Multispectral imaging film coating uniformity on-line detection system

Abstract

A multispectral imaging film coating uniformity on-line detection system comprises the following steps: carrying out full-spectrum response curve test on a standard coating sample, screening basic candidate wavelengths corresponding to different coating materials, carrying out industrial scene test on the basic candidate wavelengths, and determining the uniformity of the coating according to the basic candidate wavelengths. The comprehensive adaptive value of the basic candidate wavelength of each coating material under different thicknesses is obtained; acquiring a sensitive wave band combination of a to-be-detected coating of a current production line based on the comprehensive adaptive value, acquiring a spectral image of each sensitive wave band on the surface of the to-be-detected coating, performing adaptive fitting baseline correction preprocessing on the spectral image of each sensitive wave band, performing spectral feature mining and correlation screening on the spectral image, and acquiring a low-dimensional feature vector; and performing thickness inversion based on the low-dimensional feature vector, generating a coating thickness two-dimensional distribution heat map, performing uniformity judgment on the coating thickness two-dimensional distribution heat map, and outputting a uniformity judgment result. According to the invention, the accuracy and pertinence of coating uniformity detection are obviously improved.

Description

Technical Field

[0001] This invention relates to the field of thin film coating inspection technology, specifically to an online inspection system for thin film coating uniformity using multispectral imaging. Background Technology

[0002] In the thin-film coating process, coating uniformity directly determines the core indicators of the product, such as optical, mechanical, and electrical properties, and is a key factor affecting product yield. For example, insufficient coating uniformity in photovoltaic backsheet films can lead to localized degradation of weather resistance and shortened lifespan; uneven metal coating thickness on semiconductor wafers can affect circuit conductivity; and thickness deviations in optical functional films can cause abnormal transmittance and refractive index, failing to meet the requirements of precision optical equipment. Therefore, achieving high-precision, real-time detection of coating uniformity is of great significance to industrial production.

[0003] Existing coating uniformity testing technologies can be mainly divided into three categories: First, contact point measurement technologies, such as mechanical probe thickness gauges, which obtain single-point thickness data by contacting the coating surface with a probe. This technology has drawbacks such as low detection efficiency, easy damage to the coating surface, and inability to achieve full-coverage detection. It can only be used for sampling inspection and is prone to missing local areas with abnormal uniformity. Second, single-point non-contact spectral detection technologies, such as ellipsometrists and infrared spectral thickness gauges, although they solve the damage problem of contact detection, are still single-point measurement modes with limited detection range. They are difficult to adapt to the real-time detection needs of high-speed production lines and are greatly affected by ambient light and vibration, resulting in insufficient detection robustness. Third, offline laboratory testing technologies, such as atomic force microscopes and laser interferometers, can achieve high-precision thickness measurement, but the detection cycle is long, the operation is complex, and it cannot be linked with the production line in real time. It can only be used for post-production quality traceability and cannot achieve online correction during the production process.

[0004] To address the aforementioned technical shortcomings, existing technologies have developed online inspection solutions based on machine vision. However, traditional visual inspection often employs single-band imaging, which can only identify obvious defects through grayscale differences and cannot accurately invert coating thickness information, making it difficult to quantify uniformity indicators. While some multispectral inspection solutions introduce multi-band spectra, they suffer from insufficient band selection targeting, low efficiency of thickness inversion algorithms, and poor synchronization with production lines. Consequently, the inspection accuracy and speed cannot simultaneously meet industrial-grade requirements, and closed-loop linkage between inspection data and coating equipment cannot be achieved, making it difficult to fundamentally improve product yield. Summary of the Invention

[0005] To address the aforementioned technical problems, the present invention aims to provide an online detection system for thin film coating uniformity using multispectral imaging, comprising a monitoring center, which is communicatively connected to a coating characteristic analysis module, a data acquisition module, and a thickness inversion module. The coating characteristic analysis module is used to perform full-spectrum response curve testing on standard coating samples, screen the basic candidate wavelengths corresponding to different coating materials, and conduct industrial scenario testing on the basic candidate wavelengths to obtain the comprehensive adaptability value of the basic candidate wavelengths of each coating material under different thicknesses. The data acquisition module is used to obtain the sensitive band combination of the coating to be tested on the current production line based on the comprehensive fitness value, acquire the spectral image of each sensitive band on the surface of the coating to be tested, perform adaptive fitting baseline correction preprocessing on the spectral image of each sensitive band, and reconstruct the spectral image of each sensitive band. The thickness inversion module is used to mine spectral features and filter correlations in the reconstructed spectral image, obtain low-dimensional feature vectors, construct a thickness inversion model, perform thickness inversion based on the low-dimensional feature vectors and the thickness inversion model, generate a two-dimensional heat map of coating thickness distribution, calibrate the two-dimensional heat map of coating thickness distribution, determine the uniformity of the calibrated two-dimensional heat map of coating thickness distribution, mark the coating to be detected as qualified or unqualified, and output the thickness anomaly areas of the unqualified coating to be detected.

[0006] Furthermore, the process of performing full-spectrum response curve tests on standard coating samples and screening the basic candidate wavelengths corresponding to different coating materials includes: Obtain standard coating samples covering the target thickness range and the target coating material range, perform full-spectrum reflectance and transmission response curve tests on the standard coating samples, and obtain a test dataset. The test dataset includes the response intensity of samples with different thicknesses of different coating materials at different wavelengths. Statistical analysis was performed on the response intensity of samples with different thicknesses of each coating material in the test dataset at different wavelengths. The response intensity change corresponding to the thickness change of each coating material at different wavelengths was obtained. Based on the response intensity change, each wavelength of each coating material was sorted in ascending order to generate a wavelength queue. The top k wavelengths in the wavelength queue were selected as the base candidate wavelengths.

[0007] Furthermore, the process of conducting industrial scenario testing on the fundamental candidate wavelengths includes: Extract the processing scenario information of the target production line, and conduct scenario simulation tests on the basic candidate wavelengths of different thickness samples of each coating material based on the processing scenario information. Obtain the response intensity variation coefficient of the basic candidate wavelengths of each coating material at different thicknesses under the processing scenario information conditions. Obtain the basic candidate wavelength interval for each coating material, set the index weights corresponding to the response intensity change, response intensity coefficient of variation and basic candidate wavelength interval, and perform a weighted summation of the response intensity change, response intensity coefficient of variation and basic candidate wavelength interval for the basic candidate wavelength of each coating material at different thicknesses to obtain the comprehensive fitness value of the basic candidate wavelength of each coating material at different thicknesses.

[0008] Furthermore, the process of acquiring spectral images of the surface of the coating to be tested in various sensitive bands includes: The thickness range and material of the coating to be tested on the current production line are obtained. The comprehensive fitness value corresponding to several basic candidate wavelengths under the conditions of thickness range and coating material is obtained. The top three basic candidate wavelengths with the largest comprehensive fitness value are selected to construct a sensitive band combination. Based on the sensitive band combination, the coating to be tested is illuminated with multispectral light. The spectral images of each sensitive band on the surface of the coating to be tested are obtained and the acquisition period is marked.

[0009] Furthermore, the process of adaptively fitting baseline correction preprocessing for the spectral images of each sensitive band includes: The spectral images of each sensitive band are preprocessed for denoising. The response intensity of each pixel in the denoised spectral images of each sensitive band is extracted, and the original spectral vector of each pixel is constructed. The original spectral vector is normalized, and then adaptive fitting parameters are set. The following iterative steps are performed on the original spectral vector of each pixel: Step 1: Construct a second-order difference matrix and a weighted diagonal matrix based on the number of sensitive bands; Step 2: Construct the objective function. Based on the second-order difference matrix and the weight diagonal matrix, transform the objective function into a matrix equation form. Construct the solution function for the baseline vector based on the matrix equation form of the objective function. Obtain the baseline vector based on the solution function for the baseline vector and the original spectral vector. Step 3: Compare the baseline vector with the original spectral vector to obtain the deviation vector for each sensitive band, and perform adaptive weight update on the baseline vector based on the deviation vector for each sensitive band. Step 4: Obtain the mean square error and cumulative iteration count of the baseline vector after adaptive weight update and the previous baseline vector. If the mean square error is less than the convergence threshold or the cumulative iteration count is greater than or equal to the maximum iteration count, stop the iteration; otherwise, proceed to the next iteration. Subtract the baseline vector of each pixel after the iteration from the original spectral vector of each pixel to obtain the baseline-corrected spectral vector of each pixel. Reconstruct the spectral image of each sensitive band based on the baseline-corrected spectral vector of each pixel.

[0010] Furthermore, the process of spectral feature mining and correlation screening of the reconstructed spectral image includes: Interference features are extracted from the spectral vectors of each sensitive band corresponding to each pixel in the reconstructed spectral image to obtain multi-class interference features for each pixel. The multi-class interference features are then concatenated to obtain a high-dimensional feature vector for each pixel. Feature correlation filtering is performed on the high-dimensional feature vectors to obtain a low-dimensional feature vector.

[0011] Furthermore, the process of feature relevance screening includes: The thickness range and material combination of the coating to be tested are used as cluster centers. Standard coating samples with coating materials consistent with the cluster centers and thicknesses contained within the cluster centers are obtained. Multispectral illumination is applied to the standard coating samples according to the sensitive bands of the coating to be tested, and spectral images of each sensitive band of the standard coating samples are obtained. Interference features are extracted from the spectral images to obtain multiple types of interference features for each pixel. Correlation analysis is performed on the thickness of the standard coating samples and the multiple types of interference features for each pixel to obtain the correlation coefficients between various interference features and thickness. By setting a thickness-related threshold, interference features in the high-dimensional feature vector that have an absolute value of the correlation coefficient with the thickness that is less than the thickness-related threshold are removed, and a low-dimensional feature vector is generated.

[0012] Furthermore, the process of constructing a thickness inversion model and performing thickness inversion based on low-dimensional feature vectors and the thickness inversion model includes: A thickness inversion model is constructed. Interference features are extracted from the test dataset to obtain multiple interference features of samples with different thicknesses of different coating materials at different wavelengths. The multiple interference features are spliced ​​to obtain a high-dimensional feature vector. Feature correlation is filtered on the high-dimensional feature vector to obtain a low-dimensional feature vector. The low-dimensional feature vectors of samples with different thicknesses of different coating materials at different wavelengths are used as training data. The thickness inversion model is trained using the training data to obtain the completed thickness inversion model. The thickness range of the coating to be detected, the coating material, and the low-dimensional feature vectors of each pixel corresponding to each sensitive band are input into the thickness inversion model. The coating thickness value of each pixel is output according to the thickness inversion model, and a two-dimensional heat map of coating thickness distribution is generated based on the coating thickness value of each pixel.

[0013] Furthermore, the process of calibrating the two-dimensional thermal map of coating thickness distribution includes: The uniformity online detection period of the production line is obtained. Based on the acquisition cycle, the uniformity online detection period is divided into several identical acquisition cycles. The start timestamp of each acquisition cycle is used as the calibration time point corresponding to the acquisition cycle. At the calibration time point, a standard film thickness calibration block in the production line is selected, and the spectral images of each sensitive band on the surface of the standard film thickness calibration block are acquired. The low-dimensional feature vectors corresponding to the spectral images of each sensitive band are obtained. Based on the low-dimensional feature vectors, the coating thickness value of each pixel is obtained through the thickness inversion model. According to the coating thickness value of each pixel and the actual thickness of the standard film thickness calibration block, the deviation value of each pixel is obtained. The deviation values ​​of each pixel in the current acquisition cycle are used to calibrate the two-dimensional distribution heat map of the coating thickness.

[0014] Furthermore, the process of determining the uniformity of the two-dimensional thermal map of the calibrated coating thickness distribution includes: Global statistics are performed on the calibrated two-dimensional distribution heat map of coating thickness to obtain the mean thickness, anomaly coefficient and maximum thickness difference of the two-dimensional distribution heat map of coating thickness. The process thresholds corresponding to the mean thickness, anomaly coefficient and maximum thickness difference are preset. The mean thickness, anomaly coefficient and maximum thickness difference are compared with the corresponding process thresholds. If the mean thickness, anomaly coefficient and maximum thickness difference are all less than or equal to the corresponding process thresholds, the coating to be tested is marked as qualified. If the average thickness, abnormal coefficient, or maximum thickness difference exceeds the corresponding process threshold, the coating to be tested is marked as unqualified. The neighborhood thickness gradient of each pixel in the two-dimensional distribution heat map of the coating thickness is obtained. The neighborhood thickness gradient of each pixel is compared with the preset gradient threshold. Pixels with neighborhood thickness gradients greater than the gradient threshold are marked as abnormal pixels. The area covered by the abnormal pixels is marked as a thickness abnormal area.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. Improve the accuracy and specificity of coating uniformity detection. This invention conducts full-spectrum response curve tests on standard coating samples and, combined with industrial scenario simulations, selects the optimal combination of sensitive bands with the best overall adaptability, avoiding the limitations of fixed bands in traditional multispectral detection. This band combination can accurately capture the spectral interference characteristics of different coating materials and thicknesses, ensuring a strong correlation between the spectral signal and the coating thickness from the source. Simultaneously, based on adaptive fitting baseline correction preprocessing of the sensitive bands, baseline drift caused by non-thickness factors such as substrate material inhomogeneity, optical path offset, and environmental interference is effectively eliminated, purifying the effective interference signal directly related to the coating thickness. Subsequent feature correlation screening removes redundant features, further enhancing the accuracy of thickness inversion and ensuring that the final generated two-dimensional thickness distribution heatmap accurately reflects the coating thickness distribution.

[0016] 2. Enhance the industrial scenario adaptability and stability of the detection system. This invention introduces industrial scenario information simulation testing during the sensitive band selection stage. It comprehensively considers factors such as response intensity changes, anti-interference capabilities, and band spacing to calculate a comprehensive adaptability value, ensuring that the selected sensitive band combinations can adapt to the complex operating conditions of the production line and effectively resist interference from ambient light fluctuations and production line vibrations. Simultaneously, in the thickness thermal mapping calibration stage, calibration time points are set based on the acquisition cycle. The standard film thickness calibration block in the production line is used to obtain deviation values ​​in real time and complete the calibration, dynamically offsetting the decrease in detection accuracy caused by sensor drift, light source attenuation, and other factors. This ensures the stability and consistency of the detection system during long-term continuous operation and avoids the impact of equipment aging or environmental changes on the detection results.

[0017] 3. Achieving Quantitative Judgment and Precise Source Tracing of Coating Uniformity. This invention overcomes the limitations of traditional testing methods that rely solely on qualitative observation to determine coating uniformity. By calculating global quantitative indicators such as mean thickness, coefficient of variation, and maximum thickness difference, it achieves precise judgment of the overall coating uniformity. Simultaneously, based on spatial thickness gradient analysis, it can accurately identify abnormal areas with sudden changes in local coating thickness, clearly marking the problem locations of defective coatings. This "global judgment + local source tracing" detection mode not only provides the production line with a quantitative assessment basis for coating quality but also guides the production end to quickly locate the root cause of problems, such as nozzle blockage or uneven roller speed leading to local thickness anomalies. This allows for targeted adjustments to process parameters, achieving a quality control upgrade from "post-production rejection of defective products" to "in-process precise correction," thus helping to improve product yield. Attached Figure Description

[0018] Figure 1 This is a flowchart of an online detection system for thin film coating uniformity using multispectral imaging, according to an embodiment of this application. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] like Figure 1 As shown, a multispectral imaging online detection system for thin film coating uniformity includes a monitoring center, which is communicatively connected to a coating characteristic analysis module, a data acquisition module, and a thickness inversion module. The coating characteristic analysis module is used to perform full-spectrum response curve testing on standard coating samples, screen the basic candidate wavelengths corresponding to different coating materials, and conduct industrial scenario testing on the basic candidate wavelengths to obtain the comprehensive adaptability value of the basic candidate wavelengths of each coating material under different thicknesses. The data acquisition module is used to obtain the sensitive band combination of the coating to be tested on the current production line based on the comprehensive fitness value, acquire the spectral image of each sensitive band on the surface of the coating to be tested, perform adaptive fitting baseline correction preprocessing on the spectral image of each sensitive band, and reconstruct the spectral image of each sensitive band. The thickness inversion module is used to mine spectral features and filter correlations in the reconstructed spectral image, obtain low-dimensional feature vectors, construct a thickness inversion model, perform thickness inversion based on the low-dimensional feature vectors and the thickness inversion model, generate a two-dimensional heat map of coating thickness distribution, calibrate the two-dimensional heat map of coating thickness distribution, determine the uniformity of the calibrated two-dimensional heat map of coating thickness distribution, mark the coating to be detected as qualified or unqualified, and output the thickness anomaly areas of the unqualified coating to be detected.

[0021] It should be further explained that, in the specific implementation process, the process of conducting full-spectrum response curve tests on standard coating samples and screening the basic candidate wavelengths corresponding to different coating materials includes: Obtain standard coating samples covering the target thickness range (e.g., 1.5-90μm) and target coating material range (e.g., transparent PET base film + fluorocarbon coating for photovoltaic backsheet film, silicon base + aluminum metal coating for semiconductor wafers). The target thickness range and target coating material cover the thickness and material of the coating to be tested on the target production line. Perform full-spectrum reflectance and transmission response curve tests on the standard coating samples to obtain the test dataset. The test dataset includes the response intensity of samples with different thicknesses of different coating materials at different wavelengths. Statistical analysis was performed on the response intensities of different thicknesses of various coating materials in the test dataset at different wavelengths to obtain the thickness variation of each coating material at different wavelengths. (e.g., 0.1 μm) corresponding to the change in response intensity Based on the change in response intensity For each coating material, the wavelengths are sorted in ascending order to generate a wavelength queue (response intensity change). The larger the wavelength, the higher the ranking. The top k wavelengths (k = total number of wavelengths × 30%) in the wavelength queue are selected as the base candidate wavelengths.

[0022] It should be further explained that, in the specific implementation process, the process of conducting industrial scenario tests on the basic candidate wavelengths to obtain the comprehensive adaptability value of the basic candidate wavelengths for each coating material at different thicknesses includes: Extract the processing scenario information (ambient light, temperature fluctuation) of the target production line, and conduct scenario simulation tests on the basic candidate wavelengths of different thickness samples of each coating material based on the processing scenario information. Obtain the coefficient of variation (CV) of the response intensity of the basic candidate wavelengths of each coating material at different thicknesses under the processing scenario information conditions. Where n is the number of thickness samples of the coating material participating in the test. Let i be the thickness value of the i-th thickness sample. wavelength Below, the spectral response intensity of the i-th thickness sample, wavelength The mean response intensity of all thickness samples.

[0023] Obtain the basic candidate wavelength interval for each coating material (to avoid information redundancy caused by band overlap), set the index weights corresponding to the response intensity change, response intensity coefficient of variation, and basic candidate wavelength interval, and perform a weighted summation of the response intensity change, response intensity coefficient of variation, and basic candidate wavelength interval for each coating material at different thicknesses (the weights can be adjusted according to industry testing needs, such as increasing the sensitivity weight for high-precision testing and increasing the anti-interference weight for complex environments), and obtain the comprehensive fitness value of the basic candidate wavelength for each coating material at different thicknesses (normalize the response intensity change and basic candidate wavelength interval with positive indices, normalize the response intensity coefficient of variation with negative indices, and perform a weighted summation of the response intensity change after positive index normalization, the basic candidate wavelength interval, and the response intensity coefficient of variation after negative index normalization to obtain the comprehensive fitness value).

[0024] It should be further explained that, in the specific implementation process, the process of obtaining the sensitive band combination of the coating to be tested on the current production line based on the comprehensive fitness value, and acquiring spectral images of each sensitive band on the surface of the coating to be tested, includes: Obtain the thickness range (e.g., 5-30μm) and coating material of the coating to be tested on the current production line. Obtain the comprehensive fitness value corresponding to several basic candidate wavelengths under the conditions of thickness range and coating material. Select the top three basic candidate wavelengths with the largest comprehensive fitness value to construct a sensitive band combination. Based on the sensitive band combination, perform multispectral illumination on the coating to be tested, obtain the spectral images of each sensitive band on the surface of the coating to be tested, and mark the acquisition period.

[0025] It should be further explained that, in the specific implementation process, the process of performing adaptive fitting baseline correction preprocessing on the spectral images of each sensitive band and reconstructing the spectral images of each sensitive band includes: The spectral images of each sensitive band are pre-processed for denoising (using wavelet thresholding to remove high-frequency noise caused by ambient light and sensor noise), and each pixel in the spectral images of each sensitive band after denoising preprocessing is extracted. The response intensity, constructing each pixel The original spectral vector , , , For the sensitive band, the original spectral vector is normalized, and the response intensity is mapped to the [0,1] interval. Then, adaptive fitting parameters are set (iteration number k=20, penalty factor). Initial weight vector =[1,1,1], weight decay coefficient Convergence threshold Initial baseline vector =I, using the original spectral vector as the initial baseline, start the iterative loop), and perform the following iterative steps on the original spectral vector of each pixel: Step 1: Construct a second-order difference matrix and a weighted diagonal matrix based on the number of sensitive bands. The second-order difference matrix form of the three sensitive bands is as follows: By calculating the second-order difference of the baseline values ​​of three adjacent bands The curvature of the constraint baseline; the weight diagonal matrix. The form is , , , These represent the weight vectors for the k-th iteration of the three sensitive bands; Step 2: Construct the objective function. Based on the second-order difference matrix and the weight diagonal matrix, transform the objective function into a matrix equation form. Construct the solution function for the baseline vector based on the matrix equation form of the objective function. Obtain the baseline vector based on the solution function for the baseline vector and the original spectral vector. Wherein, the objective function is , Let be the response intensity of the i-th sensitive band. Let i be the baseline vector of the i-th sensitive band. This represents the weight vector for the i-th sensitive band in the k-th iteration. The penalty factor is k, and the iteration number is k. The function for solving the baseline vector is: ,in Let be the baseline vector for the k-th iteration. This is a weighted diagonal matrix. It is a second-order difference matrix. This is the original spectral vector; Step 3: Compare the baseline vector with the original spectral vector to obtain the deviation vector for each sensitive band. The baseline vector is adaptively weighted and updated based on the deviation vector of each sensitive band. The adaptive weight update process is as follows: ,in, Deviation vector standard deviation This is the weight decay coefficient; when This indicates that the intensity of this band is significantly higher than the smoothed baseline, belonging to a fast-fluctuating effective signal, and its weight is... Set to the minimum value (e.g.) In subsequent iterations, this point will no longer affect the baseline fitting; when This indicates that the intensity of this band is close to the smooth baseline, which is part of the baseline drift, and therefore the weights are preserved. =1, continue to fit the smooth trend of this point; It should be further explained that, in the specific implementation process, during the first iteration, the initial weight vector... =[1,1,1], the smoothing penalty term of the objective function A smooth curve that covers the original signal is forcibly fitted; this curve serves as the initial baseline. The reference point at this point is the overall trend of the original signal. Step 4: Obtain the mean squared error (MSE) and cumulative iteration count of the baseline vector after adaptive weight update compared to the previous baseline vector. If the mean square error is less than the convergence threshold or the cumulative number of iterations is greater than or equal to the maximum number of iterations, the iteration stops; otherwise, the next iteration (k=k+1) begins. Through iterative reweighting, the fitting of the baseline features is continuously strengthened, and the baseline intensity of each pixel and each sensitive band is finally obtained. Subtract the baseline vector of each pixel after the iteration from the original spectral vector of each pixel to obtain the baseline-corrected spectral vector of each pixel. Reconstruct the spectral image of each sensitive band based on the baseline-corrected spectral vector of each pixel.

[0026] Traditional baseline correction methods (such as the standard sample method) require external references. For example, baseline spectra are collected using uncoated substrate samples as a reference for subtraction; baseline anchor points on the spectral curve are manually selected (assuming these points have no effective signal) and the baseline is fitted. However, this method does not require external references or manual intervention and relies entirely on the characteristics of the signal itself. This is crucial for industrial online inspection because minute differences in substrate materials and real-time environmental changes in the production line can cause the calibration values ​​of external references to quickly become invalid. The adaptively fitted baseline can dynamically track these changes and ensure correction accuracy.

[0027] It should be further explained that, in the specific implementation process, the process of performing spectral feature mining and correlation screening on the reconstructed spectral image to obtain low-dimensional feature vectors includes: Interference features are extracted from the spectral vectors of each pixel in the reconstructed spectral image corresponding to each sensitive band. Multiple interference features are obtained for each pixel, including intensity features (peak intensity, peak-to-valley difference, and mean of the effective signal in each band), phase features (Hilbert transform of the spectral vector to extract the phase shift), and spectral features (Fast Fourier Transform (FFT) of the spectral vector to extract the dominant frequency and spectral amplitude). These multiple interference features are then concatenated to obtain a high-dimensional feature vector for each pixel. Feature correlation filtering is performed on the high-dimensional feature vectors to obtain low-dimensional feature vectors.

[0028] It should be further explained that, in the specific implementation process, the feature relevance screening process includes: The thickness range and material combination of the coating to be tested are used as cluster centers. Standard coating samples with the same material as the cluster centers and whose thickness is contained within the cluster centers are obtained. Multispectral illumination is applied to the standard coating samples according to the sensitive bands of the coating to be tested, and spectral images of each sensitive band of the standard coating samples are obtained. Interference features are extracted from the spectral images to obtain multiple types of interference features for each pixel. Correlation analysis is performed on the thickness of the standard coating samples and the multiple types of interference features for each pixel to obtain the correlation coefficients between various interference features and thickness. For example: Obtain standard coating samples covering a range of thicknesses (e.g., 5μm, 10μm, 15μm, 20μm, 25μm, 30μm, a total of 6 thicknesses). The material of the samples is completely consistent with the coating to be tested on the production line. Select 100 evenly distributed pixels for each sample (6×100=600 pixels in total, n=600). Extract 8 interference features from each pixel (shallow features: peak intensity and intensity variance of 3 bands; deep features: phase shift, peak frequency of the spectrum, and multi-band fusion amplitude). Taking the correlation coefficient r between phase offset and thickness as an example: ;in, This is the mean of the phase offset. Let be the phase offset of the i-th pixel. The average thickness. Let the thickness be the thickness of the i-th pixel. The correlation coefficients between the remaining 7 features and the thickness Y were calculated, resulting in 8 correlation coefficient values. A preset thickness correlation threshold (0.6) is used to remove interference features in the high-dimensional feature vector whose absolute value of the correlation coefficient with the thickness is less than the thickness correlation threshold, thereby generating a low-dimensional feature vector.

[0029] It should be further explained that, in the specific implementation process, the process of constructing a thickness inversion model, performing thickness inversion based on low-dimensional feature vectors and the thickness inversion model, and generating a two-dimensional heat map of coating thickness distribution includes: A thickness inversion model was constructed using an SVR model with a radial basis function (RBF) kernel function. Interference features were extracted from the test dataset to obtain multiple types of interference features of samples with different thicknesses of different coating materials at different wavelengths. The multiple types of interference features were spliced ​​to obtain a high-dimensional feature vector. Feature correlation was filtered from the high-dimensional feature vector to obtain a low-dimensional feature vector. The low-dimensional feature vectors of samples with different thicknesses of different coating materials at different wavelengths were used as training data to train the thickness inversion model and obtain the trained thickness inversion model. Input the thickness range of the coating to be detected, the coating material, and the low-dimensional feature vectors corresponding to each pixel at each sensitive wavelength band into the thickness inversion model. Output the coating thickness values of each pixel according to the thickness inversion model, and generate a two-dimensional distribution heat map H(x, y) of the coating thickness based on the coating thickness values of each pixel.

[0030] It should be further noted that in the specific implementation process, the process of calibrating the two-dimensional distribution heat map of the coating thickness includes: To offset the influence of industrial environment drift, obtain the online detection period of the uniformity of the production line. Divide the online detection period of uniformity into several identical acquisition cycles based on the acquisition period. Use the start timestamp of each acquisition cycle as the calibration time point corresponding to the acquisition cycle. Select a standard film thickness calibration block (with known true thickness and the same coating material as the coating to be detected) on the production line at the calibration time point, collect the spectral images of each sensitive wavelength band on the surface of the standard film thickness calibration block, obtain the low-dimensional feature vectors corresponding to the spectral images of each sensitive wavelength band, obtain the coating thickness values of each pixel through the thickness inversion model based on the low-dimensional feature vectors, and obtain the deviation value of each pixel (coating thickness value minus true thickness) according to the coating thickness value of each pixel and the true thickness of the standard film thickness calibration block; Obtain the deviation values of each pixel in the current acquisition cycle and calibrate the two-dimensional distribution heat map of the coating thickness (subtract the deviation value corresponding to each pixel from the coating thickness value of each pixel in the two-dimensional distribution heat map of the coating thickness).

[0031] It should be further noted that in the specific implementation process, the process of determining the uniformity of the calibrated two-dimensional distribution heat map of the coating thickness, marking the coating to be detected as a qualified state or an unqualified state, and outputting the thickness abnormal area of the coating to be detected in the unqualified state includes: Perform global statistics on the calibrated two-dimensional distribution heat map of the coating thickness, obtain the thickness mean value, abnormality coefficient (thickness standard deviation / thickness mean value), and maximum thickness difference of the two-dimensional distribution heat map of the coating thickness. Preset the process thresholds corresponding to the thickness mean value, abnormality coefficient, and maximum thickness difference. Compare the thickness mean value, abnormality coefficient, and maximum thickness difference with the corresponding process thresholds. If there are thickness mean value, abnormality coefficient, and maximum thickness difference that are all less than or equal to the corresponding process thresholds, mark the coating to be detected as a qualified state. If there are thickness mean value, abnormality coefficient, or maximum thickness difference that are greater than the corresponding process thresholds, mark the coating to be detected as an unqualified state, and obtain the neighborhood thickness gradient G(x, y) of each pixel in the two-dimensional distribution heat map of the coating thickness. , Characterize the thickness change rate between this point and neighboring pixels. Is the thickness difference between the left and right neighboring pixels of the current pixel. The thickness difference between the current pixel and its adjacent pixels is used to compare the neighborhood thickness gradient of each pixel with a preset gradient threshold (determined by process experiments, such as 0.02μm / pixel). Pixels with neighborhood thickness gradients greater than the gradient threshold are marked as abnormal pixels, and the areas covered by abnormal pixels are marked as thickness abnormal regions.

[0032] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A multispectral imaging-based online detection system for thin film coating uniformity, characterized in that, This includes a monitoring center, which has communication connections to a coating characteristic analysis module, a data acquisition module, and a thickness inversion module. The coating characteristic analysis module is used to perform full-spectrum response curve testing on standard coating samples, screen the basic candidate wavelengths corresponding to different coating materials, and conduct industrial scenario testing on the basic candidate wavelengths to obtain the comprehensive adaptability value of the basic candidate wavelengths of each coating material under different thicknesses. The data acquisition module is used to obtain the sensitive band combination of the coating to be tested on the current production line based on the comprehensive fitness value, acquire the spectral image of each sensitive band on the surface of the coating to be tested, perform adaptive fitting baseline correction preprocessing on the spectral image of each sensitive band, and reconstruct the spectral image of each sensitive band. The thickness inversion module is used to mine spectral features and filter correlations in the reconstructed spectral image, obtain low-dimensional feature vectors, construct a thickness inversion model, perform thickness inversion based on the low-dimensional feature vectors and the thickness inversion model, generate a two-dimensional heat map of coating thickness distribution, calibrate the two-dimensional heat map of coating thickness distribution, determine the uniformity of the calibrated two-dimensional heat map of coating thickness distribution, mark the coating to be detected as qualified or unqualified, and output the thickness anomaly areas of the unqualified coating to be detected.

2. The online detection system for thin film coating uniformity using multispectral imaging according to claim 1, characterized in that, The process of performing full-spectrum response curve tests on standard coating samples and screening for basic candidate wavelengths corresponding to different coating materials includes: Obtain standard coating samples covering the target thickness range and the target coating material range, perform full-spectrum reflectance and transmission response curve tests on the standard coating samples, and obtain a test dataset. The test dataset includes the response intensity of samples with different thicknesses of different coating materials at different wavelengths. Statistical analysis was performed on the response intensity of samples with different thicknesses of each coating material in the test dataset at different wavelengths. The response intensity change corresponding to the thickness change of each coating material at different wavelengths was obtained. Based on the response intensity change, each wavelength of each coating material was sorted in ascending order to generate a wavelength queue. The top k wavelengths in the wavelength queue were selected as the base candidate wavelengths.

3. The online detection system for thin film coating uniformity using multispectral imaging according to claim 2, characterized in that, The process of conducting industrial scenario testing on the basic candidate wavelengths includes: Extract the processing scenario information of the target production line, and conduct scenario simulation tests on the basic candidate wavelengths of different thickness samples of each coating material based on the processing scenario information. Obtain the response intensity variation coefficient of the basic candidate wavelengths of each coating material at different thicknesses under the processing scenario information conditions. Obtain the basic candidate wavelength interval for each coating material, set the index weights corresponding to the response intensity change, response intensity coefficient of variation and basic candidate wavelength interval, and perform a weighted summation of the response intensity change, response intensity coefficient of variation and basic candidate wavelength interval for the basic candidate wavelength of each coating material at different thicknesses to obtain the comprehensive fitness value of the basic candidate wavelength of each coating material at different thicknesses.

4. The online detection system for thin film coating uniformity using multispectral imaging according to claim 3, characterized in that, The process of acquiring spectral images of the surface of the coating to be tested in various sensitive bands includes: The thickness range and material of the coating to be tested on the current production line are obtained. The comprehensive fitness value corresponding to several basic candidate wavelengths under the conditions of thickness range and coating material is obtained. The top three basic candidate wavelengths with the largest comprehensive fitness value are selected to construct a sensitive band combination. Based on the sensitive band combination, the coating to be tested is illuminated with multispectral light. The spectral images of each sensitive band on the surface of the coating to be tested are obtained and the acquisition period is marked.

5. The online detection system for thin film coating uniformity using multispectral imaging according to claim 4, characterized in that, The process of adaptive fitting baseline correction preprocessing for spectral images in each sensitive band includes: The spectral images of each sensitive band are preprocessed for denoising. The response intensity of each pixel in the denoised spectral images of each sensitive band is extracted, and the original spectral vector of each pixel is constructed. The original spectral vector is normalized, and then adaptive fitting parameters are set. The following iterative steps are performed on the original spectral vector of each pixel: Step 1: Construct a second-order difference matrix and a weighted diagonal matrix based on the number of sensitive bands; Step 2: Construct the objective function. Based on the second-order difference matrix and the weight diagonal matrix, transform the objective function into a matrix equation form. Construct the solution function for the baseline vector based on the matrix equation form of the objective function. Obtain the baseline vector based on the solution function for the baseline vector and the original spectral vector. Step 3: Compare the baseline vector with the original spectral vector to obtain the deviation vector for each sensitive band, and perform adaptive weight update on the baseline vector based on the deviation vector for each sensitive band. Step 4: Obtain the mean square error and cumulative iteration count of the baseline vector after adaptive weight update and the previous baseline vector. If the mean square error is less than the convergence threshold or the cumulative iteration count is greater than or equal to the maximum iteration count, stop the iteration; otherwise, proceed to the next iteration. Subtract the baseline vector of each pixel after the iteration from the original spectral vector of each pixel to obtain the baseline-corrected spectral vector of each pixel. Reconstruct the spectral image of each sensitive band based on the baseline-corrected spectral vector of each pixel.

6. The online detection system for thin film coating uniformity using multispectral imaging according to claim 5, characterized in that, The process of spectral feature mining and correlation screening of the reconstructed spectral image includes: Interference features are extracted from the spectral vectors of each sensitive band corresponding to each pixel in the reconstructed spectral image to obtain multi-class interference features for each pixel. The multi-class interference features are then concatenated to obtain a high-dimensional feature vector for each pixel. Feature correlation filtering is performed on the high-dimensional feature vectors to obtain a low-dimensional feature vector.

7. The online detection system for thin film coating uniformity using multispectral imaging according to claim 6, characterized in that, The process of feature relevance screening includes: The thickness range and material combination of the coating to be tested are used as cluster centers. Standard coating samples with coating materials consistent with the cluster centers and thicknesses contained within the cluster centers are obtained. Multispectral illumination is applied to the standard coating samples according to the sensitive bands of the coating to be tested, and spectral images of each sensitive band of the standard coating samples are obtained. Interference features are extracted from the spectral images to obtain multiple types of interference features for each pixel. Correlation analysis is performed on the thickness of the standard coating samples and the multiple types of interference features for each pixel to obtain the correlation coefficients between various interference features and thickness. By setting a thickness-related threshold, interference features in the high-dimensional feature vector that have an absolute value of the correlation coefficient with the thickness that is less than the thickness-related threshold are removed, and a low-dimensional feature vector is generated.

8. The online detection system for thin film coating uniformity using multispectral imaging according to claim 7, characterized in that, The process of constructing a thickness inversion model and performing thickness inversion based on low-dimensional feature vectors includes: A thickness inversion model is constructed. Interference features are extracted from the test dataset to obtain multiple interference features of samples with different thicknesses of different coating materials at different wavelengths. The multiple interference features are spliced ​​to obtain a high-dimensional feature vector. Feature correlation is filtered on the high-dimensional feature vector to obtain a low-dimensional feature vector. The low-dimensional feature vectors of samples with different thicknesses of different coating materials at different wavelengths are used as training data. The thickness inversion model is trained using the training data to obtain the completed thickness inversion model. The thickness range of the coating to be detected, the coating material, and the low-dimensional feature vectors of each pixel corresponding to each sensitive band are input into the thickness inversion model. The coating thickness value of each pixel is output according to the thickness inversion model, and a two-dimensional heat map of coating thickness distribution is generated based on the coating thickness value of each pixel.

9. The online detection system for thin film coating uniformity using multispectral imaging according to claim 8, characterized in that, The process of calibrating the two-dimensional thermal map of coating thickness distribution includes: The uniformity online detection period of the production line is obtained. Based on the acquisition cycle, the uniformity online detection period is divided into several identical acquisition cycles. The start timestamp of each acquisition cycle is used as the calibration time point corresponding to the acquisition cycle. At the calibration time point, a standard film thickness calibration block in the production line is selected, and the spectral images of each sensitive band on the surface of the standard film thickness calibration block are acquired. The low-dimensional feature vectors corresponding to the spectral images of each sensitive band are obtained. Based on the low-dimensional feature vectors, the coating thickness value of each pixel is obtained through the thickness inversion model. According to the coating thickness value of each pixel and the actual thickness of the standard film thickness calibration block, the deviation value of each pixel is obtained. The deviation values ​​of each pixel in the current acquisition cycle are used to calibrate the two-dimensional distribution heat map of the coating thickness.

10. The online detection system for thin film coating uniformity using multispectral imaging according to claim 9, characterized in that, The process of determining the uniformity of the two-dimensional thermal map of the calibrated coating thickness distribution includes: Global statistics are performed on the calibrated two-dimensional distribution heat map of coating thickness to obtain the mean thickness, anomaly coefficient and maximum thickness difference of the two-dimensional distribution heat map of coating thickness. The process thresholds corresponding to the mean thickness, anomaly coefficient and maximum thickness difference are preset. The mean thickness, anomaly coefficient and maximum thickness difference are compared with the corresponding process thresholds. If the mean thickness, anomaly coefficient and maximum thickness difference are all less than or equal to the corresponding process thresholds, the coating to be tested is marked as qualified. If the average thickness, abnormal coefficient, or maximum thickness difference exceeds the corresponding process threshold, the coating to be tested is marked as unqualified. The neighborhood thickness gradient of each pixel in the two-dimensional distribution heat map of the coating thickness is obtained. The neighborhood thickness gradient of each pixel is compared with the preset gradient threshold. Pixels with neighborhood thickness gradients greater than the gradient threshold are marked as abnormal pixels. The area covered by the abnormal pixels is marked as a thickness abnormal area.

Images