Coating high-precision deviation detection method based on sub-pixel edge fitting

Abstract

The invention relates to the technical field of image processing, in particular to a coating high-precision deviation detection method based on sub-pixel edge fitting. The method comprises the following steps: acquiring an initial image of a coating product in a coating production process, determining a target area containing a coating edge from the initial image, and performing convolution processing on each row of scanning lines in the target area to obtain a gradient field matrix; for each row of the gradient field matrix, based on statistical characteristics of local gradient distribution, determining sub-pixel edge coordinates so as to construct an edge track point set; constructing an edge fitting model, and constructing a target function based on the edge track point set and the edge fitting model; and determining a model parameter of the trend component by minimizing the objective function, and calculating a theoretical edge position at the current moment by using the determined model parameter of the trend component so as to determine a deviation detection result of the coating edge. According to the technical scheme, a more accurate deviation detection result of the coating machine can be obtained.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and in particular to a high-precision coating deviation detection method based on sub-pixel edge fitting. Background Technology

[0002] In industries such as adhesive tape manufacturing, label printing, hygiene product material lamination, and functional film processing, coating process is a key factor in determining product quality. Coating machines such as hot melt adhesive coating machines, self-adhesive coating machines, tape coating machines, and hot melt adhesive laminating machines all work on the core principle of using a mechanical transmission system to drive substrates such as release paper, PET film, and non-woven fabric at high speed, and using a slit die or anilox roller to evenly coat the surface of the substrate with liquid adhesive.

[0003] The edge neatness of the adhesive layer and the lateral positional accuracy of the coating relative to the substrate are important indicators for measuring coating quality. If coating deviation occurs, it may cause adhesive to overflow and contaminate the roller, increasing the frequency of machine shutdown for cleaning, and may also lead to lamination misalignment, insufficient effective width of the product, or uneven winding end face.

[0004] During high-speed operation, equipment such as hot melt adhesive coating machines and tape coating machines inevitably generate complex vibration interference, such as radial runout of the coating roller, meshing frequency of the gearbox, and even tension fluctuations of the substrate itself.

[0005] The detection algorithms in related technologies cannot distinguish between high-frequency vibrations caused by mechanical factors and actual coating deviations. They may misjudge instantaneous vibration peaks as deviation signals, thereby triggering correction actuators such as servo motor push rods to make frequent adjustments, which seriously affects the straightness of the coating edge. Therefore, a detection solution is needed that can adapt to the optical characteristics of special media such as hot melt adhesives and accurately separate the actual deviation trend from mechanical vibration noise. Summary of the Invention

[0006] To obtain more accurate deviation detection results for coating machines, this application provides a high-precision deviation detection method for coating based on sub-pixel edge fitting, comprising: acquiring an initial image of the coated product during the coating production process, determining a target region containing the coating edge from the initial image, performing convolution processing on each row of scan lines within the target region to obtain a gradient field matrix reflecting the energy distribution of the physical position of the coating edge; for each row of the gradient field matrix, calculating the sub-pixel edge coordinates using the spatial moment method based on the statistical characteristics of the local gradient distribution, and constructing a set of edge trajectory points distributed along the coating direction from multiple rows of sub-pixel edge coordinates. An edge fitting model is constructed, and an objective function is built based on the edge trajectory point set and the edge fitting model. The edge fitting model is composed of a trend component that characterizes the actual deviation and a vibration component that characterizes the mechanical disturbance. The objective function includes a data fidelity term and a regularization term. The regularization term is used to suppress the attribution of high-frequency drastic changes to the trend component based on physical inertial constraints. The model parameters of the trend component are determined by minimizing the objective function. The theoretical edge position after removing the vibration component at the current moment is calculated using the determined model parameters of the trend component. The deviation detection result of the coating edge is determined using the theoretical edge position.

[0007] This avoids the interference of mechanical vibration and noise on the deviation detection of the coating machine, and obtains more accurate deviation detection results for the coating machine.

[0008] Optionally, convolution processing is performed on each scan line within the target region using a smoothing differential kernel; the smoothing differential kernel is constructed using the following formula: ,in, For smooth differential kernel, The distance variable is relative to the center of the convolution. Let be the standard deviation parameter of the Gaussian distribution. The numerical distribution of the smoothing differential kernel corresponds to the Gaussian pulse signal obtained by differentiating the S-shaped edge curve. It is a natural exponential function.

[0009] In this way, by using Gaussian derivative kernels for convolution, matched filtering can be performed on the unique S-shaped gray-scale transition characteristics of hot melt adhesive or tape layers, which can suppress image noise while enhancing the signal-to-noise ratio of edge gradients.

[0010] Optionally, based on the statistical characteristics of the local gradient distribution, the sub-pixel edge coordinates are calculated using the spatial moment method, including: searching for gradient peaks in each row of the gradient field matrix to determine integer coordinates, and extracting a local neighborhood interval centered on the integer coordinates; using the gradient distribution within the local neighborhood interval as a probability density function, calculating the zeroth moment and first moment of the gradient distribution within the local neighborhood interval; and using the ratio of the first moment to the zeroth moment of the gradient distribution within the local neighborhood interval to determine the sub-pixel edge coordinates of each row.

[0011] Optionally, before constructing the edge fitting model, the method further includes: using the absolute value of the differential change of the edge positions of two adjacent rows and the product of the position weight function as the row feature value, and using the average value of the row feature values ​​corresponding to all rows as the fluctuation feature parameter; using the fluctuation feature parameter to determine whether the current edge fluctuation is dominated by high-frequency vibration or low-frequency deviation, so as to adjust the initial value of the edge fitting model.

[0012] In this way, by pre-analyzing the spectral characteristics of edge fluctuations and adaptively adjusting the initial values ​​of the model's iterations, the algorithm avoids getting trapped in local optima and improves its versatility across different machine types, such as self-adhesive coating machines and hot melt adhesive laminating machines.

[0013] Optionally, the objective function is constructed using the following formula: ,in, Let be the objective function. The number of sub-pixel edges, This represents the observed sub-pixel edge coordinates. These represent the basic intercept parameter, slope parameter, and curvature parameter, respectively. These represent the vibration amplitude, spatial frequency, and initial phase, respectively. This represents the regularization coefficient, which is set based on the inertia of the coating machine's physical system.

[0014] In this way, the regularization term in the objective function forces the model to follow the inertial laws of the physical system, ensuring that the separated trend components are smooth and reliable, and avoiding erroneous control commands caused by overfitting.

[0015] Optionally, the method further includes: calculating the product of the curvature parameter and the row coordinate at the current time, and using the sum of the product and the inclination parameter as the deviation change rate index at the current time; the deviation change rate index is used to characterize the speed at which the edge position changes with the travel distance; and outputting a warning signal when the deviation detection result of the coating edge is within the tolerance range and the deviation change rate index continues to increase in the neighborhood time period.

[0016] Optionally, before determining the model parameters of the trend component by minimizing the objective function based on the edge trajectory point set, the method further includes: performing a multi-iteration screening process, randomly selecting a subset of points from the edge trajectory point set in each iteration, using the subset of points to determine a temporary straight line model, and counting the number of interior points in the edge trajectory point set whose distance to the temporary straight line model is less than a preset distance threshold; using the temporary straight line model with the most interior points as the benchmark linear model, determining the distance residuals from each point in the edge trajectory point set to the benchmark linear model, and marking points whose distance residuals are greater than a preset distance threshold as outliers, and removing or reducing the weight of outliers when constructing the objective function.

[0017] In this way, by using the random sampling consensus algorithm to eliminate outliers caused by glue bubbles or substrate dirt, the robustness of the algorithm in detecting deviations in harsh industrial environments is improved.

[0018] Optionally, the theoretical edge position after removing the vibration component at the current moment is calculated using the model parameters of the determined trend component, including: determining the image row coordinates corresponding to the latest acquisition moment; extracting the basic intercept parameter, slope parameter and curvature parameter from the determined model parameters, and constructing a second-order polynomial function containing only the trend component; substituting the image row coordinates as independent variables into the second-order polynomial function for calculation, and using the result as the theoretical edge position.

[0019] Optionally, the deviation detection result of the coating edge is determined using the theoretical edge position, including: obtaining the preset baseline coordinates, which are used to characterize the ideal edge position when the coated product is qualified; calculating the difference between the theoretical edge position and the preset baseline coordinates to obtain the pixel-level position deviation; obtaining the camera pixel equivalent, multiplying the pixel-level position deviation by the camera pixel equivalent to obtain the deviation detection result of the coating edge expressed in physical length units; the camera pixel equivalent is used to convert the pixel value in the image domain into the length value in the physical domain.

[0020] Optionally, the method further includes: pre-configuring corresponding algorithm parameter packages for coating products of different specifications; the algorithm parameter packages include at least the preset baseline coordinates for the specified products, the half-width parameter of the gradient calculation window, the initial value of the spatial frequency of the vibration component, and the regularization coefficient; in response to receiving production work order information, automatically retrieving the corresponding algorithm parameter packages to configure the edge fitting model and the smoothing differential kernel.

[0021] The technical solutions provided by the embodiments of this application may include the following beneficial effects: by establishing a physical model that includes trend terms and vibration terms, and using an optimization algorithm with inertial constraints to solve it, the mechanical high-frequency vibration interference unique to equipment such as hot melt adhesive coating machines can be effectively filtered out, the real low-frequency deviation trajectory can be restored from the noisy edge signals, and more accurate deviation detection results can be obtained for the coating machine, which helps to achieve more stable deviation correction control for the coating machine.

[0022] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating a high-precision coating deviation detection method based on sub-pixel edge fitting according to an exemplary embodiment;

[0024] Figure 2 This is a schematic diagram illustrating the principle of subpixel edge fitting and component separation. Figure 3 This is a schematic diagram illustrating the detection effect of vibration interference in the embodiments of this application. Detailed Implementation

[0025] First, a brief introduction to the application scenarios of the embodiments of this application will be given. The application scenarios of the embodiments of this application can be the production site of various high-precision coating and laminating equipment, such as the production line of products such as hot melt adhesive coating machine, self-adhesive coating machine, hot melt adhesive film coating machine, tape coating machine and hot melt adhesive laminating machine.

[0026] Hot melt adhesive coating machines typically apply high-viscosity adhesives to substrates at high temperatures. Due to the rheological properties of hot melt adhesives and the high-speed relative movement of the coating head, the edges are highly susceptible to hydraulic pulsation and mechanical vibration. Self-adhesive coating machines and tape coating machines, on the other hand, have high requirements for the uniformity of coating thickness and the neatness of edges. Deviation may cause adhesive seepage or adhesion in the finished tape after slitting. Therefore, it is necessary to detect deviation during the operation of the coating machine.

[0027] To address the aforementioned technical problems, this application provides a high-precision coating deviation detection method based on sub-pixel edge fitting. Figure 1 This is a flowchart illustrating a high-precision coating deviation detection method based on sub-pixel edge fitting according to an exemplary embodiment, such as... Figure 1 As shown, the method includes the following steps.

[0028] In step S101, an initial image of the coated product is obtained during the coating production process, and a target area containing the coating edge is determined from the initial image. Each scan line within the target area is convolved to obtain a gradient field matrix that reflects the energy distribution of the physical location of the coating edge.

[0029] Initial images of coated products can be obtained using an industrial line scan camera mounted at the exit of the coating machine; for hot melt adhesive coating machines or hot melt adhesive laminating machines, which are subject to high temperature, high humidity, and possible oil mist environments, the camera can be equipped with a protective cover with an air purging function.

[0030] The resolution of the line scan camera can be set to 4096 pixels or 8192 pixels, and the scanning frequency can be set to 20kHz or higher to match the belt speed of the coating machine from 150 m / min to 300 m / min.

[0031] For transparent hot melt adhesive films, a high-brightness coaxial light source can be used for lighting to enhance edge contrast by utilizing the scattering characteristics of short-wavelength light on the adhesive surface; for colored self-adhesive coatings, a white backlight with high color rendering index can be used.

[0032] After acquiring the full-frame image data, dynamic locking of the target area containing the coating edge can be performed first. Since the coating width may vary between 1000mm and 2000mm, the amount of data processed for the entire image is too large. Therefore, the image can be downsampled by 8 times first, and the Otsu method adaptive threshold algorithm can be used to quickly locate the rough boundary between the adhesive layer and the substrate.

[0033] Centered on a rough boundary line, a strip-shaped region with a width of 64 pixels on each side is extracted from the original high-resolution image as the target region. The hierarchical localization strategy can reduce the amount of data processing to less than 5% of the entire image while ensuring accuracy, and can also meet real-time requirements.

[0034] In one embodiment, convolution processing of each scan line within the target region is achieved using a smoothing differential kernel; the smoothing differential kernel is constructed using the following formula: ,in, For smooth differential kernel, The distance variable is relative to the center of the convolution. Let be the standard deviation parameter of the Gaussian distribution. The numerical distribution of the smoothing differential kernel corresponds to the Gaussian pulse signal obtained by differentiating the S-shaped edge curve. It is a natural exponential function.

[0035] In practical implementation, the formula for calculating the smoothing differential kernel is used to generate a discrete finite-length impulse response filter; for example, for hot melt adhesives with viscosities of 2000 cps to 5000 cps, the width of the edge transition zone formed on the substrate is typically about 3 to 5 pixels; to match this physical characteristic, the standard deviation parameter of the Gaussian distribution can be set. It is 1.5.

[0036] According to statistics Principles, distance variables The range of values ​​is set to The set of integers, by Substituting these values ​​into the above formula for the smoothing differential kernel, a convolution kernel vector containing 11 floating-point values ​​can be generated.

[0037] During the coating process, hot melt adhesives or tapes, due to the combined effects of surface tension and fluid viscosity, result in a meniscus with a certain curvature at the microscopic scale at their physical edges, rather than an ideal vertical sectional surface. This causes the grayscale distribution at the edge position on optical imaging sensors to exhibit a smooth S-shaped (e.g., Sigmoid function) characteristic rather than a step characteristic.

[0038] If conventional Sobel or Prewitt operators are used, these operators are essentially differential approximations and are extremely sensitive to high-frequency noise. In the hot melt adhesive coating process, fiber burrs on the substrate surface, tiny air bubbles in the adhesive, and the jitter of the lighting spot will all generate high-frequency noise, causing the differential operator to output erroneous pseudo-edge peaks.

[0039] By convolving the image signal with a kernel function, the image signal is first smoothed and denoised using a Gaussian function to filter out substrate texture and sensor thermal noise. The first derivative of the smoothed signal is then calculated to extract the edge gradient, so that the kernel function can maximize the edge response energy while minimizing the noise response.

[0040] Thus, for each row of pixel data within the target area, such as a grayscale array containing 128 pixels, a one-dimensional sliding convolution operation can be performed using the generated convolution kernel vector to obtain a new numerical array, namely a row of the gradient field matrix.

[0041] In the gradient array, the absolute value of the value represents the probability that the position is an edge. This can transform the originally blurry and noise-saturated transparent adhesive layer edge into a high signal-to-noise ratio gradient signal with more significant single-peak characteristics.

[0042] In step S102, for each row of the gradient field matrix, based on the statistical characteristics of the local gradient distribution, the sub-pixel edge coordinates are calculated using the spatial moment method, and the set of edge trajectory points distributed along the coating direction is constructed from the sub-pixel edge coordinates of multiple rows.

[0043] In one embodiment, based on the statistical characteristics of the local gradient distribution, the sub-pixel edge coordinates are calculated using the spatial moment method, including: searching for gradient peaks in each row of the gradient field matrix to determine integer coordinates, and extracting a local neighborhood interval centered on the integer coordinates; using the gradient distribution within the local neighborhood interval as a probability density function, calculating the zeroth moment and the first moment of the gradient distribution within the local neighborhood interval; and using the ratio of the first moment to the zeroth moment of the gradient distribution within the local neighborhood interval to determine the sub-pixel edge coordinates of each row.

[0044] For any nth element in the gradient field matrix To find the coarse edge point, we can first iterate through the gradient values ​​of the row and find the point with the largest value. Let the integer coordinates be... ;by Centered on the data, extract samples to the left and right respectively. For example, a pixel This forms a local neighborhood interval with a width of 7 pixels. .

[0045] Within this local interval, let the first... The coordinates of each pixel are The corresponding gradient value is Then we can continue to calculate the zeroth moment of the gradient distribution within the local neighborhood interval. The calculation formula is , This represents the total gradient energy within that local region.

[0046] Calculate the first moment , This represents the sum of the torques of the gradient energy relative to the origin, expressed as a ratio. Subpixel edge coordinates can be calculated.

[0047] The image sensor of an industrial camera is composed of discrete pixel arrays. If directly taking... As an edge position, it means that the measurement error is uniformly distributed, with a maximum error of 0.5 pixels. For precision equipment such as tape coating machines, an error of 0.5 pixels may correspond to a physical deviation of tens of micrometers, which makes it impossible to meet the needs of high-end manufacturing.

[0048] Relying solely on peak points is susceptible to interference from single-point noise, such as a bad pixel in an image sensor falling near an edge. The spatial moment method introduces statistical concepts, enabling the examination of the energy distribution within the entire neighborhood surrounding the peak point.

[0049] If the actual edge position is slightly biased to the right, the gradient value of the pixel adjacent to the right of the peak will usually be higher than that of the pixel adjacent to the left. By calculating the centroid of the gradient, this slight asymmetry can be keenly captured, and the true center position of the light signal can be restored through interpolation.

[0050] In this way, by performing the above calculations on each scan line, a series of coordinate points with higher precision can be obtained, such as theoretical precision of 0.1 pixels or even 0.05 pixels. Multiple continuously collected lines, such as 1024 continuously collected lines, corresponding to a coated product of about 20 centimeters in length, are arranged in chronological order to form a discrete time series signal, namely the edge trajectory point set.

[0051] The obtained edge trajectory point set records every jitter and drift of the coating edge at the microscale, containing rich process status information.

[0052] In step S103, an edge fitting model is constructed, and an objective function is constructed based on the edge trajectory point set and the edge fitting model.

[0053] The edge fitting model is composed of the superposition of the trend component, which characterizes the actual deviation, and the vibration component, which characterizes the mechanical disturbance. The objective function includes a data fidelity term and a regularization term. The regularization term is used to suppress the attribution of high-frequency drastic changes to the trend component based on physical inertial constraints.

[0054] In one embodiment, before constructing the edge fitting model, the product of the absolute value of the differential change of the edge positions of two adjacent rows and the position weight function can be used as the row feature value, and the average value of the row feature values ​​corresponding to all rows can be used as the fluctuation feature parameter. The fluctuation feature parameter is used to determine whether the current edge fluctuation is dominated by high-frequency vibration or low-frequency deviation, so as to adjust the initial value of the edge fitting model.

[0055] This might involve a fast pre-scan of the edge trajectory point set, calculating the absolute value of the sub-pixel coordinate difference between two adjacent rows. To eliminate the influence of lens edge distortion on the measurement results, a position weighting function is introduced. When the edge position is close to the center of the image, the weight is close to 1, and the weight gradually decreases when it is close to the edge of the image.

[0056] Calculate the weighted average As a fluctuation characteristic parameter, an empirical threshold can be preset for a specific model. ;like This means the coating machine is in the acceleration phase or there is abnormal vibration. When initializing the edge fitting model, the amplitude parameter of the vibration component can be used. The initial value is set to a larger value, such as 5 pixels; if This means that the coating machine is operating in a steady state. The initial value is set to a smaller value, such as 0.5 pixels, to focus on fitting the trend component.

[0057] Solving the edge fitting model is a nonlinear least squares optimization problem. Such problems are quite sensitive to the choice of initial values. If the initial value deviation is too large, optimization algorithms such as the Gauss-Newton method may require more iterations to converge, or may even get stuck in local minima, leading to fitting failure.

[0058] For example, under strong vibration conditions, if vibration parameters are not preset, the algorithm may incorrectly attempt to fit the vibration waveform by increasing the coefficients of higher-order polynomials, leading to overfitting. By predicting the current physical conditions, a reasonable starting point can be provided for the algorithm, thereby improving the convergence speed of the algorithm.

[0059] In this way, the coating machine can quickly adapt to the current dynamic characteristics during the start-up acceleration, deceleration stop, or constant speed production stages, ensuring the stability and reliability of the subsequent model fitting process.

[0060] In one embodiment, before determining the model parameters of the trend component by minimizing the objective function based on the edge trajectory point set, multiple iterative screening processes can be performed. In each iteration, a subset of points is randomly selected from the edge trajectory point set, and a temporary straight line model is determined using the subset points. The number of interior points in the edge trajectory point set whose distance to the temporary straight line model is less than a preset distance threshold is counted. The temporary straight line model with the most interior points is used as the baseline linear model. The distance residuals from each point in the edge trajectory point set to the baseline linear model are determined, and points whose distance residuals are greater than the preset distance threshold are marked as outliers. Outliers are removed or their weights are reduced when constructing the objective function.

[0061] The iterative selection process can be based on the principle of random sampling consensus algorithm. For example, in a trajectory set containing 1024 points, two points are randomly selected each time to determine a straight line, and then the vertical distance from the remaining 1022 points to this straight line is calculated.

[0062] If the vertical distance is less than a preset threshold, such as 2 pixels, these two points can be recorded as interior points. The iteration process can be repeated 50 to 100 times to select the line with the most interior points as the baseline. Points that are more than 3 times the standard deviation from the baseline can be identified as outliers.

[0063] The coating production environment is very complex. Unmelted particles may be mixed in the hot melt adhesive, or tiny air bubbles may be present in the adhesive solution. These defects can cause the edges to be missing adhesive or bulge when passing through the coating die. Dirt and scratches on the substrate surface may also be misidentified as edges.

[0064] These anomalies manifest as isolated, large-amplitude coordinate jumps in the image processing results. Ordinary least squares methods have the characteristic of amplifying errors, and the squared error of a single outlier will dominate the entire objective function, thus severely skewing the fitted curve and causing the system to misjudge that a serious deviation has occurred. By cleaning the data based on the principle of random sampling consensus algorithm, these pseudo-deviation data points caused by process defects can be identified and removed.

[0065] In this way, the cleaned edge trajectory point set contains only effective data reflecting real physical motion, improving the detection algorithm's anti-interference ability and reliability in harsh industrial environments.

[0066] In one embodiment, the objective function is constructed using the following computational formula: ,in, Let be the objective function. The number of sub-pixel edges, This represents the observed subpixel edge coordinates.

[0067] The first part of the objective function The data fidelity term describes the difference between the model's predicted values ​​and the actual observed values. It constitutes a trend component, used to characterize the macroscopic trend of the coating edge.

[0068] Represents the basic position intercept. Represents the angle of inclination of the edge; The curvature parameter of the edge can characterize the degree of curvature of the edge and can be determined based on the difference in curvature of the edge points. The embodiments of this application will not be described in detail here. These constitute vibration components, used to characterize high-frequency interference in mechanical systems; Represents the amplitude of vibration. It represents spatial frequency and is related to mechanical rotational speed. Represents the initial phase.

[0069] The second part of the objective function For regularization terms, The regularization coefficient is a non-negative real number; the macroscopic trajectory of the coating edge will not undergo drastic bending deformation in a short period of time, while rapid reciprocating motion usually originates from mechanical vibration or eccentricity, so a regularization term can be introduced. Penalize the curvature parameter.

[0070] The regularization coefficient can be set based on the physical inertia of the coating machine; the heavier the equipment or the greater the inertia, the better. A larger value, such as 1000, indicates a stronger inhibition of bending.

[0071] In this way, by minimizing the objective function, high-frequency or drastic changes can be attributed to vibration terms by the model. Low-frequency or gradual changes can be attributed to trend terms.

[0072] Figure 2 This is a schematic diagram illustrating the principle of sub-pixel edge fitting and component separation, as shown below. Figure 2 As shown, the microscopic principles of subpixel edge fitting and component separation are illustrated. Figure 2 The horizontal axis represents the local sampling point index, and the vertical axis represents the microscopic location coordinates. Figure 2 The origin in the equation is a discrete sampling point calculated from the gradient field using the spatial moment method. Due to the influence of possible vibrations, these discrete points exhibit random discreteness at the microscale.

[0073] Traditional linear fitting cannot explain the periodic fluctuations in data. This application constructs a superposition model that includes trend and oscillation terms. Figure 2 The dashed line represents the periodic vibration component identified by the model. It can be seen that the vibration component captures the fluctuation frequency and phase of the sampling point data. Figure 2The solid line in the figure represents the fit of the physical motion trend after removing the vibration component, which corresponds to the trend component in the objective function. The trend line smoothly reflects the true direction of the coating edge.

[0074] In step S104, the model parameters of the trend component are determined by minimizing the objective function. The theoretical edge position after removing the vibration component at the current moment is calculated using the determined model parameters of the trend component, so as to determine the deviation detection result of the coating edge using the theoretical edge position.

[0075] In one embodiment, the theoretical edge position after removing the vibration component at the current moment is calculated using model parameters of the determined trend component, including: determining the image row coordinates corresponding to the latest acquisition moment; extracting the basic intercept parameter, slope parameter, and curvature parameter from the determined model parameters, and constructing a second-order polynomial function containing only the trend component; substituting the image row coordinates as independent variables into the second-order polynomial function for calculation, and using the result as the theoretical edge position.

[0076] The specific formula for calculating a second-order polynomial function containing only trend components is: ,in, Given the row coordinates of the latest scanned row, a set of optimal parameters can be obtained by minimizing the objective function. When calculating the final output position, only the trend component parameters are used. Reconstruct the edge positions.

[0077] The actual edge locations observed actually have superimposed layers. If this vibration noise is directly sent to the controller for closed-loop control, the correction actuator, such as a stepper motor or servo hydraulic cylinder, will attempt to track and cancel this high-frequency vibration.

[0078] However, the response bandwidth of mechanical correction mechanisms is usually lower than the mechanical vibration frequency. Forcing the tracking of high-frequency signals is not only physically impossible, but it also leads to frequent reversals, overheating, and wear of the actuator, and may even cause self-excited oscillations of the system, resulting in further deterioration of coating quality.

[0079] By substituting only the trend component for calculation, the output is the ideal edge position after removing mechanical noise. The downstream controller receives a smooth, stable signal that truly reflects the substrate deviation trend, and can make accurate and gentle correction actions accordingly, thus achieving high-precision steady-state control of the coating process.

[0080] Figure 3 This is a schematic diagram illustrating the detection effect of vibration interference in an embodiment of this application, as shown below. Figure 3 As shown, Figure 3 The horizontal axis represents the sampling sequence, and the vertical axis represents the edge displacement. Figure 3 The original sensor signals in the data represent the initial edge position sequence calculated based on the gradient field matrix.

[0081] like Figure 3 As shown, due to the influence of the mechanical transmission system of the coating machine, high-frequency mechanical vibration noise is superimposed on the original sensor signal. Due to the superposition of vibration amplitude, the instantaneous signal peak exceeds the upper limit tolerance of the qualified process window. In related technologies, this situation may be misjudged as a deviation signal, thereby triggering unnecessary deviation correction actions of the coating machine's deviation correction mechanism.

[0082] Figure 3 The thick red solid line in the middle represents the denoised true edge trajectory obtained after processing using the edge fitting model described in this application. By minimizing the objective function containing the regularization term, the high-frequency drastic changes are attributed to vibration components and then stripped away.

[0083] like Figure 3 As shown, although the original sensor signal fluctuates, the fitting trend component after denoising remains within the tolerance and can be judged as normal fluctuation. The embodiments of this application can restore the real low-frequency deviation trajectory from the noisy edge signal, avoid false alarms and false control caused by mechanical vibration noise, and realize steady-state detection of coating edge.

[0084] In one embodiment, the product of the curvature parameter and the row coordinate at the current moment can also be calculated, and the sum of the product and the tilt rate parameter can be used as the deviation change rate index at the current moment. The deviation change rate index is used to characterize the speed at which the edge position changes with the travel distance. If the deviation detection result of the coating edge is within the tolerance range and the deviation change rate index continues to increase in the neighborhood time period, an early warning signal is output.

[0085] Deviation rate index The calculation formula is based on the trend function with respect to... Find the result of the first derivative: Simple position detection can only provide information on how much deviation there is at the current moment, but once deviation occurs in high-speed coating production, it is often an irreversible deterioration process.

[0086] For example, when a tension controller on the unwinding shaft malfunctions, causing uneven stress on the substrate, the absolute position of the edge may still be within the process tolerance range in the early stages of the malfunction. Millimeters could lead to non-conforming products being mistakenly identified as conforming products.

[0087] However, if the speed of lateral drift, i.e. the rate of change, begins to increase abnormally, and the alarm is only triggered and the machine is stopped when the position deviation truly exceeds the tolerance, hundreds of meters of waste have often already been generated. By monitoring the rate of deviation in real time, the control system of the coating machine has the ability to perform predictive maintenance.

[0088] When the rate of deviation is detected to be increasing continuously over a period of time, an early warning signal can be issued to the operator, prompting them to check the tension system or the status of the correction roller, which helps to reduce material loss.

[0089] In one embodiment, determining the deviation detection result of the coating edge using the theoretical edge position includes: obtaining preset baseline coordinates, which are used to characterize the ideal edge position when the coated product is qualified; calculating the difference between the theoretical edge position and the preset baseline coordinates to obtain the pixel-level position deviation; obtaining the camera pixel equivalent, multiplying the pixel-level position deviation by the camera pixel equivalent to obtain the deviation detection result of the coating edge expressed in physical length units; the camera pixel equivalent is used to convert the pixel value in the image domain into the length value in the physical domain.

[0090] For example, if the pixel equivalent of an industrial camera is determined to be 0.035 mm / pixel through calibration using a calibration board, and the preset baseline is located at the 1000th column of the image, while the calculated theoretical edge position is at the 1005.4th column, then the pixel-level deviation is +5.4 pixels, and the final output physical deviation is... Millimeters; the output results have intuitive physical meaning, can be directly understood by on-site process personnel, and can be used as feedback input to the control module of the control system.

[0091] In one embodiment, corresponding algorithm parameter packages can be pre-configured for coating products of different specifications. The algorithm parameter packages include at least the preset baseline coordinates for the specified products, the half-width parameter of the gradient calculation window, the initial value of the spatial frequency of the vibration component, and the regularization coefficient. In response to receiving production work order information, the corresponding algorithm parameter packages are automatically retrieved to configure the edge fitting model and the smoothing differential kernel.

[0092] In actual production, a coating machine often needs to be compatible with multiple product specifications. For example, when producing double-sided tape, the adhesive layer is thin and highly transparent, requiring a wide smooth differential kernel window. When producing foam tape, the edges are relatively rough, requiring a large regularization coefficient to suppress shape noise. By using parameter package management, the problems of cumbersome model changeover and reliance on manual debugging in traditional vision systems can be solved. By communicating with the factory's production execution system, the product part number can be automatically read and the corresponding underlying algorithm parameters can be loaded when the work order is switched.

[0093] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only.

[0094] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope.

Claims

1. A high-precision coating deviation detection method based on sub-pixel edge fitting, characterized in that, include: The initial image of the coated product during the coating production process is obtained, and the target area containing the coating edge is determined from the initial image. Each scan line in the target area is convolved to obtain the gradient field matrix that reflects the energy distribution of the physical location of the coating edge. For each row of the gradient field matrix, based on the statistical characteristics of the local gradient distribution, the spatial moment method is used to calculate the sub-pixel edge coordinates, and the set of edge trajectory points distributed along the coating direction is constructed from the sub-pixel edge coordinates of multiple rows. An edge fitting model is constructed, and an objective function is built based on the edge trajectory point set and the edge fitting model. The edge fitting model is composed of the superposition of the trend component used to characterize the actual deviation and the vibration component used to characterize the mechanical disturbance. The objective function includes a data fidelity term and a regularization term. The regularization term is used to suppress high-frequency drastic changes from being attributed to the trend component based on physical inertia constraints. The model parameters of the trend component are determined by minimizing the objective function. The theoretical edge position after removing the vibration component at the current moment is calculated using the determined model parameters of the trend component. The deviation detection result of the coating edge is then determined using the theoretical edge position.

2. The high-precision coating deviation detection method based on sub-pixel edge fitting according to claim 1, characterized in that, Convolution is performed on each scan line within the target region using a smoothing differential kernel; the smoothing differential kernel is constructed using the following formula: ,in, For smooth differential kernel, The distance variable relative to the convolution center, The standard deviation parameter of the Gaussian distribution, the numerical distribution of the smooth differential kernel corresponds to the Gaussian pulse signal obtained by differentiating the S-shaped edge curve, and exp is the natural exponential function.

3. The high-precision coating deviation detection method based on sub-pixel edge fitting according to claim 1, characterized in that, Based on the statistical properties of local gradient distribution, the sub-pixel edge coordinates are calculated using the spatial moment method, including: Search for gradient peaks in each row of the gradient field matrix to determine integer coordinates, and extract a local neighborhood interval centered on the integer coordinates; use the gradient distribution within the local neighborhood interval as the probability density function, and calculate the zeroth moment and first moment of the gradient distribution within the local neighborhood interval; The sub-pixel edge coordinates of each row are determined by using the ratio of the first moment to the zero moment of the gradient distribution within the local neighborhood interval.

4. The high-precision coating deviation detection method based on sub-pixel edge fitting according to claim 1, characterized in that, Before constructing the edge fitting model, the method further includes: The product of the absolute value of the differential change at the edge positions of two adjacent rows and the position weight function is used as the row feature value, and the average value of the row feature values ​​corresponding to all rows is used as the fluctuation feature parameter. By using the fluctuation characteristic parameters, it can be determined whether the current edge fluctuation is dominated by high-frequency vibration or low-frequency deviation, so as to adjust the initial value of the edge fitting model.

5. The high-precision coating deviation detection method based on sub-pixel edge fitting according to claim 1, characterized in that, The objective function is constructed using the following formula: Where J is the objective function and N is the number of sub-pixel edges. This represents the observed sub-pixel edge coordinates. These represent the basic intercept parameter, slope parameter, and curvature parameter, respectively. These represent the vibration amplitude, spatial frequency, and initial phase, respectively. This represents the regularization coefficient, which is set based on the inertia of the coating machine's physical system.

6. The high-precision coating deviation detection method based on sub-pixel edge fitting according to claim 5, characterized in that, The method further includes: Calculate the product of the curvature parameter and the current line coordinate, and use the sum of the product and the inclination parameter as the deviation change rate index at the current moment; the deviation change rate index is used to characterize the speed at which the edge position changes with the distance traveled. If the deviation detection result at the coating edge is within the tolerance range, and the deviation change rate index continues to increase within the neighborhood time period, an early warning signal will be output.

7. The high-precision coating deviation detection method based on sub-pixel edge fitting according to claim 1, characterized in that, Before determining the model parameters of the trend component by minimizing the objective function based on the edge trajectory point set, the method further includes: The process involves multiple iterations of filtering. In each iteration, a subset of points is randomly selected from the set of edge trajectory points. The subset of points is used to determine a temporary straight line model. The number of interior points in the set of edge trajectory points whose distance to the temporary straight line model is less than a preset distance threshold is counted. The temporary straight line model with the most interior points is used as the baseline linear model. The distance residuals from each point in the edge trajectory point set to the baseline linear model are determined, and points with distance residuals greater than a preset distance threshold are marked as outliers. Outliers are removed or their weights are reduced when constructing the objective function.

8. The high-precision coating deviation detection method based on sub-pixel edge fitting according to claim 1, characterized in that, Calculate the theoretical edge position at the current moment after removing the vibration component using the model parameters of the determined trend component, including: Determine the image row coordinates corresponding to the latest acquisition time; extract the basic intercept parameter, slope parameter, and curvature parameter from the determined model parameters, and construct a second-order polynomial function containing only the trend component; substitute the image row coordinates as independent variables into the second-order polynomial function for calculation, and use the result as the theoretical edge position.

9. The high-precision coating deviation detection method based on sub-pixel edge fitting according to claim 1, characterized in that, The deviation detection results of the coating edge are determined using the theoretical edge position, including: Obtain the preset baseline coordinates, which are used to characterize the ideal edge position when the coated product is qualified; calculate the difference between the theoretical edge position and the preset baseline coordinates to obtain the pixel-level position deviation; The camera pixel equivalent is obtained, and the pixel-level positional deviation is multiplied by the camera pixel equivalent to obtain the deviation detection result of the coating edge expressed in physical length units; the camera pixel equivalent is used to convert the pixel value in the image domain into the length value in the physical domain.

10. The high-precision coating deviation detection method based on sub-pixel edge fitting according to claim 1, characterized in that, The method further includes: For coated products of different specifications, corresponding algorithm parameter packages are pre-configured; the algorithm parameter packages include at least the preset baseline coordinates, half-width parameters of the gradient calculation window, initial values ​​of the spatial frequency of the vibration components, and regularization coefficients for the specified products. In response to receiving production work order information, the corresponding algorithm parameter package is automatically retrieved to configure the edge fitting model and smoothing differential kernel.

Images