Annealing kiln roller cleaning system and method based on AI visual airflow

Abstract

The invention discloses an annealing kiln roller cleaning system and cleaning method based on AI visual airflow. The system comprises an airflow monitoring and adjusting module, a roller cleanliness monitoring and cleaning module and a central AI processing and control unit which are connected with one another; the airflow monitoring and adjusting module realizes visual and quantifiable monitoring and regulation of an airflow field in the annealing kiln, and the roller cleanliness monitoring and cleaning module can be combined with ultrasonic vibration stripping and negative pressure collection by monitoring the surface cleanliness of a roller on line; the central AI processing and control unit can integrate multi-source data for intelligent analysis, decision making and instruction output, and continuous optimization of the model is completed through data feedback. By the adoption of the system or the cleaning method, online cleaning of the roller of the float glass annealing kiln is achieved, the situation that the cleaning effect and the glass quality are affected due to the fact that airflow in the annealing kiln is uncontrollable can be avoided, shutdown is not needed, the cleaning efficiency is high, and secondary pollution does not exist.

Description

Technical Field

[0001] This invention relates to the field of float glass production equipment technology, specifically to an annealing kiln roller cleaning system and cleaning method based on AI vision airflow. Background Technology

[0002] The annealing process for float glass is a crucial step in eliminating residual stress and ensuring the final product quality. The uniformity of the temperature field within the annealing furnace and the surface cleanliness of the transfer rollers directly affect the glass forming quality and production efficiency. In current production practices, two major technical problems and their interrelationship exist:

[0003] On the one hand, the airflow state inside the annealing furnace is in a black box mode, and the airflow direction and velocity are mainly coarsely adjusted by fixed air ducts and baffles, making it impossible to achieve real-time and more precise local control. Due to the thermal pressure difference, irregular lateral airflow or vortices are easily generated inside the furnace, resulting in uneven lateral temperature of the glass plates, which can easily lead to quality defects such as stress spots, warping, and optical distortion. Moreover, existing factories lack effective online monitoring methods to directly observe and quantify airflow behavior.

[0004] On the other hand, the conveyor rollers operate in high-temperature environments for extended periods, easily accumulating volatile substances such as alkali metal oxides and sulfides from detached glass sheets, forming a layer of dirt. This dirt can cause scratches on the glass sheets, uneven heat transfer, and surface vibration, severely impacting product quality. Currently, roller cleaning requires manual operation after production has stopped and the rollers have cooled, which is not only inefficient and disruptive to production continuity but also poses safety risks. Some attempts have been made to use intermittent online ultrasonic slope cleaning for dirt removal; however, performing ultrasonic slope cleaning in the unstable airflow and uneven temperature environment inside the annealing furnace still increases the risk to glass quality. Furthermore, there is a lack of online monitoring methods, making preventative intervention impossible before the dirt causes harm.

[0005] Therefore, providing an online intelligent cleaning system and method for float glass annealing furnace rollers based on AI visual airflow and ultrasonic waves has become a technical problem to be solved in this field. Summary of the Invention

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] An annealing kiln roller cleaning system based on AI visual airflow includes an airflow monitoring and adjustment module, a roller cleanliness monitoring and cleaning module, and a central AI processing and control unit. The central AI processing and control unit is connected to the airflow monitoring and adjustment module and the roller cleanliness monitoring and cleaning module, respectively.

[0008] The airflow monitoring and regulation module is used to collect airflow schlieren images and fixed-point wind speed data inside the annealing kiln, generate a dynamic airflow vector diagram, and perform precise local regulation of the airflow field inside the kiln based on the airflow vector diagram. This module includes a high-definition high-temperature resistant industrial endoscope array, a miniature high-temperature resistant wind speed and direction sensor, and a porous intelligent air duct and a miniature electric regulating valve array. The high-definition high-temperature resistant industrial endoscope array is distributed on the side walls of each key temperature zone of the annealing kiln, equipped with a wide-angle lens and a high-temperature resistant protection kit, which can stably capture airflow schlieren images in high-temperature and dusty environments. The miniature high-temperature resistant wind speed and direction sensor is sparsely arranged at key points to provide benchmark calibration data for the AI ​​vision model. The porous intelligent air duct and the miniature electric regulating valve array are respectively arranged in the space above and below the glass plate. Each miniature air outlet on the air duct corresponds to an independent miniature electric regulating valve, which can realize pixel-level adjustment of the airflow direction and airflow.

[0009] The roller cleanliness monitoring and cleaning module is used to monitor the surface cleanliness of the transmission roller in real time. When the cleanliness index exceeds a preset threshold, it triggers an ultrasonic vibration cleaning action and collects dirt simultaneously. The module includes a laser profile scanner, an ultrasonic cleaning actuator, and a waste collection device. The laser profile scanner is installed at the end or lower side of the roller and scans the outer contour of the roller at high speed. It quantitatively assesses the dirt accumulation state by analyzing the roundness error and surface roughness. The ultrasonic cleaning actuator is integrated into the roller bearing seat and generates high-frequency mechanical vibration, which is transmitted to the roller body through a waveguide rod to achieve non-contact dirt removal. The waste collection device uses a negative pressure suction nozzle set directly below the roller to promptly remove the shaken-off dirt and prevent secondary pollution.

[0010] The central AI processing and control unit, as the core of the system, communicates with the two modules mentioned above, receiving and analyzing multi-source data to generate wind field control commands and roller cleaning commands. It incorporates a convolutional neural network model based on the U-Net architecture for airflow schlieren image analysis and airflow vector diagram generation. It also integrates a multi-objective optimization controller, employing a control strategy combining fuzzy PID and local feedback to ensure the accuracy and coordination of wind field control. Simultaneously, the central AI processing and control unit stores monitoring data, control commands, and product quality inspection results, and periodically performs offline reinforcement learning training on the AI ​​model to continuously optimize the control strategy.

[0011] The above system is also an intelligent monitoring system for airflow in glass annealing furnaces.

[0012] An AI-based intelligent cleaning method for annealing kiln rollers based on the above system includes the following steps:

[0013] S1: Real-time acquisition of multi-source data. High-definition high-temperature resistant industrial endoscope array captures airflow schlieren images, laser contour scanner acquires roller outer circle contour data, and miniature high-temperature resistant wind speed and direction sensor acquires fixed-point wind speed and direction data.

[0014] S2: AI Data Processing and Analysis. The central AI processing and control unit analyzes airflow schlieren images through a convolutional neural network model, combines fixed-point wind speed data to generate dynamic airflow vector maps, and identifies abnormal airflow areas; at the same time, it analyzes roller contour data and calculates the real-time cleanliness index based on the cleanliness index model.

[0015] If there is abnormal airflow, the central AI processing and control unit drives the micro electric regulating valve to balance the air field; if the cleanliness index exceeds the standard, and the airflow vector diagram shows normal airflow, the ultrasonic cleaning actuator of the corresponding roller is triggered to run at the set power and duration, and the negative pressure suction nozzle is started to collect dirt simultaneously.

[0016] S3: Associates and stores relevant data, and regularly uses the data to train and optimize the AI ​​model to improve system performance.

[0017] The loss function of the convolutional neural network model in step S2 is a composite function of cross-entropy and Dice coefficients, expressed as: Loss = -Σ(y_true * log (y_pred)) - λ * (2 * |y_true ∩ y_pred| + ε) / (|y_true| + |y_pred| + ε), where y_true is the true label, y_pred is the predicted value, λ is the balancing weight, and ε is the smoothing term; the airflow vector map covers the entire width of the glass plate and is used to reflect the airflow direction and speed of each pixel.

[0018] The expression for the cleanliness index model in step S2 is: CI = α * (ΔR_max / ΔR_threshold) + β * (Ra_current / Ra_initial), where α and β are weighting coefficients and α+β=1, ΔR_max is the current maximum roundness error, ΔR_threshold is the roundness alarm threshold, Ra_current is the current surface roughness, and Ra_initial is the initial roughness of the clean roller.

[0019] The control strategy expression combining fuzzy PID and local feedback in step S2 is: ΔU_i =K_p * e_i + K_i * Σe_i * Δt + K_d * (e_i - e_{i-1}) / Δt + η * Σ_{j∈N (i)}(e_j - e_i), where i is the valve index, K_p, K_i, and K_d are PID parameters, η is the coupling coefficient between adjacent regions, N(i) is the set of adjacent valves of valve i, and e_i is the deviation between the airflow vector matrix and the ideal model.

[0020] In step S2, the output power of the ultrasonic cleaning actuator is adjustable from 50W to 500W. The cleaning process is carried out under normal operating conditions of the annealing furnace and does not require shutdown for cooling.

[0021] The above method is also a smart monitoring method for airflow in glass annealing furnaces.

[0022] Compared with the prior art, the beneficial effects of the present invention are:

[0023] This paper presents an online intelligent cleaning system and method for float glass annealing furnace rollers based on AI-powered visual airflow and ultrasonic waves. It deeply integrates AI vision with schlieren technology, capturing airflow schlieren images using a high-definition, high-temperature resistant industrial endoscope array. A dynamic airflow vector map is then generated using a convolutional neural network model, making the previously invisible airflow visible and quantifiable. Combined with a porous intelligent air duct and a micro-electric regulating valve array, pixel-level local control of the airflow field is achieved, effectively eliminating abnormal airflow and providing a stable environment for roller cleaning. This also solves quality problems such as glass stress spots and optical distortion caused by uneven airflow.

[0024] The surface cleanliness of the rollers is monitored in real time by a laser profile scanner. Based on the cleanliness index model, the cleaning timing is scientifically determined. Preventive cleaning can be triggered before dirt accumulates to the point of affecting product quality, without the need for machine shutdown and cooling. At the same time, ultrasonic vibration is used for non-contact dirt removal, avoiding damage to the roller surface. Simultaneously, dirt is collected by a negative pressure suction nozzle, effectively achieving online preventive cleaning of the rollers without secondary pollution.

[0025] Its central AI processing and control unit integrates airflow monitoring and roller cleaning monitoring data to achieve integrated collaborative control. It can also regularly conduct offline reinforcement learning training on the AI ​​model by associating and storing monitoring data, control commands and product quality inspection results, so as to continuously optimize the control strategy and ensure production efficiency and product quality stability. Attached Figure Description

[0026] Figure 1 is a block diagram illustrating the working principle of the system of the present invention;

[0027] Figure 2 is a schematic diagram showing the installation position of the components of the roller cleanliness monitoring and cleaning module in the annealing kiln and the connection of the control unit in an embodiment of the present invention.

[0028] Figure 3 is a flowchart of the annealing kiln roller cleaning method based on AI visual airflow in an embodiment of the present invention.

[0029] Icons: 2 - Transfer roller, 3 - Underboard intelligent air duct, 4 - Onboard intelligent air duct, 5 - Miniature electric regulating valve array, 6 - High-temperature resistant industrial endoscope, 7 - Laser profile scanner, 8 - Ultrasonic cleaning actuator, 9 - Negative pressure nozzle, 10 - Central AI processing and control unit. Detailed Implementation

[0030] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0031] As shown in Figures 1-3, an annealing kiln roller cleaning system based on AI visual airflow includes an airflow monitoring and adjustment module, a roller cleanliness monitoring and cleaning module, and a central AI processing and control unit 10.

[0032] The airflow monitoring and regulation module is used to collect airflow schlieren images and fixed-point wind speed data in the annealing kiln, generate a dynamic airflow vector diagram, and perform precise local control of the airflow field in the kiln based on the airflow vector diagram. This module includes a high-definition high-temperature resistant industrial endoscope array, a miniature high-temperature resistant wind speed and direction sensor, and a multi-hole intelligent air duct and miniature electric regulating valve array. In the airflow monitoring and regulation module, the high-definition high-temperature resistant industrial endoscope array 6 is arranged at 2.5-meter intervals in zones A, B1, and B2 of the annealing kiln. Each group includes one endoscope on the upper plate and one on the lower plate. The lens has a temperature resistance of ≥650℃, is equipped with a vortex tube cooling sleeve, has a viewing angle of ≥120°, and a resolution of 1920×1080 @ 60fps, and is used to capture airflow schlieren images in the space above and below the plate. The miniature high-temperature resistant wind speed and direction sensor is sparsely arranged at key points such as the heat spreader and important cooling zones to collect fixed-point wind speed and direction data. The multi-hole intelligent air ducts 3 and 4 are made of 310S stainless steel, with φ8mm openings in the duct wall. The miniature air outlets have a center-to-center distance of 50mm. Each air outlet corresponds to a miniature electric regulating valve 5 with a diameter of DN10. The valve stroke time is ≤1.5 seconds, and it can withstand an ambient temperature of ≥300℃. It can achieve precise adjustment of air volume and air outlet direction.

[0033] The roller cleanliness monitoring and cleaning module is used to monitor the surface cleanliness of the transmission roller in real time. When the cleanliness index exceeds a preset threshold, it triggers an ultrasonic vibration cleaning action and collects dirt simultaneously. This module includes a laser profile scanner, an ultrasonic cleaning actuator, and a waste collection device. In the roller cleanliness monitoring and cleaning module, the laser profile scanner 7 is installed 1.5 meters at the end of the roller, with a sampling frequency of 1kHz and a linear accuracy of ±3μm, and is used to measure the roller roundness error and surface roughness Ra. The ultrasonic cleaning actuator 8 is integrated into the bearing seat of the roller 2, with a working frequency of 28kHz±2kHz and an output power adjustable from 50W to 500W. It generates vibration through magnetostrictive material and transmits it to the roller 2 through a waveguide rod. The negative pressure suction nozzle 9 of the waste collection device is located directly below the roller 2 and is connected to the dust removal system to promptly remove the shaken-off dust-like dirt.

[0034] The central AI processing and control unit, as the core of the system, communicates with the two modules mentioned above, receives and analyzes multi-source data, and generates wind field control commands and roller cleaning commands. The central AI processing and control unit 10 incorporates a convolutional neural network model based on the U-Net architecture and a multi-objective optimization controller. It receives data from the endoscope 6, the laser profile scanner 7, and the wind speed sensor. It analyzes the airflow schlieren image through the convolutional neural network model and combines it with fixed-point wind speed data to generate a dynamic airflow vector diagram. By analyzing the laser profile scanning data, it calculates the real-time cleanliness index based on the cleanliness index model CI = α * (ΔR_max / ΔR_threshold) + β * (Ra_current / Ra_initial) (where α = 0.6, β = 0.4, ΔR_threshold = 50μm). The multi-objective optimization controller, based on the deviation between the airflow vector diagram and the ideal model, uses the control law ΔU_i = K_p * e_i + K_i * Σe_i * Δt + K_d * (e_i -e_{i-1}) / Δt + η. * Σ_{j∈N (i)} (e_j - e_i) Calculate the opening adjustment of each valve and drive the micro electric regulating valve 5 to operate; when the cleanliness index exceeds the preset threshold and the airflow vector diagram shows normal airflow, send a command to the ultrasonic cleaning actuator 8 to control its power and running time, and simultaneously start the negative pressure suction nozzle 9.

[0035] The aforementioned system is also an AI vision-based monitoring or control system for the airflow in an annealing kiln.

[0036] A method for cleaning annealing kiln rollers based on the above system and AI-based visual airflow includes the following steps:

[0037] S1: High-definition high-temperature resistant industrial endoscope array 6 continuously captures the airflow pattern video stream inside the annealing kiln; laser contour scanner 7 acquires the outer circle contour data of roller 2 at a frequency of 1kHz; miniature high-temperature resistant wind speed and direction sensor transmits fixed-point wind speed and direction data in real time.

[0038] S2: The central AI processing and control unit 10 performs real-time analysis of the airflow schlieren video stream using a convolutional neural network model, separates the airflow schlieren effect, calculates the optical flow vector of each pixel, generates a dynamic airflow vector map covering the entire width of the glass plate, compares it with the ideal airflow model, and identifies abnormal airflow areas; at the same time, it processes the roller profile data, calculates the roundness error ΔR_max and surface roughness Ra_current, and substitutes them into the cleanliness index model to obtain the real-time cleanliness index;

[0039] If abnormal airflow is detected, the central AI processing and control unit 10 calculates the optimal opening degree of each micro electric regulating valve 5, drives the valve to act, and adjusts the local air field until the deviation between the airflow vector diagram and the ideal model is within the allowable range. If the cleanliness index exceeds the preset threshold and the airflow vector diagram shows normal airflow, a command is sent to the ultrasonic cleaning actuator 8 of the corresponding roller 2 to set the output power and running time, start ultrasonic vibration cleaning, and simultaneously start the negative pressure suction nozzle 9 to pump the shaken dirt to the dust removal system to avoid secondary pollution.

[0040] S3: The system associates and stores airflow monitoring data, valve adjustment commands, roller cleaning data, cleaning commands, and subsequent product quality inspection results (such as stress meter data). It periodically uses this data to perform offline reinforcement learning training on the convolutional neural network model and the multi-objective optimization controller to optimize model parameters and control strategies.

[0041] This paper presents an online intelligent cleaning system and method for float glass annealing furnace rollers based on AI-powered visual airflow and ultrasonic waves. It deeply integrates AI vision with schlieren technology, capturing airflow schlieren images using a high-definition, high-temperature resistant industrial endoscope array. A dynamic airflow vector map is then generated using a convolutional neural network model, making the previously invisible airflow visible and quantifiable. Combined with a porous intelligent air duct and a micro-electric regulating valve array, pixel-level local control of the airflow field is achieved, effectively eliminating abnormal airflow and providing a stable environment for roller cleaning. This also solves quality problems such as glass stress spots and optical distortion caused by uneven airflow.

[0042] The surface cleanliness of the rollers is monitored in real time by a laser profile scanner. Based on the cleanliness index model, the cleaning timing is scientifically determined. Preventive cleaning can be triggered before dirt accumulates to the point of affecting product quality, without the need for machine shutdown and cooling. At the same time, ultrasonic vibration is used for non-contact dirt removal, avoiding damage to the roller surface. Simultaneously, dirt is collected by a negative pressure suction nozzle, effectively achieving online preventive cleaning of the rollers without secondary pollution.

[0043] Its central AI processing and control unit integrates airflow monitoring and roller cleaning monitoring data to achieve integrated collaborative control. It can also regularly conduct offline reinforcement learning training on the AI ​​model by associating and storing monitoring data, control commands and product quality inspection results, so as to continuously optimize the control strategy and ensure production efficiency and product quality stability.

[0044] The above method is also a monitoring or control method for the airflow in an annealing furnace based on AI vision.

[0045] The above system and method can also be applied to all glass production fields that use roller conveyor annealing furnaces, such as ordinary float glass, ultra-thin electronic glass, high-alumina glass, lithium aluminum silicon glass, high borosilicate glass, pharmaceutical glass and microcrystalline glass, and solar rolled glass.

[0046] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An annealing kiln roller cleaning system based on AI visual airflow, characterized in that, It includes an airflow monitoring and adjustment module, a roller cleanliness monitoring and cleaning module, and a central AI processing and control unit, which are respectively connected to the airflow monitoring and adjustment module and the roller cleanliness monitoring and cleaning module; The airflow monitoring and regulation module is used to collect airflow schlieren images and fixed-point wind speed data in the annealing kiln, generate dynamic airflow vector diagrams, and perform local regulation of the airflow field in the kiln based on the airflow vector diagrams. The roller cleanliness monitoring and cleaning module is used to monitor the surface cleanliness of the annealing kiln conveyor rollers in real time. When the cleanliness index exceeds the preset threshold, it triggers an ultrasonic vibration cleaning action and collects the dirt generated during cleaning simultaneously. The central AI processing and control unit is communicatively connected to the airflow monitoring and adjustment module and the roller cleanliness monitoring and cleaning module, respectively. It is used to receive and analyze the airflow schlieren image, fixed-point wind speed data and roller surface state data, generate wind field control commands and roller cleaning commands, and optimize the analysis model based on historical data and product quality feedback.

2. The annealing kiln roller cleaning system based on AI visual airflow according to claim 1, characterized in that, The airflow monitoring and regulation module includes: High-definition, high-temperature resistant industrial endoscope arrays are distributed on the side walls of each temperature zone of the annealing furnace, equipped with wide-angle lenses and high-temperature resistant protective kits; Miniature high-temperature resistant wind speed and direction sensor, deployed at the annealing kiln location; A multi-hole intelligent air duct and a micro electric regulating valve array are respectively arranged in the space above and below the glass plate. The intelligent air duct has multiple micro air outlets, and each air outlet or adjacent small area corresponds to an independently controlled micro electric regulating valve.

3. The annealing kiln roller cleaning system based on AI visual airflow according to claim 2, characterized in that, The high-definition, high-temperature resistant industrial endoscope array has lenses with a temperature resistance of ≥650℃, is equipped with a vortex tube cooling sleeve, has a viewing angle of ≥120°, and a resolution of 1920×1080 @ 60fps. It is used in annealing furnace zones A, B1, and B2. The endoscopes are arranged at 2.5-meter intervals, with each group containing one endoscope on the upper and one on the lower part of the plate; the valve body of the miniature electric regulating valve has a diameter of DN10, a stroke time of ≤1.5 seconds, and can withstand an ambient temperature of ≥300℃; the intelligent air duct is made of 310S stainless steel, with a φ8mm miniature air outlet on the duct wall and a center distance of 50mm between the air outlets.

4. The annealing kiln roller cleaning system based on AI visual airflow according to claim 1, characterized in that, The roller cleanliness monitoring and cleaning module includes: A laser profile scanner is installed 1.5 meters below or to the side of the roller and scans the outer profile of the roller at a frequency of 1 kHz. The ultrasonic cleaning actuator is integrated into the inside or outside contact point of the roller bearing housing, generating high-frequency mechanical vibration of 28kHz±2kHz. The waste collection device includes a negative pressure suction nozzle located directly below the roller, the negative pressure suction nozzle being connected to a dust removal system.

5. The annealing kiln roller cleaning system based on AI visual airflow according to claim 1, characterized in that, The central AI processing and control unit incorporates a convolutional neural network model based on the U-Net architecture and a multi-objective optimization controller. The convolutional neural network model is used to segment airflow schlieren images and calculate optical flow, outputting an airflow vector matrix. The multi-objective optimization controller adopts a control strategy combining fuzzy PID and local feedback to generate the opening adjustment amount of each micro electric regulating valve and the working parameters of the ultrasonic cleaning actuator.

6. A method for intelligent cleaning of annealing kiln rollers based on AI visual airflow, using the system described in any one of claims 1-5, characterized in that, Includes the following steps: S1: High-definition high-temperature resistant industrial endoscope array captures airflow schlieren images inside the annealing furnace, laser contour scanner acquires roller outer circle contour data, and miniature high-temperature resistant wind speed and direction sensor collects fixed-point wind speed and direction data. S2: The central AI processing and control unit analyzes the airflow schlieren image through a convolutional neural network model, combines it with fixed-point wind speed data to generate a dynamic airflow vector image, and identifies abnormal airflow areas; at the same time, it analyzes the roller profile data, calculates the roundness error and surface roughness, and derives the real-time cleanliness index based on the cleanliness index model. If the airflow vector diagram shows abnormal airflow, the central AI processing and control unit calculates the optimal opening of each micro electric regulating valve and drives the valve to balance the airflow field; if the cleanliness index exceeds the preset threshold, the ultrasonic cleaning actuator of the corresponding roller is triggered when the airflow vector diagram shows normal airflow, and runs at the set power and duration, while the negative pressure suction nozzle is started to collect dirt. S3: Associate and store monitoring data, control commands, and product quality inspection results, and periodically use the data to perform offline reinforcement learning training on the AI ​​model to optimize the control strategy.

7. The intelligent cleaning method for annealing kiln rollers based on AI visual airflow according to claim 6, characterized in that, The loss function of the convolutional neural network model in step S2 is a composite function of cross-entropy and Dice coefficients, expressed as: Loss = -Σ(y_true * log (y_pred)) - λ * (2 * |y_true ∩ y_pred| + ε) / (|y_true| + |y_pred| + ε), where y_true is the true label, y_pred is the predicted value, λ is the balancing weight, and ε is the smoothing term; the airflow vector map covers the entire width of the glass plate and is used to reflect the airflow direction and speed of each pixel.

8. The intelligent cleaning method for annealing kiln rollers based on AI visual airflow according to claim 6, characterized in that, The expression for the cleanliness index model in step S2 is: CI = α * (ΔR_max / ΔR_threshold) + β * (Ra_current / Ra_initial), where α and β are weighting coefficients and α+β=1, ΔR_max is the current maximum roundness error, ΔR_threshold is the roundness alarm threshold, Ra_current is the current surface roughness, and Ra_initial is the initial roughness of the clean roller.

9. The intelligent cleaning method for annealing kiln rollers based on AI visual airflow according to claim 6, characterized in that, The control strategy expression combining fuzzy PID and local feedback in step S2 is: ΔU_i = K_p* e_i + K_i * Σe_i * Δt + K_d * (e_i - e_{i-1}) / Δt + η * Σ_{j∈N (i)} (e_j - e_i), where i is the valve index, K_p, K_i, and K_d are PID parameters, η is the coupling coefficient between adjacent regions, N (i) is the set of adjacent valves of valve i, and e_i is the deviation between the airflow vector matrix and the ideal model.

10. The annealing kiln roller cleaning method based on AI visual airflow according to claim 6, characterized in that, In step S2, the output power of the ultrasonic cleaning actuator is adjustable from 50W to 500W. The cleaning process is carried out under normal operating conditions of the annealing furnace and does not require shutdown for cooling.

Images