Flatness prediction method, flatness prediction device, and electronic device

Abstract

The present application relates to the technical field of flatness detection, and provides a flatness prediction method, a flatness prediction device and electronic equipment, comprising: obtaining a feature parameter set of a workpiece to be processed before a film plating process and a device parameter set of a processing device; inputting the feature parameter set and the device parameter set into a first flatness prediction model to obtain a probability of unqualified flatness of the workpiece to be processed output by the first flatness prediction model; and in the case that the probability of unqualified flatness of the workpiece to be processed is greater than a preset probability threshold, determining that the workpiece to be processed is a workpiece to be detected for subsequent flatness detection. The present application can output unqualified products in a processing batch more accurately by predicting the probability of unqualified flatness of the workpiece to be processed, so as to facilitate the detection of unqualified products with a higher recall rate at a lower sampling rate in subsequent manual flatness detection.

Description

Technical Field

[0001] This invention relates to the field of flatness detection technology, and in particular to a flatness prediction method, a flatness prediction device, and an electronic device. Background Technology

[0002] The production process of cover glass (CG) in consumer electronics products mainly includes the following steps: CNC (Computerized Numerical Control) process, fine carving process, chemical pressure process, asymmetric reverse grinding process, and coating process. Specifically, the CNC process is used to carve the overall outline of the glass; the fine carving process is used for precision machining of functional structures such as microphone holes; the chemical pressure process is used to form a compressive stress layer on the glass surface through ion exchange, improving the glass's mechanical strength; the asymmetric reverse grinding process is used to apply reverse suction before coating to compensate for convex deformation caused by suction during the coating process; and the coating process is used to deposit an optical thin film on the glass surface to improve optical performance and touch experience.

[0003] After the aforementioned coating process is completed, the glass's color, flatness, and transmittance need to be tested. In existing technologies, flatness testing is typically performed at the end of the production process, using a laser interferometer or contact probe for manual sampling after the coating process. During manual sampling, a high sampling rate is often required to ensure a high recall rate for non-conforming products (i.e., a high detection rate of non-conforming products), leading to high labor and equipment costs. Furthermore, even with a high sampling rate, the improvement in non-conforming product recall is quite limited. Therefore, achieving a high recall rate for non-conforming glass flatness testing under lower sampling rates is a pressing technical problem that needs to be solved. Summary of the Invention

[0004] This invention provides a flatness prediction method, a flatness prediction device, and an electronic device to solve the technical problem that existing flatness detection schemes are unable to achieve a high recall rate of flatness defective products under low sampling rate conditions.

[0005] The first aspect of this invention provides a method for predicting flatness, comprising the following steps: Obtain the feature parameter set of the workpiece before the coating process, and the equipment parameter set of the processing equipment. The feature parameter set includes: feature parameters after at least one processing step before the coating process. Input the feature parameter set and the equipment parameter set into the first flatness prediction model to obtain the probability that the flatness of the workpiece to be processed is unqualified, as output by the first flatness prediction model. If the probability of the flatness of the workpiece to be processed being unqualified is greater than the preset probability threshold, the workpiece to be processed is determined as the workpiece to be inspected for subsequent flatness inspection. The first flatness prediction model is trained based on the flatness of the feature parameter set samples, equipment parameter set samples, and corresponding workpiece samples. The preset probability threshold is determined based on the recall rate of unqualified products and the recall rate of qualified products after the first flatness prediction model is trained.

[0006] According to the method for predicting flatness provided by the present invention, before inputting the feature parameter set and the equipment parameter set into the first flatness prediction model, the method further includes: Feature expansion operations are performed on a single feature parameter in the feature parameter set to obtain extended feature parameters, and / or cross-process feature expansion operations are performed on related feature parameters corresponding to different processing steps in the feature parameter set to obtain extended feature parameters. All extended feature parameters are merged into the feature parameter set.

[0007] According to the planarity prediction method provided by the present invention, at least one processing step before the coating process includes: a chemical pressure process and an asymmetric anti-grinding process, and the feature parameter set includes: a first feature parameter subset after the chemical pressure process and a second feature parameter subset after the asymmetric anti-grinding process; The first subset of characteristic parameters includes: the surface compressive stress and the depth of the compressive stress layer of the workpiece to be processed; The second subset of characteristic parameters includes: the thickness difference of each preset thickness measurement point on the workpiece before and after asymmetric back-grinding, and the height value of each preset flatness measurement point after asymmetric back-grinding. Feature expansion operations are performed on individual feature parameters from the feature parameter set to obtain expanded feature parameters, including expansion in at least one of the following ways: Perform a difference operation on the height values ​​corresponding to adjacent preset flatness measurement points, and use the distance from each preset flatness measurement point to the center point to weight the height values ​​of each preset flatness measurement point, to obtain extended feature parameters including height difference and weighted height value; For each feature parameter in the feature parameter set, the feature weight parameter is expanded according to the debugging time of the corresponding processing equipment of the cover glass to be processed; Based on the height values ​​of each preset flatness measurement point, extended characteristic parameters are obtained, including those reflecting the overall degree of curvature, the degree of torsional deformation, the direction of warping deformation, and / or the degree of local warping. Based on the difference between the average height values ​​of two groups of flatness measurement points obtained by dividing the height values ​​of each preset flatness measurement point according to different mirror directions, an extended characteristic parameter including mirror deviation in at least one mirror direction is obtained. Based on the difference between the maximum and minimum height values ​​of each preset flatness measurement point, extended characteristic parameters including the deviation of the maximum asymmetric point are obtained; Based on the correlated feature parameters corresponding to different processing steps in the feature parameter set, cross-process feature extension operations are performed to obtain extended feature parameters, including extension in at least one of the following ways: Calculate the ratio of the average surface compressive stress to the average thickness difference to obtain extended characteristic parameters across processes, including the ratio; Calculate the product of the standard deviation of the difference between the compressive stress layer depth and thickness to obtain extended characteristic parameters across processes, including the product.

[0008] According to the flatness prediction method provided by the present invention, the preset flatness measurement points are determined based on the following method: Obtain the height measurement values ​​of M original points of the workpiece sample obtained by the flatness measuring device, where M is greater than 0; Using the height measurements of M original points as sample data, and the flatness detection results as labels, a second flatness prediction model is trained. After the second flatness prediction model is trained, the feature importance weights of each output original point are sorted from largest to smallest, and the top m original points with the largest importance weights are selected as preset flatness measurement points, where m is greater than 0 and less than M.

[0009] According to the method for predicting flatness provided by the present invention, the first flatness prediction model is a binary classification model, with unqualified as the positive class label and qualified as the negative class label, and the training method is as follows: Input the feature parameter set sample and equipment parameter set sample of any workpiece sample into the first flatness prediction model. The first flatness prediction model performs multiple rounds of iterative fitting on any input workpiece sample to output the non-conformity prediction probability. Substitute the output failure prediction probability and the flatness acceptance label into the loss function. When the loss function converges, the model training is complete.

[0010] According to the flatness prediction method provided by the present invention, during the process of multiple rounds of iterative fitting of any input workpiece sample by the first flatness prediction model, sample weights are generated based on the production time and sample category of any workpiece sample, so that the sample weight of the workpiece sample that is closer to the current time is larger, and in each batch of workpiece samples at the same time, the sample weight of the unqualified positive sample is greater than the sample weight of the qualified negative sample.

[0011] According to the method for predicting flatness provided by the present invention, the loss function is obtained by weighting and summing a binary classification loss term and a physical consistency constraint penalty term according to a preset penalty term coefficient; For any workpiece sample, if the first difference between the stress layer volume energy in the corresponding feature parameter set sample and the mean stress layer volume energy of all samples exceeds a preset multiple of the standard deviation of the stress layer volume energy of all samples, the physical consistency constraint penalty term is determined by the second difference between the first difference and the preset multiple of the standard deviation of the stress layer volume energy of all samples.

[0012] According to the planarity prediction method provided by the present invention, before inputting the feature parameter set sample and the equipment parameter set sample of any workpiece sample into the first planarity prediction model, the method further includes: Group the feature parameter set samples of all workpiece samples according to different processes. Based on the recent standard deviation and historical benchmark standard deviation of the feature parameter set samples of any process, determine the stability index corresponding to any process. The recent standard deviation is the standard deviation of the feature parameter set samples of any process within the preset time period from the current time, and the historical benchmark standard deviation is the standard deviation of the feature parameter set samples of any process during the historical stable operation period. The first flatness prediction model is input into the feature parameter set sample and the equipment parameter set sample of any workpiece sample. The process of the first flatness prediction model performing multiple rounds of iterative fitting on any input workpiece sample also includes: The attention weight of the feature parameter set samples for any process is calculated based on the stability index and the corresponding preset amplification factor.

[0013] According to the flatness prediction method provided by the present invention, the preset probability threshold is determined as follows: Search within the interval [0.0, 1.0] to find the candidate probability threshold that makes the first recall rate of unqualified positive samples closest to the preset recall rate, and record the second recall rate of qualified negative samples corresponding to the first recall rate. If the second recall rate is within the preset recall rate range of negative class samples, the candidate probability threshold is determined to be the preset probability threshold. If the second recall rate is less than the lower limit of the preset recall rate interval for negative class samples, the candidate probability threshold is increased. When the candidate probability threshold is increased to the point that the second recall rate rises back to the lower limit, the increased candidate probability threshold is determined as the preset probability threshold. If the second recall rate is greater than the upper limit of the preset recall rate range for negative samples, the candidate probability threshold is reduced. When the candidate probability threshold is reduced to the point where the second recall rate returns to the upper limit, the reduced candidate probability threshold is determined as the preset probability threshold.

[0014] The method for predicting flatness according to the present invention further includes: After the first flatness prediction model is trained, the model is periodically trained and updated at preset intervals, using the feature parameter set and equipment parameter set within the period as samples, and the corresponding flatness detection results as labels. After updating the first flatness prediction model, the following is also included: If the total number of workpieces to be processed is found to be greater than the first quantity threshold, and the ratio of the number of products predicted as unqualified by the first flatness prediction model to the total number of workpieces to be processed is greater than the preset sampling rate limit, the updated first flatness prediction model will be rolled back to the next newest version of the first flatness prediction model. If the total number of workpieces to be processed is found to be greater than the second quantity threshold, and the ratio of the number of products predicted as defective by the first flatness prediction model to the total number of workpieces to be processed is less than the lower limit of the preset sampling rate, the preset probability threshold is redefined.

[0015] A second aspect of the present invention provides a flatness prediction device, comprising: The data acquisition module is used to acquire the feature parameter set of the workpiece to be processed before the coating process, and the equipment parameter set of the processing equipment. The feature parameter set includes: feature parameters after at least one processing step before the coating process. The model execution module is used to input the feature parameter set and the equipment parameter set into the first flatness prediction model to obtain the probability that the flatness of the workpiece to be processed is unqualified, as output by the first flatness prediction model. The workpiece determination module is used to determine the workpiece to be processed as the workpiece to be inspected when the probability of the flatness of the workpiece to be processed being unqualified is greater than a preset probability threshold, so as to be used for subsequent flatness inspection. The first flatness prediction model is trained based on the flatness of the feature parameter set samples, equipment parameter set samples, and corresponding workpiece samples. The preset probability threshold is determined based on the recall rate of unqualified products and the recall rate of qualified products after the first flatness prediction model is trained.

[0016] A third aspect of the present invention provides an electronic device, including a processor, a communication interface, a memory, and a communication bus. The processor, the communication interface, and the memory communicate with each other through the communication bus. The processor can call logical instructions in the memory to execute the flatness prediction method described in the first aspect.

[0017] The flatness prediction method, flatness prediction device, and electronic equipment provided by this invention, based on the feature parameter set of the workpiece before the coating process and the equipment parameter set of the processing equipment, can accurately predict the probability of whether the flatness of the workpiece is unqualified based on a trained first flatness prediction model. Then, workpieces to be inspected are selected according to a preset probability threshold, achieving early warning of workpieces to be inspected before manual flatness inspection. Since the preset probability threshold is set based on the recall rate of unqualified and qualified products, the proportion of unqualified workpieces selected according to the preset probability threshold to the total number of workpieces to be inspected in a production batch is relatively small. Furthermore, because the first flatness prediction model can accurately predict the probability of whether the flatness is qualified, it ensures that a large proportion of the predicted unqualified workpieces are truly unqualified products, thus guaranteeing a high recall rate of unqualified products. This is beneficial for subsequent manual flatness inspection, where the detected unqualified products have a high recall rate even with a lower sampling rate. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in this 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 some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating the flatness prediction method provided by the present invention.

[0020] Figure 2 This is a schematic diagram illustrating the solution of the diagonal height difference, which reflects the degree of torsional deformation, in the flatness prediction method provided by this invention.

[0021] Figure 3 This is a schematic diagram illustrating the solution of the height difference between the two ends of the warping deformation direction in the flatness prediction method provided by this invention.

[0022] Figure 4 This is a schematic diagram of the flatness prediction device provided by the present invention.

[0023] Figure 5 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0025] The flatness prediction method of the present invention can be used for flatness prediction of silicon compound workpieces, wherein silicon compound workpieces include, but are not limited to, workpieces such as glass and wafers. The following embodiments use cover glass (or other flat glass) as an example to illustrate the flatness prediction method of the present invention in detail.

[0026] The flatness prediction method of this invention, as described in the embodiments, is as follows: Figure 1 As shown, the procedure includes steps S110 to S130.

[0027] Step S110: Obtain the feature parameter set of the cover glass to be processed before the coating process, and the equipment parameter set of the processing equipment. The feature parameter set includes: feature parameters after at least one processing step before the coating process.

[0028] The feature parameter set before the coating process includes various feature parameters of the cover glass before coating and after the previous process, such as: surface pressure of the cover glass, thickness at different points and / or height value representing flatness (with the height of the center point of the cover glass as 0, the height value of each other point relative to the center point is positive if it is higher than the center point, and negative if it is lower than the center point), etc. Each feature parameter can be detected by corresponding measuring equipment.

[0029] The processing equipment may include at least one piece of equipment for the chemical pressure process, the asymmetric back-grinding process, and the coating process; that is, the equipment parameter set may be the equipment parameter set corresponding to at least one piece of equipment for the chemical pressure process, the asymmetric back-grinding process, and the coating process. Preferably, since the equipment parameters of the coating equipment have a direct impact on the flatness of the cover glass after coating, the processing equipment includes the equipment corresponding to the coating process. The equipment parameter set contains the equipment parameters of the coating equipment, which can be obtained from the parameter setting information of the coating equipment before coating. For example, the equipment parameter set includes: coating position information (row and column number of the cover glass in the coating equipment), product model of the cover glass, line number, factory number, and machine number. Of course, the processing equipment may also include the processing equipment of the preceding processes.

[0030] Step S120: Input the feature parameter set and equipment parameter set into the first flatness prediction model to obtain the probability of the flatness of the cover glass to be processed being unqualified, output by the first flatness prediction model. The first flatness prediction model is trained based on the feature parameter set samples, equipment parameter set samples, and the corresponding flatness compliance labels of the cover glass samples. It can be understood that each cover glass sample corresponds to a set of feature parameter set samples, equipment parameter set samples, and corresponding flatness compliance labels. For example, the flatness compliance label for a cover glass sample can be defined as: 1 for unqualified, and 0 for qualified.

[0031] Since the first flatness prediction model is trained based on the flatness qualification labels of the feature parameter set samples, equipment parameter set samples, and corresponding cover glass samples, a mapping relationship between the feature parameters and equipment parameters of the cover glass and the probability of flatness qualification is established during training. Therefore, in the model inference stage, it can accurately predict the probability of flatness qualification of the cover glass to be processed after coating based on the feature parameter set and equipment parameter set of the cover glass to be processed currently being produced.

[0032] Step S130: If the probability of the flatness of the cover glass to be processed being unqualified is greater than a preset probability threshold, the cover glass to be processed is determined as the cover glass to be tested for subsequent flatness testing. The preset probability threshold is determined based on the recall rate of unqualified products and the recall rate of qualified products after the first flatness prediction model has been trained.

[0033] It is understandable that the recall rate of non-conforming products reflects the proportion of cover glass samples predicted by the first flatness prediction model to be non-conforming out of the actual non-conforming cover glass samples, while the recall rate of conforming products reflects the proportion of cover glass samples predicted by the first flatness prediction model to be conforming out of the actual conforming cover glass samples. Therefore, appropriate preset probability thresholds can be set based on the recall rates of non-conforming and conforming products, so that the proportion of non-conforming cover glass samples selected according to the preset probability thresholds out of the total number of cover glass samples to be processed in a production batch is relatively small. In other words, the sampling rate during subsequent manual flatness testing is relatively small. Moreover, since the first flatness prediction model can accurately predict the probability of whether the flatness is conforming, it can ensure that a large proportion of the predicted non-conforming cover glass samples are actually non-conforming products, thus ensuring a high recall rate of non-conforming products. This is beneficial for subsequent manual flatness testing, where the detected non-conforming products have a high recall rate even with a low sampling rate.

[0034] The flatness prediction method in this embodiment utilizes the feature parameter set of the cover glass to be processed before the coating process and the equipment parameter set of the coating equipment. Based on a trained first flatness prediction model, it can accurately predict the probability that the flatness of the cover glass to be processed after coating will be unqualified. Then, it filters out the cover glass to be tested according to a preset probability threshold, realizing early warning of the cover glass to be processed before manual flatness inspection. Since the preset probability threshold is set based on the recall rate of unqualified products and the recall rate of qualified products, the number of unqualified cover glass to be processed according to the preset probability threshold accounts for the total number of cover glass to be processed in a production batch. That is, the sampling rate during subsequent manual flatness inspection is relatively small. Moreover, since the first flatness prediction model can accurately predict the probability of flatness being qualified, it can ensure that a large proportion of the unqualified cover glass to be processed is truly unqualified, that is, it ensures a high recall rate of unqualified products. This is beneficial for subsequent manual flatness inspection, where the unqualified products detected have a high recall rate even with a low sampling rate.

[0035] In some embodiments, before inputting the feature parameter set and the device parameter set into the first flatness prediction model, the method further includes: Extended feature parameters are obtained by performing feature expansion operations on individual feature parameters in the feature parameter set, and / or by performing cross-process feature expansion operations on related feature parameters corresponding to different processing steps in the feature parameter set. In particular, performing cross-process feature expansion operations on related feature parameters corresponding to different processing steps can yield coupled feature parameters that affect subsequent flatness, enabling the first flatness prediction model to learn the characteristics of coupled feature parameters during training and improving the accuracy of model prediction.

[0036] By merging all extended feature parameters into the feature parameter set, the number of feature parameters that affect the smoothness after coating is increased, which improves the prediction accuracy to some extent.

[0037] Because the chemical pressure process and the asymmetric back-grinding process before the coating process in the cover glass manufacturing process have a significant impact on the flatness of the cover glass after coating, in some embodiments, at least one processing step before the coating process includes: a chemical pressure process and an asymmetric back-grinding process, and the set of characteristic parameters includes: a first subset of characteristic parameters after the chemical pressure process and a second subset of characteristic parameters after the asymmetric back-grinding process.

[0038] The first subset of characteristic parameters includes: the surface compressive stress (CS) and the compressive stress layer depth (CT, the depth of the compressive stress layer formed by ion exchange) of the cover glass to be processed. For example, the first subset of characteristic parameters may also include: Stress layer depth to total thickness ratio (CT / TA): Characterizes the proportion of the stress layer in the thickness direction of the cover glass to be processed; Center offset of compressive stress layer (DOC): The offset distance of the center of the stress layer relative to the geometric center of the cover glass to be processed; Stress layer depth multiplied by total thickness (TA×CT): a coupling parameter that comprehensively characterizes the volume of the stress layer; The total thickness (T) of the cover glass to be processed, the thickness of each preset thickness measurement point (e.g., 9 preset thickness measurement points) on the cover glass to be processed, and the average thickness.

[0039] Each characteristic parameter in the first subset of characteristic parameters mentioned above can be detected and calculated by specialized stress testing equipment and thickness detection equipment (such as confocal thickness measurement equipment) after the chemical pressurization process is completed.

[0040] The second subset of characteristic parameters includes: the thickness difference of each preset thickness measurement point on the cover glass to be processed before and after asymmetric back-grinding, and the height value of each preset flatness measurement point (e.g., 28 preset thickness measurement points) after asymmetric back-grinding.

[0041] For example, the second subset of feature parameters may also include: The average thickness difference at each preset thickness measurement point before and after asymmetric reverse grinding; Flatness comprehensive index FAI14.

[0042] Based on this, feature expansion operations are performed on individual feature parameters from the feature parameter set to obtain expanded feature parameters, including expansion in at least one of the following ways: Single feature expansion method 1: Perform differential calculation on the height values ​​corresponding to adjacent preset flatness measurement points, and use the distance from each preset flatness measurement point to the center point to weight the height values ​​of each preset flatness measurement point, so as to obtain the extended feature parameters including the height difference and the weighted height value.

[0043] The height difference between adjacent preset flatness measurement points reflects the height (undulation) change of the cover glass to be processed at adjacent flatness measurement points, that is, the trend of flatness change at adjacent preset flatness measurement points. This allows the first flatness prediction model to pay attention to the height change trend of adjacent preset flatness measurement points during training and inference, making the flatness prediction more accurate.

[0044] Since the deformation of the central area of ​​the cover glass to be processed is usually less than that of the edge area, the flatness of the central area of ​​the cover glass to be processed is usually relatively good. Therefore, the height values ​​of each preset flatness measurement point are weighted by sampling the distance from each preset flatness measurement point to the center point. The farther the distance, the greater the weighted height value. This makes the first flatness prediction model pay more attention to the area far away from the center of the cover glass during training and inference, and obtain a more accurate flatness prediction result.

[0045] Single feature expansion method two: For each feature parameter in the feature parameter set, expand the feature weight parameter according to the debugging time of the processing equipment of the corresponding cover glass to be processed. The closer the debugging time is to the current time, the larger the feature weight parameter. For example, the feature weight parameter... The calculation formula is: ,in, and For coefficients, To adjust the time difference between the current time and the current time. The value can range from 0.8 to 0.9. The value range can be 14 to 16.

[0046] Single feature extension method three: Based on the height values ​​of each preset flatness measurement point, extended feature parameters are obtained, including those reflecting the overall degree of curvature, degree of torsion deformation, direction of warping deformation, and / or degree of local warping.

[0047] Specifically, the distance and azimuth angle from each preset flatness measurement point to the center of the glass are grouped and statistically analyzed. The average height difference between the central area and the edge area is extracted to reflect the overall degree of curvature; the height difference in the diagonal direction is extracted to reflect the degree of torsion deformation; the height difference between the two opposite ends of the glass is extracted to reflect the direction of warping deformation; and the number of local abrupt change points is extracted to reflect the severity of local warping.

[0048] The height difference between the central region and the edge region is calculated by first calculating the height average of each preset flatness measurement point in the central region, then calculating the height average of each preset flatness measurement point in the edge region, and finally calculating the height difference. The distance between the central region and the edge region is preset based on the distance to the center point.

[0049] like Figure 2 As shown, the diagonal height difference is calculated as follows: For any point A, on the straight line passing through point A and the center point O, obtain the height value of point B, which is the closest point to the straight line on the side of the center point O away from point A. The height difference between any point A and point B is the diagonal height difference. It can be understood that there will be a height difference for any point; therefore, the set of diagonal height differences reflects the degree of distortion and deformation.

[0050] like Figure 3 As shown, the height difference between the two ends of the glass is calculated as follows: In the height direction, find two points C and D at the very ends of the glass, and calculate the height difference between points C and D as one end height difference. In the width direction, find two points E and F at the very ends of the glass, and calculate the height difference between points E and F as the other end height difference. The height difference between the two ends reflects the direction of warping deformation.

[0051] The number of local abrupt change points is calculated as follows: For any preset flatness measurement point, determine whether its height value reaches a preset percentage (e.g., 90%) of its corresponding upper limit height value (different preset flatness measurement points have different upper limit height values). If it does, then any preset flatness measurement point is a local abrupt change point. The severity of local warping can be reflected by counting the number of local abrupt change points.

[0052] In the actual processing of cover glass, the preset flatness measurement points are not independently and randomly distributed, but rather exhibit a specific spatial structure constrained by processing mechanics. Directly inputting the height values ​​of each preset flatness measurement point as an independent feature into the model would prevent the model from distinguishing between different deformation modes such as overall bending and local warping. Therefore, expanding the model using single-feature extension method three to generate multiple feature parameters reflecting bending and warping is beneficial for the model to learn the characteristics of glass bending and warping during training, resulting in more accurate flatness predictions during inference.

[0053] Single feature extension method four: Based on the difference between the average height values ​​of two groups of flatness measurement points obtained by dividing the height values ​​of each preset flatness measurement point according to different mirror directions, an extended feature parameter including mirror deviation in at least one mirror direction is obtained.

[0054] Specifically, the mirroring directions include: horizontal mirroring, vertical mirroring, and diagonal mirroring.

[0055] The glass is divided into two horizontal mirror regions by the centerline along the height direction. The average height of the flatness measurement points in each horizontal mirror region is calculated, and the difference between the average heights of the two horizontal mirror regions is the horizontal mirror deviation.

[0056] The glass is divided into two vertical mirror regions by the centerline in the width direction of the glass. The average height of the flatness measurement points in each vertical mirror region is calculated, and the difference between the average heights of the two vertical mirror regions is the vertical mirror deviation.

[0057] Diagonal mirroring divides the glass into two diagonal mirror regions along its diagonal. The average height of the flatness measurement points in each diagonal mirror region is calculated. The difference between the average heights of the two diagonal mirror regions is the diagonal mirroring deviation. It can be understood that since there are two diagonals, there are two diagonal mirroring deviations.

[0058] Single feature extension method five: Based on the difference between the maximum and minimum height values ​​of each preset flatness measurement point, extended feature parameters including the deviation of the maximum asymmetric point are obtained.

[0059] Under normal processing conditions, the flatness distribution of glass should approximately satisfy mirror symmetry about the central axis. When abnormalities such as eccentricity or uneven suction occur in the processing equipment, the symmetry is first disrupted, but the overall flatness index FAI14 may still be within the acceptable range. Therefore, by expanding the feature parameters of each mirror deviation through single feature expansion method four, and by expanding the maximum asymmetric point deviation through single feature expansion method five, the model can perceive changes in height value symmetry during training and inference, thereby more accurately predicting flatness.

[0060] Based on the correlated feature parameters corresponding to different processing steps in the feature parameter set, cross-process feature extension operations are performed to obtain extended feature parameters, including extension in at least one of the following ways: Cross-process extension method one: Calculate the ratio of the average surface compressive stress to the average thickness difference, and obtain the cross-process extension characteristic parameters including this ratio.

[0061] Cross-process extension method two: Calculate the product of the standard deviation of the difference between the depth and thickness of the compressive stress layer to obtain the cross-process extension characteristic parameters including the product.

[0062] Physical analysis reveals a correlation between surface compressive stress and the average thickness difference; that is, the greater the surface compressive stress, the greater the average thickness difference.

[0063] By extending the features across different processes as described above, we obtain extended features related to flatness. This allows the model to learn the influence of feature parameters between different processes on flatness during training and inference, thereby improving prediction accuracy.

[0064] In the process of measuring the flatness of the cover glass to be processed, the flatness measuring device has 108 original points, which can theoretically be used as preset flatness measurement points. To reduce the computational load during model training and inference, in some embodiments, each preset flatness measurement point is determined based on the following method: Obtain the height measurements of M original points on the cover glass sample using a flatness measuring device. M is greater than 0, for example, M=108.

[0065] Using the height measurements of M original points as sample data, and the flatness detection results as labels, a second flatness prediction model is trained, such as a random forest model.

[0066] After the second flatness prediction model is trained, the feature importance weights of each output original point are sorted from largest to smallest, and the top m original points with the largest importance weights are selected as preset flatness measurement points, where m is greater than 0 and less than M, for example: m=28.

[0067] The greater the importance weight of the output of the second flatness prediction model, the greater the impact of the height value of the corresponding original point on the flatness. Therefore, by selecting the m points with the highest weight as the final preset flatness measurement points, the number of measurement points is reduced, the data acquisition cost and the computational load of the first flatness prediction model are reduced while ensuring prediction accuracy. This also reduces the training time of the first flatness prediction model and improves the real-time prediction speed of the first flatness prediction model.

[0068] In some embodiments, the first flatness prediction model is a binary classification model. For example, it can be a binary classification model built using the LightGBM framework, with "unqualified" as the positive class label and "qualified" as the negative class label. The training method for this first flatness prediction model is as follows: The feature parameter set and equipment parameter set of any cover glass sample are input into the first flatness prediction model. The first flatness prediction model performs multiple rounds of iterative fitting on any input cover glass sample to output the probability of non-compliance. The feature parameter set includes the original feature parameter samples obtained by directly detecting the cover glass samples, as well as the extended feature parameter samples obtained according to the aforementioned extension methods. It is understood that before performing the extension operation on the feature parameter samples in the feature parameter set, preprocessing operations such as deduplication, outlier removal, and missing value interpolation are performed on the feature parameter samples in the original feature parameter set.

[0069] Substitute the output failure prediction probability and the flatness compliance label into the loss function. When the loss function converges, the model training is complete. The loss function can be the standard binary classification cross-entropy loss function.

[0070] Furthermore, when collecting the feature parameter set samples and corresponding equipment parameter set samples of the cover glass samples, the collection strategy is as follows: all data of the cover glass samples within the first preset number of days (e.g., 30 days) closest to the current time are collected. Since in the cover glass production process, there are always more qualified products than unqualified products, in order to ensure that the unqualified positive samples and qualified negative samples in the training data maintain class balance, only positive samples with unqualified flatness detection results are collected from the cover glass samples before the first preset number of days to the nearest second preset number of days (e.g., 60 days, where the second preset number of days is greater than the first preset number of days), thereby ensuring the data freshness and class balance of the model training.

[0071] Before training the first flatness prediction model, the dataset consisting of the feature parameter set samples and the corresponding equipment parameter set samples of the collected cover glass samples is divided. The division strategy adopts the order of production time of the cover glass samples, that is, the production time of the cover glass samples in the test set is closer to the current time than that of the cover glass samples in the validation set, and the production time of the cover glass samples in the validation set is closer to the current time than that of the cover glass samples in the training set. Specifically, the division strategy according to the order of production time of the cover glass samples is as follows: The training set uses feature parameter sets and corresponding equipment parameter sets of cover glass samples whose production time is far removed from the current time, such as historical data from 48 hours ago (within 60 days).

[0072] The validation set uses the feature parameter set of cover glass samples whose production time is close to the current time and the corresponding equipment parameter set, for example, data from the last 24 to 48 hours.

[0073] The test set uses the feature parameter set of the cover glass sample whose production time is closest to the current time and the corresponding equipment parameter set, for example, data from the last 24 hours.

[0074] The training set, validation set, and test set are divided according to the above-mentioned division strategy based on the production time sequence of the cover glass samples. With sufficient samples, the first flatness prediction model trained is more adapted to the current production line equipment status and operating conditions.

[0075] In this embodiment, a binary classification model framework is used to train the first flatness prediction model, and the training is performed with non-conformity as the positive label, so that the first flatness prediction model can output the probability of non-conformity products. Finally, the non-conformity products predicted by the model can be screened out according to the probability threshold. When sampling by manual inspection, it can have a high recall rate of non-conformity products under low sampling rate.

[0076] In some embodiments, during the process of multiple rounds of iterative fitting of any input cover glass sample by the first flatness prediction model, sample weights are generated based on the production time and sample category of any cover glass sample, so that the sample weight of the cover glass sample that is closer to the current time is greater, and in each cover glass sample of the same batch at the same time, the sample weight of the unqualified positive sample is greater than the sample weight of the qualified negative sample.

[0077] Specifically, differentiated weights are assigned to each training sample, taking into account both time freshness and class scarcity (fewer defective categories). The sample weight consists of two parts: a time weight component that assigns higher weight to more recent samples (decreasing exponentially with time difference), and a class weight component that gives defective samples an additional weight bonus. The two parts are weighted and summed using proportional coefficients, then normalized to obtain the final sample weight. Specifically, the method for generating sample weights for any cover glass sample is as follows: The time weight component is determined based on the production time of any cover glass sample, such that samples closer to the current time have a larger time weight component. For example, the time weight component... Calculate using the following formula: .

[0078] Where λ is the time decay coefficient, and its value is greater than 0. , For cover glass samples Production time Recent time The time interval.

[0079] The category weight components for determining sample weights based on the category of any cover glass sample are such that the category weight component corresponding to a non-compliant sample is greater than the category weight component corresponding to a compliant sample. For example, the category weight components... Calculate using the following formula: .

[0080] in, This is the magnification factor for non-conforming product categories, with a value greater than 1. This is an indicator function; a value of 1 is assigned to defective products, and a value of 0 is assigned to acceptable products.

[0081] The time weight component and the category weight component are weighted, summed, and normalized according to a preset proportional coefficient to generate the sample weights. Specifically, the absolute weights are obtained by weighted summation according to the following formula. : .

[0082] in, A preset ratio coefficient is used to balance the freshness of time and the scarcity of a category. .

[0083] The cover glass sample was obtained by normalizing the absolute weights according to the following formula. Sample weights : .

[0084] in, For any cover glass sample absolute weight, This represents the total number of cover glass samples.

[0085] In this embodiment, the sample weights are determined by time freshness and category scarcity. This ensures that recent samples receive higher weights, enabling the model to adapt to the current production line conditions. Long-term defective samples still retain higher weights to fully utilize scarce failure mode information. As a result, compared with uniform weight training, the recall rate of defective products is improved, and the adaptability to production line condition drift is significantly enhanced.

[0086] In some embodiments, the loss function of the first flatness prediction model is obtained by weighting and summing a binary classification loss term and a physical consistency constraint penalty term according to a preset penalty term coefficient.

[0087] For any cover glass sample, if the first difference between the stress layer volume energy in the corresponding feature parameter set sample and the mean stress layer volume energy of all samples exceeds a preset multiple of the standard deviation of the stress layer volume energy of all samples, a physical consistency constraint penalty term is determined by the second difference between the first difference and the preset multiple of the standard deviation of the stress layer volume energy of all samples. Here, the stress layer volume energy is the product of the surface compressive stress and the compressive stress layer depth.

[0088] Specifically, physical consistency constraint penalty item This can be expressed using the following formula: .

[0089] Loss function of the first flatness prediction model for: .

[0090] in, This is a binary loss term. The preset penalty coefficient can take values ​​ranging from [0.5, 2]. For cover glass samples The stress layer volume energy, i.e., the cover glass sample The product of the surface compressive stress sample and the compressive stress layer depth sample. and These represent the mean and standard deviation of the stress layer volumetric energy for all cover glass samples. This is the tolerance ratio.

[0091] In this embodiment, a physical consistency constraint penalty term is introduced into the loss function, so that the model satisfies the physical laws of glass processing while fitting the data, the prediction results have physical interpretability, improve the model prediction stability, and reduce the false alarm rate.

[0092] In cover glass production, the stability of different processes varies significantly at different times. For example, the ion exchange parameters in the chemical pressurization process are usually relatively stable, while those in the asymmetric back-grinding process may fluctuate considerably due to factors such as equipment suction fluctuations and fixture wear. Conventional machine learning training treats all feature parameters equally, failing to perceive the current stable or fluctuating state of each process. This results in the first flatness prediction model's insufficient ability to perceive the sources of quality fluctuations and its inability to learn the impact of process instability on flatness. Therefore, in some embodiments, before inputting the feature parameter set sample and equipment parameter set sample of any cover glass sample into the first flatness prediction model, the following steps are also included: The characteristic parameter sets of all cover glass samples are grouped according to different processes. Based on the recent standard deviation and historical baseline standard deviation of the characteristic parameter set samples for any process, a stability index corresponding to any process is determined. The recent standard deviation is the standard deviation of the characteristic parameter set samples for any process within a preset time period (e.g., 24 hours) from the current time, and the historical baseline standard deviation is the standard deviation of the characteristic parameter set samples for any process during its historical stable operation. Specifically, the stability index... Stability Index The value range is [0,1]. The closer the value is to 0, the more drastic the fluctuations and the more unstable the process is.

[0093] The feature parameter set sample and equipment parameter set sample of any cover glass sample are input into the first flatness prediction model. During the multi-round iterative fitting process of the first flatness prediction model on any input cover glass sample, the following steps are also included: calculating the attention weight of the feature parameter set sample of any process based on the stability index and the corresponding preset amplification factor. Specifically, the attention weight is calculated for the feature parameter set sample corresponding to any process according to the stability index and the corresponding preset amplification factor. It can be calculated using the following formula: .

[0094] in, To preset the amplification factor, the degree of amplification of the weights of unstable process characteristics is controlled; this attention weight... As a feature scaling factor, it weights the feature parameter set samples of the corresponding process during training, making the model more sensitive to changes in the feature parameters of unstable processes during iteration.

[0095] In this embodiment, by calculating the stability index for each process before model training and dynamically adjusting the attention weight of the corresponding feature parameter set during training based on the stability index, the model can learn the impact of unstable feature parameter sets on flatness during training, thereby improving the final prediction accuracy of the model. Moreover, in actual deployment, when a process experiences abnormal fluctuations due to equipment aging or process parameter drift, the first flatness prediction model can automatically increase the attention weight of the corresponding process's feature parameters without manual intervention, so as to quickly adapt to the weak links of the current production line after a new round of training.

[0096] In some embodiments, the preset probability threshold is determined as follows: Search within the interval [0.0, 1.0] to find a candidate probability threshold that makes the first recall rate of unqualified positive samples closest to the preset recall rate, and record the second recall rate of qualified negative samples corresponding to the first recall rate. For example, if the preset recall rate is 85%, then a candidate probability threshold is found such that after dividing the probability of unqualified positive samples predicted by the first flatness prediction model during training (especially in the validation and testing phases) according to this candidate probability threshold, the proportion of truly unqualified samples in the final selected unqualified positive samples is closest to 85%.

[0097] If the second recall rate falls within the preset recall rate range for negative samples, the candidate probability threshold is determined as the preset probability threshold. The preset recall rate range for negative samples can be [0.75, 0.80], meaning that after the probability of unqualified positive samples predicted by the model is divided according to this candidate probability threshold, the recall rate of qualified negative samples must reach the target range.

[0098] If the second recall rate is less than the lower limit of the preset recall interval for negative samples, the candidate probability threshold is increased. When the candidate probability threshold increases until the second recall rate returns to the lower limit, the increased candidate probability threshold is set as the preset probability threshold. If the second recall rate is less than the lower limit of the preset recall interval for qualified negative samples, it means that if classified according to the candidate probability threshold, a large number of qualified negative samples are misclassified as unqualified positive samples, resulting in a high sampling rate for subsequent manual testing. For example, when the second recall rate is less than 0.75, the equivalent sampling rate would be greater than 25%. Therefore, it is necessary to increase the candidate probability threshold until the second recall rate returns to the lower limit to reduce the sampling rate for subsequent manual testing, for example, to ensure the sampling rate does not exceed 20%.

[0099] If the second recall rate is greater than the upper limit of the preset recall range for negative samples, the candidate probability threshold is reduced. When the candidate probability threshold is reduced to the point where the second recall rate falls back to the upper limit, the reduced candidate probability threshold is determined as the preset probability threshold. For example, if the second recall rate is greater than 0.80, the equivalent sampling rate will be less than 20%, indicating that the first flatness prediction model has a good prediction result. Therefore, reducing the candidate probability threshold can increase the sampling rate to 20% (not exceeding 20%), while simultaneously increasing the first recall rate. This ensures a relatively low sampling rate and a high first recall rate for unqualified positive samples.

[0100] In this embodiment, by using the above-mentioned dynamic determination method of preset probability threshold, it can be ensured that both the positive class recall rate and the negative class recall rate can meet the set requirements when the sampling rate is low.

[0101] In some embodiments, the flatness prediction method further includes: After the first flatness prediction model is trained, it is periodically trained and updated at preset intervals, using the feature parameter set and equipment parameter set within the period as samples, and the corresponding flatness detection results as labels, so that the first flatness prediction model can adapt to the current production line and equipment operating conditions.

[0102] However, updating the first flatness prediction model may result in a larger proportion of defective products in the output, without improving the recall rate of defective products. Therefore, updating the first flatness prediction model also includes: If the total number of cover glass to be processed is found to be greater than the first quantity threshold, and the ratio of the number of products predicted as unqualified by the first flatness prediction model to the total number of cover glass to be processed is greater than the preset sampling rate limit, the updated first flatness prediction model will be rolled back to the next newest version of the first flatness prediction model to reduce the sampling rate of subsequent manual inspection.

[0103] If the total number of cover glass panels awaiting processing exceeds the second quantity threshold, and the ratio of the number of products predicted as defective by the first flatness prediction model to the total number of cover glass panels awaiting processing is less than the preset sampling rate lower limit, the preset probability threshold is redefined. The fact that the ratio of the number of products predicted as defective by the first flatness prediction model to the total number of cover glass panels awaiting processing is less than the preset sampling rate lower limit indicates a high probability that some defective products are misclassified. Therefore, it is necessary to readjust the preset probability threshold, i.e., lower the preset probability threshold, to increase the sampling rate.

[0104] In this embodiment, the model is updated periodically and the output results of the updated model are monitored to ensure that while the model adapts to the drift of recent production line conditions, the data output by the model meets the requirements of low sampling rate and high recall rate of defective products for subsequent manual inspection.

[0105] The following describes the experimental results of the first flatness prediction model based on the improved LightGBM framework for flatness prediction: In the experimental deployment of the A2 production line of a cover glass factory, the experimental line for three models of cover glass, namely V1, V2 and V3, was verified over a period of one month. Through iterative optimization and anomaly handling of the first flatness prediction model, the final results of the three models of products met the requirements.

[0106] Validation results: The nonconforming product identification rate (i.e., nonconforming recall rate) increased from 60% to 85.7%, the sampling rate decreased from 60% to 19.2%, and the proportion of nonconforming products captured remained stable between 80% and 90%. Compared to the standard LightGBM model, the improved LightGBM model and dynamic threshold strategy of this invention improved the recall rate of nonconforming products by 15% to 20% while predicting fewer nonconforming products (i.e., a lower sampling rate).

[0107] The smoothness prediction device provided by the present invention is described below. The smoothness prediction device described below can be referred to in correspondence with the smoothness prediction method described above.

[0108] The flatness prediction device of this invention, such as Figure 4 As shown, it includes: The data acquisition module 410 is used to acquire the feature parameter set of the workpiece to be processed before the coating process, and the equipment parameter set of the processing equipment. The feature parameter set includes feature parameters after at least one processing step before the coating process.

[0109] The model execution module 420 is used to input the feature parameter set and the equipment parameter set into the first flatness prediction model to obtain the probability that the flatness of the workpiece to be processed is unqualified, as output by the first flatness prediction model.

[0110] The workpiece determination module 430 is used to determine the workpiece to be processed as the workpiece to be inspected when the probability that the flatness of the workpiece to be processed is greater than a preset probability threshold, so as to be used for subsequent flatness inspection.

[0111] The first flatness prediction model is trained based on the flatness of the feature parameter set samples, equipment parameter set samples, and corresponding workpiece samples. The preset probability threshold is determined based on the recall rate of unqualified products and the recall rate of qualified products after the first flatness prediction model is trained.

[0112] Figure 5 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 5 As shown, the electronic device may include: a processor 510, a communications interface 520, a memory 530, and a communication bus 540, wherein the processor 510, the communications interface 520, and the memory 530 communicate with each other via the communication bus 540. The processor 510 can call logical instructions in the memory 530 to execute a flatness prediction method, which includes: Obtain the feature parameter set of the workpiece before the coating process, and the equipment parameter set of the processing equipment. The feature parameter set includes feature parameters after at least one processing step before the coating process.

[0113] Input the feature parameter set and equipment parameter set into the first flatness prediction model to obtain the probability that the flatness of the workpiece to be processed is unqualified, as output by the first flatness prediction model.

[0114] If the probability of the flatness of the workpiece to be processed being unqualified is greater than the preset probability threshold, the workpiece to be processed is determined as the workpiece to be inspected for subsequent flatness inspection.

[0115] The first flatness prediction model is trained based on the flatness of the feature parameter set samples, equipment parameter set samples, and corresponding workpiece samples. The preset probability threshold is determined based on the recall rate of unqualified products and the recall rate of qualified products after the first flatness prediction model is trained.

[0116] Furthermore, the logical instructions in the aforementioned memory 530 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0117] On the other hand, the present invention also provides a computer program product, the computer program product comprising a computer program that can be stored on a non-transitory computer-readable storage medium, wherein when the computer program is executed by a processor, the computer is able to execute the flatness prediction method provided by the above methods, the method comprising: Obtain the feature parameter set of the workpiece before the coating process, and the equipment parameter set of the processing equipment. The feature parameter set includes feature parameters after at least one processing step before the coating process.

[0118] Input the feature parameter set and equipment parameter set into the first flatness prediction model to obtain the probability that the flatness of the workpiece to be processed is unqualified, as output by the first flatness prediction model.

[0119] If the probability of the flatness of the workpiece to be processed being unqualified is greater than the preset probability threshold, the workpiece to be processed is determined as the workpiece to be inspected for subsequent flatness inspection.

[0120] The first flatness prediction model is trained based on the flatness of the feature parameter set samples, equipment parameter set samples, and corresponding workpiece samples. The preset probability threshold is determined based on the recall rate of unqualified products and the recall rate of qualified products after the first flatness prediction model is trained.

[0121] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the flatness prediction method provided by the methods described above, the method comprising: Obtain the feature parameter set of the workpiece before the coating process, and the equipment parameter set of the processing equipment. The feature parameter set includes feature parameters after at least one processing step before the coating process.

[0122] Input the feature parameter set and equipment parameter set into the first flatness prediction model to obtain the probability that the flatness of the workpiece to be processed is unqualified, as output by the first flatness prediction model.

[0123] If the probability of the flatness of the workpiece to be processed being unqualified is greater than the preset probability threshold, the workpiece to be processed is determined as the workpiece to be inspected for subsequent flatness inspection.

[0124] The first flatness prediction model is trained based on the flatness of the feature parameter set samples, equipment parameter set samples, and corresponding workpiece samples. The preset probability threshold is determined based on the recall rate of unqualified products and the recall rate of qualified products after the first flatness prediction model is trained.

[0125] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0126] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence 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 ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0127] All actions involving the acquisition of signal information or data in this invention are carried out in compliance with the relevant data protection laws and policies of the country where the device is located, and with the authorization granted by the owner of the device.

[0128] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting flatness, characterized in that, include: Obtain the feature parameter set of the workpiece before the coating process and the equipment parameter set of the processing equipment. The feature parameter set includes: feature parameters after at least one processing step before the coating process. The feature parameter set and the equipment parameter set are input into the first flatness prediction model to obtain the probability that the flatness of the workpiece to be processed is unqualified, as output by the first flatness prediction model. If the probability of the flatness of the workpiece to be processed being unqualified is greater than a preset probability threshold, the workpiece to be processed is determined to be the workpiece to be inspected for subsequent flatness inspection. The first flatness prediction model is trained based on the flatness qualification labels of the feature parameter set samples, the equipment parameter set samples, and the corresponding workpiece samples. The preset probability threshold is determined based on the non-qualified product recall rate and qualified product recall rate of the first flatness prediction model after training.

2. The flatness prediction method according to claim 1, characterized in that, Before inputting the feature parameter set and equipment parameter set into the first flatness prediction model, the method further includes: Feature expansion operations are performed on a single feature parameter in the feature parameter set to obtain extended feature parameters, and / or cross-process feature expansion operations are performed on related feature parameters corresponding to different processing steps in the feature parameter set to obtain extended feature parameters. All extended feature parameters are merged into the feature parameter set.

3. The flatness prediction method according to claim 2, characterized in that, At least one processing step before the coating process includes: a chemical pressure process and an asymmetric anti-grind process, and the feature parameter set includes: a first feature parameter subset after the chemical pressure process and a second feature parameter subset after the asymmetric anti-grind process; The first subset of characteristic parameters includes: the surface compressive stress and the compressive stress layer depth of the workpiece to be processed; The second subset of feature parameters includes: the thickness difference of each preset thickness measurement point on the workpiece before and after asymmetric reverse grinding, and the height value of each preset flatness measurement point after asymmetric reverse grinding; Based on a single feature parameter in the feature parameter set, feature expansion operations are performed to obtain expanded feature parameters, including expansion in at least one of the following ways: Perform a difference operation on the height values ​​corresponding to adjacent preset flatness measurement points, and use the distance from each preset flatness measurement point to the center point to weight the height values ​​of each preset flatness measurement point, to obtain extended feature parameters including height difference and weighted height value; For each feature parameter in the feature parameter set, the feature weight parameter is expanded according to the debugging time of the corresponding processing equipment of the cover glass to be processed; Based on the height values ​​of each preset flatness measurement point, extended characteristic parameters are obtained, including those reflecting the overall degree of curvature, the degree of torsional deformation, the direction of warping deformation, and / or the degree of local warping. Based on the difference between the average height values ​​of two groups of flatness measurement points obtained by dividing the height values ​​of each preset flatness measurement point according to different mirror directions, an extended characteristic parameter including mirror deviation in at least one mirror direction is obtained. Based on the difference between the maximum and minimum height values ​​of each preset flatness measurement point, extended characteristic parameters including the deviation of the maximum asymmetric point are obtained; Based on the correlated feature parameters corresponding to different processing steps in the feature parameter set, cross-process feature extension operations are performed to obtain extended feature parameters, including: extension in at least one of the following ways: Calculate the ratio of the surface compressive stress to the average value of the thickness difference to obtain extended characteristic parameters across processes that include the ratio; Calculate the product of the compressive stress layer depth and the standard deviation of the thickness difference to obtain extended characteristic parameters across processes, including the product.

4. The flatness prediction method according to claim 3, characterized in that, The preset flatness measurement points are determined based on the following method: Obtain the height measurement values ​​of M original points of the workpiece sample obtained by the flatness measuring device, where M is greater than 0; Using the height measurements of M original points as sample data, and the flatness detection results as labels, a second flatness prediction model is trained. After the second flatness prediction model is trained, the feature importance weights of each output original point are sorted from largest to smallest, and the top m original points with the largest importance weights are selected as the preset flatness measurement points, where m is greater than 0 and less than M.

5. The method for predicting flatness according to any one of claims 1 to 4, characterized in that, The first flatness prediction model is a binary classification model, with "unqualified" as the positive label and "qualified" as the negative label. The training method is as follows: The feature parameter set sample and equipment parameter set sample of any workpiece sample are input into the first flatness prediction model. The first flatness prediction model performs multiple rounds of iterative fitting on any input workpiece sample to output the non-conformity prediction probability. Substitute the output failure prediction probability and the flatness acceptance label into the loss function. When the loss function converges, the model training is complete.

6. The flatness prediction method according to claim 5, characterized in that, During the process of the first flatness prediction model performing multiple rounds of iterative fitting on any input workpiece sample, sample weights are generated based on the production time and sample category of any workpiece sample, so that the sample weight of the workpiece sample that is closer to the current time is greater, and in each batch of workpiece samples at the same time, the sample weight of the unqualified positive sample is greater than the sample weight of the qualified negative sample. And / or, the loss function is obtained by weighting and summing the binary classification loss term and the physical consistency constraint penalty term according to a preset penalty term coefficient; For any workpiece sample, if the first difference between the stress layer volume energy in the corresponding feature parameter set sample and the mean stress layer volume energy of all samples exceeds a preset multiple of the standard deviation of the stress layer volume energy of all samples, the physical consistency constraint penalty term is determined by the second difference between the first difference and the preset multiple of the standard deviation of the stress layer volume energy of all samples. And / or, before inputting the feature parameter set sample and equipment parameter set sample of any workpiece sample into the first flatness prediction model, the method further includes: The feature parameter set samples of all workpiece samples are grouped according to different processes. Based on the recent standard deviation and historical baseline standard deviation of the feature parameter set samples of any process, the stability index corresponding to any process is determined. The recent standard deviation is the standard deviation of the feature parameter set samples of any process within a preset time period from the current time, and the historical baseline standard deviation is the standard deviation of the feature parameter set samples of any process during the historical stable operation period. The feature parameter set sample and equipment parameter set sample of any workpiece sample are input into the first flatness prediction model. The process of the first flatness prediction model performing multiple rounds of iterative fitting on any input workpiece sample also includes: The attention weight of the feature parameter set sample for any process is calculated based on the stability index and the corresponding preset amplification factor.

7. The method for predicting flatness according to any one of claims 1 to 5, characterized in that, The preset probability threshold is determined as follows: Search within the interval [0.0, 1.0] to find the candidate probability threshold that makes the first recall rate of unqualified positive samples closest to the preset recall rate, and record the second recall rate of qualified negative samples corresponding to the first recall rate. If the second recall rate is within the preset recall rate range of negative class samples, the candidate probability threshold is determined to be the preset probability threshold. If the second recall rate is less than the lower limit of the preset recall rate range for negative samples, the candidate probability threshold is increased. When the candidate probability threshold is increased to the point that the second recall rate rises back to the lower limit, the increased candidate probability threshold is determined as the preset probability threshold. If the second recall rate is greater than the upper limit of the preset recall rate range for negative samples, the candidate probability threshold is reduced. When the candidate probability threshold is reduced to the point where the second recall rate returns to the upper limit, the reduced candidate probability threshold is determined as the preset probability threshold.

8. The method for predicting flatness according to any one of claims 1 to 5, characterized in that, Also includes: After the first flatness prediction model is trained, the first flatness prediction model is periodically trained and updated at preset intervals, using the feature parameter set and equipment parameter set within the period as samples, and the corresponding flatness detection results as labels. After updating the first flatness prediction model, the following is also included: If the total number of workpieces to be processed is found to be greater than the first quantity threshold, and the ratio of the number of unqualified products predicted by the first flatness prediction model to the total number of workpieces to be processed is greater than the preset sampling rate limit, the updated first flatness prediction model will be rolled back to the next newest version of the first flatness prediction model. If the total number of workpieces to be processed is found to be greater than the second quantity threshold, and the ratio of the number of products predicted as defective by the first flatness prediction model to the total number of workpieces to be processed is less than the lower limit of the preset sampling rate, the preset probability threshold is redefined.

9. A flatness prediction device, characterized in that, include: The data acquisition module is used to acquire the feature parameter set of the workpiece to be processed before the coating process, and the equipment parameter set of the processing equipment. The feature parameter set includes: feature parameters after at least one processing step before the coating process. The model execution module is used to input the feature parameter set and the equipment parameter set into the first flatness prediction model to obtain the probability that the flatness of the workpiece to be processed is unqualified, as output by the first flatness prediction model. The workpiece determination module is used to determine the workpiece to be processed as the workpiece to be inspected when the probability of the flatness of the workpiece to be processed being unqualified is greater than a preset probability threshold, so as to be used for subsequent flatness inspection. The first flatness prediction model is trained based on the flatness of the feature parameter set samples, equipment parameter set samples, and corresponding workpiece samples. The preset probability threshold is determined based on the recall rate of unqualified products and the recall rate of qualified products after the first flatness prediction model is trained.

10. An electronic device, comprising a processor, a communication interface, a memory, and a communication bus, characterized in that, The processor, the communication interface, and the memory communicate with each other through the communication bus. The processor can call the logic instructions in the memory to execute the flatness prediction method according to any one of claims 1 to 8.

Images