Fresh air conditioner control system and method based on double-model safety evolution

Abstract

The invention relates to a fresh air conditioner control system and method based on double-model safety evolution. The fresh air conditioner control system comprises a credible data acquisition module, a mechanism main model controller and a shadow AI model training and verification module. A dynamic anchoring control module; a man-machine collaborative decision-making interface; and a fault isolation and model switching module. On the premise of guaranteeing control safety, the invention provides an industrial air conditioner control system architecture with evolution capability, adaptive capability and deployment flexibility, breaks through the limitation of the traditional control technology in nonlinear, multivariable and strong coupling scenes, and has relatively high engineering application value and popularization potential.

Description

Technical Field

[0001] The present invention relates to the technical field of fresh air air conditioning control, and in particular to a fresh air air conditioning control system and method based on dual-model safe evolution. Background Art

[0002] Traditional PID control algorithms are still widely used in current fresh air air conditioning systems. Although systems using advanced process control (APC) technology have emerged on the market, their penetration in small and medium-sized buildings remains limited. This method relies on manual experience for parameter tuning and struggles to adapt to the complex characteristics of system operation, such as nonlinearity, time-varying behavior, and disturbances. In practical applications, temperature control often experiences overshoot or fluctuations, especially in multivariable coupled scenarios such as coordinated temperature and humidity control. Due to limited decoupling capabilities, PID controllers are prone to system oscillation and response lag.

[0003] Furthermore, existing control systems struggle to deploy complex algorithms given limited edge computing resources. Typical embedded controllers, constrained by computing power and memory resources, struggle to run highly complex control models in real time. While traditional mechanism-based modeling approaches offer advantages such as strong interpretability and robustness, they suffer from long model construction cycles and complex parameter adjustments, making them difficult to adapt to rapidly changing scenarios and placing higher demands on project implementation and maintenance.

[0004] At the same time, existing technologies still face a conflict between reliability and security when it comes to control system upgrades and iterations. For example, during remote updates or online adjustments, control logic failures or system responsiveness interruptions can lead to serious consequences such as environmental control anomalies, equipment damage, and even service interruptions. Therefore, the industry urgently needs a control strategy that maintains system stability and security while offering greater adaptability and flexibility to meet the growing performance demands and operational challenges of intelligent air conditioning systems. Summary of the Invention

[0005] The purpose of the present invention is to overcome the deficiencies of the prior art and provide a fresh air air conditioning control system and method based on dual-model safe evolution.

[0006] To achieve the above object, the present invention adopts the following technical solutions: A fresh air air conditioning control system based on dual-model safe evolution, characterized by comprising: Trusted Data Acquisition Module: This module collects air conditioning system operating data in real time and selects high-quality samples using a trustworthiness evaluation algorithm. The trustworthiness evaluation algorithm's indicators include operating stability, control strategy stability, rationality of key variable fluctuations, and minimization of system interference. Mechanism main model controller: built based on the principles of thermodynamics and fluid mechanics, used to output control instructions when the AI ​​model has not yet been established or is being verified; Shadow AI model training and verification module: runs in parallel with the main mechanism model controller, uses trusted data to train the AI ​​model and compares and verifies the output with the main mechanism model controller in a sandbox environment; Dynamic anchoring control module: used to adaptively adjust the output weight ratio of the mechanism main model controller and the AI ​​model according to the degree of environmental disturbance or system state changes; Human-machine collaborative decision-making interface: used to display AI model suggestions, receive user instructions, and support forced switching of control modes; Fault isolation and model switching module: used to automatically disconnect the AI ​​model control path when abnormal AI model output is detected, and switch to the dominant control of the mechanism main model controller.

[0007] Furthermore, the indicators of the credibility evaluation algorithm specifically include: Operating stability: The air flow, temperature, and humidity parameters fluctuate within the set period without exceeding the threshold; Control strategy stability: no control strategy switching occurs during the sampling period, and the system does not perform parameter tuning or model updating; Reasonable fluctuation of key variables: The combustion section outlet temperature, spray section pressure, and surface temperature of the surface cooler are within the preset normal range, without sudden changes or fault-type outliers; Minimize system interference: No equipment maintenance, parameter debugging or strong external disturbance records are recorded during the sample collection period.

[0008] Furthermore, the operating logic of the shadow AI model training and verification module includes: The AI ​​model output is only used to compare with the results of the mechanism main model controller and does not directly participate in equipment control; When the prediction deviation, response delay, and energy consumption optimization ratio of the AI ​​model in the sandbox meet the preset indicators, its output is gradually introduced into the control decision.

[0009] Furthermore, the weight adjustment strategy of the dynamic anchoring control module is: When the environment is stable, increase the output weight of the AI ​​model; When external disturbances, sudden model deviations, or abnormal actuator feedback are detected, the output weight of the mechanism main model controller is increased to more than 90%; The weights are dynamically adjusted according to system response feedback to achieve a smooth transition between "mechanism main control-AI assistance" and "AI main control-mechanism guarantee".

[0010] Furthermore, the human-machine collaborative decision-making interface supports: Displays the energy saving ratio and comfort improvement ratio recommended by the AI ​​model; Provide users with action options: accept, observe, or ignore AI suggestions; Provide AI suggestion history, deviation trend and original data comparison chart; A high-privilege user can forcibly switch the control mode.

[0011] Furthermore, the triggering conditions of the fault isolation and model switching module include any one of the following: The AI ​​model output mutation exceeds the preset threshold; The deviation between the control result and the set value continues to expand; The user repeatedly rejected the AI ​​model's control suggestions; Sandbox verification indicators continue to decline.

[0012] A fresh air air conditioning control method based on dual-model safe evolution is applied to the system, comprising the steps of: Step S1: Trusted data collection and filtering Collect air conditioning system operation data in real time and select high-quality samples through credibility evaluation algorithm; Step S2: Initial control of the mechanism main model controller When the AI ​​model has not yet been established or verified, the main model controller based on the principles of thermodynamics and fluid mechanics outputs control instructions; Step S3: Shadow AI model training and sandbox evaluation Use trusted data to train AI models and run them in parallel with the main mechanism model controller in a sandbox environment; Compare the outputs of the AI ​​model and the mechanism master model controller to generate a performance profile that includes prediction deviation, response delay, and energy-saving and consumption optimization ratio; Step S4: Dynamic anchoring fusion control When the AI ​​model performance profile meets the stability threshold, it enters the low-proportion fusion control stage: The dynamic anchoring mechanism mixes the dual model outputs in a preset ratio; Real-time monitoring of fusion control effects and assessment of risk-benefit; As the confidence of the AI ​​model increases, the output weight of the AI ​​model is gradually increased; and when external disturbances, sudden changes in model deviations, or abnormal actuator feedback are detected, the output weight of the mechanism main model controller is increased to more than 90%; Step S5: Human-machine collaborative intervention and feedback Display AI model control suggestions and expected benefits to users through the human-machine collaborative decision-making interface; Receive user acceptance or rejection instructions, record intervention frequency, and use this information to evaluate the acceptability of the AI ​​model, assisting with subsequent fine-tuning and strategy updates; Step S6: Abnormal identification and safe switching If any of the following anomalies are detected, the AI ​​model will be immediately isolated and the mechanism master model controller will be switched to take the lead: The AI ​​model output mutation exceeds the preset threshold; The deviation between the control result and the set value continues to expand; The user repeatedly rejected the AI ​​model's control suggestions; Sandbox verification indicators continue to decline.

[0013] This paper proposes a dual control architecture that integrates physical and artificial intelligence models. Combining the characteristics of the multi-stage processing technology of industrial fresh air systems, it aims to address the bottleneck issues of current intelligent control systems in cold start, reliability, safety, and actual deployment. Specific technical points and beneficial effects include: 1. Dual-channel control mechanism based on parallel physical and AI models This paper designs a control structure in which the physical model leads control in the initial stages of the system, gradually introducing an AI model to assist in optimization as data accumulates. This architecture incorporates a "shadow model" mechanism, whereby the AI ​​model does not directly control the actuators in the early stages. Instead, it trains and verifies strategies through sandbox simulation, gradually increasing its participation once stability and reliability are confirmed. This mechanism ensures stable initial system operation while achieving incremental improvements in energy efficiency and control accuracy.

[0014] 2. Credible sample screening method based on industrial working conditions In view of the characteristics of industrial fresh air systems where users do not actively intervene and sensor differences are prevalent, the present invention proposes a set of high-quality sample identification criteria that focus on operational stability, identifiability of system process status, and consistency of control response. For example, by identifying the system in a constant air volume and heat source power output range for a long time, and combining the temperature and humidity change trends within the operating cycle, reliable data segments are automatically marked. This method effectively avoids the training bias caused by operational noise or sensor errors in conventional collection methods, significantly improving the training efficiency and generalization ability of AI models.

[0015] 3. Dynamic Anchoring Fusion Control Strategy This paper proposes a fusion algorithm that dynamically adjusts control weights based on the intensity of environmental disturbances, achieving a smooth transition between physical control and AI prediction. When the system detects significant disturbances (such as external climate change or process changes), control weights shift toward the physical model; when the environment is stable, the AI ​​model's weight is increased. This strategy effectively reduces system oscillations caused by model switching and enhances the robustness of the control system under complex industrial conditions.

[0016] 4. AI model launch process based on sandbox verification To ensure the stability and security of AI model control strategies in actual deployment, this paper sets up a sandbox environment to simulate and verify the AI ​​model control logic in advance, and conducts multi-level review and confirmation through a human-computer interaction interface. Only after confirmation can the actual control link be gradually entered. This process significantly reduces the uncertainty and risk of the AI ​​model going online, preventing the impact of training bias or logical errors on the actual system.

[0017] 5. Authority Gradient Mechanism for Human-Machine Collaboration This invention establishes a hierarchical authority management system that assigns different operational permissions based on user roles (e.g., operator, engineer, system administrator, etc.), limiting the scope of operations for model updates, control strategy modifications, and system parameter adjustments. This mechanism effectively prevents system anomalies caused by misoperation or unauthorized changes, providing institutional guarantees for the long-term stable operation of industrial sites.

[0018] 6. Segmented Modeling Strategy for Process Stage Coupling Given the characteristic of industrial fresh air systems, which consist of multiple process sections (combustion, spray, surface cooling, and secondary heating) connected in series, this paper proposes modeling each process section separately and employing a stage-aware mechanism to assist the AI ​​model in identifying the current process section, thereby improving the accuracy of prediction and control strategies. This approach effectively addresses the significant differences in operating conditions and strong coupling between sections, enhancing the model's generalization capabilities across the entire process.

[0019] 7. Control architecture design supporting edge deployment Considering the limited resources of industrial field controllers, this paper specifically considers the lightweight, computational efficiency, and deployment adaptability of AI models when designing the control architecture. Through model compression and inference optimization techniques, the AI ​​model can run efficiently on edge devices such as STM32 and RK3588, lowering the deployment threshold and improving the real-time performance and reliability of the control system.

[0020] In summary, the present invention provides an industrial air-conditioning control system architecture with evolutionary capabilities, adaptive capabilities, and deployment flexibility while ensuring control security. It breaks through the limitations of traditional control technology in nonlinear, multivariable, and strongly coupled scenarios, and has strong engineering application value and promotion potential. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments: Figure 1 Schematic diagram of the framework of the control system of the present invention. DETAILED DESCRIPTION

[0022] like Figure 1As shown, the present invention provides a fresh air air conditioning control system based on dual-model safe evolution, comprising: Trusted Data Collection Module: This module runs on the edge of the air conditioning system and is used to collect real-time operating data from the air conditioning system. It then selects high-quality samples using a trustworthiness evaluation algorithm. The trustworthiness evaluation algorithm's indicators include operating stability, control strategy stability, rationality of key variable fluctuations, and minimization of system interference. Mechanism main model controller: built based on the principles of thermodynamics and fluid mechanics, used to output control instructions when the AI ​​model has not yet been established or is being verified; Shadow AI model training and verification module: runs in parallel with the main mechanism model controller, uses trusted data to train the AI ​​model and compares and verifies the output with the main mechanism model controller in a sandbox environment; Dynamic anchoring control module: used to adaptively adjust the output weight ratio of the mechanism main model controller and the AI ​​model according to the degree of environmental disturbance or system state changes; Human-machine collaborative decision-making interface: used to display AI model suggestions, receive user instructions, and support forced switching of control modes; Fault isolation and model switching module: used to automatically disconnect the AI ​​model control path when abnormal AI model output is detected, and switch to the dominant control of the mechanism main model controller.

[0023] It should be noted that the sandbox environment is an isolated operating environment used to verify and evaluate AI model control strategies so that it will not affect the actual system and ensure that its output logic and behavior meet security and stability requirements.

[0024] Specifically, the indicators of the credibility evaluation algorithm include: Operating stability: Fluctuations in air flow, temperature, and humidity parameters within a continuous set period do not exceed the threshold, for example, within 30 minutes, the above parameters do not exceed ±5%. This is to avoid data collection during the system's dynamic adjustment process and ensure that the collected data can reflect the characteristics under stable control.

[0025] Control strategy stability: No control strategy switching occurs during the sampling period, and the system does not perform parameter tuning or model updates. This ensures that the data accurately reflects the response relationship under the current strategy and avoids interference with the training samples caused by "strategy drift."

[0026] Reasonable fluctuations in key variables: The combustion section outlet temperature, spray section pressure, and surface temperature of the condenser are within the preset normal range, with no sudden changes or fault-type outliers. This ensures that the model learns the response characteristics under normal operating conditions and avoids mislearning abnormalities as "target behaviors."

[0027] Minimize system interference: No equipment maintenance, parameter adjustment, or strong external disturbances are recorded during the sample collection period. This is to ensure that human or uncontrollable interference is eliminated and the quality of the sampled data is guaranteed.

[0028] The data must meet the above-mentioned indicators before entering the AI ​​model training process to ensure that what the model learns is effective behavior under normal working conditions.

[0029] The operating logic of the shadow AI model training and verification module includes: The AI ​​model output is only used to compare with the results of the mechanism main model controller and does not directly participate in equipment control; When the prediction deviation, response delay, and energy consumption optimization ratio of the AI ​​model in the sandbox meet the preset indicators, its output is gradually introduced into the control decision.

[0030] The above operating logic can ensure that the AI ​​model's control capabilities are controllable, observable, and reversible.

[0031] The weight adjustment strategy of the dynamic anchor control module is: When the environment is stable, increase the output weight of the AI ​​model; When external disturbances, sudden model deviations, or abnormal actuator feedback are detected, the output weight of the mechanism main model controller is increased to more than 90%; Dynamically adjust the weights based on system response feedback to achieve a smooth transition between "mechanism master control - AI assistance" and "AI master control - mechanism guarantee". Dynamic weight allocation is achieved through the following pseudo-code logic: content_change_detected: control_output = 90% * physical_output + 10% * ai_output Else if ai_model_confidence_high: control_output = 30% * physical_output + 70% * ai_output Else: control_output = 50% * physical_output + 50% * ai_output The human-machine collaborative decision-making interface supports: Displays the energy saving ratio and comfort improvement ratio recommended by the AI ​​model; Provide users with action options: accept, observe, or ignore AI suggestions; Provide AI suggestion history, deviation trend and original data comparison chart; High-authority users can forcefully switch the control mode (AI / mechanics / combined).

[0032] The triggering conditions of the fault isolation and model switching module include any of the following: The AI ​​model output mutation exceeds the preset threshold; The deviation between the control result and the set value continues to expand; The user repeatedly rejected the AI ​​model's control suggestions; Sandbox verification indicators continue to decline.

[0033] At the same time, the system automatically records the event and provides a prompt in the log to assist in subsequent problem tracing and model correction.

[0034] A fresh air air conditioning control method based on dual-model safe evolution includes the following steps: Step S1: Trusted data collection and filtering Collect air conditioning system operation data in real time and select high-quality samples through credibility evaluation algorithm; Step S2: Initial control of the mechanism main model controller When the AI ​​model has not yet been established or verified, the main model controller based on the principles of thermodynamics and fluid mechanics outputs control instructions; Step S3: Shadow AI model training and sandbox evaluation Use trusted data to train AI models and run them in parallel with the main mechanism model controller in a sandbox environment; Compare the outputs of the AI ​​model and the mechanism master model controller to generate a performance profile that includes prediction deviation, response delay, and energy-saving and consumption optimization ratio; Step S4: Dynamic anchoring fusion control When the AI ​​model performance profile meets the stability threshold, it enters the low-proportion fusion control stage: The dynamic anchoring mechanism mixes the dual model outputs in a preset ratio (e.g. 30% AI + 70% mechanism); Real-time monitoring of fusion control effects and assessment of risk-benefit; As the confidence of the AI ​​model increases, the output weight of the AI ​​model is gradually increased; and when external disturbances, sudden changes in model deviations, or abnormal actuator feedback are detected, the output weight of the mechanism main model controller is increased to more than 90%; Step S5: Human-machine collaborative intervention and feedback Display AI model control suggestions and expected benefits to users through the human-machine collaborative decision-making interface; Receive user acceptance or rejection instructions, record intervention frequency, and use this information to evaluate the acceptability of the AI ​​model, assisting with subsequent fine-tuning and strategy updates; Step S6: Abnormal identification and safe switching If any of the following anomalies are detected, the AI ​​model will be immediately isolated and the mechanism master model controller will be switched to take the lead: The AI ​​model output mutation exceeds the preset threshold; The deviation between the control result and the set value continues to expand; The user repeatedly rejected the AI ​​model's control suggestions; Sandbox verification indicators continue to decline.

[0035] The above describes a specific embodiment of the present invention, but those skilled in the art should understand that this is only an example. Those skilled in the art can make various changes or modifications to this embodiment without departing from the principles and essence of the present invention, but these changes and modifications will fall within the scope of protection of the present invention.

Claims

1. A fresh air air conditioning control system based on dual-model safe evolution, characterized by: include: Trusted Data Acquisition Module: This module collects air conditioning system operating data in real time and selects high-quality samples using a trustworthiness evaluation algorithm. The trustworthiness evaluation algorithm's indicators include operating stability, control strategy stability, rationality of key variable fluctuations, and minimization of system interference. Mechanism main model controller: built based on the principles of thermodynamics and fluid mechanics, used to output control instructions when the AI ​​model has not yet been established or is being verified; Shadow AI model training and verification module: runs in parallel with the main mechanism model controller, uses trusted data to train the AI ​​model and compares and verifies the output with the main mechanism model controller in a sandbox environment; Dynamic anchoring control module: used to adaptively adjust the output weight ratio of the mechanism main model controller and the AI ​​model according to the degree of environmental disturbance or system state changes; Human-machine collaborative decision-making interface: used to display AI model suggestions, receive user instructions, and support forced switching of control modes; Fault isolation and model switching module: used to automatically disconnect the AI ​​model control path when abnormal AI model output is detected, and switch to the dominant control of the mechanism main model controller.

2. The fresh air air conditioning control system based on dual-model safe evolution according to claim 1 is characterized in that: The indicators of the credibility evaluation algorithm specifically include: Operating stability: The air flow, temperature, and humidity parameters fluctuate within the set period without exceeding the threshold; Control strategy stability: no control strategy switching occurs during the sampling period, and the system does not perform parameter tuning or model updating; Reasonable fluctuation of key variables: The combustion section outlet temperature, spray section pressure, and surface temperature of the surface cooler are within the preset normal range, without sudden changes or fault-type outliers; Minimize system interference: No equipment maintenance, parameter debugging or strong external disturbance records are recorded during the sample collection period.

3. The fresh air air conditioning control system based on dual-model safety evolution according to claim 1 is characterized in that: The operating logic of the shadow AI model training and verification module includes: The AI ​​model output is only used to compare with the results of the mechanism main model controller and does not directly participate in equipment control; When the prediction deviation, response delay, and energy consumption optimization ratio of the AI ​​model in the sandbox meet the preset indicators, its output is gradually introduced into the control decision.

4. The fresh air air conditioning control system based on dual-model safe evolution according to claim 1 is characterized in that: The weight adjustment strategy of the dynamic anchor control module is: When the environment is stable, increase the output weight of the AI ​​model; When external disturbances, sudden model deviations, or abnormal actuator feedback are detected, the output weight of the mechanism main model controller is increased to more than 90%; The weight is dynamically adjusted according to the system response feedback to achieve a smooth transition between "mechanism main control-AI assistance" and "AI main control-mechanism guarantee".

5. The fresh air air conditioning control system based on dual-model safety evolution according to claim 1 is characterized in that: The human-machine collaborative decision-making interface supports: Displays the energy saving ratio and comfort improvement ratio recommended by the AI ​​model; Provide users with action options: accept, observe, or ignore AI suggestions; Provide AI suggestion history, deviation trend and original data comparison chart; A high-privilege user can forcibly switch the control mode.

6. The fresh air air conditioning control system based on dual-model safety evolution according to claim 1 is characterized in that: The triggering conditions of the fault isolation and model switching module include any of the following: The AI ​​model output mutation exceeds the preset threshold; The deviation between the control result and the set value continues to expand; The user repeatedly rejected the AI ​​model's control suggestions; Sandbox verification indicators continue to decline.

7. The fresh air air conditioning control method based on dual-model safe evolution according to claim 1, applied to the system according to any one of claims 1 to 6, characterized in that: Including steps: Step S1: Trusted data collection and filtering Collect air conditioning system operation data in real time and select high-quality samples through credibility evaluation algorithm; Step S2: Initial control of the mechanism main model controller When the AI ​​model has not yet been established or verified, the main model controller based on the principles of thermodynamics and fluid mechanics outputs control instructions; Step S3: Shadow AI model training and sandbox evaluation Use trusted data to train AI models and run them in parallel with the main mechanism model controller in a sandbox environment; Compare the outputs of the AI ​​model and the mechanism master model controller to generate a performance profile that includes prediction deviation, response delay, and energy-saving and consumption optimization ratio; Step S4: Dynamic anchoring fusion control When the AI ​​model performance profile meets the stability threshold, it enters the low-proportion fusion control stage: The dynamic anchoring mechanism mixes the dual model outputs in a preset ratio; Real-time monitoring of fusion control effects and assessment of risk-benefit; As the confidence of the AI ​​model increases, the output weight of the AI ​​model is gradually increased; and when external disturbances, sudden changes in model deviations, or abnormal actuator feedback are detected, the output weight of the mechanism main model controller is increased to more than 90%; Step S5: Human-machine collaborative intervention and feedback Display AI model control suggestions and expected benefits to users through the human-machine collaborative decision-making interface; Receive user acceptance or rejection instructions, record intervention frequency, and use this information to evaluate the acceptability of the AI ​​model, assisting with subsequent fine-tuning and strategy updates; Step S6: Abnormal identification and safe switching If any of the following anomalies are detected, the AI ​​model will be isolated immediately and the control will be switched to the main model controller: The AI ​​model output mutation exceeds the preset threshold; The deviation between the control result and the set value continues to expand; The user repeatedly rejected the AI ​​model's control suggestions; Sandbox verification indicators continue to decline.

Images