A method and system for detecting defects on a film surface

Abstract

The application provides a film surface defect detection method and system, relates to the technical field of defect detection, and fundamentally improves detection performance by introducing a defect confidence evaluation mechanism based on morphology and energy distribution and constructing an intelligent secondary verification link. The mechanism can accurately distinguish real physical defects with regular morphology and concentrated energy from pseudo defects caused by electromagnetic interference, film instantaneous wrinkles and the like, effectively solves the high false alarm rate problem caused by the dependence of the prior art on single signal strength while retaining high detection rate for low contrast defects. In addition, the judgment model based on physical characteristics has natural robustness to background noise generated in high-speed motion, and the adjustable confidence threshold gives the system high practical flexibility, enabling it to adapt to complex industrial environments and different quality control standards, significantly enhancing the accuracy, reliability and intelligence level of the entire detection system.

Description

Technical Field

[0001] This invention relates to the field of defect detection technology, specifically to a method and system for detecting defects on thin film surfaces. Background Technology

[0002] In advanced manufacturing fields such as flexible electronics, high-end optical films, and power battery separators, the surface quality of functional thin film materials is a key factor determining the performance, reliability, and yield of the final product. As production processes evolve towards higher speeds (linear speeds generally exceeding 100 meters per minute) and greater precision, unprecedented challenges arise for the online, real-time, and high-precision detection of submicron to micron-level defects on the film surface (such as pinholes, scratches, particle agglomeration, and gels). To address the motion blur problem caused by high-speed movement, a high-brightness line light source combined with a short exposure time is typically used. However, while this improves temporal resolution, it also reduces the signal-to-noise ratio, making the detection of weak defects with low contrast extremely difficult. More seriously, during high-speed roll-to-roll transport, flexible films inevitably experience out-of-plane three-dimensional geometric deformations, such as jitter, wrinkles, and warping, due to factors such as tension fluctuations, airflow disturbances, and guide roller mechanical tolerances. These deformations manifest as drastic changes in grayscale and distortions in geometric shapes in two-dimensional images. The intensity of the false feature signals they produce is even far greater than that of real defects, constituting the fundamental bottleneck of the high false alarm rate in current machine vision inspection technology.

[0003] The prior art, with publication number CN119224001B, entitled "A Method and System for Detecting Surface Defects in Thin Films," specifically relates to the field of thin film defect detection technology. It introduces two sets of light sources with different directions and a multi-frequency light source into the detection system, and combines this with a camera to collect data on the reflection intensity and response differences of the thin film surface from multiple angles. It extracts and analyzes the emission intensity fluctuation characteristics and light source response deviation characteristics to evaluate the accuracy of the detection system in identifying real defects. The accuracy is divided into three levels: high, medium, and low, and dynamic processing is performed according to these levels. Especially for the medium accuracy level, by predicting the accuracy change trend, the light source intensity and angle are adjusted in real time to improve detection reliability, significantly reduce the false texture false alarm rate, and reduce the problem of false alarms caused by complex backgrounds and lighting changes. This achieves efficient and stable detection under different production conditions.

[0004] In the high-speed production of high-end thin-film materials such as flexible OLEDs, micron-level functional defects on the surface (such as microbubbles, particle agglomeration, scratches, etc.) can seriously affect product performance and yield. Existing detection methods either struggle to achieve clear imaging at high speeds, suffer from interference from three-dimensional deformation due to film vibration and wrinkles, or have low signal-to-noise ratios in single detection modes (such as image only), resulting in high rates of missed detections and false alarms. This invention aims to achieve high-precision, robust detection, classification, and traceability of minute defects on the surface of high-speed moving thin films by fusing data from three modes: structured light three-dimensional geometric correction, high-frequency triboelectric signals, and high-resolution images, and by introducing intelligent analysis algorithms.

[0005] The information disclosed in the background section above is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] In view of the shortcomings of the prior art, the present invention provides a method and system for detecting defects on thin film surfaces to solve the problems mentioned in the background art.

[0007] The objective of this invention is achieved as follows: a method for detecting defects on thin film surfaces, comprising the following steps:

[0008] Step S1: Simultaneously freeze the high-speed motion of the thin film, and divide the high-speed motion image of the thin film obtained in each freezing into multiple preset regions to be tested;

[0009] Collect stripe deformation data of each preset area to be tested on the thin film to reconstruct the three-dimensional geometric deformation data of each preset area to be tested;

[0010] Based on the reconstructed three-dimensional geometric deformation data, the high-speed motion image of the thin film is reverse-mapped to the planar coordinate system to generate a geometrically corrected image by structured light.

[0011] Simultaneously collect the triboelectric charge signals of each preset area to be tested;

[0012] Step S2: Perform bandpass filtering, baseline drift removal and dynamic normalization on the triboelectric charge signals collected in each preset area to be tested to obtain smooth and normalized triboelectric characteristics;

[0013] Step S3: Perform grayscale contrast normalization on the images of each preset area to be tested after structured light geometric correction to obtain image features;

[0014] Step S4: Based on the sliding window temporal difference and adaptive weight algorithm, fuse the triboelectric features and image features to generate fused features for each preset region to be tested;

[0015] Step S5: Adaptively calculate the judgment threshold based on the fusion features, and locate the defect area and output the coordinates through connected component analysis.

[0016] Furthermore, the film is a flexible OLED composite film strip;

[0017] The high-speed movement of the frozen film is synchronized by using an ultra-short pulse LED lighting unit, a global shutter camera, a structured light projector, an encoder trigger module, and a nano-piezoelectric triboelectric acquisition module on the surface of the guide roller.

[0018] The stripe deformation of each preset area to be measured is obtained through a structured light projector;

[0019] The process of reconstructing the three-dimensional geometric deformation data of each preset region to be tested and generating the structured light geometrically corrected image further includes: First, based on the acquired stripe deformation image, the stripe modulation degree of each pixel is calculated; then, the stripe modulation degree is mapped through a preset power function to generate a pixel-level confidence map; finally, when performing inverse mapping to generate the geometrically corrected image, the confidence map is used to perform weighted adaptive smoothing processing on the original three-dimensional geometric deformation data, thereby suppressing reconstruction noise and correction artifacts introduced by low-quality stripe regions.

[0020] The high-speed motion of the synchronously frozen film further includes: dynamically adjusting the illumination incident angle and polarization state of the ultrashort pulse LED illumination unit according to the local change gradient of the real-time acquired triboelectric charge signal through a preset nonlinear mapping function, thereby maximizing the optical imaging contrast between the defect and the film background in the preset area to be tested where there is a functional micro-agglomeration defect.

[0021] Furthermore, the triboelectric charge signal after baseline drift removal is processed, and a transient impact factor is generated by calculating the ratio of its first-order difference to the local standard deviation in the time dimension. This factor is used to quantify the sharp pulse characteristics in the triboelectric signal caused by hard point defects or transient events.

[0022] Furthermore, the structured light geometrically corrected images of each preset test area are processed to generate image features. The processing includes: first, based on the corrected height map, the original high-speed motion image is reverse-mapped to generate a planar corrected image that eliminates three-dimensional geometric distortion; then, the planar corrected image is contrast-normalized through a linear transformation based on pre-calibrated minimum and maximum gray values ​​to generate normalized image features; finally, the inherent image noise level of the system is estimated by calculating the gray standard deviation in a defect-free static thin film region.

[0023] Furthermore, after performing grayscale contrast normalization on the geometrically corrected image, the geometrically corrected image is further processed by calculating the local binary mode value of each pixel and statistically analyzing the information entropy of this value in the local region to generate a texture entropy feature map, which is used to quantify the texture complexity caused by surface microstructure anomalies in the image.

[0024] Furthermore, a sliding window temporal difference and adaptive weighting algorithm is used to fuse triboelectric features and image features to generate fused features for each preset region to be tested. The fusion process includes: first, calculating the difference between normalized image features and normalized triboelectric features in the time dimension to obtain their instantaneous changes; then, calculating the instantaneous signal-to-noise ratio of the two features by dividing their respective instantaneous changes by their estimated noise standard deviation, and dynamically generating a fusion dynamic weight based on the ratio of the signal-to-noise ratios; finally, using the dynamic weight to perform a weighted average of the two original features, and smoothing it through a first-order infinite impulse response filter to obtain the final fused features.

[0025] Furthermore, the process of generating fusion features further includes: after calculating the dynamic weights, an additional co-enhancement factor is calculated, which is obtained by nonlinearly multiplying the triboelectric features and the image features; finally, when calculating the fusion features, the co-enhancement factor is used as an independent enhancement term and weighted and superimposed on the weighted average calculated by the dynamic weights, thereby nonlinearly amplifying the region where the two features respond synchronously.

[0026] Furthermore, a judgment threshold is adaptively calculated based on the fusion features, and the defect region is located and its coordinates are output through connected component analysis. The judgment process includes: first, the mean and standard deviation of the fusion features are statistically analyzed within a sliding time window; then, based on the statistical results, a dynamic segmentation threshold is calculated using the method of "mean + k6 times standard deviation", and upper and lower limits are set for this threshold to ensure its stability; finally, the fusion feature map is binarized using this threshold, and the defect region is identified using a standard 8-connected component labeling algorithm. After removing noise spots with too small an area, the center coordinates of the remaining regions are mapped to the physical coordinate system and output.

[0027] Furthermore, after locating the defect region through connected component analysis, each located defect region is further processed: First, the circularity of the region is calculated to quantify its shape regularity; then, the offset between the centroid of the fused feature and its geometric center within the region is calculated to quantify the uniformity of its energy distribution; through a preset nonlinear weighting formula, the circularity and centroid offset are combined to generate a final defect confidence score; finally, the confidence score is compared with a preset confidence threshold, and only defect regions with confidence scores higher than the threshold are output as the finally confirmed defects.

[0028] Let R be the connected component of each candidate defect segmented. Calculate the confidence score of each candidate defect connected component R. With a pre-set defect confidence threshold Compare; if Greater than If the candidate defect connected component R is confirmed as a real defect, then... Not greater than If the candidate defect connected domain R is determined to be noise or a non-critical anomaly, it will be removed from the results.

[0029] A thin film surface defect detection system, the method for performing the thin film surface defect detection method, comprising:

[0030] Data acquisition and processing module: used to synchronously freeze the high-speed motion of the film and divide the high-speed motion image of the film obtained in each freezing process into multiple preset areas to be tested;

[0031] Collect stripe deformation data of each preset area to be tested on the thin film to reconstruct the three-dimensional geometric deformation data of each preset area to be tested;

[0032] Based on the reconstructed three-dimensional geometric deformation data, the high-speed motion image of the thin film is reverse-mapped to the planar coordinate system to generate a geometrically corrected image by structured light.

[0033] Simultaneously collect the triboelectric charge signals of each preset area to be tested;

[0034] Triboelectric feature acquisition module: used to perform bandpass filtering, baseline drift removal and dynamic normalization processing on the triboelectric charge signals collected in each preset area to be tested, to obtain smooth and normalized triboelectric features;

[0035] Image feature acquisition module: used to perform grayscale contrast normalization processing on the images of each preset area to be tested after structured light geometric correction to obtain image features;

[0036] Fusion module: used to fuse the triboelectric features and image features based on sliding window temporal difference and adaptive weight algorithm to generate fused features for each preset region to be tested;

[0037] Defect determination and location module: It is used to adaptively calculate the determination threshold based on the fusion features, locate the defect area through connected component analysis and output the coordinates.

[0038] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0039] By introducing a defect confidence assessment mechanism based on morphology and energy distribution, an intelligent secondary verification step is constructed, fundamentally improving detection performance. This mechanism can accurately distinguish between real physical defects with regular morphology and concentrated energy and false defects caused by electromagnetic interference, instantaneous wrinkles in thin films, etc. While maintaining a high detection rate for low-contrast defects, it effectively solves the problem of high false alarm rate caused by existing technologies relying on a single signal intensity. In addition, this judgment model based on physical characteristics has natural robustness to background noise generated during high-speed motion, while the adjustable confidence threshold gives the system extremely high practical flexibility, enabling it to adapt to complex and ever-changing industrial environments and different quality control standards, significantly enhancing the accuracy, reliability, and intelligence level of the entire detection system. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0041] Figure 1 This is a schematic diagram of the overall method flow of the present invention.

[0042] Figure 2 This is a block diagram of the overall system modules of the present invention. Detailed Implementation

[0043] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0044] Example 1:

[0045] Please see Figure 1 and Figure 2 The present invention provides a technical solution:

[0046] A method for detecting defects on thin film surfaces, comprising the following steps:

[0047] Step S1: Simultaneously freeze the high-speed motion of the thin film, and divide the high-speed motion image of the thin film obtained in each freezing into multiple preset regions to be tested;

[0048] Collect stripe deformation data of each preset area to be tested on the thin film to reconstruct the three-dimensional geometric deformation data of each preset area to be tested;

[0049] Based on the reconstructed three-dimensional geometric deformation data, the high-speed motion image of the thin film is reverse-mapped to the planar coordinate system to generate a geometrically corrected image by structured light.

[0050] Simultaneously collect the triboelectric charge signals of each preset area to be tested;

[0051] Further explanation: The film is a flexible OLED composite film strip;

[0052] The high-speed movement of the frozen film is synchronized by using an ultra-short pulse LED lighting unit, a global shutter camera, a structured light projector, an encoder trigger module, and a nano-piezoelectric triboelectric acquisition module on the surface of the guide roller.

[0053] The stripe deformation of each preset area to be measured is obtained through a structured light projector;

[0054] The image after structured light geometric correction is a result of confidence-weighted 3D reconstruction based on fringe modulation and adaptive geometric correction; the reasons for choosing this technical approach are as follows:

[0055] During high-speed motion, due to film vibration and changes in surface material (such as high reflectivity or translucency), the quality of the acquired structured light fringes will locally degrade, causing standard 3D reconstruction algorithms to generate noise or serious errors in these areas. These errors will be directly transmitted to the geometric correction stage, generating erroneous image textures (artifacts), which are easily misjudged as real defects by subsequent algorithms. This embodiment fundamentally solves this problem by evaluating data quality at the reconstruction source and adaptively correcting it.

[0056] Traditional methods typically treat all reconstructed 3D points as equally reliable. This approach assigns a "confidence score" to each 3D data point and uses this score to guide the subsequent geometric correction process, forming an intelligent processing chain of "perceived quality - adaptive correction".

[0057] Stripe modulation density is a classic metric for measuring signal quality in the field of structured light and is easy to calculate. Using it as a confidence weight to guide an adaptive smoothing filtering process is logically sound and can be fully implemented in real time using modern FPGAs or GPUs.

[0058] 1) The process of reconstructing the three-dimensional geometric deformation data of each preset region to be tested and generating the structured light geometrically corrected image further includes: First, based on the acquired stripe deformation image, the stripe modulation degree of each pixel is calculated; then, the stripe modulation degree is mapped through a preset power function to generate a pixel-level confidence map; finally, when performing inverse mapping to generate the geometrically corrected image, the confidence map is used to perform weighted adaptive smoothing processing on the original three-dimensional geometric deformation data, thereby suppressing reconstruction noise and correction artifacts introduced by low-quality stripe regions; the specific implementation details are as follows:

[0059] For the calculation of stripe modulation degree: For the stripe image I(i,j) collected in the preset area to be tested, calculate its local stripe modulation degree M(i,j); where (i,j) are pixel coordinates;

[0060] 1.1) The local window contrast method is used to calculate the stripe modulation of each pixel. This method quantifies the sharpness and contrast of the stripes at a given pixel location by calculating the maximum and minimum gray values ​​within the pixel's neighborhood. The results are then normalized; the numerical quantization is represented as follows: ;

[0061] Where M(i,j) is the stripe tone of pixel coordinate (i,j); It is a local calculation window centered at pixel coordinates (i,j). In this embodiment... Use a size of 5×5 pixels; It is a local calculation window The maximum pixel grayscale value within; It is a local calculation window The minimum pixel grayscale value within; These are normal numbers; the values ​​used in this embodiment are as follows. Used to prevent the denominator It equals zero;

[0062] The stripe modulation index M(i,j) is converted into a confidence score C(i,j) through a power function, generating a confidence map with the same size as the original image;

[0063] 1.2) Confidence scores are generated by performing a power transformation on the stripe modulation. This transformation uses an adjustable exponential parameter to non-linearly amplify or reduce the influence of the original modulation, thus providing greater flexibility in defining what constitutes "credible" data. The confidence score C(i,j) is represented as: ;

[0064] Where C(i,j) is the confidence score of the three-dimensional data at pixel coordinate (i,j); M(i,j) comes from the stripe modulation degree calculated in the previous step; β is the confidence adjustment index, which is a constant greater than 0. In this example, β=2.

[0065] Since the range of M(i,j) is [0,1], the output range of C(i,j) is also strictly limited to [0,1].

[0066] When the output of C(i,j) approaches 0: the current application state of the technical goal is "the 3D data is extremely unreliable"; this occurs in areas with blurred fringes, oversaturation, or occlusion. The trend of the technical goal "precise geometric correction" is "tends to ignore the original 3D data of the point and rely more on the information of its reliable neighborhood for interpolation repair";

[0067] When the output of C(i,j) approaches 1: the current technical objective is "highly reliable 3D data". This occurs in areas with clear stripes and high contrast. The trend of the technical objective "precise geometric correction" is "completely trusting the original 3D data of this point and preserving its reconstructed geometric details".

[0068] It should be noted that the approach intervals in this application are set as [0.8,1) or (0,0.3], respectively;

[0069] An output approaching 0 indicates that the value is within the interval (0, 0.3]. The meaning of "output approaching" in the following steps is the same as above, and the values ​​[0.8, 1] and (0, 0.3] are specifically adjusted by the expert group based on actual usage scenarios.

[0070] By limiting the confidence score C(i,j) to the interval [0,1], it can be used as an ideal, standardized weighting factor directly for subsequent weighted average calculation, ensuring the stability of the algorithm and seamless connection between different modules.

[0071] Reasoning for parameter β: When β > 1, the function curve is concave, meaning only high fringe modulation M(i,j) can achieve a high confidence score C(i,j). This indicates that the system has very stringent requirements for data quality and gives lower confidence to medium- and low-quality fringes. When 0 < β < 1, the function curve is convex, and the system is more "tolerant" of medium- and low-quality fringes. Therefore, by adjusting β, the "confidence" standard can be flexibly defined based on the reflective properties and noise level of the actual thin film material.

[0072] Using 3D geometric deformation data Before performing the reverse mapping, the height map is smoothed based on the confidence score C(i,j) to obtain the corrected height map. .

[0073] A weighted average filtering method guided by confidence scores is employed. The final height value of each pixel is a weighted sum of its original reconstructed height and the average height of its neighborhood. The core principle is that the weights are entirely determined by the pixel's own confidence score. The numerical quantization standard is as follows:

[0074]

[0075] Among them, the average height of the neighborhood The calculation formula is: ;

[0076] in, In the local calculation window The arithmetic mean of all original height values ​​within the range; It is a local calculation window Total number of pixels within; k1 and It is a general coordinate pair that represents the currently accessed local computation window during the summation process. The coordinates of any pixel within the range;

[0077] As C(i,j) approaches 1, 1-C(i,j) approaches 0. The system fully trusts the original 3D reconstruction result of the current pixel, preserving all geometric details.

[0078] As C(i,j) approaches 0, 1-C(i,j) approaches 1. The system completely distrusts the original 3D data of the current pixel, instead replacing it with the average height of its surrounding neighbors, thus effectively "smoothing out" the outlier caused by noise or error.

[0079] When C(i,j) is between (0,1), It is a smooth transition between the original value and the neighborhood average.

[0080] This formula is ingeniously designed, treating C(i,j) as a dynamic "switch" or "harmonyr." Instead of simply setting a threshold to discard bad pixels, it achieves a smooth transition from "complete trust" to "complete repair." This ensures that while repairing noise, no new, abrupt edges are created at the boundary between trusted and untrusted regions, resulting in a smooth final height map. It is both clean and natural.

[0081] At the source of the defect detection process—the geometric correction stage—an intelligent "quality inspection gate" has been established. By quantifying the reliability of 3D reconstruction and adaptively preprocessing the 3D data accordingly, the image quality after structured light geometric correction can be greatly improved. Final output... The image not only eliminates geometric distortions caused by film jitter and wrinkles, but more importantly, it removes artifacts caused by reconstruction errors that are easily confused with real defects.

[0082] Introducing adaptive lighting control based on real-time triboelectric feedback;

[0083] Functional micro-agglomeration defects on flexible OLED films (such as ITO particles and small organic molecule agglomerations) have extremely low optical contrast with the substrate material, making imaging very difficult under normal lighting. Furthermore, these defects often cause significant changes in local charge distribution. This solution utilizes triboelectricity to guide optical imaging, directly addressing the core challenge of "unclear visibility."

[0084] In traditional machine vision systems, the angle and polarization state of the illumination parameters are usually fixed once set. This example constructs an "electro-optical" real-time closed-loop feedback system that deeply couples sensing information from two completely different physical modalities.

[0085] 2) The synchronous freezing of the high-speed film further includes: dynamically adjusting the illumination incident angle and polarization state of the ultrashort pulse LED illumination unit according to the local change gradient of the real-time acquired triboelectric charge signal through a preset nonlinear mapping function, thereby maximizing the optical imaging contrast between the defect and the film background in the preset test area with functional micro-agglomeration defects. The specific implementation details are as follows:

[0086] Real-time calculation of the local spatial gradient of the triboelectric charge signal Q(x,t) along the direction of film motion. This gradient reflects the degree of drastic change in charge distribution and is a strong indicator of potential functional defects;

[0087] 2.1) The central difference method is used to calculate the gradient of the triboelectric charge signal in the spatial dimension. To enhance sensitivity to subtle changes and suppress noise, a normalization factor based on the signal's own amplitude is introduced, and the Sigmoid function is used to smoothly map the gradient value to the (0,1) interval, resulting in a standardized gradient strength index; specifically expressed as: ;

[0088] in, Q(x,t) is the normalized triboelectric gradient at position x and time t; Q(x,t) is the original triboelectric charge signal collected at position x and time t; Δx is the spatial step size of the differential calculation, which is equal to the distance between adjacent electrodes. It is the average value of the triboelectric charge signal of all channels at the current moment, used for dynamic normalization; It is a positive number to prevent the denominator from being zero. It is the gain coefficient of the sigmoid function, used to adjust the sensitivity to gradients. The larger the value, the steeper the function curve and the more drastic the response to small gradient changes; The output range is limited to the interval (0,1);

[0089] When the output approaches 0, it indicates that the triboelectric charge distribution in the current test area is uniform, corresponding to a normal state where the film surface is smooth and the material is homogeneous. At this time, the urgency of the technical objective "defect identification" is lowest, and the system should use conventional lighting parameters.

[0090] When the output approaches 1, it indicates a drastic change in the triboelectric charge distribution in the current test area. This is caused by defects such as functional micro-agglomerations. At this point, the urgency of the technical objective "defect identification" is highest, and the system needs to immediately adjust the illumination parameters to enhance the imaging contrast of this area.

[0091] 2.2) Based on the normalized frictional gradient The illumination incident angle θ(t) and polarization state P(t) are adjusted in real time.

[0092] The adjustment of the incident illumination angle is linearly related to the tribological gradient; that is, the greater the gradient, the more the angle deviates from the normal value, in order to enhance surface texture using grazing light. The adjustment of the polarization state is controlled by an exponentially decaying function with the tribological gradient as input. The greater the gradient, the smaller the polarization state rotation angle, aiming to suppress substrate reflection and highlight defect scattering by approaching orthogonal polarization.

[0093] For the angle of incidence of illumination : ;

[0094] Illumination polarization state rotation angle : ;

[0095] in, It is the real-time adjusted incident angle of the lighting, in degrees; This is the incident angle of the basic lighting. In this example, the initial value is set to 45° for use when there are no abnormalities. This is the maximum incident angle adjustment range; It is the polarization state rotation angle after real-time adjustment, in degrees. This is the maximum polarization rotation angle, which is initially set to 90° in this embodiment to achieve orthogonal polarization. It is the attenuation coefficient for polarization adjustment, which controls the response speed of the polarization state to the gradient;

[0096] For the angle of incidence of illumination :when When increasing from 0 to 1, from linearly increase to . The increase led to This increases the contrast of the surface's tiny bumps or depressions, which is consistent with the principles of optical imaging.

[0097] The added technical effect is the detection of potential defect signals; The added technical effect is that the lighting becomes a more angled, grazing light;

[0098] For polarization state rotation angle :when As the exponent increases from 0 to 1, the exponent term... Monotonically decreasing from 1 to .therefore, from Smoothly decrease to near 0°;

[0099] when When approaching 0, Approaching The illumination light is nearly orthogonal to the analyzer in front of the camera, which can effectively suppress specular reflection on the smooth thin film substrate and obtain a uniform dark field.

[0100] when When it approaches 1, As the angle approaches 0°, the illumination light becomes more parallel to the analyzer direction, allowing more light scattered and depolarized by the defects to pass through, thus making the defects stand out against the dark background.

[0101] In this embodiment, when a defect is detected, the illumination is quickly adjusted, sacrificing some background suppression in exchange for maximizing the enhancement of the defect signal.

[0102] When the film is flat, the system uses conventional illumination to ensure overall imaging quality. Once an anomaly is detected in the triboelectric charge signal, the system instantly adjusts the illumination parameters of the area within microseconds. Using grazing light and polarization technology, the optical features of the object being detected are amplified. Compared with fixed illumination, this method can improve the signal-to-noise ratio of specific types of defects by more than an order of magnitude, thereby improving the sensitivity and reliability of detection.

[0103] Step S2: Perform bandpass filtering, baseline drift removal and dynamic normalization on the triboelectric charge signals collected in each preset area to be tested to obtain smooth and normalized triboelectric characteristics;

[0104] Further explanation: Baseline drift removal aims to eliminate low-frequency noise or DC bias, i.e., baseline drift, caused by factors such as changes in ambient temperature, long-term sensor aging, or slow changes in the overall electrostatic level of the thin film; the characterization formula is as follows:

[0105] ; ;

[0106] in, This is the raw triboelectric charge signal acquired at position x and time t; it is the input data without any processing.

[0107] This is the baseline value of the signal estimated at time t. This value is recursively calculated using an exponential moving average (EMA) filter, representing the slow trend of the signal's change.

[0108] It is the baseline tracking smoothing factor, a constant between 0 and 1, used to control the "memory" or "inertia" of baseline updates.

[0109] The tribocharge signal after bandpass filtering and baseline shifting is characterized as... ;

[0110] It is the triboelectric charge signal after baseline correction; it is obtained by subtracting the estimated baseline from the original signal.

[0111] The calculation formula is essentially a first-order low-pass filter used to extract the DC and very low-frequency components of the signal as a baseline. Parameters The value of is crucial:

[0112] when When it approaches 1, such as 0.99, the new Mainly from the previous moment This makes the signal baseline value update very slowly, which can effectively track and filter out extremely low-frequency drifts, while not mistakenly filtering out real defect signals with slightly higher frequencies as part of the baseline.

[0113] Subtracting this slowly changing baseline from the original triboelectric charge signal achieves the core effects of "drift removal" and "high-pass filtering." The processed signal... The mean value will fluctuate around zero, eliminating measurement deviations caused by different basic electrostatic levels in different time periods or batches of products, and providing a stable and reliable data basis for subsequent normalization and threshold judgment.

[0114] After removing baseline drift, this step is used to dynamically and adaptively track the upper and lower boundaries of the signal, providing a dynamic range for subsequent normalization processing.

[0115] ;

[0116] ;

[0117] in, It is the dynamic maximum boundary updated at time t. It is the dynamic minimum boundary updated at time t. This is the boundary forgetting factor. It is a constant less than 1 used to introduce a decay mechanism. After correction at all spatial locations x at time t The instantaneous maximum value. The corrected signal at all spatial locations x at time t. The instantaneous minimum value.

[0118] The normalization range of the system in this embodiment is not calibrated all at once, but dynamically adjusted based on the real-time signal of the thin film flowing through the sensor. If the current batch of thin film material or the production environment causes an increase in the overall signal amplitude, and It will automatically adjust accordingly to ensure the normalization effect remains effective.

[0119] Boundary forgetting factor Its purpose is to prevent the system from being constantly affected by a historical extreme outlier. When no new peak appears, the old one... It will be multiplied by each time step This causes it to decay slowly towards zero. This means that if a large peak occurs and then doesn't reappear for a long time, the system will gradually "forget" it and adjust the boundaries back to the normal fluctuation range of the current signal. This greatly enhances the system's robustness and adaptability to current operating conditions.

[0120] The signal is scaled to a standard fixed range, and then a final smoothing process is performed to improve the signal-to-noise ratio.

[0121] ;

[0122] ;

[0123] in, It is the triboelectric charge signal after dynamic boundary normalization; its value is linearly mapped to the [0,1] interval.

[0124] It is the smooth and normalized triboelectric characteristic of the final output;

[0125] This is the final feature smoothing factor. It's a constant between 0 and 1, used to control the smoothness of the final output features. (This example demonstrates...) The initial value is 0.2;

[0126] Explanation of the smoothing step: It is an exponential moving average (EMA) low-pass filter. The goal is to filter out residual high-frequency random noise in the normalized signal;

[0127] parameter Trade-offs: The initial value is 0.2, meaning the current normalized signal... The weight of the smoothed value is 20%, while the weight of the historical smoothed value is 80%. This is a relatively strong smoothing.

[0128] The advantage is that it can effectively suppress random noise spikes, making the final output characteristic curve smoother and more stable, thereby reducing false alarms caused by noise.

[0129] The downside is that excessive smoothing can slightly blur the edges of real defect signals or reduce the peak value of very narrow pulse signals. Therefore, The choice of is a key trade-off between noise suppression and detail preservation. A value of 0.2 indicates that feature stability is prioritized in this application scenario.

[0130] 2) For step S2: Introduce transient impact feature extraction based on signal morphology; the reason for choosing this technical solution is as follows:

[0131] For scratches from hard particles or the rupture of tiny bubbles, S2 manifests as an extremely brief and sharp transient impact in the triboelectric signal. Traditional smoothing processes weaken or even erase these crucial transient characteristics.

[0132] By introducing signal morphology analysis, particularly transient impact characteristics, the analysis of triboelectric signals is extended from the one-dimensional "amplitude" domain to a two-dimensional space that includes "morphological" information. This allows the system to not only sense "how large the charge is," but also "how fast and sharp the charge change is," thereby distinguishing defects caused by different physical factors.

[0133] The calculation of transient impacts is based on simple difference and ratio operations, which requires little computation and can be integrated into existing FPGA real-time processing workflows. Following step S2, "obtaining smooth and normalized triboelectric features," a supplementary analysis is performed on the same data source to extract features from another dimension.

[0134] 2.1) The triboelectric charge signal after baseline drift removal The process involves calculating the ratio of the first-order difference to the local standard deviation in the time dimension to generate a transient impact factor, which is used to quantify the sharp pulse characteristics in the triboelectric signal caused by hard particle defects or transient events. The specific implementation details are as follows:

[0135] Transient impact factor calculation: after obtaining Then, the transient impact factor at each sampling point is calculated. .

[0136] An adaptive normalized kurtosis approximation method is employed. This method quantifies the "sharpness" of the signal at the current moment by calculating the ratio of the instantaneous change (first-order difference) of the signal to the degree of fluctuation (local standard deviation) of the signal within a short historical window. To enhance robustness and standardize the results, the hyperbolic tangent function is used to smoothly map this ratio to the (-1,1) interval.

[0137] Transient impact factor Numerical quantization is represented as: ;

[0138] Among them, local standard deviation The calculation formula is:

[0139] ;

[0140] yes The mean within the same window;

[0141] in, It is the transient impact factor at position x and time t; It is the time step, i.e., the sampling period; It is a time window before time t. The local standard deviation within the range reflects the average fluctuation level of the recent signal; It is the size of the time window for calculating the local standard deviation; It is a positive integer to prevent the denominator from being zero; This is the transient impact sensitivity coefficient, a positive constant, as shown in this example. The value is 0.5;

[0142] It is the hyperbolic tangent function.

[0143] The output range is limited to the interval (-1, 1);

[0144] When the output approaches 0: This indicates the rate of change of the current triboelectric charge signal. Much smaller than its recent average volatility level The current application status of the technical target is "stable or gradually changing signal". This corresponds to a film surface with uniform material or a defect type such as large-area gradually changing oil stains. At this time, the technical target "sharp defect identification" has no response.

[0145] When the output approaches +1 or -1: This indicates that the rate of change of the current triboelectric charge signal is much greater than its recent average fluctuation level, meaning a very sharp pulse has occurred. The current technical target application status is "high-frequency transient impact detected." This could be caused by events such as scratches from hard particles or the instantaneous rupture of tiny bubbles. In this case, the technical target's "sharp defect identification" response reaches its strongest, and the system should pay close attention to this signal.

[0146] For parameters Reasoning: The system's definition of "sharp" was controlled. The larger the value, the more likely it is that even if the instantaneous change in the triboelectric charge signal is only slightly larger than the local fluctuation, The value will also quickly saturate to ±1, making the system more sensitive to minor shocks. The smaller the value, the more instantaneous changes are required to achieve this. As the system approaches saturation, its response to shocks becomes more sluggish. This can be addressed by adjusting... It can optimize the detection capability for defects of specific shapes based on the main defect types on the production line.

[0147] This step focuses on capturing the "shape" or "kurtosis" information of the signal. These two types of information are complementary. For example, a large but gently sloping oil slick will produce a high-amplitude but... A signal whose value approaches 0; while a tiny scratch from a metal particle, though the total charge changes little, will produce a signal with a low amplitude but... A sharp signal with a value close to ±1. Through subsequent step S4 fusion, the system can distinguish between these two completely different types of defects, greatly enriching the basis for defect classification and improving the level of intelligence in detection.

[0148] Step S3: Perform grayscale contrast normalization on the images of each preset area to be tested after structured light geometric correction to obtain image features;

[0149] Further explanation: 3.1) The structured light geometrically corrected images of each preset area to be tested are processed to generate image features; the processing includes: first, based on the corrected height map, the original high-speed motion image is reverse-mapped to generate a planar corrected image that eliminates three-dimensional geometric distortion; then, the planar corrected image is contrast-normalized through a linear transformation based on pre-calibrated minimum and maximum gray values ​​to generate a normalized image feature with a value range in the range [0,1]; finally, the inherent image noise level of the system is estimated by calculating the gray standard deviation in the defect-free static thin film region; the specific implementation is as follows:

[0150] 3.11) Based on the corrected height map reconstructed by the structured light binocular vision system in step S1. For the raw images synchronously acquired by the global shutter camera A pixel-by-pixel reverse planar mapping is performed. This process repositions pixels distorted by film jitter and wrinkles in the original image back to their proper positions in an ideal planar coordinate system, thereby generating a planar corrected image with geometric distortion eliminated. x and t represent position and time, respectively; and Characterizes pixel grayscale values;

[0151] These are raw, uncorrected images of high-speed motion of thin films. The image data was directly captured by a 4K global shutter line scan camera.

[0152] The resulting planar corrected image after geometric correction. The content of this image is... The same, but the spatial arrangement of its pixels has been based on Adjustments were made to eliminate the visual distortion caused by three-dimensional deformation;

[0153] 3.12) The linear normalization (Min-Max-Scaling) method is adopted; the gray value of each pixel in the corrected image is subtracted from a preset minimum gray value, and then divided by the preset maximum gray range (the difference between the maximum gray value and the minimum gray value), thereby linearly mapping the original gray values ​​of any range to the standardized interval [0,1].

[0154] The normalized image features are represented as follows:

[0155] ;

[0156] in, It is the normalized image feature or normalized contrast at position x and time t, and its value range is limited to the interval [0,1].

[0157] : The pixel grayscale value at position x and time t of the image after the previous geometric correction;

[0158] This is the minimum grayscale reference value. This value is pre-calibrated by taking a stable value, or slightly below the average grayscale value, from a large number of samples taken from a defect-free, ideal film surface area before production. It represents the "darkest" normal state in the system.

[0159] This is the maximum grayscale reference value. This value is pre-calibrated by sampling known severe defects that cause the strongest optical response, taking the average of their grayscale peak values ​​or a stable upper limit. It represents the "brightest" anomalous state occurring in the system.

[0160] While the system is idling or running on a confirmed defect-free thin film, a series of static planar calibration images are acquired. Within one or more fixed regions of these images, the standard deviation of all pixel grayscale values ​​is calculated, and the average or stable value of this standard deviation is recorded as the system image noise level. .

[0161] This is the standard deviation of the image noise level. This value quantifies the inherent random fluctuation amplitude of the image signal under conditions of no defects and uniform surface material. It is mainly determined by factors such as the thermal noise of the camera sensor, readout noise, and slight non-uniformity of the illumination system.

[0162] The high-speed movement of a flexible film inevitably produces out-of-plane jitter and wrinkles, which are reflected in the original image. The defects manifest as localized stretching, compression, and distortion. Without correction, a straight scratch defect will appear curved, a circular point defect will appear elliptical, and even the film's own wrinkles and shadows will be misidentified as defects. Geometric correction, by utilizing precise three-dimensional data, "irons" the image, ensuring that any shape and size in the image accurately reflects the physical condition of the film surface. This reduces false alarms caused by geometric deformation and guarantees the accuracy of defect location and dimensional measurement.

[0163] The effect of contrast normalization: The core function of this step is to eliminate the effects of lighting variations and camera gain fluctuations, and to provide a standardized input for subsequent processing. In actual production, ambient light, LED light source aging, power supply fluctuations, etc., all cause the overall image to be either too bright or too dark. By mapping grayscale values ​​to a fixed [0,1] range, The features are no longer affected by these global, slowly varying brightness factors, but only reflect the relative brightness of the pixel relative to the "standard darkest" and "standard brightest". This enhances the robustness of the algorithm.

[0164] The technical advantages of noise estimation: The calculation provides a crucial "benchmark" for the system to determine whether a signal is a "meaningful change" or "random noise".

[0165] 3.13) Further explanation: Texture entropy feature extraction based on Local-Binary-Pattern (LBP) is introduced; the reasons for choosing this technical solution are as follows:

[0166] Simple grayscale contrast normalization It can only reflect the brightness information of pixels and is not sensitive to defects with similar brightness to the background but abnormal texture structure (such as minor scratches, pinholes, and subtle texture changes caused by uneven local coating). LBP is an extremely powerful texture description operator that can effectively capture the local spatial structure information of an image. Combined with the concept of "entropy", it can quantify the complexity and disorder of the texture, thereby accurately identifying such defects.

[0167] This approach elevates image analysis from pixel-level grayscale analysis to region-level texture structure analysis. It adds a dimension orthogonal to grayscale values ​​to image features, enabling the system to not only see "brightness" and "darkness," but also to "touch" the "roughness" or "smoothness" of surfaces. This multi-dimensional feature extraction significantly enhances the ability to identify complex defects.

[0168] 3.14) After performing grayscale contrast normalization on the geometrically corrected image, the geometrically corrected image is further processed... The process involves calculating the local binary pattern value of each pixel and statistically analyzing the information entropy of this value within the local region to generate a texture entropy feature map. This map is used to quantify the texture complexity in the image caused by surface microstructure anomalies. The specific implementation details are as follows:

[0169] Image after geometric correction Calculate the texture entropy of each pixel. .

[0170] The method for calculating local region information entropy is adopted: First, within a neighborhood window surrounding each pixel, the local binary pattern (LBP) values ​​of all pixels are calculated; then, the probability distribution of these LBP values ​​within this window is statistically analyzed; finally, according to the definition of information entropy, the entropy value of this probability distribution is calculated, which is used as a measure of the texture complexity of the center pixel; the numerical quantization is as follows:

[0171]

[0172] Among them, probability The calculation is as follows:

[0173]

[0174] It is the standard LBP operator, and for an 8-neighborhood, its computation is as follows:

[0175]

[0176] in, It is the texture entropy at position (x,t); The entropy calculation window is centered at (x,t); L is the total number of patterns in LBP coding. For standard 8-neighborhood LBP, . It is in the window Within, the proportion of pixels with an LBP value of k4. It is the entropy calculation window. The total number of pixels within. It is the LBP value of pixel (i,j); It is the grayscale value of the center pixel. It is the gray value of the nth neighboring pixel. These are normal numbers; the values ​​used in this embodiment are as follows. This prevents the log function from receiving a zero input;

[0177] The output range is in Between L=256, i.e., [0, 8], normalize it to [0, 1]: .

[0178] The closer the output is to 0, the more uniform the LBP mode is within the local window. The current technical target application state is "highly consistent texture". This corresponds to a smoother and more uniform area on the film surface, which is a more ideal defect-free state. At this time, the technical target "texture defect recognition" has no response.

[0179] The closer the output is to 1, the richer and / or more chaotic the LBP patterns within the local window, and the smaller the probability difference between each pattern. The current technical target application state is "highly disordered / complex texture". This indicates that the area is more prone to microstructural anomalies, such as pitting caused by uneven coating, fine mesh scratches, or foreign matter contamination. At this time, the technical target "texture defect recognition" response reaches its strongest.

[0180] For entropy calculation window Reasoning The size determines the scale of texture analysis. Smaller... Sensitive to small, isolated texture anomalies (such as pinholes). Larger ones... This allows for better capture of large-scale, low-frequency texture variations (such as unevenly coated stripes). By selecting appropriate... The size makes the feature most sensitive to defect types at a specific scale, thus justifying the choice of this parameter.

[0181] This step not only identifies a dark spot but also determines whether it is a "smooth-edged circle" or an "irregularly shaped, rough-edged foreign object." This ability to differentiate is crucial for subsequent defect classification. and By combining these features, an image descriptor more powerful than a single grayscale feature can be constructed, thereby significantly improving the detection performance for complex and low-contrast defects.

[0182] Step S4: Based on the sliding window temporal difference and adaptive weight algorithm, fuse the triboelectric features and image features to generate fused features for each preset region to be tested;

[0183] Further explanation: 4.1) Based on a sliding window temporal difference and adaptive weighting algorithm, triboelectric features and image features are fused to generate fused features for each preset region to be tested. The fusion process includes: first, calculating the difference between normalized image features and normalized triboelectric features in the time dimension to obtain their instantaneous changes; then, calculating the instantaneous signal-to-noise ratio of the two features by dividing their respective instantaneous changes by their estimated noise standard deviation, and dynamically generating a fusion dynamic weight based on the ratio of the signal-to-noise ratio; finally, using this dynamic weight to perform a weighted average of the two original features, and smoothing it through a first-order infinite impulse response filter to obtain the final fused features; the specific implementation details are as follows:

[0184] The difference in the time dimension is represented as:

[0185] ;

[0186] in: Normalized image features The instantaneous change at time t. Normalized triboelectric characteristics The instantaneous change at time t. Δt is the time step, i.e., the time interval between two adjacent data acquisitions;

[0187] The generation of fusion weights is characterized as follows: ;

[0188] in, It is the instantaneous signal-to-noise ratio of the image features at point (x,t); It is the instantaneous signal-to-noise ratio of the triboelectric characteristics at point (x,t); : In terms of system image noise level ; It is the standard deviation of the background noise of the triboelectric signal, calculated in a defect-free area. The standard deviation is obtained; These are the dynamic weights used for fusion at point (x,t), with a value range in the range [0,1]. The fusion features are calculated and represented as follows:

[0189] ;

[0190] in, It is the final output fused feature value, with a value range in the interval [0,1]. It is the smoothing coefficient, in this example. A value of 0.7 indicates that 70% of the current fusion result comes from historical values ​​and 30% comes from the current weighted average, which is used to enhance the temporal continuity of features and suppress sudden noise.

[0191] Calculate fusion eigenvalues ​​while the system is idling or running on a confirmed defect-free thin film. The standard deviation in the static region is obtained. .

[0192] Step 4.1) The fusion method is "competitive," meaning that when the signal-to-noise ratio of one feature is much higher than that of the other, the latter's information is completely ignored. For translucent gels, only weak signals are generated optically and triboelectrically, but both signals occur simultaneously. Under competitive fusion, these defects are missed because both features have low signal-to-noise ratios. The "cooperative enhancement" mechanism in this embodiment specifically addresses this type of "dual weak signal" problem.

[0193] By introducing a synergistic factor, when two features change synchronously, even if they are both very weak, they can be greatly enhanced. This is a smarter and more physically intuitive method of fusion.

[0194] In another embodiment, it is further explained that: 4.2) the process of generating fusion features further includes: after calculating the dynamic weights, an additional co-enhancement factor is calculated, which is obtained by nonlinearly multiplying the triboelectric features and image features; finally, when calculating the fusion features, this co-enhancement factor is used as an independent enhancement term and weighted and superimposed on the weighted average calculated by the dynamic weights, thereby nonlinearly amplifying the region where the two features respond synchronously; the specific implementation is as follows:

[0195] After dynamic weight calculation, the collaborative enhancement factor is calculated in parallel. ;

[0196] Specifically, a feature cross-multiplication method based on power transformation is adopted. This method first performs power transformation on the normalized triboelectric features and image features respectively to adjust their respective sensitivity curves. Then, the two transformed feature values ​​are multiplied to obtain a co-response value. Finally, the sigmoid function is used to smoothly map the co-response value to the (0,1) interval to form a standardized co-enhancement factor.

[0197] Synergistic enhancement factor The numerical quantization is as follows:

[0198] ;

[0199] in, It is the synergistic enhancement factor at position (x,t), with a value range of (0,1); These are the normalized image features obtained from S3; The smoothed normalized triboelectric characteristics obtained from S2; , The power-law adjustment exponents for both image features and triboelectric features are constants greater than 0; in this embodiment, they are all set to 2. This is the gain coefficient of the Sigmoid function, used to adjust the intensity of the synergistic enhancement. In this embodiment, the initial value is 10. This is the activation threshold for the cooperative response, used to suppress the multiplicative effect of low-level noise; in this embodiment, the value is set to 0.1.

[0200] The output range is limited to (0,1);

[0201] When the output approaches 0: it indicates that the product of the image features and the triboelectric features is much smaller than the threshold. This occurs when at least one eigenvalue approaches 0. The current application state of the technical target is "asynchronous feature or single feature response." That is, there is only an optical signal or only a triboelectric signal, or both are very weak. At this time, the technical target "cooperative defect enhancement" has no response, and the system degenerates into the original competitive fusion.

[0202] When the output approaches 1: it indicates that the product of the image features and the triboelectric features is much greater than the threshold. This is only when... and This can only occur when both sides possess a non-negligible amplitude. The current application state of the technical target is "characteristic synchronous high response." This strongly indicates the existence of a defect in this region that manifests both optically and electrically. At this point, the response of the technical target "cooperative defect enhancement" reaches its strongest.

[0203] parameter , Reasoning: When , When the value is greater than 1, the transformed value will only increase significantly when the original eigenvalue is large. This effectively raises the threshold for triggering collaborative enhancement, making it sensitive only to strong synchronization signals. , When the value is less than 1, even if the original eigenvalue is small, the transformed value will be amplified, making the system more sensitive to weak synchronization signals. By adjusting... , It can precisely control the intensity of "dual weak signals" to be amplified.

[0204] The original fusion formula was further modified, and the following was introduced: Specifically, a three-part weighted sum method is adopted; the final fusion feature consists of three parts: the first part is the memory term of historical fusion features; the second part is the competitive fusion term based on signal-to-noise ratio; and the third part is the collaborative enhancement term. A main fusion weight is used to balance the contributions of competitive fusion and collaborative enhancement.

[0205] The modified fusion feature values ​​are quantized as follows:

[0206]

[0207] Among them, competitive integration items It is the original weighted average: ; It is a new, enhanced fusion feature; It is the original fusion part based on signal-to-noise ratio competition; It is a synergistic enhancement factor from the previous step; The co-enhancement weights are fixed hyperparameters, as shown in this example. A value of 0.3 indicates that the collaborative enhancement items account for 30% of the information;

[0208] This embodiment addresses the question of "whether to believe information when all sources are weak but point to the same conclusion." It does so by introducing... The system is now able to detect subtle defects that would be drowned out by noise in a single sensor mode, but have high confidence because they occur simultaneously in two independent modes.

[0209] Step S5: Adaptively calculate the judgment threshold based on the fusion features, and locate the defect area and output the coordinates through connected component analysis.

[0210] 5.1) Based on the fusion features, an adaptive threshold is calculated, and the defect region is located and its coordinates are output through connected component analysis. The judgment process includes: first, calculating the mean and standard deviation of the fusion features within a sliding time window; then, based on the statistical results, calculating a dynamic segmentation threshold using the method of "mean + k6 times the standard deviation", and setting upper and lower limits for this threshold to ensure its stability; finally, using this threshold to binarize the fusion feature map, and identifying the defect region using a standard 8-connected component labeling algorithm, after removing noise spots with excessively small areas, mapping the center coordinates of the remaining regions to the physical coordinate system and outputting them; the specific implementation details are as follows:

[0211] The adaptive threshold is designed as follows: ;

[0212] in, is the mean value of all fused features within a time window before the current moment t ; is the standard deviation of all fused features within the same time window ; is the threshold sensitivity coefficient, i.e., a multiple of the standard deviation. In this embodiment, k6 = 3

[0213] is the adaptive threshold finally used for segmentation

[0214] : a function that returns min if v < min, returns max if v > max, and returns v otherwise. Here, v represents ;

[0215] , are the minimum and maximum limits of the adaptive threshold respectively. In this embodiment , , which are used to prevent the threshold from being too low or too high due to drastic changes in the background

[0216] The binarization is expressed as: ; thereby generating a binary defect map M

[0217] Apply the 8-connected component labeling algorithm to the binary defect map M, group the connected defect pixels into the same region, and calculate the area of each region; eliminate all connected components with an area smaller than the preset threshold

[0218] Calculate the centroid of each remaining connected component, convert its pixel coordinates to physical world coordinates through the pre-calibrated camera parameters, and then output

[0219] It is further described in another embodiment: introducing a confidence evaluation based on the morphology and energy distribution of the defect region. The reasons for choosing this technical solution are as follows

[0220] Traditional binary segmentation only answers the question of "is it a defect", but does not answer the question of "the probability of being a true defect". Random noise blobs may be marked as defects because they accidentally exceed the threshold, resulting in false alarms. This solution analyzes the shape and internal energy distribution of each candidate defect region after segmentation to calculate a "confidence score" for it, thereby effectively distinguishing between real defects with typical physical morphologies and random, irregular noise blobs

[0221] 5.2) After locating the defect region through connected component analysis, each located defect region is further processed: First, the circularity of the region is calculated to quantify its shape regularity; then, the offset between the centroid of the fused feature and its geometric center within the region is calculated to quantify the uniformity of its energy distribution; using a preset nonlinear weighting formula, the circularity and centroid offset are combined to generate a final defect confidence score; finally, this confidence score is compared with a preset confidence threshold, and only defect regions with confidence scores higher than the threshold are output as the finally confirmed defects; specific implementation details include:

[0222] Let R be the segmented candidate defect connected component, and calculate the confidence score of candidate defect connected component R. Specifically, a confidence assessment model based on multi-feature weighting is adopted. This model first calculates two independent geometric and energy features of the candidate region: circularity and normalized centroid offset. Then, these two features are linearly weighted by their respective weighting factors to obtain a comprehensive score. Finally, to make the score more penalizing of low-quality features, this linear weighted sum is used as an exponent and mapped to the (0,1) interval through an exponential decay function to obtain the final confidence score.

[0223]

[0224] Among them, roundness and normalized centroid offset The calculation is as follows:

[0225] ; ;

[0226] in, It is the energy center of mass. It is the geometric center; It is the confidence score of the candidate defect connected component R, with a value range of (0,1]. A(R) represents the circularity of region R, with a value in the range (0,1], where 1 represents a perfect circle. A(R) is the area, and L(R) is the perimeter. It is the normalized centroid offset. It is made by dividing the Euclidean distance between the energy center of mass and the geometric center by the radius of the equivalent area circle, making it a dimensionless offset index that is independent of size. It is the weight of the centroid offset, in this example A value of 0.4 indicates that roundness accounts for 60% of the weight, and offset accounts for 40%. This is the confidence decay coefficient, used to adjust the rate at which the confidence level decreases. (In this embodiment...) The value is 2;

[0227] When the output approaches 0: it indicates that the exponent term is large, i.e. or Or both are very large; this corresponds to a region with an irregular shape or extremely uneven energy distribution. The current application status of the technical target is "the probability that the candidate region is noise is high". The changing trend of the technical target "reliable defect output" is "suppress the output of this region, or mark it as a low-priority item to be re-inspected";

[0228] An output equal to 1 is only possible when the exponent is 0. This requires C(R) = 1, and The condition is satisfied while being zero. The current application status of the technical objective is "the candidate region has an ideal defect morphology". This conforms to the physical morphology of typical defects (such as bubbles, oil droplets, and round particles). The changing trend of the technical objective "reliable defect output" is "highly confident that the region is a real defect, and output it first";

[0229] The confidence score of each candidate defective connected component R. With a pre-set defect confidence threshold Compare. If Greater than If the candidate defect's connected component R is identified as a real defect, its relevant information (such as physical coordinates, area, confidence score, etc.) is retained and added to the final defect report list; if Not greater than If the candidate defect connected domain R is determined to be noise or a non-critical anomaly, it will be removed from the results.

[0230] Defect confidence threshold It is a parameter that can be set by the user according to the production quality requirements. Its value range is between (0,1), and in this embodiment, it is set to 0.5. It represents the minimum confidence level required for the system to determine a candidate region as a "real defect".

[0231] when When set to a higher value, that is Above 0.8, the system's defect judgment criteria become stricter; only candidate regions with regular shapes and uniform energy distribution will be reported. The technical effect is that the system's false alarm rate will be extremely low, and the output defect reports will be extremely reliable. The cost is that some irregularly shaped real defects will be missed, leading to an increase in the false negative rate.

[0232] when When set to a lower value, that is... Below 0.3, the system's criteria for defect judgment become more lenient; even areas with slightly irregular shapes or uneven energy distribution will be reported as long as their confidence level is higher than 0.3. The technical benefit is enhanced system detection capability, enabling the capture of a wider variety of defects. The trade-off is that some larger noise clusters may be misclassified as defects, leading to an increased false alarm rate.

[0233] The existence of this feature provides a crucial and intuitive "quality control valve" for the final output of this invention. It is not a fixed parameter, but rather an interface that allows operators to flexibly adjust it according to the quality standards of the current production batch. This design transforms complex morphological and energy distribution analysis results into a simple, manually adjustable decision threshold, greatly enhancing the practicality and adaptability of this invention in real-world industrial applications and demonstrating the rationality of this screening step design.

[0234] By using an adjustable confidence threshold In the final decision-making process, the system effectively filters out a large number of irregularly shaped false defects caused by random noise, electromagnetic interference, and other factors, thus significantly reducing the false alarm rate. The output is no longer a simple list of coordinates, but a defect report with added "credibility" and intelligent verification. This not only improves the reliability of the detection results but also provides richer and more valuable data support for subsequent process analysis and quality control.

[0235] for Reasoning: The value of reflects prior knowledge. If the main defect type on the production line is oil droplets, then should be reduced. Increase the weight of roundness. If the main defect is a trailing scratch, energy is concentrated at the head, the center of mass is significantly off-center, but the overall shape is not very regular, then a more complex shape description is needed, or adjustments are required. The weight. The adjustability of this algorithm allows it to adapt to the defect characteristics of different production lines;

[0236] By comprehensively evaluating the morphology and energy distribution of candidate defects, the system can effectively filter out a large number of irregularly shaped false defects caused by random noise, electromagnetic interference, and other factors, thereby significantly reducing the final false alarm rate. The output is no longer a simple list of coordinates, but a defect report with "credibility" and intelligent verification. This not only improves the reliability of the detection results but also provides richer and more valuable data support for subsequent process analysis and quality control.

[0237] Example 2:

[0238] To verify the effectiveness of the defect confidence assessment and screening method based on morphology and energy distribution described in this invention, the following experiment was conducted. The test object was a high-speed (60 m / min) lithium-ion battery separator (PP / PE composite material, 16 μm thick) production line. This production line has deployed a dual-modal online inspection system integrating a 4K resolution global shutter linear scanning camera and a triboelectric sensor array. This system can execute steps S1 to S4 of the invention in real time, generate a fused feature map, and complete the preliminary defect candidate region segmentation based on an adaptive threshold. The core objective of this embodiment is to verify the superior ability of the confidence assessment and screening module described in step S5.2 in distinguishing between real physical defects and systematic noise / pseudo-defects. Six typical detection event samples were prepared for the experiment. These samples can all trigger alarms in the initial segmentation stage. The specific samples include: real defects introduced by the process (standard bubbles, metal particles, gels), and pseudo-defects or non-critical defects caused by external interference or transient material states (random electromagnetic interference, transient film wrinkles, minor scratches). The experimental process will execute the confidence calculation and screening procedures described in this invention on the six initially located candidate regions. First, the system automatically extracts the area A(R) and perimeter L(R) of each candidate defect connected region R, and calculates its circularity C(R). Simultaneously, the system calculates the geometric center of the region. With the energy centroid based on the fusion feature F(x,t) distribution This leads to the normalized centroid offset. In this embodiment, the parameter in the confidence calculation formula is set as: confidence decay coefficient. The centroid offset weight is 2.0. The confidence score is 0.4. This means that morphological regularity (reflected by roundness) accounts for 60% of the weight in the evaluation, while energy distribution uniformity (reflected by centroid offset) accounts for 40%. The confidence score for each candidate region is then calculated. Then, the system will use a preset, relatively strict defect confidence threshold. The final decision is made. Only areas with a confidence score higher than 0.5 are identified as "real defects" requiring alerts and recording; the rest are automatically filtered out by the system and judged as noise or non-critical events. By recording and comparing the intermediate parameters and final judgment results of these six typical samples, this embodiment aims to quantitatively demonstrate that the method of the present invention can effectively improve the accuracy of defect detection and significantly reduce the false alarm rate caused by complex operating conditions.

[0239] The following table details the data and results of this experiment:

[0240]

[0241] As can be seen from the data in the table above, the beneficial effects of this invention are extremely significant:

[0242] Accurately distinguishing between real and spurious defects: The initial fusion feature peaks of all six samples were significantly higher than the detection threshold, indicating that they would all be classified as defects. However, this invention successfully distinguished them by calculating confidence scores. Standard bubbles and metal particles, due to their regular shapes and uniform energy distribution, achieved confidence scores as high as 0.86 and 0.78, respectively. Although the gel's morphology was not very regular, its physical solidity ensured a relatively concentrated energy distribution, and its confidence score of 0.56 was still higher than the threshold. These three were all real process defects of concern and were accurately identified.

[0243] Effective suppression of system noise and transient interference: "Random electromagnetic interference" samples are essentially single-point, sudden, high-intensity noise from triboelectric signals, appearing as a region with extremely concentrated energy but no actual optical counterpart on the fused feature map. This causes a significant deviation between its energy centroid and geometric center; despite a high initial signal peak, its confidence score is only 0.11, successfully classifying it as noise. "Instantaneous film wrinkles" and "minor scratches," due to their highly irregular shapes, have confidence scores below the threshold of 0.5 even with small energy shifts, and are effectively filtered out. This avoids frequent false alarms caused by non-permanent material deformation or non-critical defects.

[0244] In summary, this embodiment demonstrates that the defect confidence assessment and screening method proposed in this invention establishes a more intelligent and robust decision-making mechanism by comprehensively and quantitatively evaluating the morphological and energy distribution characteristics of candidate defects. It no longer relies solely on a single signal strength threshold, but instead performs an effective "secondary verification" of the defect's "physical authenticity." This maintains a high detection rate for genuine defects while significantly reducing the false alarm rate caused by various complex interference factors, thereby significantly improving the reliability and intelligence level of the entire online detection system.

[0245] Further explanation: After locating the defect region through a threshold, the corresponding unfused raw triboelectric features, image features, and 3D geometric deformation data within the defect region are further extracted. A Support Vector Machine (SVM) classifier is then used to analyze these multimodal features to achieve automatic defect type classification. The specific implementation details are as follows:

[0246] Defect region feature extraction: For each located candidate defect connected component R:

[0247] For the image feature vector: extract the statistical features (mean, variance, skewness, kurtosis) and texture features (contrast, energy, correlation, homogeneity based on the gray-level co-occurrence matrix) of the gray-level image within the candidate defect connected domain R. A total of 8 features are extracted.

[0248] For triboelectric feature vectors: extract the amplitude features (peak value, mean value) and waveform morphology features (pulse width, rise time) of the triboelectric signal within the corresponding region of the candidate defect connected domain R. A total of 4 features are extracted.

[0249] Extract the height features (maximum height, average height) and shape features (volume, surface area, curvature) of the three-dimensional geometric deformation data within the connected domain R of the candidate defect. A total of 4 features are extracted.

[0250] The above feature vectors are concatenated into a 16-dimensional composite feature vector.

[0251] Numerical quantization: An SVM classifier with a radial-basis-function (RBF) kernel was used. Optimal parameters were determined through grid search: penalty coefficient C = 10, kernel parameter gamma = 0.01.

[0252] The training set for the classifier is pre-labeled manually and contains typical defect categories such as “bubble (Class1)”, “particle (Class2)”, “scratch (Class3)”, and “oil stain (Class4)”.

[0253] Output: For each detected defect, output its physical coordinates and its category label.

[0254] This step, building upon "defect detection," leaps to "defect type identification." Different types of defects exhibit varying combinations of optical, electrical, and geometric characteristics. For example, "bubbles" typically display prominent geometric protrusions and weak triboelectric signals, while "metal particles" show less geometric deformation but possess exceptionally sharp triboelectric signals. The SVM classifier can effectively learn and distinguish the subtle differences in these multimodal feature combinations, achieving high-precision classification. This provides crucial information for downstream quality control and process analysis, making it an important link in achieving intelligent manufacturing and logically a necessary extension of the detection step.

[0255] Furthermore, the automatically classified defect types and their occurrence time and location information are correlated with historical process parameters (including but not limited to film tension, guide roller speed, ambient temperature and humidity, and coating liquid viscosity) synchronously obtained from the Manufacturing Execution System (MES). Through principal component analysis (PCA) and correlation matrix calculation, key process parameters that are significantly related to specific defect types are identified to achieve root cause tracing of defects.

[0256] Create a time-series database to store each defect record. For a specific defect category, filter out all defect records of that category.

[0257] Extract the corresponding set of process parameter vectors.

[0258] Principal Component Analysis (PCA): PCA is used to reduce the dimensionality of the multidimensional process parameter vectors and extract the principal components with a contribution rate exceeding 95%. This step aims to eliminate collinearity among process parameters and identify key comprehensive trends in change, solving the problem that simple correlation analysis cannot handle the effects of multivariate coupling (meeting requirement 2.1).

[0259] Correlation calculation: Calculate the Pearson correlation coefficient between the occurrence frequency of this defect type and each extracted principal component.

[0260] Generate a root cause analysis report listing the main process components and their constituent original process parameters that are most correlated with the frequency of "bubble" defects. For example, the report states: "The frequency of 'bubble' defects is strongly positively correlated with 'main component 1' (mainly composed of excessively low film tension and excessively high coating viscosity) by 0.85."

[0261] PCA (Precision Process Analysis) allows for the extraction of the true causes of defects from numerous interrelated process parameters, preventing engineers from making blind adjustments based on experience. For example, discovering a strong correlation between "scratching" defects and the rotational speed fluctuations of a specific guide roller can directly guide equipment maintenance. This forms a complete "detection-analysis-feedback" closed loop.

[0262] A thin film surface defect detection system, the method being used to perform the thin film surface defect detection method, comprising:

[0263] Data acquisition and processing module: used to synchronously freeze the high-speed motion of the film and divide the high-speed motion image of the film obtained in each freezing process into multiple preset areas to be tested;

[0264] Collect stripe deformation data of each preset area to be tested on the thin film to reconstruct the three-dimensional geometric deformation data of each preset area to be tested;

[0265] Based on the reconstructed three-dimensional geometric deformation data, the high-speed motion image of the thin film is reverse-mapped to the planar coordinate system to generate a geometrically corrected image by structured light.

[0266] Simultaneously collect the triboelectric charge signals of each preset area to be tested;

[0267] Triboelectric feature acquisition module: used to perform bandpass filtering, baseline drift removal and dynamic normalization processing on the triboelectric charge signals collected in each preset area to be tested, to obtain smooth and normalized triboelectric features;

[0268] Image feature acquisition module: used to perform grayscale contrast normalization processing on the images of each preset area to be tested after structured light geometric correction to obtain image features;

[0269] Fusion module: used to fuse the triboelectric features and image features based on sliding window temporal difference and adaptive weight algorithm to generate fused features for each preset region to be tested;

[0270] Defect determination and location module: It is used to adaptively calculate the determination threshold based on the fusion features, locate the defect area through connected component analysis and output the coordinates.

[0271] It should be noted that all calculation formulas in this application employ regression analysis, including but not limited to machine learning algorithms, to deeply analyze the collected parameters and identify their natural trends and interrelationships. Specialized software, such as Python's Scikit-learn library or the R language, is used to automatically generate mathematical models that match the data. Then, cross-validation and other methods are used to objectively evaluate the model performance, and continuous feedback and optimization are combined to ensure that the created formulas truly reflect the inherent laws of the data, thereby guaranteeing their effectiveness and accuracy. In all calculation formulas in this application, the parameters in each formula undergo dimensionless processing within a consistent range to ensure that different physical quantities are compared on the same scale; dimensionless processing techniques include, but are not limited to, min-max-normalization and Z-score standardization.

[0272] The technical solution of this invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random-access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of various embodiments of this invention.

[0273] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0274] The above description of the embodiments is only for the purpose of helping to understand the method and core ideas of the present invention. It should be noted that those skilled in the art can make several improvements and modifications to the present invention without departing from the principles of the present invention, and these improvements and modifications also fall within the protection scope of the claims of the present invention.

Claims

1. A method of detecting defects on a film surface, characterized by, The specific steps include: Step S1: synchronously freeze the high-speed movement of the film, and divide a plurality of preset areas to be measured in each frozen high-speed movement image of the film; Collect the stripe deformation of each preset area to be measured on the film to reconstruct three-dimensional geometric deformation data of each preset area to be measured; Based on the reconstructed three-dimensional geometric deformation data, the high-speed movement image of the film is reversely mapped to a plane coordinate system to generate a structure light geometrically corrected image; Synchronously collect the triboelectric charge signals of each preset area to be measured; Step S2: perform band-pass filtering, baseline drift removal and dynamic normalization processing on the triboelectric charge signals collected in each preset area to be measured to obtain smooth and normalized triboelectric characteristics; Step S3: perform gray contrast normalization processing on the structure light geometrically corrected image of each preset area to be measured to obtain image characteristics; Step S4: fuse the triboelectric characteristics and the image characteristics based on a sliding window time sequence difference and an adaptive weight algorithm to generate fusion characteristics of each preset area to be measured; Step S5: adaptively calculate a decision threshold based on the fusion characteristics, and locate a defect area and output coordinates through connected domain analysis; The film is a flexible OLED composite film strip; An ultrashort pulse LED illumination unit, a global shutter camera, a structure light projector, an encoder trigger module and a nanometer piezoelectric triboelectric collection module on the surface of a guide roller are used to synchronously freeze the high-speed movement of the film; The stripe deformation of each preset area to be measured is obtained by the structure light projector; The process of reconstructing the three-dimensional geometric deformation data of each preset area to be measured and generating the structure light geometrically corrected image further includes: first, calculating the fringe modulation of each pixel point based on the collected stripe deformation image; then, mapping the fringe modulation through a preset power function to generate a pixel-level confidence map; finally, when performing reverse mapping to generate a geometrically corrected image, the confidence map is used to perform weighted adaptive smoothing processing on the original three-dimensional geometric deformation data, thereby suppressing the reconstruction noise and correction artifacts introduced by low-quality stripe areas; The synchronous freezing of the high-speed movement of the film further includes: according to the local variation gradient of the real-time collected triboelectric charge signals, dynamically adjusting the illumination incident angle and polarization state of the ultrashort pulse LED illumination unit through a preset nonlinear mapping function, thereby maximizing the optical imaging contrast between the defect and the film background for the preset area to be measured with functional micro-agglomerate defects.

2. The method of claim 1, wherein: The triboelectric charge signal after the baseline drift removal is processed by calculating the ratio of the first-order difference and the local standard deviation in the time dimension to generate a transient impact factor to quantify the sharp pulse characteristics in the triboelectric signal caused by hard point defects or transient events.

3. The method of claim 2, wherein: The image feature is generated by processing the structured light geometry corrected image of each preset region to be tested; the processing includes: first, according to the corrected height map, the original high-speed motion image is inversely mapped to generate a plane correction image eliminating three-dimensional geometric distortion; then, through a linear transformation based on the minimum and maximum gray value of the pre-calibration, the contrast of the plane correction image is normalized to generate a normalized image feature; finally, the system inherent image noise level is estimated by calculating the gray standard deviation of the defect-free static film area.

4. The method of claim 3, wherein: After the gray contrast normalization processing of the structured light geometry corrected image, the geometry corrected image is further processed, the local binary pattern value of each pixel point of the image is calculated, and the information entropy of the value in the local area is counted to generate a texture entropy feature map, which is used to quantify the texture complexity caused by the surface microstructure abnormality in the image.

5. The method of claim 4, wherein: The friction electricity feature and the image feature are fused based on the sliding window time difference and the adaptive weight algorithm to generate the fusion feature of each preset region to be tested. The fusion process includes: first, the difference in the time dimension of the normalized image feature and the normalized friction electricity feature is calculated respectively to obtain the instantaneous change amount; Then, the instantaneous signal-to-noise ratio of the two kinds of features is calculated by dividing the respective instantaneous change amount by the estimated noise standard deviation, and a fused dynamic weight is dynamically generated according to the ratio of the signal-to-noise ratio; finally, the dynamic weight is used to weight and average the two original features, and a first-order infinite impulse response filter is used for smoothing to obtain the final fusion feature.

6. The method of claim 5, wherein: The process of generating the fusion feature further includes: after calculating the dynamic weight, an additional synergistic enhancement factor is calculated, which is obtained by nonlinearly multiplying the friction electricity feature and the image feature; finally, in the calculation of the fusion feature, the synergistic enhancement factor is taken as an independent enhancement term and is weighted and superimposed on the weighted average value calculated by the dynamic weight, so that the region responding synchronously to the two kinds of features is nonlinearly amplified.

7. The method of claim 6, wherein: The adaptive decision threshold is calculated based on the fusion feature, and the defect area is located and the coordinates are output through connected domain analysis; the decision process includes: first, the mean and standard deviation of the fusion feature are counted in the sliding time window; then, based on the statistical result, a dynamic segmentation threshold is calculated by the method of "mean+k6 times standard deviation", and the upper and lower limits of the threshold are set to ensure its stability; finally, the threshold is used to binarize the fusion feature map, and the defect area is identified through the standard 8-connected domain labeling algorithm, and after removing the noise spots with small area, the center coordinates of the remaining area are mapped to the physical coordinate system and output.

8. The method of claim 7, wherein: After locating the defect regions by connected component analysis, each of the located defect regions is further processed: first, the circularity of the region is calculated to quantify the regularity of its shape; then, the offset of the centroid of the fusion feature in the region from its geometric center is calculated to quantify the uniformity of its energy distribution; through a preset nonlinear weighting formula, the circularity and the centroid offset are combined to generate a final defect confidence score; finally, the confidence score is compared with a preset confidence threshold, and only the defect regions with a confidence score higher than the threshold are output as the final confirmed defects; Let each segmented candidate defect connected domain be denoted as R, and let the confidence score of each candidate defect connected domain R be denoted as C(R) be compared with a preset defect confidence threshold value If C(R) > Cth, the candidate defect connected domain R is confirmed as a real defect If C(R) > Cth, the candidate defect connected domain R is confirmed as a real defect If C(R) > Cth, the candidate defect connected domain R is confirmed as a real defect If C(R) > Cth, the candidate defect connected domain R is confirmed as a real defect If C(R) > Cth, the candidate defect connected domain R is confirmed as a real defect 9. A thin film surface defect detection system characterized by: The method is used for performing the film surface defect detection method of any one of claims 1-8, comprising: A data acquisition and processing module is configured to freeze the high-speed movement of the film synchronously and divide a plurality of preset regions to be measured in each frozen high-speed movement image of the film; Striped deformation of each preset region to be measured on the film is collected to reconstruct three-dimensional geometric deformation data of each preset region to be measured; Based on the reconstructed three-dimensional geometric deformation data, the high-speed movement image of the film is reversely mapped to a plane coordinate system to generate a structure light geometric corrected image; Synchronous collection of triboelectric charge signals of each preset region to be measured is performed; A triboelectric feature acquisition module is configured to perform band-pass filtering, baseline drift removal and dynamic normalization processing on the collected triboelectric charge signals of each preset region to be measured to obtain smooth and normalized triboelectric features; An image feature acquisition module is configured to perform grayscale contrast normalization processing on the structure light geometric corrected image of each preset region to be measured to obtain image features; A fusion module is configured to fuse the triboelectric features and the image features based on a sliding window time difference and an adaptive weight algorithm to generate fusion features of each preset region to be measured; A defect judgment and positioning module is configured to adaptively calculate a judgment threshold based on the fusion features, locate defect regions through connected component analysis, and output coordinates.

Images