameter optimization based on the time-domain response characteristics of the system.

[0019] The present invention verifies the PID parameter adjustment result through a feedback mechanism and further adjusts the parameters of the PID controller according to the verification result. The process variable is the output response of the system, and the method is used to automatically adjust the PID controller of the industrial automation control system.

[0020] The present invention also provides a system for tuning PID control parameters, including:

[0021] A processing unit configured to receive the curve data of the process variable of the industrial system controlled by the PID controller changing with time;

[0022] A curve data processing module configured to convert the curve data into prompt word content containing process variable eigenvalue;

[0023] A large language model module configured to generate PID parameter adjustment suggestions based on the prompt word content;

[0024] A PID controller configured to automatically adjust the parameters of the PID controller according to the adjustment suggestions.

[0025] The present invention also provides an intelligent PID tuning system based on multi-module collaboration, including:

[0026] A process system module for executing the industrial system controlled by the PID controller to generate dynamic response time series data;

[0027] A feature extraction module communicatively coupled with the process system module and configured to extract the control system quality index from the time series data;

[0028] An effect judgment module that receives the output signal of the feature extraction module and generates a tuning decision signal based on a preset control standard;

[0029] A prompt word construction module that, in response to the tuning decision signal of the effect judgment module, encodes the dynamic feature vector into a structured natural language instruction;

[0030] A large model module that receives the output of the prompt word construction module and generates a PID parameter adjustment instruction through a pre-trained generative language model and feeds it back to the process system module;

[0031] Among them, the output end of the process system module is communicatively coupled to the input end of the feature extraction module, and the decision signal output end of the effect judgment module is respectively connected to the parameter output port and the trigger port of the prompt word construction module, forming a closed-loop iterative optimization architecture.

[0032] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0033] Through the technical route of "feature semanticization → model reasoning → closed-loop verification", the present invention achieves breakthroughs in aspects such as model adaptability, real-time performance, multi-objective optimization, and human-computer interaction, providing a new generation of basic tools for industrial control intelligence.

[0034] (1) Completely break the model dependence and achieve true model-free tuning: Through the dynamic feature extraction subsystem, directly extract an 8-dimensional feature vector such as overshoot (OS%), effective oscillation times (Nosc), and adjustment time (ts) from the time-domain response curve (such as step response), without any transfer function or state-space model, solving the pain point of difficult modeling of complex objects in fields such as petroleum and metallurgy, and expanding the application boundary of PID control.

[0035] (2) Multi-objective dynamic optimization, breaking through the bottleneck of mutual exclusion of performance indicators: Based on the 8-dimensional feature vector (overshoot, oscillation times, IAE, etc.), develop a configurable termination rule engine and an adaptive prompt strategy. Users can dynamically select feature dimensions as optimization goals (such as stopping when "overshoot < 5% and steady-state error ≤ 0.1%"), guiding the adjustment direction of the large model. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is a schematic diagram of the system module of the present invention.

[0037] Figure 2 It is a schematic diagram of the method of the present invention under the embodiment of the present invention. Figure 2 Compared with Figure 1 The difference is that Figure 1 The process system module of refers to the entity controlled by the controller, while Figure 2 is the simulation module adopted in the embodiment, used for simulating and generating data.

[0038] Figure 3 It is the comparison of the system performance before and after optimization in Embodiment 1.

[0039] Figure 4 It is the comparison of the system performance before and after optimization in Embodiment 2.

[0040] Figure 5 It is the comparison of the system performance before and after optimization in Embodiment 3.

[0041] Figure 6It is a comparison of the system performance before and after the optimization of Example 4 (non-preferred example).

[0042] Figure 7 It is a comparison of the system performance before and after the optimization of Comparative Example 1. Detailed implementation manners

[0043] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be described in detail below with reference to specific embodiments. Specific embodiments are described below to simplify the present invention. However, it should be recognized that the present invention is not limited to the described embodiments, and various modifications of the present invention are possible without departing from the basic principles, and these equivalent forms also fall within the scope defined by the appended claims of this application.

[0044] As Figure 1 and Figure 2 shown, it schematically shows the schematic diagrams of the respective modules of the intelligent PID tuning system based on multi-module collaboration of the present invention. The system mainly includes a process system module, a feature extraction module, an effect judgment module, a prompt word construction module, and a large model module; wherein, the output end of the process system module is communicatively coupled to the input end of the feature extraction module, and the decision signal output end of the effect judgment module is respectively connected to the parameter output port and the trigger port of the prompt word construction module, forming a closed-loop iterative optimization architecture.

[0045] The process system module is used to execute the industrial system controlled by the PID controller to generate dynamic response time series data; the feature extraction module is communicatively coupled to the process system module and is configured to extract the control system quality indicators included in the time series data; the effect judgment module receives the output signal of the feature extraction module and generates a tuning decision signal based on the preset control standard; the prompt word construction module responds to the tuning decision signal of the effect judgment module and encodes the dynamic feature vector into a structured natural language instruction; the large model module receives the output of the prompt word construction module and generates a PID parameter adjustment instruction through the pre-trained generative language model and feeds it back to the process system module.

[0046] Based on the foregoing system, the PID parameter tuning method based on the large language model of the present invention can be implemented according to the following steps:

[0047] S1. Collect the process variable time series data of the industrial system controlled by the PID controller (such as industrial process variables such as temperature, pressure, and flow rate), and extract the control system quality indicators of the process variables (overshoot, oscillation amplitude, decay ratio, oscillation period, time constant, settling time, steady-state error, and / or integral absolute error (IAE), etc.).

[0048] S2. Compare and analyze the quality indicators of the control system with the preset thresholds. If the preset threshold range is met, no adjustment is made. If the preset threshold range is exceeded (for example, an overshoot exceeding 10% is determined to be abnormal), a prompt word including the quality indicators of the control system and their natural language descriptions is generated.

[0049] The generation of the prompt word can be specifically as follows: Numerically quantify the characteristic quantities representing the dynamic response of the system through the feature extraction module, and generate corresponding natural language state descriptions based on the preset threshold intervals; Generate a structured prompt word input by the prompt word construction module.

[0050] S3. Input the prompt word into the large language model to generate PID parameter adjustment suggestions (suggested values for the proportional coefficient Kp, integral gain Ki, and derivative gain Kd).

[0051] S4. Optimize the PID parameters according to the adjustment suggestions, perform a closed-loop control verification experiment, and collect the time series data of the process variables after optimization.

[0052] S5. Verify whether the quality indicators of the optimized control system meet the preset threshold requirements. If not, regenerate the prompt word based on the current quality indicators of the control system and return to step S3 until the quality indicators of the control system meet the preset threshold requirements.

[0053] In another embodiment of the present invention, in order to adapt to the above method, a system for tuning PID control parameters is further provided, including: a processing unit, a curve data processing module, a large language model module, and a PID controller.

[0054] The processing unit is configured to receive the curve data of the process variables changing with time of the industrial system controlled by the PID controller; the curve data processing module is configured to convert the curve data into prompt word content including the characteristic values of the process variables; the large language model module is configured to generate PID parameter adjustment suggestions based on the prompt word content; the PID controller is configured to automatically adjust the parameters of the PID controller according to the adjustment suggestions.

[0055] The following further demonstrates the invention in combination with specific embodiments:

[0056] Example 1: Tuning of PID parameters for a simulation system (llama3.1 - 8b)

[0057] In this example, PID parameter tuning is performed for a first-order pure time-delay model system, and the system transfer function is:

[0058]

[0059] The initial PID controller parameter settings are the proportional gain K p = 1.0, integral gain Ki = 0.1 s -1 , differential gain K d = 0.01 s. In the step response test, the system exhibits significant dynamic performance defects (such as Figure 3 (a) in): the overshoot reaches 38.89%, far exceeding the 10% threshold required by industrial control standards; the response curve shows continuous oscillation, the decay ratio between the first and second peaks is 0.76, and the oscillation period is 273.01 seconds; the settling time is 418.02 seconds, the steady-state error is 0.00%, and the integral absolute error (IAE) is 92.13. Overall, it does not meet the requirements, indicating that the control accuracy needs to be further optimized.

[0060] Based on the above dynamic characteristics, structured natural language instructions are generated and input into the pre-trained large language model. The input content includes background description, demonstration cases, current state description, and tasks. Prompt words are generated and input into the large language model. The structure of the prompt words is as follows:

[0061]

Background description

[0062]

Case demonstration

[0063]

Current state description

[0064]

Task

[0065] After the large language model analyzes the input data, the output parameter adjustment suggestions are: Kp: 1.85; Ki: 0.027; Kd: 10.21

[0066] After applying the new parameters, the system performance is significantly improved (such as Figure 3 (b) in): the overshoot drops to 5.16%, a decrease of 86.7%; the steady-state error stabilizes at 0.00%; the integral absolute error (IAE) decreases from 92.13 to 62.40, and the control accuracy increases by 32.3%. The optimized step response curve shows a smooth convergence characteristic, meeting the expected industrial process control standards.

[0067] Example 2: PID parameter tuning of the simulation system (ChatGPT-4o)

[0068] In this example, PID parameter tuning is performed for a first-order pure time-delay model system, and the system transfer function is:

[0069]

[0070] The initial PID controller parameters are set as the proportional gain K p = 1.0, the integral gain K i = 0.1 s -1 , and the derivative gain K d = 0.01 s. In the step response test, the system exhibits significant dynamic performance defects (as shown in Figure 4 (a) of): the overshoot reaches 57.72%, far exceeding the 10% threshold required by industrial control standards; the response curve shows severe oscillations, the decay ratio between the first and second peaks is 0.75, and the oscillation period is 679.03 s; the settling time is 1761.09 s, the steady-state error is 3.49%, the integral absolute error (IAE) is 381.60, and there is a 2.50% residual that does not converge, resulting in seriously insufficient overall control accuracy.

[0071] Based on the above dynamic characteristics, structured natural language instructions are generated to input into a pre-trained large language model. The input content includes background description, demonstration cases, current state description, and tasks. Prompt words are generated and input into the large language model. The structure of the prompt words is as follows:

[0072]

Background description

[0073]

Case demonstration

[0074]

Current state description

[0075]

Task

[0076] After the large language model analyzes the input data, the output parameter adjustment suggestions are: Kp: 2.11; Ki: 0.012; Kd: 30.15

[0077] After applying the new parameters, the system performance is significantly improved (as shown in Figure 4 (b) of): the overshoot drops to 8.48%, a decrease of 85.3%; the integral absolute error (IAE) decreases from 381.60 to 345.96, and the control accuracy is improved by 40.7%. The steady-state error of the optimized step response curve is only 0.06%, and the optimized step response curve shows a smooth convergence characteristic, meeting the expected industrial process control standards.

[0078] Example 3: PID Parameter Tuning of the Simulation System (deepseek-r1)

[0079] In this example, PID parameter tuning is performed for a first-order pure delay model system, and the system transfer function is as follows:

[0080]

[0081] The initial PID controller parameters are set as proportional gain K p = 1.0, integral gain K i = 0.1 s -1 and derivative gain K d = 0.01 s. In the step response test, there are obvious defects in the system dynamic performance (as shown in (a) in Figure 5 ): the overshoot is 39.38%, exceeding the allowable range of industrial standards; the response curve shows periodic oscillations, the decay ratio of the first and second peaks is 0.76, and the oscillation period is 626.03 s; the settling time is 991.05 s, the steady-state error is 0.25%, and the integral absolute error (IAE) is 226.26, so the control accuracy needs to be improved.

[0082] Based on the above dynamic characteristics, structured natural language instructions are generated and input into the pre-trained large language model. The input content includes background description, demonstration cases, current state description, and tasks. A prompt is generated and input into the large language model, and the prompt structure is as follows:

[0083]

Background description

[0084]

Case demonstration

[0085]

Current state description

[0086]

Task

[0087] After the large language model analyzes the input data, the output parameter adjustment suggestions are: Kp: 6.11; Ki: 0.061; Kd: 44.23

[0088] After applying the new parameters, the system performance is significantly optimized (as shown in Figure 5In (b) of it: the overshoot is reduced to 8.42%, a decrease of 78.6%; the integral absolute error (IAE) is reduced from 226.26 to 122.64, and the steady-state error is stabilized at 0.00%. The optimized step response curve shows a smooth convergence characteristic, meeting the expected industrial process control standards.

[0089] Example 4 (non-preferred example): PID parameter tuning of the simulation system - reducing the characteristic quantities to three dimensions (llama3.1-8b)

[0090] In this example, PID parameter tuning is carried out for a first-order pure delay model system, and the system transfer function is:

[0091]

[0092] The initial PID controller parameter settings are the proportional gain K p = 1.0, the integral gain K i = 0.1 s -1 , the derivative gain K d = 0.01 s. In the step response test, the system shows significant dynamic performance defects (such as Figure 6 in (a) of it: the overshoot reaches 38.89%, far exceeding the 10% threshold required by the industrial control standard; the response curve shows continuous oscillation, the decay ratio of the first and second peaks is 0.76, and the oscillation period is 273.01 s; the settling time is 418.02 s, the steady-state error is 0.00%, and the integral absolute error (IAE) is 92.13. Overall, it does not meet the requirements, indicating that the control accuracy needs to be further optimized.

[0093] Based on the above dynamic characteristics, structured natural language instructions are generated to input into the pre-trained large language model. The input content includes background description, demonstration cases, current state description, and tasks, and the prompt words are generated and input into the large language model. The prompt word structure is as follows:

[0094]

Background description

[0095]

Case demonstration

[0096]

Current state description

[0097]

Task

[0098] After the large language model analyzes the input data, the output parameter adjustment suggestions are: Kp: 3.78; Ki: 0.019; Kd: 54.150

[0099] After applying the new parameters, the system performance is significantly improved (as shown in (b) of Figure 6 ): the overshoot is reduced to 0%; the integral absolute error (IAE) is reduced from 92.13 to 81.07, the control accuracy is increased by 12.1%, and the steady-state error is stabilized at 0.00%. The optimized step response curve shows a smooth convergence characteristic, meeting the expected industrial process control standards.

[0100] Comparative Example 1: PID parameter tuning of the simulation system - reducing the description related to characteristic quantities (llama3.1 - 8b)

[0101] In this embodiment, PID parameter tuning is performed for a first-order pure time-delay model system, and the system transfer function is:

[0102]

[0103] The initial PID controller parameter settings are proportional gain K p = 1.0, integral gain K i = 0.1s -1 , derivative gain K d = 0.01s. In the step response test, the system shows significant dynamic performance defects (as shown in (a) of Figure 7 ): the overshoot reaches 38.89%, far exceeding the 10% threshold required by the industrial control standard; the response curve shows continuous oscillation, the decay ratio of the first and second peaks is 0.76, and the oscillation period is 273.01 seconds; the settling time is 418.02 seconds, the steady-state error is 0.00%, and the integral absolute error (IAE) is 92.13. Overall, it does not meet the requirements, indicating that the control accuracy needs to be further optimized.

[0104] Based on the above dynamic characteristics, structured natural language instructions are generated and input into the pre-trained large language model. The input content includes background description, demonstration cases, current state description, and tasks, and prompts are generated and input into the large language model. The prompt structure is as follows:

[0105]

Background description

[0106]

Case demonstration

[0107]

Current Status Description

[0108]

Task

[0109] After the large - language model analyzes the input data, the output parameter adjustment suggestions are: Kp: 2.06; Ki: 0.037; Kd: 8.430

[0110] After applying the new parameters, the system performance has been significantly improved (such as Figure 7 in (b) of []): The overshoot has dropped to 0%; the integral absolute error (IAE) has decreased from 92.13 to 68.82, the control accuracy has increased by 25.3%, and the steady - state error has stabilized at 0.00%. The optimized step - response curve shows a smooth convergence characteristic, meeting the expected industrial process control standards.

[0111] The above - described embodiments only represent several implementation manners of the present invention. Their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the patent of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the appended claims.

Claims

1. A PID parameter tuning method based on a large language model, characterized in that, Including the following steps: S1. Collect the process variable time series data of the industrial system controlled by the PID controller, and extract the control system quality indicators of the process variables; S2. Compare and analyze the control system quality indicators with the preset thresholds. If they exceed the preset threshold range, generate a prompt word including the control system quality indicators and their natural language descriptions; S3. Input the prompt word into the large language model to generate PID parameter adjustment suggestions; S4. Optimize the PID parameters according to the adjustment suggestions, perform a closed-loop control verification experiment, and collect the optimized process variable time series data; S5. Verify whether the optimized control system quality indicators meet the preset threshold requirements. If not, regenerate the prompt word based on the current control system quality indicators and return to step S3 until the control system quality indicators meet the preset threshold requirements.

2. The method according to claim 1, wherein The control system quality indicators include at least three characteristic quantities representing the system's dynamic response.

3. The method according to claim 1, wherein The characteristic quantities representing the system's dynamic response include overshoot, oscillation amplitude, decay ratio, oscillation period, time constant, settling time, steady-state error, and / or integral absolute error (IAE).

4. The method according to claim 1, wherein The generation of the prompt word including the control system quality indicators and their natural language descriptions includes: Numerically quantifying the characteristic quantities representing the system's dynamic response through a feature extraction module, and generating corresponding natural language state descriptions based on the preset threshold intervals; generating a structured prompt word input by a prompt word construction module.

5. The method according to claim 1, wherein In step S3: The parameter adjustment suggestions output by the large language model include the suggested values of the proportional coefficient Kp, integral gain Ki, and derivative gain Kd.

6. The method according to claim 1, characterized in that: The large language model is selected from pre-trained language models including but not limited to at least one of the following architectures: ChatGPT-4o, Llama3.1-8b, or DeepSeek-R1 based on the Transformer architecture. The model establishes a mapping relationship between the PID parameter adjustment strategy and the control system quality indicators through a transfer learning mechanism to achieve multi-dimensional PID parameter optimization based on the system's time-domain response characteristics.

7. The method according to claim 1, further comprising the step of: verifying the PID parameter adjustment result through a feedback mechanism, and further adjusting the parameters of the PID controller according to the verification result.

8. The method according to claim 1, wherein the process variable is the output response of the system, and the method is used to automatically adjust the PID controller of the industrial automation control system.

9. A system for tuning PID control parameters, characterized in that Including: A processing unit configured to receive the curve data of the process variables of the industrial system controlled by the PID controller changing with time; A curve data processing module configured to convert the curve data into prompt word content including process variable characteristic values; A large language model module configured to generate PID parameter adjustment suggestions based on the prompt word content; A PID controller configured to automatically adjust the parameters of the PID controller according to the adjustment suggestions.

10. An intelligent PID tuning system based on multi-module collaboration, characterized in that Including: A process system module for executing the industrial system controlled by the PID controller to generate dynamic response time series data; A feature extraction module, communicatively coupled with the process system module, configured to extract control system quality indicators from the time series data; An effect judgment module, receiving the output signal of the feature extraction module, generating a tuning decision signal based on a preset control standard; A prompt word construction module, in response to the tuning decision signal of the effect judgment module, encoding the dynamic feature vector into a structured natural language instruction; A large model module, receiving the output of the prompt word construction module, generating a PID parameter adjustment instruction through a pre-trained generative language model and feeding it back to the process system module; Wherein, the output end of the process system module is communicatively coupled with the input end of the feature extraction module, and the decision signal output end of the effect judgment module is respectively connected to the parameter output port and the trigger port of the prompt word construction module, forming a closed-loop iterative optimization architecture.

Images