Vacuum coating quality intelligent monitoring method based on artificial intelligence

Abstract

The invention relates to an intelligent vacuum coating quality monitoring method based on artificial intelligence, which comprises the following steps: collecting process parameters in real time through a coating equipment sensor, dynamically monitoring vacuum degree change, deposition rate deviation and temperature gradient, and standardizing data flow to obtain a quantitative trend curve of process parameter fluctuation; combining the defect microcosmic feature description set with gas flow fluctuation and power supply voltage jitter to obtain a correlation analysis result of process parameters and defect forms, and calibrating defect types according to a defect feature library to obtain an updated defect classification basis; and inputting real-time new surface image data through the updated defect classification basis, performing defect detection for illumination reflection differences and stripe deflection angles, obtaining a preliminary defect classification result, and adjusting classification weight parameters.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to an intelligent monitoring method for vacuum coating quality based on artificial intelligence. Background Art

[0002] Traditional inspection methods often rely on manual observation or simple image processing technology. This makes it difficult to adapt to the diversity and dynamic nature of defect morphology, especially when process parameters change. This lack of flexibility in adapting to environmental changes often leads to inaccurate inspection results and even misses key issues. During the coating process, when process conditions fluctuate, the manifestation of defects will change accordingly. More importantly, this dynamic change will further affect the system's ability to judge defect types, leading to an increase in the misjudgment rate and difficulty in achieving precise control in actual production. Focusing on the specific challenges, the dynamic changes in defect morphology become the primary difficulty. Due to the instability of process parameters, defect characteristics such as distribution density or directionality will show non-uniformity. Traditional fixed rules have difficulty capturing these subtle differences. The complexity of this characteristic change brings another closely related problem: when identifying different defect types, the system has difficulty adjusting its internal judgment criteria based on real-time changes, resulting in missed detection or misreporting of certain defects. Therefore, how to dynamically adjust the identification method of defect characteristics and effectively control the misjudgment rate of different defect types when process parameter fluctuations cause changes in surface quality has become a key issue that needs to be solved urgently. Summary of the Invention

[0003] In order to solve the problems existing in the above-mentioned prior art, the purpose of this application is to provide an intelligent monitoring method for vacuum coating quality based on artificial intelligence.

[0004] The present application discloses an artificial intelligence-based intelligent monitoring method for vacuum coating quality, comprising the following steps: Collect coating equipment process parameters, monitor vacuum degree, deposition rate and temperature gradient, standardize data streams, and generate quantitative trend curves of process parameter fluctuations; identifying abnormal process stability intervals according to the quantitative trend curve and generating a clear surface image after denoising; Gridding and segmenting the clear surface image, identifying defect distribution characteristics, analyzing pinhole density concentration and fringe deflection angle non-uniformity, and generating a local defect feature data set; performing pattern matching on the pinhole distribution pattern and the stripe morphology features in the local defect feature data set to mark the abnormal defect category; Extracting the spatial coordinates of the abnormal defect category labeling results, obtaining the defect boundary sharpness and local texture complexity, and generating a defect microscopic feature description set; Correlation analysis is performed on the defect microscopic feature description set with gas flow fluctuation and power supply voltage jitter to calibrate the defect type and generate an updated defect classification basis; By using the updated defect classification basis, inputting new surface image data, detecting the difference in illumination reflection and the stripe deflection angle, generating a preliminary defect classification result, and adjusting the classification weight parameter; Based on the adjusted classification weight parameters, the defect type of the new surface image data is re-determined, the pinhole density concentration and defect boundary sharpness are analyzed, and a final defect classification list is generated.

[0005] Furthermore, the process parameters of the coating equipment are collected, vacuum degree, deposition rate and temperature gradient are monitored, data streams are standardized, and a quantitative trend curve of process parameter fluctuations is generated, including: Collect multi-point pressure data in the vacuum chamber, calculate the vacuum gradient distribution, mark the pressure values in time series, and generate a time domain characteristic sequence of vacuum fluctuations; Calculating the correlation coefficient between the temperature gradient and the vacuum gradient based on the time domain characteristic sequence of the vacuum degree fluctuation and the temperature field data, and generating a process parameter coupling relationship matrix; Using the weight coefficients of the process parameter coupling relationship matrix, the deposition rate data is corrected, a moving average value within a continuous time window is calculated, and a corrected deposition rate deviation curve is generated; The modified deposition rate deviation curve is subjected to piecewise linear fitting to extract slope variation characteristics, and a quantitative trend curve of process parameter fluctuation is generated by combining the time domain characteristic sequence of vacuum degree fluctuation and temperature gradient distribution data.

[0006] Furthermore, the identifying of abnormal process stability intervals based on the quantitative trend curve and generating a clear surface image after denoising includes: Calculate the slope values of adjacent data points in the quantitative trend curve, detect data points whose slope values and parameter amplitudes continuously exceed the threshold through a sliding window, determine the start and end time points of the abnormal interval, and generate an abnormal period sequence; According to the abnormal time period sequence, combined with the substrate position information and the coating rate, the spatial coordinates of the coating area corresponding to the abnormal time period are calculated; The reflected light intensity values are collected for the spatial coordinates of the coating area, converted into grayscale values to generate an original image, and processed by median filtering and Gaussian filtering to generate a clear surface image after denoising.

[0007] Furthermore, the gridding segmentation clears the surface image, identifies the defect distribution characteristics, analyzes the pinhole density concentration and the fringe deflection angle non-uniformity, and generates a local defect feature data set, including: The clear surface image is segmented using the watershed algorithm and divided into sub-regions according to the grid size. The number and position coordinates of defects in each sub-region are counted to generate a defect spatial distribution matrix. Calculate the sub-region defect density value based on the defect spatial distribution matrix, analyze the distance between defect points, and determine the pinhole density concentration value; For the sub-area where the pinhole density concentration value exceeds the threshold, the direction of the linear defect is detected, the deflection angle variance is calculated, and combined with the abnormal interval time nodes, a local defect feature data set containing the pinhole density concentration and the stripe deflection angle non-uniformity is generated.

[0008] Furthermore, the pattern matching of the pinhole distribution pattern and the stripe morphology features in the local defect feature data set to mark the abnormal defect category includes: Extract the pinhole spatial coordinate sequence from the local defect feature data set, calculate the direction vector and distance distribution standard deviation, measure the stripe length, width and curvature, and generate the geometric morphology feature vector; Based on the geometric morphology feature vector, similarity with historical feature vectors is calculated in the defect feature library, and parameter ranges of historical defect cases are retrieved; Based on the parameter range of the historical defect cases, the pinhole density concentration and fringe deflection angle values are compared to mark the abnormal defect category.

[0009] Furthermore, the extraction of the spatial coordinates of the abnormal defect category labeling results, obtaining the defect boundary sharpness and local texture complexity, and generating a defect microscopic feature description set include: Extract the coordinates of the defect area based on the abnormal defect category labeling results, control the microscope to move to the target position, adjust the focus in the vertical direction, and obtain a depth scanning image sequence; The grayscale gradient value of the defect edge is calculated for the depth scanning image sequence, the maximum value is selected as the boundary sharpness, the grayscale distribution regularity is analyzed, the texture complexity value is generated, and the defect microscopic feature description set is generated in combination with the focal length position information.

[0010] Furthermore, the defect microscopic feature description set is correlated with the gas flow fluctuation and the power supply voltage jitter for analysis, the defect type is calibrated, and an updated defect classification basis is generated, including: Align the boundary sharpness and texture complexity in the defect micro-feature description set with the gas flow data time series to generate a process parameter fluctuation feature matrix; Clustering parameter samples according to the process parameter fluctuation characteristic matrix and establishing a mapping relationship table between flow fluctuation and voltage jitter and defect type; According to the comparison between the mapping relationship table and the defect feature library, new categories are added to generate an updated defect classification basis.

[0011] Furthermore, the updated defect classification basis is used to input new surface image data, detect light reflection differences and stripe deflection angles, generate preliminary defect classification results, and adjust classification weight parameters, including: Based on the updated defect classification criteria, the grayscale difference of the new surface image pixels is calculated, the abnormal reflection intensity area is identified, the linear feature angle is extracted, and the feature data to be detected is generated; According to the comparison between the characteristic data to be detected and the characteristic range of the defect type, the defect type and severity are marked, and a preliminary defect classification result is generated; According to the distribution ratio of the preliminary defect classification results, the defect type determination threshold weight parameter is adjusted.

[0012] Furthermore, the defect type of the new surface image data is re-determined based on the adjusted classification weight parameters, the pinhole density concentration and defect boundary sharpness are analyzed, and a final defect classification list is generated, including: Calculating the comprehensive judgment score of the new surface image feature value and the weight coefficient according to the adjusted classification weight parameter, reclassifying the defect types, and generating a revised defect distribution map; For the pinhole-type area in the corrected defect distribution map, density values and boundary sharpness are calculated to generate a final defect classification list.

[0013] The advantage of the intelligent monitoring method for vacuum coating quality based on artificial intelligence described in the present application is that it collects process parameters in real time and dynamically monitors their changes, identifies abnormal process stability intervals and locates corresponding coating areas, performs optical scanning and image processing on abnormal areas, uses regional segmentation algorithms to identify local defect distribution characteristics, combines historical data for comparative analysis of defect types, further obtains defect microscopic characteristics through high-magnification microscopic imaging, establishes a correlation analysis model between process parameters and defect morphology, and finally, based on the updated defect classification basis, performs defect detection and classification on new surface image data, and continuously optimizes classification weight parameters. The present invention realizes accurate identification and classification of defects in the coating process, improves the accuracy and efficiency of defect detection, and provides effective technical support for coating quality control. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 This is the process of an intelligent monitoring method for vacuum coating quality based on artificial intelligence described in this application Figure 1 ; Figure 2 This is the process of an intelligent monitoring method for vacuum coating quality based on artificial intelligence described in this application Figure 2 . DETAILED DESCRIPTION

[0015] like Figure 1-Figure 2As shown, the present application describes an artificial intelligence-based intelligent monitoring method for vacuum coating quality, comprising: like Figure 1-Figure 2 As shown, S101, the process parameters are collected in real time through the coating equipment sensor, the vacuum degree change, deposition rate deviation and temperature gradient are dynamically monitored, the data stream is standardized, and a quantitative trend curve of the process parameter fluctuation is obtained.

[0016] Furthermore, in step S101, the sensor collects pressure data at multiple points in the vacuum chamber, obtains the vacuum gradient distribution through differential calculation, performs time series marking on the pressure value of each sampling point, calculates the deviation amplitude according to the preset pressure reference value, and obtains the time domain characteristic sequence of the vacuum fluctuation; Based on the time domain characteristic sequence of vacuum fluctuations and the temperature field data collected by the temperature sensor in the coating chamber, the Pearson correlation coefficient between the temperature gradient and the vacuum gradient is calculated. If the correlation coefficient exceeds the preset threshold, it is determined that a coupling effect exists, and the process parameter coupling relationship matrix is obtained; The weight coefficient in the process parameter coupling relationship matrix is used to perform weighted correction on the film growth rate collected in real time by the deposition rate sensor. The corrected deposition rate deviation curve is obtained by taking the moving average of the rate data in a continuous time window. The corrected deposition rate deviation curve is subjected to piecewise linear fitting, and the slope change characteristics of each segment are extracted. Combined with the vacuum degree time domain characteristic sequence and temperature gradient distribution data, the maximum and minimum values of each parameter are normalized to obtain the quantitative trend curve of process parameter fluctuations.

[0017] Specifically, during the actual operation of the coating equipment, the pressure distribution in the vacuum chamber is not uniform.

[0018] Specifically, in step S101, by arranging multiple pressure sensors at different positions in the cavity, such as installing capacitive pressure sensors at the top, bottom, side walls and near the target material of the cavity, the pressure value at each point can be monitored in real time. These sensors obtain pressure data at a sampling frequency of 100 Hz. The pressure value of each sampling point is accompanied by an accurate timestamp. The spatial gradient distribution of the vacuum degree can be obtained by calculating the pressure difference between adjacent measuring points.

[0019] For example, when the top pressure is 5×10^-3Pa and the bottom pressure is 8×10^-3Pa, the gradient value reflects the non-uniformity of the vacuum degree, which directly affects the uniformity of thin film deposition.

[0020] It should be noted that there is a close coupling relationship between vacuum fluctuations and temperature field changes. The temperature sensors in the coating chamber usually use thermocouples or infrared temperature measurement methods and are distributed in key locations such as the substrate stage, chamber wall and target. When the vacuum level fluctuates, the thermal conductivity characteristics of the gas molecules will change accordingly, resulting in a redistribution of the temperature field. The calculation process of the Pearson correlation coefficient involves the ratio of the covariance of two sets of time series data to the product of their respective standard deviations. When the calculated correlation coefficient is high, it indicates that the vacuum level change and the temperature gradient change are significantly synchronized. This coupling relationship matrix records the degree of mutual influence between different process parameters and provides a quantitative basis for subsequent parameter correction.

[0021] In one possible implementation, real-time monitoring of deposition rate typically uses a quartz crystal microbalance sensor. Its operating principle is based on the characteristic that the crystal oscillation frequency decreases with increasing mass. The raw deposition rate data is subject to interference from vacuum fluctuations and temperature changes. Therefore, it needs to be corrected based on the weight coefficients in the coupling relationship matrix. The weighted correction process multiplies the raw rate value by the corresponding correction factor, which is determined by the deviation of vacuum and temperature. The moving average is calculated using a fixed time window. For example, averaging the data within the previous 10 seconds can effectively filter out high-frequency noise and obtain a smoother rate change curve.

[0022] Preferably, when performing piecewise linear fitting on the corrected deposition rate deviation curve, the segmentation points can be adaptively determined according to the changing characteristics of the curve. The slope of each segment represents the speed of rate change within the time period. The normalization process maps the parameters of different dimensions to a unified interval from 0 to 1, so that the three different physical quantities of vacuum degree, temperature gradient and deposition rate can be compared and analyzed in the same coordinate system. The obtained quantitative trend curve comprehensively reflects the overall stability of the coating process and provides reliable data support for process optimization and quality control.

[0023] In one embodiment, the vacuum degree deviation calculation formula in step S101 can be expressed as: , in, It represents the vacuum degree deviation of the i-th sampling point at time t, represents the pressure value of the i-th sampling point at time t, P 基准 Indicates the preset pressure reference value.

[0024] like Figure 1-Figure 2As shown, S102, according to the quantitative trend curve, the abnormal interval of process stability is identified, and the corresponding coating area is located in combination with the time node. If the parameter exceeds the threshold, an optical scanner is used to collect an image of the abnormal area and filter it to obtain a clear surface image after denoising.

[0025] Furthermore, in step S102, a slope value is obtained by dividing the difference between adjacent data points in the quantitative trend curve by the time interval, and a sliding window is used to detect data points where the slope value and the parameter amplitude continuously exceed the preset threshold value, and the start time point and the end time point of the process stability abnormal interval are determined to obtain an abnormal period sequence; If there is an interval exceeding the threshold in the abnormal time period sequence, the starting and ending time points of the interval are used, combined with the substrate position information and coating rate recorded by the coating equipment, and the spatial coordinates of the coating area corresponding to the abnormal time period are calculated through the correspondence between time and position to determine the scope of the target area to be detected; For the target area to be detected, the optical scanner collects the reflected light intensity value of each point in the area in a raster scanning manner, converts the light intensity value into a grayscale value to form the original image, removes the impulse noise through median filtering, and then uses Gaussian filtering for smoothing to obtain a clear surface image after denoising.

[0026] Specifically, in coating process monitoring, quantitative trend curves record the changes in parameters such as vacuum level, temperature and deposition rate over time.

[0027] Specifically, in step S102, the slope value is calculated by dividing the parameter difference between two adjacent sampling points by the sampling time interval. When the sampling interval is 0.1 second, if the deposition rate is 2.5nm / s at a certain moment and 3.2nm / s at the next moment, the slope of this period is 7nm / s². This slope reflects the rate of change of process parameters. A sudden change in slope often indicates the occurrence of process abnormalities.

[0028] It should be noted that the sliding window detection method plays a key role in anomaly identification. The window size is usually set to include 20-30 consecutive data points. When the slope value or parameter amplitude of more than 60% of the data points in the window exceeds the preset threshold, the period is determined to be an abnormal interval; The abnormal period sequence not only records the start and end time of the abnormality, but also contains information about the severity of the abnormality. This time mark provides the basis for subsequent spatial positioning.

[0029] In one possible implementation, the mapping relationship between time and space depends on the operating parameters of the coating equipment. In the magnetron sputtering equipment, the substrate usually passes through the sputtering area at a constant speed in the range of 0.5-2.0m / min. When an abnormal period is detected from 10:15:30 to 10:16:45, based on the substrate transmission speed of 1.2m / min, it can be calculated that the length of the coating area corresponding to this period is 1.5 meters. Combined with the initial position information of the substrate in the equipment, the spatial coordinates of the abnormal area can be accurately located. This time-space correspondence makes subsequent fixed-point detection possible.

[0030] Preferably, the optical scanner uses a linear array CCD camera in conjunction with a precision mobile platform to achieve raster scanning. The scanning path covers the entire target area in a serpentine trajectory. The row spacing and column spacing are determined according to the required spatial resolution, usually 0.1-0.5 mm. The reflected light intensity of each scanning point is converted into a voltage signal through photoelectric conversion, and then converted into an 8-bit or 16-bit grayscale value through analog-to-digital conversion. The original image often contains noise caused by equipment vibration or ambient light interference.

[0031] For example, median filtering removes noise by replacing the grayscale value of each pixel with the median grayscale value of all pixels in its neighborhood. For a 3×3 filter window, the grayscale values of 9 pixels are sorted and the median value is taken as the new value of the center pixel. This method is particularly suitable for removing salt and pepper noise. The subsequent Gaussian filter further smoothes the image through weighted averaging. The weights are determined according to a two-dimensional Gaussian distribution, where the center pixel has the largest weight and the weights of surrounding pixels decrease with increasing distance. After these two steps of filtering, the obtained clear surface image can accurately reflect the true morphological characteristics of the coating layer, providing a reliable basis for subsequent defect analysis.

[0032] like Figure 1-Figure 2 As shown, S103, the clear surface image is gridded using a regional segmentation algorithm to identify local defect distribution characteristics, and the non-uniformity of pinhole density concentration and fringe deflection angle is obtained based on the abnormal interval time node analysis to obtain a local defect feature data set.

[0033] Furthermore, in step S103, the clear surface image is segmented using a watershed algorithm. The segmented image is divided into multiple sub-regions according to a preset grid size. The defect points in each sub-region are identified by comparing the grayscale value of each pixel with a preset threshold. The number and position coordinates of defects in each sub-region are counted to obtain a defect spatial distribution matrix. According to the defect spatial distribution matrix, the number of defect points in each sub-region is calculated and divided by the sub-region area to obtain the defect density value. For each defect point, its nearest neighbor defect point is found and the distance is calculated. If the average distance between defect points in a sub-region is less than a preset threshold, it is determined that the pinhole density concentration in this region is high, and the pinhole density concentration value of each sub-region is obtained; For the sub-regions where the pinhole density concentration value exceeds the preset threshold, Hough transform is used to detect the direction of linear defects in the image, and the deflection angle of each linear defect relative to the horizontal direction is calculated. The discreteness of the angle distribution is measured by calculating the variance of all deflection angles. Each sub-region is time-series marked in combination with the time nodes of the abnormal interval to obtain a local defect feature dataset that includes the pinhole density concentration and the non-uniformity of the stripe deflection angle.

[0034] Specifically, the application of watershed algorithm in image segmentation is based on the principle of topography, which regards the image grayscale value as the terrain height.

[0035] Specifically, in step S103, the algorithm first searches for a local minimum point as a seed point, and then simulates the process of rising water levels. When "water flows" from different seed points meet, a watershed is formed, that is, a regional boundary. In the coating surface image, the normal area presents a uniform grayscale distribution, while the defective area appears as a local grayscale anomaly due to differences in optical properties. Grid division usually adopts a fixed size, such as dividing a 1024×1024 pixel image into 32×32 sub-regions, each sub-region contains 32×32 pixels.

[0036] It should be noted that the identification of defect points depends on the reasonable setting of the grayscale threshold. The pinhole defect in the coating layer appears as a dark spot under reflected light, and its grayscale value is significantly lower than that of the normal area. By statistically analyzing the grayscale histogram of the entire image, a value lower than 70% of the average grayscale value is selected as the dark spot judgment threshold. The number and coordinates of defects in each sub-area are recorded in the defect space distribution matrix. The rows and columns of the matrix correspond to the grid positions of the image, and the element values contain the defect information of the area.

[0037] In one possible implementation, the calculation of defect density needs to take into account the actual physical size. If each sub-region corresponds to an actual area of 2mm×2mm on the substrate and 8 defect points are detected, the defect density is 2 / mm². The nearest neighbor distance calculation process is to traverse the other defect points in the same sub-region for each defect point, calculate the Euclidean distance, and find the minimum value. When the nearest neighbor distance of multiple defect points is less than 0.5mm, it indicates that the defects are clustered rather than randomly distributed. This clustering phenomenon is often related to local process anomalies.

[0038] For example, the principle of Hough transform detection of straight lines is to map points in image space to parameter space. In coating defect detection, stripe defects are usually caused by uneven sputtering of the target material or unstable substrate movement. The Hough transform converts each edge point in the image into a curve in the parameter space. The intersection of multiple curves corresponds to the straight line parameters in the original image. The angle of the detected straight line relative to the horizontal direction is the deflection angle. Under normal circumstances, the stripes should be parallel to the direction of substrate movement, and the deflection angle is close to 0 degrees.

[0039] Preferably, the calculation of the angle variance reflects the consistency of the stripe directions. If the deflection angles of the five stripes detected are 2 degrees, 3 degrees, 15 degrees, 18 degrees, and 20 degrees, respectively, the calculated variance value is large, indicating that the stripe directions are unevenly distributed. This non-uniformity is associated with process fluctuations in a specific time period. By mapping the defect features of each sub-area to the abnormal time nodes, a feature dataset with temporal and spatial correlation is formed.

[0040] In one embodiment, the defect density calculation formula in step S103 can be expressed as:

[0041] in, Expressed as the defect density of the j-th sub-region, Expressed as the number of defect points, Expressed as the sub-region area.

[0042] like Figure 1-Figure 2 As shown, S104, the pinhole distribution pattern and stripe morphology characteristics in the data set are identified, pattern matching is performed in combination with the defect feature library, and the pinhole density aggregation and stripe deflection angle are compared with historical data. If the feature parameters are out of range and are distributed in a clustered manner, they are marked as abnormal defect categories.

[0043] Furthermore, in step S104, the spatial coordinate sequence of the pinholes is extracted from the local defect feature data set, the angle between the line connecting adjacent pinholes and the horizontal direction is calculated to form a direction vector, the standard deviation of the statistical distance distribution is used to determine the pinhole distribution pattern, and the length, width and curvature of the stripes are measured. These values are combined to obtain a geometric morphological feature vector containing distribution characteristics and morphological parameters; Based on the numerical components in the geometric morphological feature vector, the sum of the products of the current vector and the corresponding components of the historical feature vector is calculated in the pre-established defect feature library and divided by the product of the vector modulus to obtain a similarity value. The historical defect cases whose similarity exceeds the preset threshold and their parameter ranges of pinhole density and fringe angle are retrieved. The retrieved parameter range of the historical defect cases is used as the judgment basis to compare the current pinhole density concentration value and the fringe deflection angle value. If both feature parameters exceed the historical parameter range and the standard deviation of the pinhole spatial distribution is less than the preset threshold, the area is marked as an abnormal defect category.

[0044] Specifically, in the spatial distribution analysis of coating defects, the calculation of direction vectors reveals the arrangement pattern of defects.

[0045] Specifically, in step S104, two adjacent pinhole defect points are selected and the angle between their connecting line and the horizontal reference line is calculated. When the connecting angles of multiple adjacent pinholes are similar, such as concentrated in the range of 15 to 20 degrees, it indicates that the defects tend to be arranged linearly. This arrangement is often related to scratches on the target surface or periodic vibrations of the substrate conveying mechanism. The standard deviation of the distance distribution quantifies the uniformity of the defect distribution. When the standard deviation is less than 1 mm, it indicates that the defects are highly concentrated, and when it is greater than 5 mm, it indicates a random distribution.

[0046] It should be noted that the measurement of stripe morphological parameters involves multiple dimensions. The stripe length is obtained by tracing the continuity of the pixels at the edge of the defect, and the width is taken as the maximum span perpendicular to the stripe direction. The degree of curvature is quantified using a curve fitting method, fitting the stripe centerline into a quadratic curve. The curve coefficient reflects the degree of curvature. These numerical parameters are combined to form a characteristic vector. For example, a typical characteristic vector may include: pinhole density 2.5 / mm², average spacing 0.8mm, main direction angle 18 degrees, stripe length 12mm, width 0.3mm, and curvature coefficient 0.05.

[0047] In one possible implementation, a defect feature library is constructed based on a large amount of historical production data. Each historical record contains not only the geometric characteristics of the defect but also the process parameters associated with the defect. When calculating cosine similarity, the current feature vector is compared component by component with the vectors in the library. For example, the current vector is [2.5, 0.8, 18, 12, 0.3, 0.05], and the historical vector is [2.3, 0.9, 16, 11, 0.35, 0.04]. The sum of the products of the components is 2.5 × 2.3 + 0.8 × 0.9 + ... = 35.6, which is then divided by the product of the two vector moduli to obtain a similarity of 0.92. For example, the similarity threshold is usually set at 0.85. Historical cases exceeding this value are considered to have a similar formation mechanism to the current defect. The retrieved historical cases show that the normal parameter range corresponding to similar defects is: pinhole density 1.5-2.0 / mm², fringe deflection angle 0-10 degrees. The current detection value obviously exceeds these ranges, especially the pinhole density reaches 2.5 / mm² and the deflection angle reaches 18 degrees, indicating that the process is abnormal.

[0048] Preferably, clustering is judged using the statistical characteristics of spatial distribution to calculate the standard deviation of all pinhole position coordinates. When the standard deviation in the X direction is 0.6 mm and the standard deviation in the Y direction is 0.7 mm, both of which are less than the preset threshold of 1.0 mm, it is determined to be a clustered distribution. This highly clustered defect distribution, combined with characteristic parameters that exceed the normal range, strongly indicates that there are systematic process problems in this area.

[0049] like Figure 1-Figure 2 As shown, S105, extracting spatial coordinates based on the abnormal defect category marking result, performing a high-magnification microscopic imaging depth scan on the marked area, obtaining subtle morphological data of defect boundary sharpness and local texture complexity, and obtaining a defect microscopic feature description set.

[0050] Furthermore, in step S105, the center coordinates and boundary coordinates of the defect area are extracted based on the abnormal defect category labeling result, the microscope stage is controlled to move to the target position through the coordinate mapping relationship, and the focus is adjusted at fixed intervals along the vertical direction using a high-magnification objective lens to obtain a depth scanning image sequence containing multiple focal planes; The grayscale gradient value is calculated for the defect edge pixels of each image in the depth scanning image sequence, and the maximum gradient value is selected as the boundary sharpness of the focal plane. At the same time, the grayscale distribution relationship of adjacent pixel pairs in different directions and distances is statistically analyzed to calculate the regularity of grayscale changes and obtain a value reflecting the local texture complexity. The boundary sharpness and texture complexity values of each focal plane are used in combination with the focal length position information to construct the three-dimensional morphological distribution of the defect. The maximum boundary sharpness, the mean texture complexity and the depth range parameters are combined to obtain a description set containing the microscopic morphological characteristics of the defect.

[0051] Specifically, the precise positioning of the microscope stage depends on the establishment of a coordinate mapping relationship.

[0052] Specifically, in step S105, when the abnormal defect marking result shows that a defect is located at the 256th row and the 512th column of the image coordinate system, these pixel coordinates need to be converted into the physical coordinates of the stage. Assuming that each pixel corresponds to an actual size of 10 microns on the substrate, the actual position of the defect is 2.56 mm in the X direction and 5.12 mm in the Y direction from the origin of the substrate. The stage is driven by a stepper motor and moves to the target position with an accuracy of 0.1 microns to ensure that the defect area is located in the center of the objective lens field of view.

[0053] It should be noted that the implementation of deep scanning requires continuous adjustment of the focal length along the Z-axis. The depth of field of a high-magnification objective lens is usually only a few microns, while coating defects may have high fluctuations. The position of the objective lens is controlled by a piezoelectric ceramic driver, and it moves at intervals of 2 microns each time. Eleven images with different focal planes are collected within a depth range of 20 microns. Each image records the clear details of a specific depth layer, while information at other depths appears blurred. This multi-focal plane imaging method can fully capture the three-dimensional morphological information of the defect.

[0054] In one possible implementation, the calculation of boundary sharpness is based on image gradient theory. For each pixel point on the defect edge, the grayscale difference between it and the eight adjacent pixels is calculated, and the maximum difference is taken as the gradient value of the point. When the defect edge is clear, the grayscale difference on both sides of the boundary is obvious, and the gradient value can reach more than 200, while the gradient value of the blurred edge is usually less than 50. By counting the gradient values of all pixels on the entire edge contour, the maximum value is taken as the boundary sharpness index of the focal plane.

[0055] For example, the quantification of texture complexity involves the analysis of the regularity of grayscale distribution. Within a local window of 5×5 pixels, the frequency of occurrence of different grayscale levels and the grayscale transition probability between adjacent pixels are counted. The grayscale distribution in uniform texture areas is concentrated and the transition probability is regular, while the grayscale distribution in complex texture areas is discrete and the transition probability is random. By calculating the standard deviation of the grayscale distribution and the entropy value of the transition probability, a texture complexity value between 0 and 1 is obtained. The larger the value, the more complex the texture.

[0056] Preferably, the construction of the three-dimensional morphological distribution requires the integration of information from multiple focal planes. According to the boundary sharpness value of each focal plane, the clarity of the defect contour at that depth layer is determined. The focal plane with the greatest sharpness corresponds to the actual height position of the defect surface. By recording the focal length values when the maximum sharpness is reached at different positions, the height distribution map of the defect can be reconstructed. Combined with the mean texture complexity of each focal plane, a characteristic parameter group including the boundary sharpness peak, average texture complexity and height variation range is formed. These microscopic characteristic parameters accurately describe the subtle morphology of the defect and provide a quantitative basis for in-depth analysis of the cause of the defect.

[0057] like Figure 1-Figure 2As shown, S106, the defect microscopic feature description set is combined with the gas flow fluctuation and the power supply voltage jitter to obtain the correlation analysis result between the process parameters and the defect morphology, and the defect type is calibrated according to the defect feature library to obtain an updated defect classification basis.

[0058] Furthermore, in step S106, the edge sharpness value and texture complexity in the defect microscopic feature description set are time-series aligned with the gas flow data recorded during the coating process, and the Pearson correlation coefficient between the two is calculated. At the same time, the jitter frequency and amplitude parameters are extracted from the voltage signal collected from the power supply output end to obtain a process parameter fluctuation feature matrix containing the flow fluctuation amplitude, voltage jitter parameters, and defect characteristic values; Based on the numerical combinations in the process parameter fluctuation feature matrix, a K-means clustering algorithm is used to group samples with similar parameters into the same category. The correspondence between process parameters and defect morphology in each category is analyzed, and a mapping table is established between defect types corresponding to flow fluctuation amplitude ranges and defect types corresponding to voltage jitter frequency ranges. The mapping relationship table is compared with the existing classification standards in the defect feature library. If a new process parameter combination is found to produce a defect morphology that is not defined in the feature library, the parameter combination and its corresponding defect feature are added to the feature library as a new category, while retaining the original category definition to obtain an updated defect classification basis.

[0059] Specifically, timing alignment is the key basis for realizing correlation analysis between process parameters and defect characteristics.

[0060] Specifically, in step S106, the gas flow sensor of the coating equipment records the argon flow value at a frequency of 100 Hz, and the acquisition time of the defect microscopic characteristics corresponds to the moment when the substrate passes through a specific position. Through timestamp matching, the gas flow state when a certain defect occurs can be determined.

[0061] For example, when a blurred defect with a boundary sharpness value lower than 50 is detected, the flow rate data 5 seconds before and after the corresponding moment is reviewed and it is found that the flow rate suddenly changes from 50 sccm to 35 sccm within 2.5 seconds. This rapid change has a temporal correspondence with the blurred defect boundary phenomenon.

[0062] It should be noted that the calculation of the Pearson correlation coefficient can quantify the degree of linear correlation between two variables. In this application, the boundary sharpness values of a series of defects are used as one set of data, and the gas flow fluctuation amplitude at the corresponding moment is used as another set of data. By calculating the covariance of the two sets of data and dividing it by the product of their respective standard deviations, a correlation coefficient between -1 and 1 is obtained. When the coefficient is close to -0.8, it indicates that the greater the flow fluctuation, the more blurred the boundary, showing a strong negative correlation. The voltage jitter parameters are extracted by performing spectral analysis on the power supply output signal to identify the main jitter frequency components and their amplitudes.

[0063] In one possible implementation, the K-means clustering algorithm requires a predetermined number of clusters. Based on historical experience, coating defects can typically be categorized into 3-5 categories. The algorithm first randomly selects K samples as initial cluster centers. It then calculates the Euclidean distance from each sample to each center, assigning the sample to the closest category. The cluster centers are then iteratively updated until convergence is achieved. For example, after clustering, it was found that the first category contained samples with flow fluctuations less than 5% and stable voltage, corresponding to normal coating, the second category had flow fluctuations of up to 15%, corresponding to pinhole defects, and the third category had voltage jitter frequencies above 1kHz, corresponding to streak defects.

[0064] For example, the mapping table is based on the statistical patterns of a large number of samples. When the flow rate fluctuation amplitude is within the range of 10-20 sccm, 80% of the samples show an increase in pinhole density. When the voltage jitter frequency exceeds 500 Hz and the amplitude exceeds 5V, 90% of the samples exhibit streak defects. This statistical pattern forms a corresponding relationship between process parameter ranges and defect types. In actual applications, a new phenomenon was also discovered. When flow fluctuations and voltage jitters occur simultaneously, spiral defects are generated, which is a type not recorded in the original feature library.

[0065] Preferably, the update of the defect classification basis follows the conservative principle. The original classification standards have been verified for a long time and have high reliability, so they are retained unchanged. Newly discovered defect types need to be repeatedly verified multiple times before they can be officially added to the feature library. The update process includes defining the feature parameter range of the new category, setting the judgment threshold, and establishing association rules with process parameters.

[0066] like Figure 1-Figure 2 As shown, S107, using the updated defect classification basis, inputting real-time new surface image data, performing defect detection based on illumination reflection differences and stripe deflection angles, obtaining preliminary defect classification results, and adjusting classification weight parameters.

[0067] Furthermore, in step S107, the newly acquired surface image is preprocessed using the threshold parameters defined in the updated defect classification basis. The difference between the grayscale value of each pixel in the image and the mean value of the surrounding pixels is calculated as the reflection intensity value. Areas where the reflection intensity value exceeds the normal range are identified as potential defect points. At the same time, edge detection is used to extract linear features and calculate their angles relative to the horizontal direction, thereby obtaining feature data to be detected including reflection intensity distribution and fringe angles. The feature data to be detected is compared with the feature range of each defect type in the classification basis. If the reflection intensity value is lower than the pinhole threshold, it is marked as a pinhole defect. If the detected stripe angle deviates from the reference angle by more than a preset range, it is marked as a stripe defect. If both features are met, it is marked as a mixed defect. The degree of deviation is determined by calculating the difference between the feature value and the threshold and divided into three levels: mild, moderate, and severe. The preliminary classification results including the defect type, specific location coordinates and severity rating are obtained; The proportion of each defect type and the distribution ratio of each severity level in the preliminary classification results are used as the characteristic statistical value of the current batch, which are compared with the historical average value obtained by cumulative calculation of multiple batches in advance, and the judgment threshold weight parameter of the corresponding defect type is adjusted according to the deviation ratio.

[0068] Specifically, the calculation of the reflection intensity value is based on the diffuse reflection characteristics in optical principles.

[0069] Specifically, in step S107, on the uniformly coated surface, the reflection intensity of the incident light presents a regular distribution, and the grayscale value of each pixel should be similar to that of its neighboring pixels. A 5×5 window is selected during calculation, and the average grayscale value of the central pixel and the surrounding 24 pixels is calculated. The difference between the two is the reflection intensity deviation. The deviation in a normal coated area is usually within ±10 gray levels, while the pinhole defect causes light scattering due to deep depression, and the deviation can reach more than -50 gray levels. This significant negative deviation becomes an important feature for identifying pinhole defects.

[0070] It should be noted that the detection of stripe angles involves the precise extraction of edge directions. The Sobel operator is used to calculate the gradients in the horizontal and vertical directions respectively, and the local direction angle of the edge is obtained by the inverse tangent value of the gradient direction. For stripe defects, the edges present a continuous linear distribution, and the direction angles of multiple edge points should remain consistent. When the direction angles of more than 10 consecutive edge points deviate from the horizontal reference by more than 15 degrees, they are judged as abnormal stripes. The reference angle is usually set to 0 degrees, corresponding to the ideal horizontal stripe direction, which is consistent with the movement direction of the substrate in the coating equipment.

[0071] In one possible implementation, the severity grading criteria are based on the quantified degree of feature deviation, taking pinhole defects as an example: When the reflection intensity deviation is between -30 and -50 gray levels, it is classified as mild, indicating that the defect depth is shallow; Deviations between -50 and -80 are considered moderate, and the defects have affected the functionality of the coating. If it exceeds -80, it is severe and may cause local failure of the coating layer; The classification of streak defects is based on the degree of angular deviation: 5-15 degrees is mild; 15-30 degrees is moderate; Over 30 degrees is severe; The severity of a mixed defect is determined by the more severe of the two features.

[0072] For example, the establishment of historical averages requires long-term data accumulation. In actual production, the number and severity distribution of various defects will be counted after each batch of coating is completed. Assume that in the past 100 batches, the average proportion of pinhole defects was 15%, of which 60% were mild, 30% were moderate, and 10% were severe. When the inspection results of a new batch show that the proportion of pinhole defects reaches 25% and the severe proportion rises to 20%, it indicates that there may be abnormalities in the current process.

[0073] Preferably, the weight parameter is adjusted using a dynamic feedback mechanism. When the actual detection rate of a certain type of defect is continuously higher than the historical average, the judgment threshold of this type of defect is appropriately lowered to improve the detection sensitivity. For example, if the pinhole defect rate of three consecutive batches exceeds 1.5 times the historical average, the reflection intensity threshold for pinhole judgment will be adjusted from -30 to -25, so that more boundary cases can be identified as defects. This adaptive adjustment mechanism can respond to changes in process conditions in a timely manner, while ensuring detection accuracy and improving sensitivity to abnormal situations, thereby achieving more accurate defect classification and quality control.

[0074] In one embodiment, the reflection intensity deviation calculation formula in step S107 can be expressed as: , in, Expressed as the reflection intensity deviation, It is represented by the grayscale value at (x, y), Ω: 5×5 neighborhood (excluding the center point), (u, v) represents the pixel coordinate index within the neighborhood Ω, It is expressed as the average gray value of the neighborhood Ω.

[0075] like Figure 1-Figure 2 As shown, S108, based on the adjusted classification weight parameters, the defect type is re-determined for the new surface image data, and a secondary analysis is performed on the pinhole density concentration and defect boundary sharpness to obtain a final defect classification list, record the feature data of the suspected misjudgment area, and obtain a supplementary feature data set.

[0076] Furthermore, in step S108, based on the adjusted classification weight parameters, the feature value of each pixel in the new surface image data is multiplied by the weight coefficient of the corresponding type, the pinhole type weight is multiplied by the reflection intensity deviation value, and the stripe type weight is multiplied by the angle deviation value, and the sum is calculated to obtain a comprehensive judgment score. The defect type is reclassified based on the updated threshold value to obtain a revised defect distribution map. For the areas marked as pinholes in the corrected defect distribution map, the number of pinholes per unit area is counted and divided by the area to obtain the density value. The standard deviation of the distance between adjacent pinholes is calculated to determine the degree of aggregation. The maximum grayscale difference between adjacent pixels at the defect edge is extracted as the boundary sharpness index. The final defect classification list containing defect type, location coordinates, density aggregation, and boundary sharpness is obtained. The areas in the final defect classification list whose boundary sharpness values are less than the preset lower limit or greater than the preset upper limit are marked. If the comprehensive judgment score of a certain area is in the middle range of the difference between the thresholds of two adjacent defect types, the area is defined as a suspected misjudgment area, and its complete image features and parameter values are extracted to obtain a supplementary feature dataset containing the features of the suspected misjudgment area.

[0077] Specifically, the application of weight coefficients reflects the differences in the importance of different defect types in the current production batch.

[0078] Specifically, in step S108, when the production line detects frequent pinhole defects, the pinhole type weight coefficient is adjusted from the initial 1.0 to 1.3, while the stripe type weight coefficient remains at 0.9. For a certain pixel point, if its reflection intensity deviation is -40 grayscale levels and the angle deviation is 8 degrees, the comprehensive judgment score is calculated as: 1.3×40+0.9×8=59.2; This score reflects the comprehensive degree to which the point has both defect characteristics. When the score exceeds the set threshold of 50, the pixel is marked as a defect point.

[0079] It should be noted that the calculation of defect density needs to take into account the actual physical size conversion. In image processing, a 100×100 pixel window is selected as the statistical unit, corresponding to an actual area of 1mm×1mm on the substrate. The number of pixels marked as pinhole defects in the statistical window is counted. Assuming that 25 defect points are detected, the density value is 25 / mm². The degree of aggregation is determined based on the distribution characteristics of the nearest neighbor distance. The distance from each defect point to its nearest defect point is calculated to obtain a set of distance data. The standard deviation of this set of data is calculated to quantify the uniformity of the distribution. A standard deviation less than 2 pixels indicates high aggregation, and greater than 10 pixels indicates random distribution.

[0080] In one possible implementation, the abnormality judgment of edge sharpness is based on statistical distribution laws. By analyzing a large number of normal defect samples, it is found that the edge sharpness values are generally normally distributed with a mean of 120 and a standard deviation of 20. Therefore, sharpness values less than 80 or greater than 160 are defined as abnormal. Too low sharpness may be caused by image blur or excessive defect depth, while too high sharpness may be the result of noise interference or excessive edge enhancement. These abnormalities may lead to misjudgment of defect type. For example, the identification of suspected misjudgment areas focuses on the fuzzy areas of the classification boundaries. Assume that the judgment threshold for pinhole defects is 55, and the threshold for stripe defects is 45, with a difference of 10. When the comprehensive judgment score of a certain area falls within the range of 48-52, it is very close to the threshold of both defect types, and the classification result is uncertain. This boundary situation often corresponds to complex defects or new defect forms, and its characteristic data needs to be specially recorded.

[0081] Preferably, the construction of the supplementary feature data set not only includes the original image data of the suspected misjudgment area, but also records the multi-dimensional feature parameters of the area, including reflection intensity distribution histogram, edge direction statistics, texture complexity index, contrast with the surrounding normal area, etc. These detailed feature descriptions provide valuable samples for subsequent classification rule optimization. By accumulating sufficient supplementary feature data, new defect patterns can be discovered, the classification system can be improved, and the accuracy and adaptability of the overall detection can be improved.

[0082] Those skilled in the art can make various other corresponding changes and deformations based on the technical solutions and concepts described above, and all of these changes and deformations should fall within the scope of protection of the claims of this application.

Claims

1. An intelligent monitoring method for vacuum coating quality based on artificial intelligence, characterized in that: include: Collect coating equipment process parameters, monitor vacuum degree, deposition rate and temperature gradient, standardize data streams, and generate quantitative trend curves of process parameter fluctuations; identifying abnormal process stability intervals according to the quantitative trend curve and generating a clear surface image after denoising; Gridding and segmenting the clear surface image, identifying defect distribution characteristics, analyzing pinhole density concentration and fringe deflection angle non-uniformity, and generating a local defect feature data set; performing pattern matching on the pinhole distribution pattern and the stripe morphology features in the local defect feature data set to mark the abnormal defect category; Extracting the spatial coordinates of the abnormal defect category labeling results, obtaining the defect boundary sharpness and local texture complexity, and generating a defect microscopic feature description set; Correlation analysis is performed on the defect microscopic feature description set with gas flow fluctuation and power supply voltage jitter to calibrate the defect type and generate an updated defect classification basis; By using the updated defect classification basis, inputting new surface image data, detecting the difference in illumination reflection and the stripe deflection angle, generating a preliminary defect classification result, and adjusting the classification weight parameter; Based on the adjusted classification weight parameters, the defect type of the new surface image data is re-determined, the pinhole density concentration and defect boundary sharpness are analyzed, and a final defect classification list is generated.

2. The method for intelligent monitoring of vacuum coating quality based on artificial intelligence according to claim 1, characterized in that: Generating a quantitative trend curve of process parameter fluctuations specifically includes: Collect multi-point pressure data in the vacuum chamber, calculate the vacuum gradient distribution, and perform time-series marking on the pressure values to generate a time domain feature sequence; According to the time domain characteristic sequence and temperature field data, the correlation coefficient between the temperature gradient and the vacuum gradient is calculated to generate the process parameter coupling relationship matrix; The weight coefficient of the matrix is used to correct the deposition rate data, and the moving average value in the continuous time window is calculated to generate the deposition rate deviation curve; The deviation curve is fitted piecewise linearly to extract the slope change characteristics, and the quantitative trend curve is generated by combining the time domain feature sequence and temperature gradient distribution data.

3. The method for intelligent monitoring of vacuum coating quality based on artificial intelligence according to claim 1, characterized in that: The generating of a clear surface image after denoising specifically includes: Calculate the slope values of adjacent data points in the quantitative trend curve, and use the sliding window to detect points that continuously exceed the threshold to determine the abnormal period sequence; Calculate the spatial coordinates of the corresponding coating area based on the abnormal time period sequence and substrate position information; The reflected light intensity values are collected according to the spatial coordinates to generate the original image, and a clear surface image is generated through median filtering and Gaussian filtering.

4. The method for intelligent monitoring of vacuum coating quality based on artificial intelligence according to claim 1, characterized in that: The generating of the local defect feature data set specifically includes: The watershed algorithm is used to segment the clear surface image into grid sub-regions, and the number of defects and location coordinates are counted to generate a spatial distribution matrix; Calculate the defect density value of the sub-region according to the matrix, and analyze the distance between defect points to determine the pinhole density concentration; For the sub-regions where the aggregation exceeds the threshold, the direction of the linear defect is detected and the variance of the deflection angle is calculated, and a data set is generated by combining the time nodes of the abnormal interval.

5. The method for intelligent monitoring of vacuum coating quality based on artificial intelligence according to claim 1, characterized in that: The marking abnormal defect categories specifically include: Extract the pinhole coordinate sequence from the data set to calculate the direction vector and distance distribution standard deviation, measure the stripe length, width and curvature to generate the geometric morphological feature vector; Calculate the similarity between the feature vector and the historical features in the defect feature library and retrieve the parameter range of the historical defect cases; Compare the current pinhole density concentration and fringe deflection angle values. If they exceed the historical range and are clustered, they are marked as abnormal defects.

6. The method for intelligent monitoring of vacuum coating quality based on artificial intelligence according to claim 1, characterized in that: The generating of the defect microscopic feature description set specifically includes: Control the microscope to perform depth scanning according to the coordinates of the abnormal defect to obtain a multi-focal plane image sequence; Calculate the maximum value of the grayscale gradient of the defect edge of each image as the edge sharpness, and analyze the regularity of the grayscale distribution to generate the texture complexity value; The three-dimensional morphological distribution is constructed by combining the boundary sharpness, texture complexity and focal length position of each focal plane to form a microscopic feature description set.

Images