Method for detecting surface defects of steel strip by fusing steam parameter adjustment and image recognition

Abstract

The present application relates to the technical field of surface defect detection, and discloses a steel strip surface defect detection method combining steam parameter adjustment and image recognition, which acquires steel strip surface state parameters in real time to calculate initial steam parameters for configuration; uses a water-cooled roller to pre-cool the steel strip in sections and sprays laminar wet steam to form a condensation development pattern; acquires a development image, extracts an abnormal image slice and a standard time domain feature sequence, inputs a target detection model introducing a conditional batch normalization mechanism, synchronously outputs a defect result, a global condensation quality and a local condensation uniformity index, and accordingly constructs a long and short cycle decoupling dual-mode control, adjusts gas phase parameters through fast cycle adjustment, adjusts cooling water flow through slow cycle adjustment, and realizes parameter closed loop updating. The present application amplifies low contrast defect features from a physical layer through physical development means, combines with a multi-modal target detection algorithm, avoids the limitation that a pure visual method only relies on static weak gray difference, and reduces the missed detection rate of low contrast defects.

Description

Technical Field

[0001] This invention relates to the field of surface defect detection technology, specifically to a method and system for detecting surface defects in steel strips that integrates steam parameter adjustment and image recognition. Background Technology

[0002] Surface defect detection is a crucial aspect of quality control in the steel and metallurgical industry, especially for high-value-added products such as cold-rolled tin-plated steel strip, galvanized sheet, and stainless steel strip, where surface quality directly determines product grade and market price. Traditional surface defect detection methods mainly include manual visual inspection, eddy current testing, magnetic flux leakage testing, and machine vision-based optical inspection. Among these, automatic inspection methods based on optical imaging have become the mainstream configuration in industrial production lines due to their advantages such as non-contact operation, high speed, and online operation. In recent years, with the development of deep learning technology, target detection algorithms based on convolutional neural networks (such as the YOLO series) have been widely applied in the field of steel strip surface defect detection. Through training on a large number of labeled defect images, they can automatically identify and locate common defects such as scratches, dents, roll marks, and oxide scale.

[0003] Currently, there are two main technical approaches for online detection of surface defects in steel strips. One approach is based on pure vision-based deep learning, which uses high-resolution linear or area array cameras to acquire images of the steel strip surface and then employs improved target detection models such as YOLO and SSD for defect identification and classification. The other approach is based on physical development using vapor condensation. This method utilizes the differences in condensation rates and adhesion states of water vapor in different surface regions (defective and non-defective areas) to create condensation patterns in defective areas that differ from normal areas, thus magnifying low-contrast defects into visually identifiable features.

[0004] However, existing surface defect detection methods are still insufficient for the high-speed, heavy-load, and high-temperature complex conditions of continuous annealing production lines for cold-rolled tin-plated steel strips, and the following technical problems remain: First, pure visual deep learning methods have limited ability to detect low-contrast defects, such as minor scratches, roller marks, and oil residues. This is because the grayscale difference between the defect and the background is extremely small, making it difficult to reliably extract defect features from the original image even with complex neural network structures, resulting in a high false negative rate. Second, while steam condensation development can enhance defect contrast at a physical level, existing solutions heavily rely on manual experience to set steam parameters, failing to adaptively adjust according to the dynamic changes in the steel strip surface condition. This leads to inconsistent condensation effects and makes online closed-loop control difficult. Therefore, a steel strip surface defect detection method integrating steam parameter adjustment and image recognition is urgently needed to solve these problems. Summary of the Invention

[0005] To address the problems in related technologies, this invention provides a method for detecting surface defects in steel strips that integrates steam parameter adjustment and image recognition, thereby overcoming the aforementioned technical problems in existing related technologies.

[0006] To solve the aforementioned technical problem, the present invention is achieved through the following technical solution: In a first aspect, embodiments of the present invention provide a method for detecting surface defects in steel strips that integrates steam parameter adjustment and image recognition, specifically including: Step S100: Real-time acquisition of surface state parameters of the steel strip, and calculation of initial steam parameter configuration vector based on the surface state parameters; Step S200: A contact water-cooled roller with a surface treated by femtosecond laser micro-nano texturing is used to pre-cool the steel strip in sections according to the pre-cooling temperature difference in the initial steam parameter configuration, and the steam generator is controlled to spray wet steam onto the surface of the steel strip to form a condensation development pattern. Step S300: The condensation development pattern is acquired across the entire frame by the linear array of cameras in the main detection channel. An online dynamically calibrated threshold multiplier is introduced to perform adaptive threshold segmentation to extract abnormal image slices. At the same time, the sub-pixel edge coordinates acquired by the linear array are reused by the area array of cameras in the auxiliary dynamic channel for deviation compensation. Temporal local images under fixed spatial nodes are acquired and output as standard time domain feature sequences after being mapped by the Gaussian kernel temporal resampling interpolation algorithm. Step S400: Input the abnormal image slices and the standard time domain feature sequence into the conditional batch normalization mechanism and the target detection model with trainable edge enhancement convolution to perform defect detection. Train the multi-task loss function based on uncertainty dynamic weight allocation and output the defect confidence, global condensation quality score and local condensation uniformity index simultaneously. Step S500: Construct a long- and short-cycle decoupled dual-mode control based on the global condensation quality score and the local condensation uniformity index. The fast-cycle loop adjusts the low-inertia steam parameters, while the slow-cycle loop adjusts the zoned cooling water flow of the water-cooled roller based on the steady-state decoupling matrix and the incremental multivariable proportional-integral-derivative control algorithm. The initial steam parameter configuration vector is updated in a closed loop.

[0007] As a preferred embodiment of the steel strip surface defect detection method integrating steam parameter adjustment and image recognition described in this invention, step S100 includes: Real-time reading of current production status parameters and apparent temperature of steel strip collected by infrared thermometer; A heat transfer compensation formula is constructed. By extracting a comprehensive thermal resistance compensation coefficient that includes thermal diffusivity and contact thermal resistance, and combining the real-time production thickness of the steel strip, the real-time running line speed, and the temperature difference between the actual measured temperature and the set temperature of the internal circulating cooling water, the apparent temperature is corrected to the true surface temperature of the steam sprayed surface. The initial steam parameter configuration vector is constructed by steam temperature, steam dryness, steam flow rate per unit width, precooling temperature difference and nozzle distance. The deviation is adjusted by combining the pre-calibrated basic parameter vector with the real-time operating condition deviation vector and the deviation compensation gain matrix. The initial steam parameter configuration vector is obtained by boundary truncation through the amplitude limiting function. The deviation compensation gain matrix is ​​based on the system identification method of open-loop step response experiment to avoid closed-loop data cyclic dependence.

[0008] As a preferred embodiment of the steel strip surface defect detection method integrating steam parameter adjustment and image recognition described in this invention, the step of performing zoned precooling of the steel strip based on the precooling temperature difference in the initial steam parameter configuration includes: Based on the steel strip's density, specific heat capacity, thickness, width, linear velocity, and actual surface temperature, as well as the target pre-cooling temperature, the physical properties of the cooling water, and the heat exchange efficiency coefficient of the water-cooled roller system calibrated by inversely solving the heat balance equation in the open-loop state, the basic cooling water volume flow rate required to reach the target pre-cooling temperature is calculated. A control dead zone is set at the steel strip temperature at the water-cooled roll outlet. When the temperature deviates from the target pre-cooling temperature and exhibits a steady-state drift trend exceeding the control dead zone, a discrete position proportional-integral control algorithm is used to calculate and superimpose the corrected flow rate to the basic cooling water volume flow rate. The proportional and integral gain coefficients of the control algorithm are pre-tuned using an open-loop single-step response test curve. The transient temperature deviation remaining in the control dead zone after the water-cooled roller precooling is used as a feedforward disturbance and input into the fast circulation loop. The steam generator then performs adaptive compensation by fine-tuning the steam dryness and flow rate.

[0009] As a preferred embodiment of the steel strip surface defect detection method integrating steam parameter adjustment and image recognition described in this invention, step S300, which extracts abnormal image slices and outputs a standard time-domain feature sequence, includes: Adaptive threshold segmentation is performed line by line on the full-frame image. Offline benchmark calibration is combined with real-time global grayscale standard deviation online calibration to generate a dynamic threshold multiplier. Abnormal pixel intervals are determined based on the dynamic threshold multiplier and local sliding window statistics. Independent connected components are extracted to generate abnormal image slices. Extract image data synchronously acquired by the line scan camera array, calculate sub-pixel level edge coordinates based on the first-order gradient of grayscale and parabolic fitting algorithm, and calculate the actual lateral physical coordinates in combination with the physical spatial resolution of the line scan camera. Based on this, perform lateral deviation dynamic compensation and pixel coordinate repositioning and cropping on the field of view target area of ​​the auxiliary dynamic channel, and output local images in the time series. The actual physical time of the target area reaching each area array camera is calculated based on the pulse discrete integral solution of the main encoder. The high-dimensional spatial features of the local image are extracted to form the actual time domain feature matrix. The Gaussian kernel time domain resampling interpolation algorithm is used to linearly map it into the standard time domain feature sequence in real time.

[0010] As a preferred embodiment of the steel strip surface defect detection method integrating steam parameter adjustment and image recognition described in this invention, the construction and evaluation of the target detection model in step S400 includes: A parallel dual-branch feature extraction module is designed at the input end. The first branch extracts the global low-frequency features of water film uniformity, and the second branch is cascaded with trainable edge-enhancing convolution to extract the high-frequency features of droplet and defect edges. The trainable edge-enhancing convolution is set as a discrete difference operator matrix during the initialization phase, and its weights are dynamically updated during the training phase based on the gradient of the loss function through backpropagation. The initial steam parameter configuration vector is dimensionlessly standardized and used as a conditional input to the conditional batch normalization module of the backbone network. The scaling factor and offset of the normalization layer are dynamically generated through learnable projection weights. A parallel condensation quality assessment sub-network is added next to the detection head, which splices and fuses the multi-scale spatial feature map with the flattened standard time domain feature sequence to simultaneously output the global condensation quality score and the local condensation uniformity index. During the model training phase, a learnable log-variance parameter based on homoscedasticity uncertainty is introduced to adaptively and dynamically allocate weights to the multi-task loss function, which includes target detection loss and condensation quality regression loss.

[0011] As a preferred embodiment of the steel strip surface defect detection method integrating steam parameter adjustment and image recognition described in this invention, the judgment logic after outputting the evaluation index in step S400 includes: Pre-set high global condensation quality threshold, low global condensation quality threshold, high local condensation uniformity threshold, and low local condensation uniformity threshold; If the current global condensation quality score is greater than or equal to the high global condensation quality threshold and all components of the local condensation uniformity index are greater than or equal to the high local condensation uniformity threshold, then the condensation quality is determined to be good and the test result is reliable. If the current global condensation quality score is less than the low global condensation quality threshold or any component of the local condensation uniformity index is less than the low local condensation uniformity threshold, the condensation quality is determined to be unqualified, triggering an alarm in the characterization system and performing feedback adjustment. If the current evaluation index falls between the two aforementioned judgment conditions, the current state is recorded and the poor condensation area is locked according to the spatial distribution of the local condensation uniformity index. The locked information is fed back for subsequent adjustment while the detection result is output.

[0012] As a preferred embodiment of the steel strip surface defect detection method integrating steam parameter adjustment and image recognition described in this invention, the step of determining the thresholds of each item in the judgment logic includes: Based on a static calibration dataset with absolutely true labels, image samples covering the full range of scores are generated. A mapping function between the global condensation quality score and the expected value of the comprehensive performance evaluation of target detection is fitted. The lower limit of the compliance evaluation required by the process specification is used as the lower limit condition to solve the high global condensation quality threshold. The lowest tolerance evaluation lower limit is used as the upper limit condition to solve the low global condensation quality threshold. A mapping function between the local condensation uniformity index value and the expected value of the false positive rate is established using a defect-free background calibration subset. The upper limit of the base false positive rate under uniform condensation state is used as the lower limit condition to solve for the high local condensation uniformity threshold. The critical threshold of the false positive rate rejection when the system is unusable due to interference is used as the upper limit condition to solve for the low local condensation uniformity threshold.

[0013] As a preferred embodiment of the steel strip surface defect detection method integrating steam parameter adjustment and image recognition described in this invention, the long and short cycle decoupled dual-mode control in step S500 includes: The fast circulation loop is set in seconds to perform high-frequency regulation and control of the steam phase parameters that do not include the pre-cooling temperature difference. When the global condensation quality score is too low or the continuous decrease exceeds the limit, the regulation is triggered, prioritizing the increase of steam flow rate, the reduction of steam dryness and the fine adjustment of steam temperature. The slow loop is set with a time interval of ten seconds, and a weighted moving average filter formula with an exponentially decaying memory structure is introduced to remove high-frequency oscillation characteristics and solve the steady-state uniformity error corresponding to each partition. Among them, the exponential decay forgetting factor is strictly mapped by the inherent thermal time constant of the water-cooled roller system; when the steady-state uniformity error of any zone exceeds the adjustment trigger threshold, the control law combining the static steady-state decoupling matrix and the incremental proportional integral derivative control algorithm is used to uniformly solve the cooling water flow correction vector of each zone in order to eliminate the thermal coupling interference between adjacent cooling zones. The decoupling matrix and control parameters are pre-tuned during open-loop testing.

[0014] As a preferred embodiment of the steel strip surface defect detection method integrating steam parameter adjustment and image recognition described in this invention, the step of determining the adjustment trigger threshold in the slow circulation loop includes: The real-time thickness and real-time running linear speed of the steel strip during the production process are obtained to characterize the real-time area heat flux under the current working conditions, and the nominal parameters are extracted to characterize the reference area heat flux. The relative drift is obtained by comparing the real-time area heat flux with the reference area heat flux, and the system’s preset reference control dead zone is dynamically adjusted nonlinearly by combining the thermal inertia sensitivity adjustment coefficient. When the real-time area heat flux is higher than the reference area heat flux, the reference control dead zone is narrowed nonlinearly; when the real-time area heat flux is lower than the reference area heat flux, the reference control dead zone is widened nonlinearly. The dead zone value after nonlinear dynamic scaling is combined with physical and process boundaries to determine the adjustment trigger threshold.

[0015] In a second aspect, embodiments of the present invention provide a steel strip surface defect detection system that integrates steam parameter adjustment and image recognition, including: a parameter acquisition and calculation module, used to collect steel strip surface state parameters in real time, and calculate an initial steam parameter configuration vector as a physical reference by combining the deviation compensation gain matrix of open-loop calibration; The physical development control module is used to perform slow-response zoned precooling compensation on the steel strip through a water-cooled roller with surface micro-nano textured according to the precooling temperature difference in the initial steam parameter configuration, and to spray wet steam onto the surface to form a condensation development pattern. The multimodal image acquisition module is used to acquire the developed pattern across the entire frame and extract abnormal image slices based on the dynamic threshold multiplier. It reuses the sub-pixel edge coordinates of the line array camera to perform deviation compensation and capture temporal local images. It outputs a standard time domain feature sequence based on the Gaussian kernel time domain resampling interpolation algorithm. The defect detection and evaluation module is used to run the target detection model, which includes trainable edge-enhancing convolution and conditional batch normalization mechanism, for inference. It is trained using a multi-task loss function with adaptive weight allocation based on uncertainty parameters, and outputs defect confidence, global condensation quality score and local condensation uniformity index. The dual-mode decoupling feedback module is used to run a fast circulation loop to adjust the gas phase steam parameters according to the evaluation index, and to run a slow circulation loop to adjust the cooling water flow rate based on the steady-state decoupling matrix and a multivariable incremental algorithm.

[0016] The present invention has the following beneficial effects: 1. First, the present invention constructs physical boundary conditions by using a water-cooled roll with surface micro-nano texture and laminar wet steam, enabling the steam to form a differential condensation development pattern in regions with different surface characteristics, enhancing the visual features of low-contrast defects at the physical level. Meanwhile, the line array camera in the main channel pre-screens abnormal slices across the full width, and the sequential local images obtained by the area array camera in the auxiliary channel are input into an improved YOLO-Cond deep learning model. By using a parallel dual-branch structure and a conditional batch normalization mechanism, the global feature of the uniformity of the dynamically condensing water film in the physical state is combined with the high-frequency feature of the droplet / defect edge. Through the mutual promotion of physical development means and multi-modal object detection algorithms, feature extraction is based on the dynamic condensation and evaporation evolution process, avoiding the limitation of pure vision methods relying only on static weak gray-scale differences, thereby reducing the missed detection rate of low-contrast defects.

[0017] 2. In the present invention, while the deep learning model outputs the defect detection result, it simultaneously calculates the quantified global condensation quality score and local condensation uniformity index, and uses the quantified index as a feedback source to trigger a decoupled dual-mode control with short and long cycles: the fast loop is used to adjust the gas-phase parameters with low thermal inertia to cope with short-term high-frequency disturbances, and the slow loop is used to adjust the cooling water flow rate in the partition of the water-cooled roll with large heat capacity characteristics to eliminate the systematic deviation caused by the lateral thermal drift of the steel strip. Through this control strategy that links algorithm evaluation indicators with actuators with different physical characteristics, the system can adaptively shield the interference caused by the fluctuations in the production line conditions, maintain the stability of the condensation development pattern, and achieve online closed-loop control of surface defect detection.

[0018] 3. First, the present invention pre-cools the surface of the steel strip by using a femtosecond laser micro-nano textured water-cooled roll, and then sprays wet steam according to the adaptively optimized steam parameters, making the defect area and the normal area produce a significant condensation difference pattern, amplifying the visual features of the defects at the physical level, and solving the problem that it is difficult for existing pure vision methods to reliably extract low-contrast defect features from the original image. Meanwhile, the lightweight ROI pre-screening architecture only sends abnormal local image slices into the deep learning model, and directly determines the non-abnormal areas as qualified, enabling the system to quickly locate the abnormal areas across the full width within an extremely short control cycle, and focusing limited computing resources on truly suspicious areas. The two reinforce each other. Spraying wet steam creates high-quality images, and ROI pre-screening can efficiently screen abnormal areas, enabling the entire detection link to meet the real-time requirements of high-speed production lines while ensuring accuracy, solving the contradiction between the high missed detection rate of low-contrast defects and the computing power bottleneck of massive data, and improving the detection sensitivity of low-contrast defects and the real-time performance of the system.

[0019] Of course, it is not necessary for any product implementing the present invention to simultaneously achieve all the above-mentioned advantages. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] To more clearly illustrate the technical solutions of the embodiments of the invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the invention. For those skilled in the art, the drawings can be obtained from these drawings without creative effort.

[0021] Figure 1 The flowchart is for the steel strip surface defect detection method provided by the present invention.

[0022] Figure 2 This is a flowchart illustrating step S400 provided by the present invention.

[0023] Figure 3 A flowchart of the steel strip surface defect detection system provided by the present invention.

[0024] Figure 4 This is a schematic diagram of the steel strip surface defect detection process provided by the present invention. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] Example 1

[0027] When facing the high-speed, heavy-load, high-temperature and complex working conditions of a continuous annealing production line for cold-rolled tin-plated steel strip, the existing surface defect detection methods are not perfect enough. Pure vision systems still cannot reliably detect low-contrast defects, such as minor scratches, roller marks, and oil residues. Traditional steam development methods rely on human experience and cannot be closed-loop adjusted.

[0028] To solve the above technical problems, such as Figure 1 and Figure 4 As shown, Embodiment 1 of the present invention provides a method for detecting surface defects in steel strips that integrates steam parameter adjustment and image recognition. Specifically, Embodiment 1 takes the exit section of a continuous annealing production line for cold-rolled tin-plated steel strips as an example. By constructing a closed-loop architecture encompassing steam pretreatment, multimodal imaging, lightweight ROI pre-screening, deep learning detection, and long / short cycle decoupled feedback, high-precision detection of low-contrast defects under complex operating conditions is achieved. The specific steps include: S100: Real-time acquisition of surface state parameters such as steel type, thickness, surface temperature, surface roughness, and production line speed of steel strip; calculation of initial steam parameter configuration vector; output of initial steam parameter configuration as physical reference for subsequent precooling and spraying control.

[0029] S200: A contact water-cooled roller with a femtosecond laser micro-nano textured surface is used to pre-cool the steel strip in sections according to the pre-cooling temperature difference output in step S100. The steam generator is controlled to spray wet steam onto the surface of the steel strip according to the initial steam parameters to form a condensation development pattern for multimodal image acquisition.

[0030] S300: The condensation development image formed in step S200 is acquired in full-frame by the linear array camera array of the main detection channel, and lightweight ROI pre-screening is implemented in the FPGA to extract abnormal image slices; at the same time, the temporal local image under fixed spatial nodes is acquired by the area array camera array of the auxiliary dynamic channel based on encoder pulse longitudinal spatial locking and lateral deviation dynamic compensation triggering, and the standard time domain feature sequence is output after dynamic time warp mapping.

[0031] S400: The abnormal image slices output in step S300 are fused with the standard time-domain feature sequence and input into the improved YOLO model with conditional batch normalization mechanism for defect detection. The defect confidence score, global condensation quality score (CQS), and local condensation uniformity index (CUI) are output simultaneously.

[0032] S500: Based on the real-time feedback of CQS and CUI output in step S400, construct a long-short cycle decoupled dual-mode control. The fast cycle loop adaptively fine-tunes the low-inertia steam parameters, while the slow cycle loop adjusts the water cooling water flow rate of the water-cooled roller zone based on the weighted moving average error formula and nonlinear adaptive dead zone. The closed-loop update of the steam parameter configuration vector in step S100 is then performed.

[0033] The S600 displays the defect confidence level and condensation quality index in real time on the industrial control computer interface. It triggers an asynchronous rejection signal to the control system for unqualified defects and stores all element detection data in the database for offline self-learning optimization.

[0034] In this invention, the initial steam parameter configuration vector constructed in step S100 serves as the physical benchmark for pre-cooling and spraying control in step S200; the condensation development pattern formed in step S200 is a prerequisite for multimodal image acquisition in step S300; the standard time-domain feature sequence and pre-screened image slices output in step S300 form the data basis for deep learning detection in step S400; the global condensation quality score (CQS) and local condensation uniformity index (CUI) calculated in step S400 are the core parameters for feedback adjustment in step S500; the closed-loop updated steam parameter configuration vector in step S500 is fed back to steps S100 and S200, forming a closed-loop link of parameter configuration, steam spraying, image acquisition, quality assessment, and feedback optimization. These steps work synergistically, overcoming the limitations of traditional steam development methods that rely on manual experience and have fixed parameters that cannot adapt to production fluctuations. This achieves closed-loop adaptive adjustment between steam parameters and detection accuracy, improving the detection sensitivity of low-contrast defects and the system's robustness to complex operating conditions.

[0035] Furthermore, to better illustrate the technical solution of Embodiment 1 of the present invention, a method for detecting surface defects in steel strips that integrates steam parameter adjustment and image recognition is described in detail, specifically including the following: First, in actual continuous production, parameters such as steel strip grade, thickness, surface roughness, and inlet atmosphere temperature will change with production plan switching, and fixed steam parameters cannot adapt to such dynamic changes. In this embodiment, step S100, by collecting steel strip surface state parameters in real time and calculating the initial steam parameter configuration, specifically includes the following sub-steps: S110. Read the current production status parameters in real time in the production line process control system, including steel grade code. (e.g., MR T-4CA, MR T-5BA, etc.), steel strip thickness Production line speed The annealing process exit temperature is monitored. An infrared thermal imager (accuracy ±0.5℃) and a laser triangulation roughness meter (measuring range 0.1μm-2.0μm) are installed in front of the inspection station to collect the surface temperature of the steel strip in real time. and surface roughness .

[0036] S120. Considering the transient temperature gradient between the upper and lower surfaces of the steel strip after water-cooled roller contact precooling, this temperature gradient is closely related to the heat flux density, steel strip thickness (thermal resistance), running speed (contact time), and the temperature difference between the hot and cold sources (thermodynamic driving force). A heat transfer compensation formula coupling kinematics, geometry, and thermodynamic driving variables is established, and the apparent temperature collected by the infrared thermometer is used as the basis for this compensation. Corrected to the actual temperature of the steam-sprayed surface: ;in, This represents the actual surface temperature of the side of the steel strip that is not in contact with the water-cooled roller (the steam-sprayed surface); The actual measured temperature is obtained by an infrared thermometer at the outlet side of the water-cooled roller. The set temperature for the internal circulating cooling water of the contact-type water-cooled roller; The thickness of the steel strip is measured in real-time, in meters (m). The real-time operating speed of the production line, in m / s; This is the comprehensive thermal resistance compensation coefficient, which includes factors such as thermal diffusivity and contact thermal resistance, and its unit is s / m².

[0037] As an optional embodiment, the comprehensive thermal resistance compensation coefficient The calibration was pre-determined using an offline calibration method. Specific calibration steps included: constructing an experimental platform containing a simulated water-cooled roller, maintaining the contact wrap angle and surface micro / nano-textured state consistent with the industrial environment; and attaching high-precision armored thermocouples to the steam-sprayed surface of the experimental steel strip (i.e., the surface of the non-contact water-cooled roller) to represent the actual surface temperature. The absolute reference measuring end; at the same time, an infrared thermal imager of the same model as the one used on site is set up at the water-cooled roller outlet to record the actual measured temperature. Set the cooling water temperature. To maintain a constant value, within the physical boundaries allowed by the process, multiple steel strip thickness calibration nodes with a gradient distribution are selected. With running linear velocity calibration node Combination. The experimental setup was operated under various parameter combinations. Once heat conduction reached a steady state, readings from the sheathed thermocouples were simultaneously acquired. Readings from infrared thermal imager A total of [number] samples were collected. Group steady-state operating condition data (of which) Establish an error objective function based on the least squares method, and solve for the comprehensive thermal resistance compensation coefficient that minimizes the sum of squares of the global true measurement errors. The fitting formula is as follows: .

[0038] S130, Set the initial steam parameter configuration vector ,in For steam temperature, Steam dryness Steam flow rate per unit width To pre-cool the temperature difference, Where is the nozzle distance. The initial steam parameter configuration vector is determined by a deviation compensation formula based on the Jacobian matrix: ;in, For steel type The basic parameter vector (5×1 dimension) was pre-calibrated through offline experiments; This is the real-time operating condition deviation vector. , , These are the nominal reference values; This is the bias compensation gain matrix (Jacobi matrix, dimension 5×3). For the amplitude limiting function, and These are the physical safety lower and upper limits for each parameter, respectively, to prevent output loss of control due to sensor malfunctions or extreme operating conditions.

[0039] As an optional embodiment, the deviation compensation gain matrix Based on the system identification method using open-loop step response experiments during the equipment calibration phase, the specific tuning steps are as follows: During the equipment calibration phase, the closed-loop feedback loop of the control system is disconnected, maintaining open-loop operation. This is tailored to the current steel grade. Set the baseline operating parameters, adjust the system to steady-state condensation and achieve the global condensation quality score, and record the baseline surface state vector at this point. With the basic steam parameter vector Keeping all other operating parameters constant, inject a step disturbance of a single variable into the input of the operating parameters. ( (These correspond to surface temperature, roughness, and linear velocity, respectively). After the condensation pattern deviates from the baseline state, the steam parameters are optimally adjusted in an open-loop state. ( Until the global condensation quality score is restored, record the steam parameter adjustment increments after steady-state reconstruction. The deviation compensation gain matrix is ​​solved using the linear derivative principle. elements in .

[0040] Specifically, for example, setting a baseline operating condition. For: surface temperature At 45℃, roughness 0.8μm, linear velocity 2.0 m / s; Basic steam parameter vector The values ​​are: [110, 0.90, 2.5, -15, 80] (corresponding to steam temperature, steam dryness, steam flow rate per unit width, precooling temperature difference, and nozzle distance, respectively). A single-variable step disturbance experiment is performed: when the linear velocity is generated... With a step increase in steam speed (m / s), the contact time shortens, leading to incomplete condensation. To restore condensation quality, experimentally determined steam parameter increments include: increasing steam flow rate. g / (s·m) and increase the precooling temperature difference ℃. Based on this, the related elements in the third column of the matrix can be calculated: , After traversing all univariate perturbation experiments, a complete 5×3 dimensional bias compensation gain matrix was constructed. : In actual online testing, let the current operating condition deviation vector be... (That is, surface temperature is 2℃ higher, roughness is 0.1μm lower, and velocity is 0.2m / s higher), through matrix multiplication. The compensation vector is calculated as follows: The system will combine the compensation vector with... The parameters are superimposed, and a limiting function is input to complete the boundary truncation, outputting the final initial steam parameter configuration vector. .

[0041] Furthermore, directly spraying steam onto the high-temperature steel strip can easily lead to rapid evaporation of the steam, preventing the formation of a continuous condensation film. However, excessive precooling can result in an excessively thick water film, causing image blurring. Additionally, the transverse temperature distribution of the steel strip may be uneven, with faster heat dissipation at the edges and better insulation in the center, requiring zoned precooling control. In this embodiment, step S200 uses a contact-type water-cooled roller for precooling and femtosecond laser micro-nano texturing to avoid increasing thermal resistance. Subsequently, wet steam is sprayed according to the configured parameters, specifically including the following sub-steps: The S210 contact-type water-cooled roller has a high thermal conductivity chromium layer (thermal conductivity > 90 W / (m·K)) on its surface, and the wrap angle between the steel strip and the water-cooled roller is 180°. To address the transient response hysteresis of the high-heat-capacity water-cooled roller, a pre-cooling strategy combining feedforward control based on heat balance and slow-cycle feedback compensation is adopted. Specifically, this includes calculating the basic cooling water volumetric flow rate required to reach the target pre-cooling temperature based on the real-time input of the steel strip's physical properties and motion parameters from the production line. ;in, The volumetric flow rate of the basic cooling water, with dimensions of ; The density of the steel strip; The specific heat capacity of the steel strip; For the thickness of the steel strip; The width of the steel strip; The linear velocity of the steel strip; This represents the actual temperature of the steel strip at the inlet of the water-cooled roller. The target pre-cooling temperature, and , It is a negative value; and These are the density and specific heat capacity of the cooling water, respectively. The inlet and outlet temperature difference designed for the cooling water system; The heat exchange efficiency coefficient of the water-cooled roller system, and as an optional embodiment, the heat exchange efficiency coefficient of the water-cooled roller system. The determination steps are as follows: Set a constant basic flow rate to drive the water-cooled roller to run. After the system reaches thermal equilibrium steady state, collect the actual inlet and outlet cooling water temperature difference and the steady-state inlet and outlet temperature difference of the steel strip. Solve the equation by the inverse equation of the heat balance equation. Its actual physical meaning characterizes the effective heat exchange rate in the heat conduction process of the system. The typical value range is calibrated to be 0.70 to 0.85.

[0042] The dead zone for controlling the steel strip temperature at the water-cooled roll outlet is set to ±2℃~±3℃. When the average temperature fed back by the outlet infrared temperature measurement array deviates from... Furthermore, when a steady-state drift trend exceeding the control dead zone is observed, the slow-cycle feedback loop calculates and superimposes the corrected flow rate based on the discrete-position PI (proportional-integral) control algorithm. : In the formula, This is the sequence number of the discrete-time sampling sequence (the sampling period is set to the order of ten seconds). This represents the steady-state temperature drift error for the current sampling period. This is the average temperature currently measured. This is the proportional gain coefficient; This represents the integral gain coefficient. To address the system's large thermal capacity and hysteresis characteristics, the derivative term is omitted to avoid actuator oscillations caused by high-frequency temperature measurement noise. Control parameters and Based on the pre-executed open-loop single-step response test curve, the parameters were tuned using the Ziegler-Nichols empirical formula, ensuring that the parameter calculation process was completely independent of the online closed-loop feedback link. The calculated corrected flow rate... With basic traffic The superimposed output is executed. The transient temperature deviation remaining in the control dead zone after the water-cooled roller precooling is used as a feedforward disturbance and input into the fast cycle loop of step S500. The steam generator, with a faster response speed, performs adaptive compensation by fine-tuning the steam dryness and flow rate.

[0043] To avoid increased contact thermal resistance caused by a hydrophobic coating with low thermal conductivity, femtosecond laser micro-nano texturing was performed on the surface of the high thermal conductivity chromium layer of the water-cooled roller to construct a biomimetic superhydrophobic structure without reducing thermal conductivity. For example, the specific preparation process and parameters are as follows: the water-cooled roller coated with a high thermal conductivity chromium layer is clamped in a multi-axis linkage laser processing machine tool; a ytterbium-doped fiber femtosecond laser is used as the processing light source, with the laser wavelength set to 1030 nm, pulse width to 250 fs, repetition frequency to 500 kHz, and average power to 15 W; the laser beam is controlled by a galvanometer scanning system to perform orthogonal grid scanning on the chromium layer surface, with the scanning speed set to 1500 mm / s and the grid line spacing to 25 μm, inducing the formation of a dual roughness structure on the surface composed of micron-scale grids and nano-scale particles; after processing, the water-cooled roller is placed in a vacuum heating furnace and held at an ambient temperature of 150°C for 2 hours to complete the adsorption modification of low surface energy substances on the surface. After the above preparation process, the static water droplet contact angle on the surface of the water-cooled roller is greater than 150°, and the roll-off angle is less than 10°. This biomimetic superhydrophobic texture prevents early condensation of ambient water vapor on the roller surface during pre-cooling. Simultaneously, a scraper and air knife are installed to blow away any physically adhering trace amounts of moisture as the steel strip leaves the water-cooled roller.

[0044] S220. After steam is ejected from the nozzle, it mixes with the surrounding air, causing it to cool down and partially condense. The actual steam state reaching the steel strip surface deviates from the generator outlet parameters. Simultaneously, the high-speed movement of the steel strip generates boundary layer airflow, interfering with effective contact between the steam and the surface. To address these issues, in this embodiment, the steam generator produces saturated steam with set parameters, which is then transported to a wide, flat nozzle through an insulated pipe. A flow-rectifying grid is installed at the nozzle outlet to ensure that the steam flows uniformly and at a low speed across the steel strip surface. To compensate for heat loss during transport, a secondary heating belt is added at the nozzle inlet to ensure that the outlet steam temperature fluctuation does not exceed a reasonable range (e.g., ±1℃). Steam dryness is monitored in real-time by an online capacitive sensor and controlled in a PID closed-loop manner with the set value. The spraying time employs either continuous spraying mode (for continuous production lines) or pulse spraying mode (for offline sampling inspection).

[0045] Special operating condition handling: When the steel strip surface temperature is too low (e.g., <40℃) or the ambient humidity is too high (e.g., relative humidity >85%), the system automatically reduces the steam flow rate or increases the dryness to prevent excessive condensation. When the steel strip surface temperature is too high (e.g., >70℃), the system temporarily skips steam spraying, uses only the original surface image for detection, and issues an alarm prompting the operator to check the preceding cooling process.

[0046] Full-width steel strip inspection requires 100% coverage, and a single area scan camera may not be able to meet the wide field of view. Simultaneously, the auxiliary dynamic channel needs to record condensation and evaporation images of the same area at different delays, but variations in steel strip speed can cause misalignment between the actual time interval and the preset nodes. In this embodiment, step S300 acquires condensation and development images across the entire width using a line scan camera array, and performs lightweight ROI pre-screening in the FPGA to extract abnormal image slices. Simultaneously, a temporal local image at a fixed spatial node is acquired using an area scan camera array based on encoder pulse spatial locking triggering, and a standard time-domain feature sequence is output after time warp mapping. Specific implementation steps include: S310, the main inspection channel uses a line scan camera array for full-frame acquisition and abnormal image slice extraction: the line frequency of the three line scan cameras is synchronously driven by the main encoder pulse. For each pulse output by the encoder, all line scan cameras simultaneously expose one line. The line frequency and production line speed satisfy the following kinematic matching formula: ;in, For line scan camera, the dimension is _____. (Right now ); The real-time operating speed of the production line, with dimensions of ; The longitudinal spatial displacement step of the steel belt corresponding to a single pulse of the encoder, with dimensions of Images from adjacent cameras are cropped at the boundaries and registered with physical coordinates, then stitched together to form a full-frame image.

[0047] In the FPGA preprocessing unit, adaptive thresholding is performed line-by-line on the full-frame image. To overcome the substrate noise drift caused by the dynamic changes in the surface condition and condensation effect of the steel strip, the fixed anomaly detection factor is replaced with a dynamic threshold multiplier. The criteria for determining local abnormal pixels are as follows: In the formula, This is the grayscale value of the current pixel; and These represent the mean and standard deviation of grayscale values ​​within a preset-sized sliding window centered on the current pixel. Dynamic threshold multiplier. The benchmark was determined by a combination of offline benchmark calibration and online dynamic calibration. The specific steps included: acquiring a standard condensation development image of a defect-free surface for the current steel grade, and statistically extracting the standard deviation of the benchmark grayscale. Based on the preset upper limit of pixel-level false positive rate control, the baseline threshold multiplier is calibrated. Extract the global grayscale standard deviation of the currently acquired image frame. Characterize the real-time basis noise drift and calculate the dynamic threshold multiplier: In the formula, This is the sensitivity compensation coefficient, used to adjust the severity of the threshold's response to background noise; This is a limiting function; and These are the absolute lower bound for preventing false alarms and the absolute upper bound for preventing missed detections. After identifying abnormal pixels, morphological dilation and connected component labeling are performed on the abnormal pixel intervals across multiple consecutive rows. The minimum bounding rectangle of each independent connected component is extracted to generate a two-dimensional spatial coordinate set of candidate defect regions. Based on the spatial coordinate set, a bounding box is cropped by expanding the original full-frame image outwards by a preset pixel width, generating an abnormal image slice containing the complete defect outline and local background. Background image data that does not contain abnormal pixel intervals is directly discarded. The generated abnormal image slice is transmitted to the GPU via the bus as the target detection input data for step S400.

[0048] The S320 auxiliary dynamic channel employs a two-dimensional dynamic tracking strategy combining a large-view array of area-array cameras with longitudinal hardware locking and lateral software compensation. To overcome lateral deviation interference during the steel strip's operation, the physical field of view (FOV) of each area-array camera is designed as a redundant region encompassing both the actual size of the target area and the maximum allowable lateral deviation. Specific implementation steps include: Longitudinal spatial lock trigger: A small rectangular area on the surface of the steel strip is set as the observation target area. When the target area passes through the reference zero point position where the steam nozzle is located, the current pulse accumulation value of the main encoder is recorded. Simultaneously record the reference lateral physical coordinates of the steel strip edge at this time. When the longitudinal displacement of the target area reaches the preset physical position of each area array camera, the corresponding camera is triggered to expose, and the calculation of the... Trigger pulse count value of the platform array camera: ;in, For the first The trigger pulse count value of the camera; For the first The physical distance between the optical axis of the camera and the steam nozzle; The spatial resolution calibration constant of the main encoder; This is a floor function to adapt to the discrete triggering mechanism of digital hardware.

[0049] To avoid introducing additional hardware and ensure absolute time synchronization, edge extraction is performed using images from a linear scan camera array that are exposed synchronously with the main encoder in the main detection channel. Exposure time of table array camera Extract the corresponding image row data synchronously acquired by the linear scan camera array. Define a local search interval near the theoretical edge pixel coordinates. Calculate the first-order gradient of grayscale Find the integer-pixel edge coordinates with the largest absolute gradient value within the specified interval. A parabolic fitting algorithm is used to solve for sub-pixel level edge coordinates. : ;in, These are integer pixel edge coordinates, with units of pixels. This represents the first-order gradient of the image's grayscale, measured in grayscale levels. The trigger time is calculated. Real-time lateral physical coordinates of the steel strip edge : ;in, The physical horizontal absolute coordinates of the starting point of the field of view of the linear array camera are in mm. denoted as the lateral physical spatial resolution of the line scan camera, measured in mm / pixel.

[0050] Lateral Deviation Dynamic Compensation and ROI Repositioning: In the first After the stage array camera is triggered by the aforementioned pulse hardware to expose and acquire a full-field image, the observed target area needs to be dynamically cropped within the image domain. Let the... Taiwanese camera exposure time The real-time lateral physical coordinates of the steel strip edge, measured by an edge detection algorithm or an external wide-field linear scan camera, are: Then in the first... In the image coordinate system of the table array camera, calculate the compensated horizontal pixel coordinates of the center of the observation target area: ;in, The actual lateral pixel coordinates of the target area center after deviation compensation; For the first The reference pixel coordinates of the target area center, calibrated by the table array camera under ideal conditions without deviation (when the steel strip is theoretically centered); For triggering time Real-time lateral physical coordinates of the steel strip edge; The reference lateral physical coordinates of the steel strip edge at the initial moment of the target area; For the first The horizontal physical spatial resolution of a tabletop array camera.

[0051] Finally, the system uses the corrected pixel coordinates Centered on the target area (with the vertical coordinates strictly locked by hardware, requiring no software compensation), the image is cropped according to the standard pixel size of the observed target area. This method outputs multiple frames of local images strictly aligned to the same physical region over time, eliminating spatial misalignment and providing data assurance for reliable mapping of standard time-domain features.

[0052] S330. To address the physical time misalignment issue caused by dynamic fluctuations in linear velocity leading to the arrival of the observation target area at each camera node, the actual time is calculated based on pulse discrete integral, and the temporal local image features are aligned to the standard time domain using a Gaussian kernel-based time-domain resampling interpolation algorithm. Specific implementation steps include: Since the spatial physical distance of the area array camera is fixed, the observation target area moves from the reference physical zero point where the steam nozzle is located to the... The actual physical time of the physical location of the optical axis of the table array camera The time varies dynamically with the linear velocity. The system uses the time interval between adjacent pulses from the main encoder to discretize and accumulate the time to calculate the actual time. In the formula, Starting from the reference physical zero point, the target area reaches the [number]th [unit]. The actual physical time at the location of the table array camera; To trigger the The total pulse count of the camera, measured in pulse number ( Physical distance (for spatial resolution) For the first The first to the second The actual time length of each encoder pulse interval.

[0053] No. After the platform-array camera completes the target area image cropping based on the pixel coordinates corrected in step S320, the output local image is input into a convolutional neural network for spatial feature extraction. Since each cropping step is physically aligned with the same observation target area, the extracted high-dimensional spatial feature vector... (dimension is) In a physical sense, this means that the region is... The state representation at any given moment. Platform (in this embodiment) Area scan camera in actual time vector The extracted feature vectors are stacked in chronological order to construct the actual time-domain feature matrix. The spatial dimension of this matrix is The temporal correlation of the implicit feature extraction of its row index.

[0054] Subsequent object detection models require standard time series data. Feature inputs (e.g., taking values ​​relative to the end time of steam spraying) , (etc.). To solve the actual time vector Compared with standard time series To address the non-alignment problem, this paper abandons nonlinear alignment strategies that require changing the time axis length and adopts a Gaussian kernel-based time-domain resampling interpolation algorithm to transform the actual time-domain feature matrix. Linear mapping to the standard time-domain feature matrix: In the formula, To reconstruct and align the feature matrix to the standard time point, the dimension is... ; The Gaussian kernel resampling weight matrix has a dimension of . Its first Line number Column weight elements Determined by the following formula: In the formula, For the preset first A standard time point; For the first The actual time point of exposure of the table array camera; These are the Gaussian kernel time bandwidth control parameters. After time-domain alignment and reconstruction, the output is a standard time-domain feature sequence. It is transmitted in real time to step S400.

[0055] Furthermore, the S400 employs the YOLO-Cond deep learning model to perform target detection and condensation quality assessment of steel strip surface defects using pre-screened anomaly image slices and mapped standard time-domain features. The YOLO-Cond model utilizes an improved YOLOv8 model tailored to the characteristics of condensation development images. Building upon the standard YOLOv8 model, it includes multi-branch parallel processing of the input layer, conditional normalization of steam parameters, and a condensation quality assessment branch, such as... Figure 2 As shown, it specifically includes: S410. The image after condensation and development contains low-frequency information about the overall uniformity of the condensate film and high-frequency detail information about droplet edges and defect contours. Traditional YOLO models using a single convolutional kernel struggle to effectively extract both types of features simultaneously. In this embodiment, a parallel dual-branch feature extraction module is designed at the input end to separate and process the input feature map, specifically including: The first branch uses a large-size convolutional kernel for local receptive field expansion and downsampling to extract global low-frequency features of water film uniformity: the input image is fed into the convolutional kernel with a kernel size of [missing information]. Step size is The convolutional layers are connected sequentially to a batch normalization layer and an average pooling layer to output a global low-frequency feature map. .

[0056] The second branch employs a cascade of small-sized convolutions and trainable edge-enhancing convolutions to extract high-frequency features from droplet and defect edges: first, it passes through a convolution kernel with a kernel size of... Step size is The standard convolutional layer is used for dimensionality reduction; then it is fed into a trainable edge-enhancing convolutional layer. The specific implementation process of this edge-enhancing convolutional layer is as follows: constructing a weight matrix containing the horizontal direction. Weight matrix in the vertical direction The depth of the separable convolutional layers is determined during the model initialization phase. and The spatial core parameters are forcibly initialized to standard nonlinear discrete difference operator matrices: Regarding boundary handling, a width of [missing information] is applied to the input feature map. Zero-padding of pixels ensures consistent spatial resolution before and after convolution operations, eliminating boundary artifacts. Feature map The formula for calculating the forward propagation through this convolutional layer is: In the formula, This is the output high-frequency edge feature map; Represents a two-dimensional cross-correlation convolution operation; For the set minimum positive constant (take ).

[0057] During the backpropagation phase of model training, the frozen state of the traditional Sobel operator weights is released, and the matrix is ​​set. and The gradient participates in the backpropagation update. The weight matrix is ​​based on the overall loss function. The gradient is dynamically updated iteratively: ; In the formula, For network learning rate; This represents the iteration cycle number. The connection and difference between this method and the traditional fixed Sobel operator are as follows: In the initial stage, the model retains the prior inductive bias of the standard difference operator's high sensitivity to edge contours; in the training stage, the algorithm allows the weights to deviate from the initial strictly integer parameters under the data-driven backpropagation mechanism, adaptively adjusting the weights of the local receptive field in each direction to remove spurious high-frequency noise generated by the reflection of the condensate film, thereby outputting a true high-frequency feature map with a high correlation to the defect structure. Finally, the global low-frequency feature map output from the first branch is... High-frequency feature map of the second branch output Assembled according to channel dimensions, via The standard convolutional layer fuses cross-channel information and reduces its dimensionality before feeding it into the backbone network.

[0058] S420. Since changes in steam parameters affect the brightness, contrast, and texture features of the condensation pattern, the model should be able to adaptively adjust the feature extraction method according to the current steam parameters. Directly inputting steam parameters with different physical dimensions into the normalization layer will lead to an imbalance in the gradient distribution. Therefore, the steam parameters are first standardized without dimensions: Let the original steam parameter vector of the input data be... The statistical mean of each component in the training set and standard deviation Pre-stored. The standardized dimensionless parameters are: In the formula, To prevent the removal of the zero constant.

[0059] Conditional Batch Normalization (CBN) modules are introduced after each basic convolutional block in the backbone network to normalize the steam parameters. As a conditional input. For the feature map... The data from each channel were normalized to zero mean and unit variance to obtain dimensionless features. Then, the scaling factor and offset of the normalization layer are dynamically generated using the following mapping formula, and the affine transformation is completed: ; ; In the formula, and Based on the affine parameters; and These are the conditional projection weights of the network parameters for each steam parameter component.

[0060] To ensure effective convergence of the CBN module and resolve data coverage and backpropagation issues, a complete training strategy and configuration scheme are set as follows: The base scaling factor is... Static initialization to constants Basic offset Static initialization to constants Conditional projection weights and Initialized as a matrix with all elements equal to zero. This initialization mechanism ensures that in the first forward propagation batch at the start of training, the CBN layer mathematically degenerates strictly into a standard batch normalization layer, maintaining the stability of the initial feature distribution and preventing gradient vanishing or explosion. To avoid training blind spots in the infinitely continuous steam parameter space, the Latin hypercube sampling (LHS) method is used to generate uniformly distributed parameter combination nodes within the physical safety upper and lower limits of the five-dimensional steam parameters, driving the production equipment to acquire the condensation and development image dataset corresponding to the operating conditions. This transforms random sampling in high-dimensional space into hierarchical orthogonal sampling, achieving efficient coverage of the entire distribution of steam parameters with a limited data scale. The comprehensive loss function of the target detection network is... Treated as a scalar, the CBN module is directly attached to the backpropagation computation graph. According to the chain rule, the loss function... For projection weights The formula for calculating the partial derivative is as follows: loss function For projection weights The formula for calculating the partial derivative is as follows: The gradient values ​​mentioned above are accumulated and collected in each batch, and the optimizer synchronously updates each dimensionless weight.

[0061] Training configuration example and convergence judgment: The AdamW optimization algorithm is used, and the initial learning rate is set to [value missing]. A cosine annealing learning rate decay strategy with a period of 50 is configured. The batch size is set to 32, and the weight decay coefficient is set to [value missing]. Monitor the trajectory of the total loss function on the validation set. When the total loss on the validation set has not decreased for 15 consecutive iterations, and the average gradient norm of the CBN module parameter updates converges to a certain value... At this scale, the early shutdown mechanism is triggered, and the system determines that the conditional mapping relationship between the model feature extraction and the steam parameters has reached the optimal convergence steady state.

[0062] S430. A parallel condensation quality assessment subnetwork is added as a side branch to the output of the target detection network. The input of this subnetwork includes the multi-scale spatial feature map output from the backbone network and the standard time-domain feature sequence generated by mapping in step S330. The specific structure and processing logic are as follows: The multi-scale spatial feature map set output from the backbone network neck contains static spatial distribution information of the condensation pattern. Global average pooling is used to compress the spatial feature maps at each scale into a one-dimensional static spatial feature vector. The standard time-domain feature sequence output by step S330 Its dimensions are (in The number of standard time points, (This refers to the number of feature channels), whose physical meaning characterizes the dynamic evolution of the condensate film in the observed target area over a fixed time series. Flattening it along the time dimension forms a one-dimensional dynamic time-domain feature vector. Its dimensions are .Will and The data are concatenated along the channel dimension and fused to form a comprehensive feature vector representing the spatiotemporal condensation state. This vector is then fed into a multilayer perceptron consisting of three fully connected layers (with 384, 64, and 9 nodes respectively). Finally, a scalar global condensation quality score (CQS) is output synchronously through a sigmoid activation function (within a range of values). ) and the vector form of the local condensation uniformity index CUI (dimension 1) value range CUI is directly output from the fully connected layer of the quality evaluation subnetwork, and its training labels are obtained by manually scoring the uniformity of the condensation pattern.

[0063] Overall loss function during model training Loss due to target detection Global score regression loss With local homogeneity regression loss Composition. Target detection loss This is a dimensionless term that includes bounding box regression loss (such as CIoU Loss), classification loss, and confidence loss. Using mean squared error loss to evaluate global quality predictions With real labels Deviation between: ; Local homogeneity prediction vector is evaluated using mean square error loss. Compared with the real label vector Deviation between: .

[0064] Due to significant differences in optimization objectives and gradient scales between target detection and condensation quality assessment tasks, using fixed weights can lead to one task dominating backpropagation. This application introduces a learnable weight allocation mechanism based on homoscedastic uncertainty, dynamically balancing the loss by having the network autonomously learn the inherent noise of each task. The overall loss function with dynamic weights is described below. Defined as: In the formula, , , These are the learnable log-variance parameters for the corresponding tasks.

[0065] Initialization during the initial training phase The initial weights of each task are 1. During the backpropagation process, The trainable network parameters are updated with gradient descent. For a given task (such as object detection)... The model is in the early stages of convergence, resulting in a very large loss value, which will cause the model to spontaneously increase. To reduce the proportion of this loss in the total gradient (i.e., assign a smaller weight). To avoid gradient explosion or overfitting to a single task; at the same time, the regularization term in the second half of the equation The increased parameters are penalized to prevent the log-variance from approaching infinity. This mechanism achieves dynamic smoothing and adaptive balancing of three different loss gradients throughout the entire training cycle without manual intervention.

[0066] S440, In the inference stage, the pre-screened abnormal image slices from step S310 are input into the YOLO-Cond model, and simultaneously... Input the condensation quality assessment branch. The model outputs the defect confidence score, CQS, and CUI.

[0067] Judgment logic: Set a high global condensation quality threshold Low global condensation mass threshold and high local condensation uniformity threshold Low local condensation uniformity threshold ,and , ; like and All components This indicates that the condensation quality is good and the test results are reliable. like or any Quantity This indicates that the condensation quality is unqualified, triggering an alarm in the characterization system and entering the feedback adjustment step S500; like or individual Components in The state between these two values ​​indicates that the condensation quality is average. The system records this state but still outputs the detection result, and simultaneously... The spatial distribution locks in the lateral areas with poor condensation, which can then be used for targeted adjustments.

[0068] As an optional embodiment, the threshold , , , Based on a statically calibrated dataset with absolutely accurate labels (jointly confirmed by high-magnification offline optical microscopy and human experts), the expected value mapping function between condensation quality scores and objective evaluation indicators for target detection (F1 score and false positive rate) can be established. Specifically, this includes: S441. By adjusting the steam parameters in the calibration platform, in the static calibration dataset... Generate coverage Full range Image samples. Statistics in various predictions. The overall target detection performance under the given values ​​is fitted to... Value Mapping function with expected F1 score The formula for determining the global condensation mass threshold is as follows: ; ; in, The estimated global condensation quality score variable is continuously distributed; For a given Values Under the given conditions, the mathematical expectation function of the F1 score (F1-Score) obtained by the detection model on the calibration dataset; The lower limit of the F1 score required to meet the process specifications (e.g., set to 0.95). This is the lowest acceptable F1 score lower bound for industrial environments (e.g., set to 0.80). This indicates that the infimum of the set is taken. This indicates taking the supremum of a set.

[0069] S442. Localized uneven condensation (such as the presence of large areas of unevaporated water film boundaries) mainly leads to background noise being misjudged as defects, thus increasing the false positive rate (FPR). A calibration subset of the defect-free background is used to establish... Value Mapping function with expected false positive rate The formula for determining the threshold of localized condensation uniformity is as follows: ; ; in, The variable is a continuously distributed index for predicting the uniformity of local condensation. In the given Values Under the condition of [condition], the mathematical expectation function of the false positive rate generated by the detection model on the defect-free calibration dataset; The upper limit of the false positive rate of the substrate under uniform condensation state (background noise tolerance, such as 0.01). This is the critical threshold for the false positive rate (rejection line, such as 0.15) when the system becomes completely unusable due to severe interference from condensation stains.

[0070] Further, step S500 adjusts the steam parameters based on the real-time feedback of the CQS and CUI output from the model. Because the surface state of the steel strip may change abruptly, such as a sudden increase in rolling oil or abnormalities in the upstream cooling water, leading to a decrease in steam condensation efficiency, traditional PID control is ill-suited to this nonlinear, time-varying process. Simultaneously, different actuators possess vastly different thermal inertia. While the temperature, dryness, and flow rate of the steam generator can be rapidly adjusted, the water-cooled roller, as a high-heat-capacity device, requires several seconds to tens of seconds to stabilize to a new thermal equilibrium state after changes in cooling conditions. Driving all actuators with the same control cycle can easily lead to integral saturation and control oscillation in the slow-response channel. To address these issues, this step employs a long-short cycle decoupled dual-mode control for feedback adjustment and parameter optimization, specifically including the following implementation steps: S510, Fast Cycle Control Loop: Controlled object: Steam parameter vector Excluding pre-cooling temperature difference.

[0071] Control cycle: CQS / CUI evaluation and parameter adjustment are performed once every second (e.g., 1 second); if the CQS drops by more than 0.2 in a single frame for two consecutive frames, adjustment is triggered immediately.

[0072] Control objective: When CQS is low, prioritize increasing the steam flow rate. For example, increase the step size by 0.1 g / (s·m); if increasing the flow rate three times consecutively is ineffective, try reducing the steam dryness. For example, adjust the step size by -0.01; if this is still ineffective, fine-tune the steam temperature. For example, a step size of ±1℃ should be used to ensure that the temperature does not exceed the reasonable range (e.g., 105-120℃); nozzle distance Adjust only when all CUI components are <0.5 and adjusting other parameters has no effect.

[0073] The S520 slow-circulation control loop, based on a discrete-time series and multivariate decoupling algorithm, performs low-frequency regulation control of the cooling water flow rate in the water-cooled roller zones of the precooling unit. To eliminate systematic deviations caused by transverse thermal drift of the steel strip and overcome thermal coupling interference between adjacent cooling zones, the specific control steps are as follows: The discrete sampling period of the slow-circulation loop is set. (Ten-second range). A weighted moving average filter formula with an exponentially decaying memory structure is introduced to remove high-frequency oscillation characteristics, specifically targeting water-cooled rollers. Each independent cooling zone is used to calculate the steady-state uniformity error vector for each zone. . No. Error of each partition The calculation is as follows: In the formula, This is the baseline value for the desired local condensation uniformity index; The first output of the object detection model The physical region of the partition in history The measured value of the shot; The length of the sliding time window; This is the extracted dimensionless steady-state error. The exponentially decaying forgetting factor is whose value is strictly mapped by the thermodynamic inherent time delay of the physical object, and its calculation formula is: In the formula, The sampling period; is the thermal time constant of the water-cooled roller system.

[0074] Incremental PID control incorporating a decoupling matrix: There is a lateral heat conduction effect between adjacent cooling zones (i.e., adjusting the flow rate in one zone will cause a change in the temperature field of adjacent zones). When the error vector... The absolute value of any component exceeds the dynamic adjustment trigger threshold. At that time, a control law combining static decoupling matrix and incremental PID is adopted to uniformly calculate the cooling water flow correction vector for each zone. : In the formula, This is the adjustment increment vector of the cooling water volumetric flow rate in each zone during the current control cycle, with dimensions of... ; dimensionless The steady-state decoupling matrix; This is the proportionality coefficient. The integral coefficient is... These are differential coefficients, and all three have dimensions. The calculated vector It is directly superimposed onto the basic flow command of the solenoid valve in the corresponding partition for execution. In extreme cases, if the partitions are physically isolated and uncoupled, then... Degenerates into an identity matrix, and the self-consistent dimensionality reduction of a multivariable system is... An independent single-loop incremental PID.

[0075] To avoid the circular dependency paradox of parameter tuning caused by the closed-loop control itself generating historical data, the thermal time constant of the water-cooled roller is... Decoupling matrix and PID control parameters ( All settings were completed during the open-loop commissioning phase of the equipment installation process. Decoupling matrix calibration: This cuts off closed-loop feedback to maintain steady-state equipment operation. This is done sequentially for each zone. Cooling water input step flow disturbance Record all partitions Changes in indicators after reaching a new steady state Define the steady-state thermal gain matrix. elements Extracting the diagonal elements from the response to construct a dimensionless normalized coupling matrix. (its elements) The final steady-state decoupling matrix is ​​statically locked as follows: .

[0076] PID and Inherent Time Delay Tuning: For the main diagonal dynamic response curve of the single-zone step experiment above (i.e., the open-loop time-domain curve corresponding to the change in this zone for the step in this zone), the time required to reach 63.2% of the steady-state response amplitude is calculated as the thermal time constant. The apparent dead zone lag time and maximum response slope are extracted and directly substituted into the Cohen-Coon open-loop tuning formula to calculate the globally constant... Numerical value.

[0077] As an optional embodiment, set The dynamic adjustment trigger threshold for lateral condensation uniformity deviation is determined as follows: The real-time thickness and linear velocity of the steel strip during production are obtained as characteristic parameters, representing the real-time area heat flux through the water-cooled roller under the current operating conditions. The nominal reference thickness and nominal reference linear velocity are extracted to represent the reference area heat flux of the system under standard operating conditions. The real-time area heat flux is compared with the reference area heat flux to obtain the relative drift, and combined with a pre-calibrated thermal inertia sensitivity adjustment coefficient, the preset reference control dead zone of the system is nonlinearly dynamically adjusted: when the real-time area heat flux is higher than the reference area heat flux, the reference control dead zone is nonlinearly narrowed; when the real-time area heat flux is lower than the reference area heat flux, the reference control dead zone is nonlinearly widened. Finally, the scaled dead zone value is input into a limiting function, and boundary truncation is performed based on the sensor noise physical lower limit and the process allowable deviation upper limit, outputting the current control cycle value. .

[0078] S530, Special Operating Condition Rules Amendment: Rule 1: When CQS is below the minimum global condensation mass threshold (e.g., 0.5) and when fast cycle adjustment is ineffective, the system will issue an emergency alarm and automatically reduce the production line speed until the condensation quality is restored.

[0079] Rule 2: When a concentrated occurrence of oil-related defects is detected, switch the fast cycle to oil-enhanced mode, for example, by setting the steam temperature. 115℃, steam dryness It is 0.88.

[0080] Rule 3: When the defect confidence is generally low but the CQS is good, record the batch of images for incremental offline model training.

[0081] S540, offline self-learning and parameter optimization of historical data: All detection data (including production parameters and actual issued steam parameter vectors) are processed. The quality evaluation index (CQS / CUI) is stored in the database. To achieve cross-cycle self-evolution of initial steam parameters, a Bayesian optimization algorithm is used to optimize the deviation compensation gain matrix. The offline iterative update is implemented as follows: A single complete steel coil production cycle is designated as the data aggregation node. Once the production of a single steel strip coil is completed and all quality inspection data is stored on the tray, the offline parameter update program is triggered. The decision variable is defined as the deviation compensation gain matrix. To ensure the safety of the physical system during the iterative process, a search space limit is preset based on the process safety boundary calibrated during the equipment commissioning period. Construct the evaluation matrix The objective function for overall performance throughout the entire roll production cycle This function is built upon a global condensation quality score and a penalty mechanism based on the frequency of fast cycle intervention. In the formula, This represents the total number of valid image frames acquired throughout the entire cycle of the steel coil. For the first Global condensation quality score for frame output; This represents the total number of times the fast loop is triggered to perform adaptive compensation during this production cycle. The penalty coefficient for fast-cycle intervention.

[0082] Treating the actual physical production process as a black box system, and establishing a gain matrix Mapping to the objective function Gaussian process proxy model. Extract historical production cumulative evaluation dataset. For any gain matrix to be explored within the search space... The surrogate model outputs the mean of the predicted objective function. With variance : ; In the formula, This is a column vector of historical observations; The covariance matrix among historical samples; This is the column vector of covariance between the sample to be tested and the historical samples; The Matern5 / 2 kernel function was selected to balance smoothness and local mutations. The system observation noise variance; The identity matrix is ​​used. Expected Improvement (EI) is chosen as the acquisition function to balance the exploration of regions with high predictive mean with the search for regions with high predictive variance. The EI calculation formula is as follows: ; In the formula, The optimal objective function value in the current historical dataset; To explore the equilibrium constant (normally set to 0.01); and These are the cumulative distribution function and probability density function of the standard normal distribution, respectively. These are dimensionless standardized variables. Numerical optimization algorithms (such as the L-BFGS-B algorithm) are used in the predefined search space. Maximize the inner acquisition function to find the optimal strategy: The obtained The parameters are overwritten into the execution memory of the process control system and used as the initial deviation compensation gain matrix execution parameters for the next steel coil production cycle, thereby completing the closed-loop self-learning optimization of the parameters. The statistical mean required by the CBN module in the model... with standard deviation Weighted updates are performed synchronously using a sliding window mechanism based on the latest batch of image data for that period.

[0083] Finally, step S600 displays the output results in real time on the industrial control computer interface: current steam parameters, CQS / CUI. For defects judged as unqualified (confidence level > 0.7 and defect area exceeding the threshold), the system sends a rejection signal to the production line control system via PLC, triggering subsequent automatic sorting of defective steel coils. Each inspection record is stored in an HDF5 format file, including: coil number, timestamp, steel type, thickness, speed; steam parameter vector. (Actual values); Condensation Quality Evaluation Index (CQS, CUI); Compressed package of abnormal image slices from the linear array camera and four local images from the area array camera (for offline analysis and model iteration).

[0084] As an optional embodiment, the steam generator can be replaced with an ultrasonic atomizing device to produce micron-sized water mist instead of wet steam, which is suitable for heat-sensitive products such as coated steel sheets.

[0085] As an optional implementation: For offline sampling inspection scenarios, a lightweight model is run using a handheld steam gun, a portable industrial camera, and a laptop computer to achieve mobile inspection. In this case, multi-channel dynamic imaging is simplified to manually capturing different moments with a single camera and a mobile phone stopwatch, and the time warp mapping can be simplified to linear interpolation.

[0086] Example 2

[0087] As a second embodiment of the present invention, such as Figure 3 As shown in Example 1, this example also discloses a steel strip surface defect detection system that integrates steam parameter adjustment and image recognition. Specifically, it includes a parameter acquisition and calculation module, a physical development control module, a multimodal image acquisition module, a defect detection and evaluation module, a dual-mode decoupling feedback module, and a result output and rejection module. The specific implementation is as follows: Parameter acquisition and calculation module: Used to collect surface state parameters of the steel strip in real time, and calculate the initial steam parameter configuration vector based on the surface state parameters. As a physical reference, this module communicates with the production line's process control system to read the steel grade code of the current steel strip. ,thickness and production line speed Simultaneously, the infrared thermal imager and laser triangulation roughness tester, installed in front of the inspection station, are connected to the hardware to collect the apparent temperature of the steel strip in real time. With surface roughness The module internally runs an algorithm program based on heat transfer compensation, which integrates thermal resistance compensation coefficients. Correct the apparent temperature to the true surface temperature. Furthermore, this module combines the current operating condition deviation vector. The deviation compensation gain matrix determined by open-loop step response calibration Calculate and output the steam temperature Steam dryness Steam flow rate per unit width Pre-cooling temperature difference and nozzle distance Initial steam parameter configuration vector .

[0088] Physical development control module: used to configure the precooling temperature difference based on the initial steam parameters. The steel strip is pre-cooled in sections using water-cooled rollers, and wet steam is sprayed onto the surface to form a condensation and development pattern. Specifically, this module includes a high-heat-capacity contact water-cooled roller and a steam generator. The surface of the water-cooled roller is coated with a high-thermal-conductivity chromium layer and undergoes femtosecond laser micro-nano texturing to create a biomimetic superhydrophobic structure to prevent early condensation of ambient moisture. The water-cooled roller operates based on the volumetric flow rate of the feedforward cooling water. With the pre-calibrated heat exchange efficiency coefficient Control commands were issued to slowly pre-cool the steel strip to the target pre-cooling temperature. The corrected flow rate is calculated by combining the discrete position PI control algorithm. Steady-state drift compensation is performed; the steam generator is connected to a wide, flat nozzle with a secondary heating belt and a rectifier grid, and continuously sprays wet steam onto the pre-cooled steel strip surface according to the set gas phase parameters, so that the low-contrast defect area presents a high-contrast physical condensation difference pattern.

[0089] Multimodal image acquisition module: used to acquire full-area condensation development patterns and extract anomalous image slices, as well as trigger the acquisition of temporal local images at fixed spatial nodes to output standard time-domain feature sequences. In practice, this module is divided into a main detection channel and an auxiliary dynamic channel. The main detection channel is determined by the encoder line frequency. The system consists of a synchronously driven array of multiple linear scan cameras. The full-frame image data stream is fed into the FPGA preprocessing unit, where a dynamic threshold multiplier is introduced. The adaptive threshold segmentation and connected component labeling algorithm is used for lightweight ROI pre-screening to crop out abnormal image slices; the auxiliary dynamic channel includes multiple large-view field-array cameras, and the trigger pulse count value is recorded using longitudinal pulse hardware locking. Simultaneously, the linear array image of the main detection channel is reused to extract subpixel-level edge coordinates. To calculate real-time lateral physical coordinates This module completes lateral software deviation compensation and ROI relocation. Subsequently, it executes a Gaussian kernel-based temporal resampling interpolation algorithm, using a resampling weight matrix... Eliminate the time misalignment caused by linear velocity fluctuations, and synchronize the actual physical time of each area array camera. Align the acquired time-series images and output a standard time-domain feature sequence. .

[0090] Defect detection and evaluation module: used to compare abnormal image slices with standard time-domain feature sequences. The input target detection model performs inference and outputs defect confidence, global condensation quality score (CQS), and local condensation uniformity index (CUI). In practice, this module is deployed in an industrial control computer equipped with a high-performance GPU. Internally, it runs an improved YOLO-Cond deep learning network, extracting features through a parallel dual-branch feature extraction module. The high-frequency branch employs a feature extraction module with… and Trainable edge-enhanced convolutions with weights; and using a conditional batch normalization (CBN) mechanism to utilize standardized parameters. The scaling factor and offset are dynamically generated. This module ultimately utilizes a learnable log-variance parameter constructed based on homoscedasticity uncertainty through the evaluation subnetwork of the detection head and its side branches. Dynamically balancing multi-task loss function Simultaneously calculate the defect category, confidence level, and objective scores of CQS and CUI used to characterize the current physical condensation state.

[0091] Dual-mode decoupled feedback module: Used to run a fast loop to adjust gas-phase steam parameters and a slow loop to adjust cooling water flow based on evaluation indicators, and to update the initial steam parameter configuration vector in a closed loop. In practical implementation, this module decouples the control of actuators with different thermal inertia. A fast-cycle loop (second-level) directly sends PLC commands to control the steam generator to cope with high-frequency environmental disturbances; a slow-cycle loop (ten-second-level) extracts the exponentially decaying forgetting factor. The nonlinear dynamic adjustment trigger threshold calculated by combining area heat flux A static steady-state decoupling matrix is ​​adopted. Compared with incremental multivariable PID control law (including gain) ), uniformly calculate the cooling water flow correction vector for each zone This is to mitigate the systematic lateral thermal drift deviation between the edge and center of the steel strip and eliminate thermal coupling interference.

[0092] The results output and rejection module displays the inspection results, triggers rejection signals for non-conforming defects, and stores the data in a database for offline optimization. In practice, this module renders the defect location box and refreshes the CQS / CUI trend chart in real time on the industrial computer monitor. When the defect confidence level exceeds a set threshold and the condensation quality assessment is reliable, an asynchronous rejection signal is sent to the production line control system via the industrial communication bus. Simultaneously, all-element inspection data is packaged and stored in the database. After the production of a single roll of steel strip is completed, a Bayesian optimization algorithm is triggered to construct a Gaussian process surrogate model to predict the mean of the objective function. With variance By maximizing the expected value, the acquisition function is improved. Solve for the optimal strategy Thus, the deviation compensation gain matrix is ​​achieved. The cross-cycle closed-loop self-learning evolution.

[0093] Since the above system embodiment and the method embodiment of Embodiment 1 are based on the same inventive concept, all operational logic, special working condition processing rules, and technical details in the method embodiment are applicable to the system embodiment. To maintain brevity, these will not be repeated here.

[0094] The preferred embodiments of the invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention.

Claims

1. A method for detecting surface defects in steel strips that integrates steam parameter adjustment and image recognition, characterized in that, include: Step S100: Real-time acquisition of surface state parameters of the steel strip, and calculation of initial steam parameter configuration vector based on the surface state parameters; Step S200: A contact water-cooled roller with a surface treated by femtosecond laser micro-nano texturing is used to pre-cool the steel strip in sections according to the pre-cooling temperature difference in the initial steam parameter configuration, and the steam generator is controlled to spray wet steam onto the surface of the steel strip to form a condensation development pattern. Step S300: The condensation development pattern is acquired across the entire frame by the linear array of cameras in the main detection channel. An online dynamically calibrated threshold multiplier is introduced to perform adaptive threshold segmentation to extract abnormal image slices. At the same time, the sub-pixel edge coordinates acquired by the linear array are reused by the area array of cameras in the auxiliary dynamic channel for deviation compensation. Temporal local images under fixed spatial nodes are acquired and output as standard time domain feature sequences after being mapped by the Gaussian kernel temporal resampling interpolation algorithm. Step S400: Input the abnormal image slices and the standard time domain feature sequence into the conditional batch normalization mechanism and the target detection model with trainable edge enhancement convolution to perform defect detection. Train the multi-task loss function based on uncertainty dynamic weight allocation and output the defect confidence, global condensation quality score and local condensation uniformity index simultaneously. Step S500: Construct a long-short cycle decoupled dual-mode control based on the global condensation quality score and the local condensation uniformity index. The fast circulation loop adjusts the low-inertia steam parameters, while the slow circulation loop adjusts the zoned cooling water flow of the water-cooled roller based on the steady-state decoupling matrix and the incremental multivariable proportional-integral-derivative control algorithm. The initial steam parameter configuration vector is updated in a closed loop. Based on the precooling temperature difference in the initial steam parameter configuration, the steel strip is precooled in zones, including: Based on the steel strip's density, specific heat capacity, thickness, width, linear velocity, and actual surface temperature, as well as the target pre-cooling temperature, the physical properties of the cooling water, and the heat exchange efficiency coefficient of the water-cooled roller system calibrated by inversely solving the heat balance equation in the open-loop state, the basic cooling water volume flow rate required to reach the target pre-cooling temperature is calculated. A control dead zone is set at the steel strip temperature at the water-cooled roll outlet. When the temperature deviates from the target pre-cooling temperature and exhibits a steady-state drift trend exceeding the control dead zone, a discrete position proportional-integral control algorithm is used to calculate and superimpose the corrected flow rate to the basic cooling water volume flow rate. The proportional and integral gain coefficients of the control algorithm are pre-tuned using an open-loop single-step response test curve. The transient temperature deviation remaining in the control dead zone after the water-cooled roller is pre-cooled is used as a feedforward disturbance and input into the fast circulation loop. The steam generator then performs adaptive compensation by fine-tuning the steam dryness and flow rate. The construction and evaluation of the target detection model in step S400 includes: A parallel dual-branch feature extraction module is designed at the input end. The first branch extracts the global low-frequency features of water film uniformity, and the second branch is cascaded with trainable edge-enhancing convolution to extract the high-frequency features of droplet and defect edges. The trainable edge-enhancing convolution is set as a discrete difference operator matrix during the initialization phase, and its weights are dynamically updated during the training phase based on the gradient of the loss function through backpropagation. The initial steam parameter configuration vector is dimensionlessly standardized and used as a conditional input to the conditional batch normalization module of the backbone network. The scaling factor and offset of the normalization layer are dynamically generated through learnable projection weights. A parallel condensation quality assessment sub-network is added next to the detection head, which splices and fuses the multi-scale spatial feature map with the flattened standard time domain feature sequence to simultaneously output the global condensation quality score and the local condensation uniformity index. During the model training phase, a learnable log-variance parameter based on homoscedasticity uncertainty is introduced to adaptively and dynamically allocate weights to the multi-task loss function, which includes target detection loss and condensation quality regression loss. The judgment logic after outputting the evaluation index in step S400 includes: Pre-set high global condensation quality threshold, low global condensation quality threshold, high local condensation uniformity threshold, and low local condensation uniformity threshold; If the current global condensation quality score is greater than or equal to the high global condensation quality threshold and all components of the local condensation uniformity index are greater than or equal to the high local condensation uniformity threshold, then the condensation quality is determined to be good and the test result is reliable. If the current global condensation quality score is less than the low global condensation quality threshold or any component of the local condensation uniformity index is less than the low local condensation uniformity threshold, the condensation quality is determined to be unqualified, triggering an alarm in the characterization system and performing feedback adjustment. If the current evaluation index falls between the two aforementioned judgment conditions, the current state is recorded and the poor condensation area is locked according to the spatial distribution of the local condensation uniformity index. The locked information is fed back for subsequent adjustment while the detection result is output.

2. The method for detecting surface defects in steel strips by integrating steam parameter adjustment and image recognition according to claim 1, characterized in that, Step S100 includes: Real-time reading of current production status parameters and apparent temperature of steel strip collected by infrared thermometer; A heat transfer compensation formula is constructed. By extracting a comprehensive thermal resistance compensation coefficient that includes thermal diffusivity and contact thermal resistance, and combining the real-time production thickness of the steel strip, the real-time running line speed, and the temperature difference between the actual measured temperature and the set temperature of the internal circulating cooling water, the apparent temperature is corrected to the true surface temperature of the steam sprayed surface. The initial steam parameter configuration vector is constructed by steam temperature, steam dryness, steam flow rate per unit width, precooling temperature difference and nozzle distance. The deviation is adjusted by combining the pre-calibrated basic parameter vector with the real-time operating condition deviation vector and the deviation compensation gain matrix. The initial steam parameter configuration vector is obtained by boundary truncation through the amplitude limiting function. The deviation compensation gain matrix is ​​based on the system identification method of open-loop step response experiment to avoid closed-loop data cyclic dependence.

3. The method for detecting surface defects of steel strips integrating steam parameter adjustment and image recognition according to claim 1, characterized in that, Step S300 involves extracting abnormal image slices and outputting a standard time-domain feature sequence, including: Adaptive threshold segmentation is performed line by line on the full-frame image. Offline benchmark calibration is combined with real-time global grayscale standard deviation online calibration to generate dynamic threshold multipliers. Abnormal pixel intervals are determined based on dynamic threshold multipliers and local sliding window statistics. Independent connected components are extracted to generate abnormal image slices. Extract image data synchronously acquired by the line scan camera array, calculate sub-pixel level edge coordinates based on the first-order gradient of grayscale and parabolic fitting algorithm, and calculate the actual lateral physical coordinates in combination with the physical spatial resolution of the line scan camera. Based on this, perform lateral deviation dynamic compensation and pixel coordinate repositioning and cropping on the field of view target area of ​​the auxiliary dynamic channel, and output local images in the time series. The actual physical time of the target area reaching each area array camera is calculated based on the pulse discrete integral solution of the main encoder. The high-dimensional spatial features of the local image are extracted to form the actual time domain feature matrix. The Gaussian kernel time domain resampling interpolation algorithm is used to linearly map it into the standard time domain feature sequence in real time.

4. The method for detecting surface defects of steel strips integrating steam parameter adjustment and image recognition according to claim 1, characterized in that, The steps for determining the thresholds in step S400 include: Based on a static calibration dataset with absolutely true labels, image samples covering the full range of scores are generated. A mapping function between the global condensation quality score and the expected value of the comprehensive performance evaluation of target detection is fitted. The lower limit of the compliance evaluation required by the process specification is used as the lower limit condition to solve the high global condensation quality threshold. The lowest tolerance evaluation lower limit is used as the upper limit condition to solve the low global condensation quality threshold. A mapping function between the local condensation uniformity index value and the expected value of the false positive rate is established using a defect-free background calibration subset. The upper limit of the base false positive rate under uniform condensation state is used as the lower limit condition to solve for the high local condensation uniformity threshold. The critical threshold of the false positive rate rejection when the system is unusable due to interference is used as the upper limit condition to solve for the low local condensation uniformity threshold.

5. The method for detecting surface defects of steel strips integrating steam parameter adjustment and image recognition according to claim 1, characterized in that, The long-short cycle decoupled dual-mode control in step S500 includes: The fast circulation loop is set in seconds to perform high-frequency regulation and control of the steam phase parameters that do not include the pre-cooling temperature difference. When the global condensation quality score is too low or the continuous decrease exceeds the limit, the regulation is triggered, prioritizing the increase of steam flow rate, the reduction of steam dryness and the fine adjustment of steam temperature. The slow loop is set with a time interval of ten seconds, and a weighted moving average filter formula with an exponentially decaying memory structure is introduced to remove high-frequency oscillation characteristics and solve the steady-state uniformity error corresponding to each partition. Among them, the exponential decay forgetting factor is strictly mapped by the inherent thermal time constant of the water-cooled roller system; when the steady-state uniformity error of any zone exceeds the adjustment trigger threshold, the control law combining the static steady-state decoupling matrix and the incremental proportional integral derivative control algorithm is used to uniformly solve the cooling water flow correction vector of each zone in order to eliminate the thermal coupling interference between adjacent cooling zones. The decoupling matrix and control parameters are pre-tuned during open-loop testing.

6. The method for detecting surface defects in steel strips by integrating steam parameter adjustment and image recognition according to claim 5, characterized in that, The steps for determining the trigger threshold in the slow loop include: The real-time thickness and real-time running linear speed of the steel strip during the production process are obtained to characterize the real-time area heat flux under the current working conditions, and the nominal parameters are extracted to characterize the reference area heat flux. The relative drift is obtained by comparing the real-time area heat flux with the reference area heat flux, and the system’s preset reference control dead zone is dynamically adjusted nonlinearly by combining the thermal inertia sensitivity adjustment coefficient. When the real-time area heat flux is higher than the reference area heat flux, the reference control dead zone is narrowed nonlinearly; when the real-time area heat flux is lower than the reference area heat flux, the reference control dead zone is widened nonlinearly. The dead zone value after nonlinear dynamic scaling is combined with physical and process boundaries to determine the adjustment trigger threshold.

7. A steel strip surface defect detection system integrating steam parameter adjustment and image recognition, employing the method described in any one of claims 1 to 6, characterized in that, include: The parameter acquisition and calculation module is used to collect the surface state parameters of the steel strip in real time, and calculate the initial steam parameter configuration vector as a physical reference by combining the deviation compensation gain matrix of the open-loop calibration. The physical development control module is used to perform slow-response zoned precooling compensation on the steel strip through a water-cooled roller with surface micro-nano textured according to the precooling temperature difference in the initial steam parameter configuration, and to spray wet steam onto the surface to form a condensation development pattern. The multimodal image acquisition module is used to acquire the developed pattern across the entire frame and extract abnormal image slices based on the dynamic threshold multiplier. It reuses the sub-pixel edge coordinates of the line array camera to perform deviation compensation and capture temporal local images. It outputs a standard time domain feature sequence based on the Gaussian kernel time domain resampling interpolation algorithm. The defect detection and evaluation module is used to run the target detection model, which includes trainable edge-enhancing convolution and conditional batch normalization mechanism, for inference. It is trained using a multi-task loss function with adaptive weight allocation based on uncertainty parameters, and outputs defect confidence, global condensation quality score and local condensation uniformity index. The dual-mode decoupling feedback module is used to run a fast circulation loop to adjust the gas phase steam parameters according to the evaluation index, and to run a slow circulation loop to adjust the cooling water flow rate based on the steady-state decoupling matrix and a multivariable incremental algorithm.

Images