Online detection and feedback control method for electroplated layer based on machine learning

Abstract

The invention discloses an online detection and feedback control method for an electroplated layer based on machine learning, particularly relates to the field of electroplating process monitoring, and is used for solving the problem that thickness deviation and local defects of the electroplated layer are difficult to quantitatively evaluate in time and correspond to process adjustment. The method comprises the following steps of: acquiring an electroplated workpiece surface image and a corresponding process operation record on an electroplating production line to construct a state input vector, and outputting an electroplated layer state description quantity and sensitivity information which represent overall thickness deviation and local abnormal risk by utilizing an electroplated layer state estimation model; and forming an appearance deviation degree and process adjustment reliability quantification result by combining workpiece image features and historical process adjustment records, and determining an adjustment coefficient for correcting subsequent workpiece process parameters according to the quantification result.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of electroplating process monitoring, and more particularly, to an online detection and feedback control method for electroplated layers based on machine learning. BACKGROUND

[0002] In the electroplating production line, especially in the case where both appearance and performance are required, the enterprise has begun to install cameras and image processing devices in addition to the traditional electroplating equipment to automatically inspect the surface of the electroplated parts or test pieces. In some published documents, it has been proposed to use machine vision and deep learning to identify the bright and dark, color, texture and surface defects of the plated surface, and to divide the workpieces into levels such as good appearance, poor appearance or existence of certain defects, to replace manual visual inspection. Such systems are usually arranged after the electroplating process to detect the workpieces that have completed electroplating, and can continuously give the appearance results of each workpiece, which seems to have the ability of online detection and has preliminarily introduced machine learning and image recognition technology.

[0003] However, it can be found from these existing visual detection schemes that there is still a clear gap between the detection results and process control. Most of the existing systems only output judgment conclusions such as qualified or unqualified, high or low appearance level, and which type of surface defect exists, and the control end only gets a discrete state signal, but not a set of continuous process quantities that can be directly used to adjust the electroplating current, time, additive ratio or stirring mode. The detailed information about the thickness variation, local unevenness and possible early abnormalities of the electroplated layer contained in the image has not been converted into a quantitative estimate of the physical state of the electroplated layer, and further formed a stable path from the image results to derive specific process adjustment instructions. That is, online visual detection and electroplating process control are two separate systems, the former only tells the operator whether the workpiece appearance is good or not, and the latter still mainly relies on experience to determine how to adjust the process for the next batch. This detection and control disconnection makes it difficult for the existing electroplating detection means based on machine vision and machine learning to truly support an online detection and feedback control method for electroplated layers based on machine learning, which is the specific technical problem to be solved by the present application.

[0004] To solve the above problems, a technical solution is provided. SUMMARY

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide an online detection and feedback control method for electroplated layers based on machine learning. This method constructs a state input vector by acquiring images of the electroplated workpiece surface and corresponding process operation records on the electroplating production line. It then uses an electroplated layer state estimation model to output electroplated layer state description quantities and sensitivity information representing overall thickness deviation and local anomaly risks. Combining workpiece image features and historical process adjustment records, it forms a quantitative result of the degree of appearance deviation and the reliability of process adjustments. Based on the quantitative result, it determines adjustment coefficients to correct subsequent workpiece process parameters, thereby solving the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: S1. Acquire online observation data related to the electroplating layer formation process on the electroplating production line, and integrate the surface image of the workpiece after electroplating with the process operation record of the corresponding electroplating section according to the workpiece identification to obtain the original observation data. S2 inputs the original observation data into the feature screening and reconstruction module, and screens out image features reflecting changes in the appearance of the electroplating layer and process features reflecting the strength of the electroplating process according to the preset discrimination criteria. The screened features are then reorganized and numerically standardized to form a state input vector for electroplating state analysis. S3. Input the state input vector into the electroplating layer state estimation model. The model outputs an electroplating layer state description quantity that represents the overall thickness deviation trend and local anomaly risk, and outputs the sensitivity information of the degree of influence of each process feature on the electroplating layer state description quantity. S4. Based on the electroplating layer state description quantity and sensitivity information, construct two quantitative results from the current workpiece image and historical process adjustment records, respectively, reflecting the degree of appearance deviation and the reliability of process adjustment. Use the reliability of process adjustment to shape the mapping relationship between the degree of appearance deviation and the adjustment range of process parameters, and determine the adjustment coefficient for correcting the adjustable process parameter settings of subsequent workpieces based on the mapping relationship.

[0007] Furthermore, step S1 includes the following: The workpiece number, entry time, and exit time are recorded to form the time boundary of the electroplating interval. Process operation records within the electroplating interval are collected at fixed time intervals. The time alignment threshold interval is determined based on the distance between the exit position and the image acquisition position and the range of conveying speed, and the electroplated surface images are selected accordingly. The sharpness threshold interval is calculated based on the calibration image, and the electroplated surface images are selected accordingly. The workpiece number, electroplating interval time boundary, process operation records, and electroplated surface images selected by the time alignment threshold interval and sharpness threshold interval are integrated into the original observation dataset.

[0008] Furthermore, step S2 includes the following: The grayscale reference of the electroplated surface image in the original observation data is adjusted, and three image candidate features are constructed: total texture, texture difference between the central and edge regions, and ratio of bright and dark areas. Three process candidate features are constructed for the electroplating current time series: current integral, high current duration, and current jump amplitude, thus forming a set of image candidate features and a set of process candidate features.

[0009] Furthermore, step S2 also includes the following: In the calibrated workpiece set, a monotonic consistency index is calculated for each candidate feature in the image candidate feature set and the process candidate feature set based on the reference appearance index. The formal features are selected by constructing a monotonic consistency threshold interval through sorting. The formal features are then subjected to linear numerical normalization based on the minimum and maximum values ​​of the calibrated samples. The normalized formal image features and formal process features are connected in a fixed order to form a state input vector, which is then input into the electroplating layer state estimation model.

[0010] Furthermore, step S3 includes the following: Thickness measurements are performed at multiple measurement locations on the surface of the workpiece in the training workpiece set. An overall thickness deviation descriptor is constructed based on the average value of the measured thickness. Local risk descriptors are constructed by sorting the thickness deviations and selecting extreme deviation locations according to the local risk proportion threshold. The overall thickness deviation descriptor and the local risk descriptor are combined to form the target value of the electroplating layer state descriptor.

[0011] Furthermore, step S3 also includes the following: The state input vector and the target value of the electroplating state descriptor are input into the electroplating state estimation model. The model parameters are trained by minimizing the sum of the absolute values ​​of the differences between the overall thickness deviation descriptor and the predicted and target values ​​of the local risk descriptor. The upper and lower bounds of the target value of the electroplating state descriptor are statistically analyzed in the training workpiece set, and the model output is linearly normalized to obtain the dimensionless electroplating state descriptor.

[0012] Furthermore, step S3 also includes the following: In the state input vector, the set of process feature component indices is determined in a predetermined order. The perturbation amplitude is set according to the value range of the process feature components in the training workpiece set and the perturbation ratio coefficient. Positive and negative perturbations are applied to a single process feature component around the state input vector to obtain two sets of dimensionless electroplating layer state description quantities. Finite difference change rates are constructed for the two sets of output values, and the maximum value of the absolute value of the change rate of the overall thickness deviation description quantity and the local risk description quantity is taken as the original sensitivity value. Then, linear normalization is performed on the upper and lower bounds of the original sensitivity value obtained from training statistics to generate process feature sensitivity information.

[0013] Furthermore, step S4 includes the following: The morphological deviation index is obtained by taking the maximum value of the dimensionless overall thickness deviation descriptor and the dimensionless local risk descriptor. The image deviation index is obtained by taking the maximum value of the normalized feature of the texture difference between the center and the edge regions and the normalized feature of the ratio of light and dark areas. Finally, the total appearance deviation index is obtained by taking the maximum value of the morphological deviation index and the image deviation index.

[0014] Furthermore, step S4 also includes the following: In terms of process feature index, pairing records of workpieces before and after adjustment are extracted from historical process adjustment records. Process adjustment amplitude threshold and appearance deviation change threshold are introduced, and significant adjustment events are screened by combining dimensionless process feature sensitivity threshold. Based on the number of successful events and the number of failed events in the significant adjustment events, an initial index of process adjustment reliability is constructed. Then, the smoothing coefficient calculated based on the number of events and the smoothing control constant is used to shrink the initial index of process adjustment reliability to obtain the process adjustment reliability index.

[0015] Furthermore, step S4 also includes the following: Using the overall index of appearance deviation as the independent variable, a basic adjustment intensity is constructed through squaring. A power exponent parameter is constructed using the process adjustment reliability index and the shape adjustment coefficient. The basic adjustment intensity is then exponentially operated on according to the power exponent parameter to generate an adjustment coefficient between zero and one. Finally, the adjustment coefficient is multiplied by the maximum adjustment range preset for each process feature to determine the adjustment range of the process parameters of the subsequent workpiece on the corresponding process feature.

[0016] The technical effects and advantages of the online detection and feedback control method for electroplated layers based on machine learning in this invention are as follows: This invention unifies the electroplated surface image and the corresponding electroplating interval process operation record into a state input vector under the same workpiece identification. Then, the electroplating layer state estimation model outputs two electroplating layer state description quantities: overall thickness deviation trend and local anomaly risk. This transforms the electroplating layer quality from the previous judgment of pass or fail and offline sampling inspection to a numerical characterization of thickness and defect risk that can be continuously obtained on the production line. This addresses the problem that it is difficult to quantify and assess the original electroplating layer thickness deviation and local defects in a timely manner, forming a clear channel from online acquisition to state quantity output, reducing the reliance on manual experience to interpret images and scattered process curves.

[0017] This invention introduces process feature sensitivity information into the state input vector and constructs a quantitative result of process adjustment reliability based on historical process adjustment records. The quantitative result representing the degree of appearance deviation and the process adjustment reliability are jointly introduced into a power function-form mapping relationship. This allows the parameter adjustment range corresponding to the same degree of appearance deviation to have adaptive differences under different process features and at different times. Thus, in the feedback control loop, it can provide sufficient adjustment in time for serious deviations, and automatically reduce the adjustment intensity on process features with unstable historical adjustment effects. This reduces over-adjustment and under-adjustment caused by trial and error adjustments, and improves the uniformity of electroplating layer thickness and the suppression effect of local defects. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating an online detection and feedback control method for electroplated layers based on machine learning, according to the present invention. Detailed Implementation

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

[0020] Example 1: Figure 1 This invention presents an online detection and feedback control method for electroplated layers based on machine learning, comprising: S1. Acquire online observation data related to the electroplating layer formation process on the electroplating production line, and integrate the surface image of the workpiece after electroplating with the process operation record of the corresponding electroplating section according to the workpiece identification to obtain the original observation data.

[0021] S2 inputs the original observation data into the feature screening and reconstruction module. Based on the preset discrimination criteria, it screens out image features that reflect changes in the appearance of the electroplated layer and process features that reflect the strength of the electroplating process. The screened features are then reorganized and numerically standardized to form a state input vector for electroplating state analysis.

[0022] S3. Input the state input vector into the electroplating layer state estimation model. The model outputs an electroplating layer state description quantity that represents the overall thickness deviation trend and local anomaly risk, and outputs the sensitivity information of the degree of influence of each process feature on the electroplating layer state description quantity.

[0023] S4. Based on the electroplating layer state description quantity and sensitivity information, construct two quantitative results from the current workpiece image and historical process adjustment records, respectively, reflecting the degree of appearance deviation and the reliability of process adjustment. Use the reliability of process adjustment to shape the mapping relationship between the degree of appearance deviation and the adjustment range of process parameters, and determine the adjustment coefficient for correcting the adjustable process parameter settings of subsequent workpieces based on the mapping relationship.

[0024] When constructing an online detection and feedback control method for electroplating layers suitable for machine learning analysis on an electroplating production line, it is necessary to form complete and consistent original observation data at the level of a single workpiece. The original observation data needs to simultaneously include electroplating interval time information, process operation records, and post-electroplating surface image information, with clear temporal relationships and stable image quality. If the time is misaligned or the image is blurry, even with subsequent machine learning processing, it is difficult to draw reliable conclusions about the electroplating layer status. Therefore, in step S1, a detailed design is performed around four categories of objects: workpiece number, electroplating interval time boundaries, time alignment threshold range, and clarity threshold range, forming an original observation dataset with clearly defined threshold ranges and acquisition criteria.

[0025] 1-1 Record of workpiece number and electroplating interval time boundary.

[0026] Each workpiece entering the electroplating zone is assigned a unique workpiece number. At the entrance, a detection device records the time the workpiece enters the electroplating zone, which is designated as the entrance time. At the exit, another detection device records the time the same workpiece leaves the electroplating zone, which is designated as the exit time. The entrance time is earlier than the exit time; both constitute the electroplating zone time boundary for the workpiece. All subsequent process operation records and surface images related to the workpiece are indexed based on the workpiece number and the electroplating zone time boundary.

[0027] 1-2 Establish the foundation for collecting and aligning process operation records with time.

[0028] Within the electroplating zone, process quantities involved in the formation of the electroplating layer are collected according to a preset fixed sampling time step. At each sampling, the sampling time and the set of process quantities are simultaneously recorded. All sampled items are arranged in chronological order to form a complete process operation record sequence. This sequence covers a period before the entry point and a period after the exit point of the electroplating zone. Subsequently, for a specific workpiece, a segment between the entry and exit times is extracted from this sequence as the corresponding electroplating process operation record for that workpiece, ensuring that the process operation record corresponds to the workpiece's electroplating process in terms of time.

[0029] 1-3 The range and acquisition logic of the time alignment threshold interval.

[0030] An image acquisition device is installed along the conveying direction at the exit of the electroplating zone. The physical distance between the exit detection position and the image acquisition position is measured during device installation and forms a fixed distance. The conveying device is configured with a maximum and a minimum allowable conveying speed, where the maximum speed is greater than the minimum speed and the minimum speed is greater than zero. The running time of the workpiece moving from the exit detection position to the image acquisition position has a definite relationship with the aforementioned distance and the actual conveying speed. During actual operation, the running time varies within a finite range as the conveying speed changes between the maximum and minimum allowable speeds. By dividing the fixed distance by the maximum and minimum allowable speeds respectively, two time values ​​are obtained, with the time value corresponding to the maximum speed being smaller and the time value corresponding to the minimum speed being larger. The minimum time value is used as the lower time alignment threshold, and the maximum time value is used as the upper time alignment threshold. The time alignment threshold range is formed by the lower and upper time alignment thresholds, constituting a closed time interval with both a lower and upper bound. For any workpiece, as long as the conveyor speed varies within the allowable range, the travel time from the exit detection position to the image acquisition position will inevitably fall within this time alignment threshold range. The trend range of the time alignment threshold range changes with the conveyor speed setting. When the maximum allowable conveyor speed increases or the minimum allowable conveyor speed decreases, the time alignment threshold range shrinks overall towards the shortest time direction; when the maximum allowable conveyor speed decreases or the minimum allowable conveyor speed increases, the time alignment threshold range expands overall towards the longest time direction. Through this construction method, the time alignment threshold range remains consistent with the physical distance of the equipment and the process speed constraints under any speed setting.

[0031] 1-4 Image filtering of post-electroplated surfaces based on time alignment thresholds.

[0032] A surface image acquisition event is triggered when the workpiece enters the field of view of the image acquisition device, and the image acquisition time is recorded. For a given workpiece, the exit time and the upper and lower bounds of the time alignment threshold range are known. The lower bound of the image acquisition time is obtained by adding the lower time alignment threshold to the exit time, and the upper bound of the image acquisition time is obtained by adding the upper time alignment threshold to the exit time. When the image acquisition time is greater than or equal to the lower bound and less than or equal to the upper bound, the acquired surface image after electroplating is considered to be time-correlated with the workpiece and is recorded as a valid time-aligned image. When the image acquisition time is earlier than the lower bound or later than the upper bound, the image is considered to possibly correspond to another workpiece or an incorrect position of the same workpiece and is recorded as an invalid time-aligned image. The time alignment threshold range acts as a filter here. The lower and upper time alignment thresholds together limit the allowable offset range between the image acquisition time and the workpiece exit time, eliminating both premature acquisition before the workpiece enters the field of view and delayed acquisition after the workpiece leaves the field of view.

[0033] The range and acquisition logic of the resolution threshold interval 1-5.

[0034] After the image acquisition device is installed and the focus and lighting conditions are adjusted, a batch of workpieces with stable appearance quality are selected as calibration workpieces under actual electroplating production conditions. Electroplated surface images are acquired on these calibration workpieces to form a calibration image set. For each calibration image in the calibration image set, by traversing all pixels within the image area, the absolute value of the grayscale difference between each pixel and its horizontally adjacent pixels, as well as the absolute value of the grayscale difference between each pixel and its vertically adjacent pixels, are calculated. All these absolute grayscale difference values ​​in the entire image are then summed to obtain the sharpness evaluation value of that calibration image. The sharpness evaluation value reflects the total grayscale variation of edges and textures in the image. A higher sharpness evaluation value indicates richer spatial grayscale variation, while a lower value indicates smoother spatial grayscale variation. By sorting all sharpness evaluation values ​​in the calibration image set, the lowest sorted sharpness evaluation value is taken as the lower threshold, and the highest sorted sharpness evaluation value is taken as the upper threshold, forming a sharpness threshold range. The sharpness threshold range is a closed interval from the lower sharpness threshold to the upper sharpness threshold. The lower sharpness threshold is not less than zero, and the upper sharpness threshold is not less than the lower sharpness threshold. The trend range of the sharpness threshold range varies with optical configuration and electroplating surface characteristics. When the lens focus position shifts or the illumination brightness is insufficient, the sharpness evaluation value of the calibrated image generally shifts towards a smaller value, and the lower and upper sharpness thresholds decrease overall. When the lens focus accuracy improves or the illumination uniformity improves, the sharpness evaluation value of the calibrated image generally shifts towards a larger value, and the lower and upper sharpness thresholds increase overall. If there are individual extreme samples in the calibrated image set, such as samples with abnormal surface reflection or partial occlusion, the first and last few samples in the sorted sequence can be removed after sorting. Then, the minimum and maximum values ​​in the intermediate sequence can be selected as the lower and upper sharpness thresholds, making the sharpness threshold range closer to the actual sharpness range under long-term operation of the electroplating production line.

[0035] 1-6 Online surface image sharpness evaluation and screening logic.

[0036] For each online electroplated surface image that has passed the time alignment criteria, the sharpness evaluation is performed using the same calculation method as the calibration image. For each pixel in the image, the absolute value of the grayscale difference between it and its horizontally adjacent pixels and the absolute value of the grayscale difference between it and its vertically adjacent pixels are calculated. All these absolute grayscale difference values ​​across the entire image are then summed to obtain the sharpness evaluation value for the online image. This sharpness evaluation value is compared with a lower and upper sharpness threshold. When the sharpness evaluation value is greater than or equal to the lower sharpness threshold and less than or equal to the upper sharpness threshold, the online image is considered to be within the same reasonable range of sharpness as the calibration image and is marked as a valid sharpness image. When the sharpness evaluation value is less than the lower sharpness threshold or greater than the upper sharpness threshold, the sharpness of the online image is considered to deviate from the calibration range and is marked as an invalid sharpness image. An image validity flag is recorded in the workpiece record. The sharpness threshold range acts as a quality gate here, excluding blurry images with excessively smooth grayscale changes and noisy images with abnormally drastic grayscale changes from the usable image set by limiting the upper and lower bounds of the threshold.

[0037] 1-7 Original observation dataset construction and threshold usage results.

[0038] After completing time alignment and sharpness filtering, for each workpiece with both time-aligned and sharpness-validated images, the workpiece number, electroplating interval time boundaries, electroplating process segments extracted from the process operation record sequence, and the corresponding post-electroplating surface image are integrated to form a raw observation record. All raw observation records are aggregated according to workpiece number to form a raw observation dataset. Each record in the raw observation dataset contains electroplating interval time boundaries defined by entry and exit times, a process operation record subsequence arranged chronologically, and a valid post-electroplating surface image filtered by both time alignment and sharpness threshold intervals. The time alignment threshold interval provides the upper and lower bounds of the allowable offset of the image acquisition time relative to the workpiece exit time, while the sharpness threshold interval provides the upper and lower bounds of the allowable variation in image grayscale intensity. These two thresholds together ensure that the image information in the raw observation dataset and the process operation record have a unified standard in terms of time and quality, which is beneficial for the stable implementation of subsequent feature filtering, electroplating layer state estimation, and feedback control.

[0039] Step S1 establishes clear constraints between workpiece number, electroplating interval time boundary, process operation record, time alignment threshold interval, and sharpness threshold interval. This ensures that each record in the original observation dataset points to a unique workpiece on the time axis and has a unified data composition, eliminating the risk of mismatch between process operation records and post-electroplating surface images from the source. The time alignment threshold interval is derived from the physical distance between the exit position and the image acquisition position, as well as the conveyor speed range. It maintains stable time association rules even under scenarios with changing conveyor speeds, ensuring that the post-electroplating surface image always corresponds to the same workpiece at the actual electroplating completion position. The sharpness threshold interval is derived from the spatial grayscale variation characteristics of the calibration image. It limits the sharpness evaluation value to the calibration range, effectively eliminating blurry and noisy images, and ensuring that the image data entering the subsequent feature selection and state estimation stages has a uniform quality level. Through the joint selection of the time alignment threshold interval and the sharpness threshold interval, the original observation dataset forms dual constraints in the time dimension and image quality dimension, improving the reliability and traceability of the input data for online detection of electroplating layers, and providing a stable data foundation for machine learning-based electroplating layer state modeling and feedback control.

[0040] Step S1 has already generated raw observation data records at the individual workpiece level. Each record includes the workpiece number, the process operation record within the electroplating interval, and the post-electroplating surface image filtered by time alignment and clarity. In step S3, the electroplating layer state estimation model needs to read state input vectors with fixed structures, uniform numerical ranges, and clear physical meanings; otherwise, it will be difficult to achieve stable convergence during training and online inference, and it will be difficult to output sensitive information with engineering interpretation significance. However, the images in the raw observation data are two-dimensional grayscale arrays, and the process operation records are sequences with varying time lengths. Directly inputting these into the electroplating layer state estimation model will lead to inconsistent feature dimensions and mixed units. Therefore, step S2 needs to start from the raw observation data and, through a series of feature construction, feature filtering, and numerical normalization processes, transform the raw observation data into state input vectors for electroplating state analysis.

[0041] 2-1 Description of the original observation data objects.

[0042] The original observation dataset formed in step S1 contains multiple workpieces, each uniquely identified by a workpiece number. For each workpiece, the entry and exit times of the electroplating zone have been recorded, with the entry time preceding the exit time. Within the time range between the entry and exit times, process parameters such as electroplating current and bath temperature are collected at fixed time intervals, forming a process operation record sequence arranged chronologically. Simultaneously, after the workpiece leaves the electroplating zone, post-electroplating surface images are acquired through time alignment and sharpness filtering, ensuring uniform image size and grayscale resolution. Step S2, based on this, does not change the number of workpieces or time boundaries in the original observation data, but only constructs numerical features within each record.

[0043] 2-2 Construction of a zero-reference grayscale image.

[0044] To mitigate the impact of ambient lighting variations and overall exposure differences, a grayscale baseline is first constructed for the post-plating surface image of each workpiece. Specifically, the grayscale values ​​of all pixels in the entire image are summed, and this sum is divided by the total number of pixels to obtain an average grayscale value, which represents the overall brightness level of the workpiece. Then, this average grayscale value is subtracted from the grayscale value of each pixel in the image to obtain a zero-baseline grayscale image. In the zero-baseline grayscale image, the grayscale value represents the offset from the average brightness of the entire image; a value greater than zero indicates that the local brightness is higher than the overall level, and a value less than zero indicates that the local brightness is lower than the overall level. This processing ensures that the subsequently constructed image features reflect local structural differences more accurately, without being overly affected by overall brightness variations.

[0045] 2-3 Image candidate feature construction.

[0046] In a zero-reference grayscale image, spatial grayscale variations are described to represent the texture and edge structure information of the electroplated layer surface. Specifically, in the horizontal direction, the difference between the zero-reference grayscale values ​​of two adjacent pixels in each row is calculated, and the absolute value of the difference is taken. In the vertical direction, the difference between the zero-reference grayscale values ​​of two adjacent pixels in each column is calculated, and the absolute value of the difference is taken. The absolute values ​​of all horizontal and vertical differences in the entire image are summed to obtain the first image candidate feature. A larger value for this feature indicates more frequent grayscale variations and more edges and fine structures in the entire image; a smaller value indicates a smoother surface texture of the electroplated layer.

[0047] To characterize the texture differences between the central and edge regions of a workpiece, during equipment calibration, the workpiece body is divided into a central region and regions near the workpiece edge and near the fixture, based on its position and outline within the field of view. These two regions remain unchanged during subsequent operation. In the zero-reference grayscale image, the above difference calculation process is repeated within both the central and edge regions. The absolute values ​​of all differences are accumulated within the central region to obtain the central region texture quantity, and the absolute values ​​of all differences are accumulated within the edge regions to obtain the edge region texture quantity. Subtracting the edge region texture quantity from the central region texture quantity yields the second image candidate feature. When this feature is positive and its absolute value increases, it indicates that the total texture quantity in the central region is increasing relative to the edge region. When this feature is negative and its absolute value increases, it indicates that the total texture quantity in the edge region is increasing relative to the central region, thus reflecting the differences in the electroplating deposition state in different regions.

[0048] Regarding brightness distribution, a brightness threshold range is introduced to quantitatively describe the area proportion of locally bright or locally dark areas. This brightness threshold range is obtained during the calibration phase using a set of calibrated workpieces. Specifically, in the zero-reference grayscale images of multiple workpieces with stable electroplating quality, the zero-reference grayscale values ​​of all pixels are collected. These grayscale values ​​are arranged in ascending order. After removing some extreme samples at the beginning and end, the smallest zero-reference grayscale value from the middle sample set is taken as the lower brightness threshold, and the largest zero-reference grayscale value from the middle sample set is taken as the upper brightness threshold. The lower brightness threshold is not greater than zero, and the upper brightness threshold is not less than zero. During online processing, in the zero-reference grayscale image, for each pixel, if the zero-reference grayscale value is less than the lower brightness threshold, the pixel is included in the dark pixel set; if the zero-reference grayscale value is greater than the upper brightness threshold, the pixel is included in the bright pixel set. The number of dark pixels and bright pixels are counted. The number of dark pixels is subtracted from the number of bright pixels, and this difference is divided by the total number of pixels to obtain the third image candidate feature. The third image candidate feature, when close to zero, indicates that the areas of bright and dark regions are similar; a positive value with increasing absolute value indicates an increase in the proportion of bright areas; and a negative value with increasing absolute value indicates an increase in the proportion of dark areas. These three indicators form a set of image candidate features that can reflect the intensity of the electroplated layer's texture, the differences in texture between different regions, and the distribution of light and dark areas.

[0049] 2-4 Construction of process candidate features.

[0050] In the process operation log, the electroplating current time series characterizes the change in charge flux throughout the electroplating process, directly affecting the overall thickness of the electroplated layer and local defects. Therefore, electroplating current-related features are given priority in the candidate feature construction. The process operation log has already collected electroplating current values ​​and corresponding sampling times at fixed time intervals.

[0051] First, an integral characteristic of the electroplating current is constructed. Specifically, between the entry and exit times of the workpiece electroplating zone, for each sampling time interval, the electroplating current value within that time interval is multiplied by the corresponding time interval to obtain the charge amount for that time interval. Then, the charge amounts of all sampling time intervals within the electroplating zone are summed to obtain a total charge amount index. The total charge amount represents the cumulative charge input of the entire workpiece during the electroplating process and has a direct relationship with the overall thickness of the electroplated layer.

[0052] Secondly, to measure the duration of excessively high current regions during electroplating, an upper current threshold needs to be constructed. This upper current threshold is obtained during the calibration phase by collecting all current sampling values ​​from the electroplating current time series of multiple workpieces. These current values ​​are arranged in ascending order, and after removing samples from the top and bottom of the sorted sequence, a value is selected from the middle portion of the samples as the upper current threshold. The upper current threshold lies in the middle of the overall range of current sampling values; it is neither equal to the absolute maximum nor the absolute minimum value of the current sample. During online processing, within the electroplating interval of the workpiece, for each sampling time interval, if the electroplating current value is greater than the upper current threshold, the time interval of that interval is included in the high current duration. All time intervals that meet the condition are accumulated to obtain the high current duration characteristic. The larger the high current duration characteristic value, the longer the electroplating process lasts above the upper current threshold, and the more significant the impact on local over-thickness or overheating defects.

[0053] Furthermore, to describe the abrupt changes in electroplating current during the later stages of the electroplating process, the latter half of the electroplating interval is divided into several consecutive time sub-intervals of equal length. For each time sub-interval, the products of all electroplating current samples within that sub-interval and their corresponding time intervals are summed, and this sum is then divided by the total time of the sub-interval to obtain the average current value for that sub-interval. Subsequently, the difference between the average current values ​​of each pair of adjacent time sub-intervals is calculated, and the absolute value of the difference is taken. The maximum absolute difference among all the absolute differences between adjacent sub-intervals is used as the current jump characteristic. The current jump characteristic reflects the abrupt change in current regulation during the later stages of electroplating; a larger value indicates a more pronounced current adjustment behavior, which may correspond to a rapid increase or decrease in current during the electroplating operation. Through the current integral characteristic, the high current duration characteristic, and the current jump characteristic, a set of candidate process characteristics reflecting the intensity level and dynamic changes of the electroplating process is formed.

[0054] 2-5 Monotonic Consistency Indicators and Feature Selection Logic.

[0055] The number of image and process candidate features is limited, but the correlation between different features and the actual state of the electroplated layer varies. Therefore, a reference appearance index needs to be introduced during the calibration stage to evaluate the monotonic relationship between each candidate feature and the reference appearance, and to select the features more suitable as state input. To this end, several batches of workpieces are selected as a calibration workpiece set at the initial production stage or during process adjustment. Each calibration workpiece, in addition to having original observation data, also obtains a reference appearance index. This index is derived from, for example, offline film thickness test results or manual quality grades. After unified conversion, a larger value indicates better appearance quality, and a smaller value indicates worse appearance quality.

[0056] For each of the above candidate features, a monotonic consistency index is constructed. Specifically, any two different calibrated workpieces are enumerated in the calibrated workpiece set. The magnitude relationship of their reference appearance indices is compared, and the value relationship of the same candidate feature on these two workpieces is also compared. If there is a strict difference in the magnitude of the reference appearance indices in a pair of workpieces, and the value of this candidate feature on these two workpieces also shows a difference in the same direction, then it is determined that the candidate feature and the reference appearance indices change in the same direction on this workpiece pair. The number of times the change direction is consistent across all calibrated workpiece pairs is accumulated to obtain the consistency count. This consistency count is then divided by the total number of calibrated workpiece pairs to obtain a monotonic consistency index with a value between zero and one. A value of zero indicates that the change direction is completely opposite across all workpiece pairs; a value of one indicates that the change direction is completely consistent across all workpiece pairs; and a value between zero and one indicates that some workpiece pairs exhibit consistent direction.

[0057] The monotonic consistency indices of all candidate features are aggregated and sorted in ascending order to form an ordered sequence. Directly selecting the feature with the largest or smallest value as the final feature is susceptible to interference from a few outliers; therefore, a monotonic consistency threshold interval is introduced. The logic for obtaining the monotonic consistency threshold interval is as follows: in the ordered sequence, several feature indices at the very beginning and several at the very end are deleted, leaving a middle section as the candidate interval. Within this middle interval, the smallest monotonic consistency index is taken as the lower bound threshold, and the largest monotonic consistency index is taken as the upper bound threshold. The lower bound threshold is greater than or equal to zero, and the upper bound threshold is less than or equal to one; the lower bound threshold is not greater than the upper bound threshold. For each candidate feature, if its monotonic consistency index falls between the lower and upper bound thresholds, it is retained as a formal feature. If its monotonic consistency index is less than the lower bound or greater than the upper bound, the ranking relationship between the feature and the reference appearance is considered unstable, and it is not included in the state input vector. This intermediate-range screening method ensures that the final selected features are neither obviously irrelevant nor completely dependent on a few extremely consistent samples.

[0058] 2-6 Feature numerical specification and state input vector construction.

[0059] After selecting the formal features, numerical standardization is required to ensure the electroplating state estimation model can accept input within a unified numerical range. Numerical standardization is based on a set of calibrated workpieces. For each formal feature, the minimum and maximum eigenvalues ​​are found within the calibrated workpiece set. For any formal feature, if the maximum eigenvalue is strictly greater than the minimum eigenvalue, then for any workpiece, the minimum eigenvalue is subtracted from its eigenvalue. This difference is then divided by the difference between the maximum and minimum eigenvalues, thus mapping the original numerical range to zero to one. Specifically, when the original eigenvalue equals the minimum in the calibrated set, the standardized value is zero; when the original eigenvalue equals the maximum, the standardized value is one; and when the original eigenvalue falls between these two values, the standardized value is between zero and one. If the minimum eigenvalue of a formal feature in the calibrated workpiece set equals the maximum, it indicates that this feature is indistinguishable across the calibrated samples. In this case, the standardized value of this feature is uniformly set to zero for all workpieces, and this feature does not produce numerical distinctions between different workpieces in the state input vector.

[0060] After performing the above numerical normalization operation on all formal image features and formal process features one by one, the normalized feature values ​​are arranged in a fixed order for each workpiece, resulting in a numerical vector. This numerical vector contains the normalized values ​​of all formal image features and all formal process features, with a fixed number of elements and a consistent arrangement order across all workpieces. This numerical vector is the state input vector, serving as the input to the electroplating layer state estimation model in step S3. The state input vector is uniformly limited to a value between zero and one, and its elements correspond to specific physical quantities such as texture intensity differences, brightness distribution, total current integral, high current duration, and current change amplitude. This facilitates model training and subsequent interpretation of sensitivity information from a physical perspective.

[0061] Through step S2, the electroplated surface image and process operation record in the original observation data are transformed into a state input vector with a fixed structure in the dimension of workpiece number. Image candidate features are constructed by zero reference gray level, texture intensity, regional difference, and brightness ratio. Process candidate features are constructed by current integral, high current duration, and current jump amplitude. Then, monotonic consistency index is calculated by combining reference appearance index and formal features are selected within the monotonic consistency threshold range. At the same time, the minimum and maximum values ​​of the calibration samples are used to perform linear numerical standardization to ensure that the state input vector of the electroplating layer state estimation model is consistent in dimension, numerical range, and physical meaning. This solves the problem of inconsistent dimensions, mixed dimensions, and difficult interpretation of feature contributions caused by the direct splicing of image information and process information in the background technology. It improves the correspondence between the electroplating layer state description and sensitivity information and the actual appearance changes and the strength of the electroplating process, and establishes a stable and traceable feature space for subsequent online detection and feedback control.

[0062] Step S2 has compressed the post-plating surface image and process operation record of each workpiece into a state input vector with a fixed structure and uniform numerical range. Step S3 needs to assign a pair of electroplating layer state descriptors to each workpiece based on these state input vectors, which can reflect the overall thickness deviation trend and local anomaly risk, and calculate the degree of influence of each process feature on this pair of state descriptors, which will be used as the basis for determining "which process quantity is more useful" in subsequent feedback control. Therefore, step S3 needs to explain how to construct the target value of the electroplating layer state descriptor from offline thickness or defect information, how to train the electroplating layer state estimation model, and how to construct process feature sensitivity information through finite difference method.

[0063] 3-1 Preparation of training workpiece set and thickness measurement data.

[0064] During the production debugging or stable operation phase, a representative set of workpieces is selected from the electroplating production line to form a training workpiece set. For each workpiece in the training workpiece set, the state input vector formed in step S2 is retained, and the thickness of the workpiece surface is measured in the offline inspection stage. During thickness measurement, the workpiece surface is divided into multiple measurement positions according to rules, with each measurement position corresponding to a thickness measurement value. The measurement positions cover the main body area and edge area of ​​the workpiece, and the number of measurements is sufficient to reflect the thickness distribution. In this way, each training workpiece has both a state input vector and a set of thickness measurement values ​​arranged by position. For some workpieces where only defect annotations can be obtained but complete thickness measurements cannot be obtained, defect statistics can be embedded instead of absolute thickness deviation in subsequent descriptions, but the processing logic of the thickness version is described as the main thread.

[0065] 3-2 Construction of overall thickness deviation from descriptive quantity.

[0066] For any training workpiece, the first step is to construct an index to measure the overall thickness uniformity. The process involves: averaging the thickness values ​​at all measurement locations; subtracting this average thickness from the thickness value at each measurement location to obtain the thickness deviation at each location; taking the absolute value of each deviation; summing the absolute deviations of all measurement locations; and finally dividing this sum by the number of measurement locations. The result is defined as the overall thickness deviation descriptor. When the overall thickness deviation descriptor is close to zero, it indicates that the thickness values ​​at each measurement location are concentrated near the average thickness, resulting in a more uniform thickness distribution; when the overall thickness deviation descriptor is far from zero, it indicates that the thickness deviation from the average value at multiple measurement locations has increased, leading to increased thickness non-uniformity.

[0067] 3-3 Construction of local risk descriptors and logic for obtaining thresholds.

[0068] The overall thickness deviation descriptor reflects the deviation under a multi-point average, while local defects often appear in a limited number of locations with large deviations. Therefore, a separate local risk descriptor needs to be constructed. Specifically, after calculating the thickness deviation, the absolute thickness deviation of each measurement location is used to sort all absolute thickness deviations from smallest to largest, resulting in an ordered deviation sequence. Then, a local risk percentage threshold is introduced. This threshold is a decimal between zero and one, representing the proportion of locations considered as candidate locations for extreme local deviations. The value of the local risk percentage threshold is determined through engineering experience; for example, it can be selected within the range of 5% to 20%. Historical data is used to observe the consistency between the local risk descriptor and the actual defect statistics when different percentages are selected, and a set of percentages that maintain a stable matching relationship across different batches of data is chosen as the threshold.

[0069] After obtaining the local risk percentage threshold, this percentage is multiplied by the number of measurement locations and rounded down to obtain the number of measurement points involved in the local risk calculation. Then, from the sorted sequence of absolute thickness deviations, the entries corresponding to the largest number of measurement points are selected; this set represents the measurement locations with the most severe thickness deviations. The average of the absolute thickness deviations at these locations is then calculated to obtain the local risk descriptor. An increase in the local risk descriptor indicates that a few locations have thickness deviations far exceeding the average thickness, making the local risk more prominent; a value close to zero indicates that the largest deviation values ​​in the sorted sequence are also close to zero, and the phenomenon of extreme local deviations is not obvious.

[0070] 3-4 The formation method of the target value of the electroplating layer state description quantity.

[0071] For each training workpiece, the overall thickness deviation descriptor and the local risk descriptor together constitute a binary numerical pair, which is defined in this invention as the target value of the electroplating state descriptor. The overall thickness deviation descriptor is used to mark the uniformity level of the electroplating layer across the entire workpiece, while the local risk descriptor is used to mark the degree of risk of significant deviation or defects in local areas of the electroplating layer. For training workpieces that only have defect annotations but not continuous thickness measurements, a similar approach is used to construct the target value of the electroplating state descriptor: for example, two indicators can be constructed using the number of defects, the proportion of defects to the total area, and the degree of brightness anomaly in the defect area. These indicators can then be scaled down to the numerical range corresponding to the thickness deviation through linear mapping, and the two indicators can be concatenated into a binary numerical pair as the target value of the electroplating state descriptor. Through this arrangement, each workpiece in the training workpiece set corresponds to a state input vector and a target value of the electroplating state descriptor, providing paired samples for subsequent training of the electroplating state estimation model.

[0072] 3-5 Training process of electroplating layer state estimation model.

[0073] In terms of model structure, the electroplating layer state estimation model is designed as a multi-layer nonlinear mapping structure. The state input vector has been numerically compressed into a fixed interval, and each component has a clear physical meaning. The state input vector is input into the first layer's combination of linear and nonlinear transformations to obtain intermediate feature representations. These representations are then passed through several similar intermediate layers, and finally, the output layer provides a vector containing two components. These two components are used to estimate the overall thickness deviation descriptor and the local risk descriptor, respectively. The entire model contains a set of trainable parameters, which control the coefficients of the linear transformations and the bias terms of the nonlinear transformations in each layer.

[0074] During the training phase, for each workpiece in the training workpiece set, its state input vector is first fed into the electroplating layer state estimation model to obtain two descriptive components of the model output. These components are then compared with the corresponding components in the target value of the electroplating layer state descriptive quantity. For the overall thickness deviation component, the absolute value of the difference between the model output value and the target value is calculated. Similarly, for the local risk component, the absolute value of the difference between the model output value and the target value is calculated. The two absolute values ​​are then added together to obtain the error value for this training workpiece. For all workpieces in the training workpiece set, their respective error values ​​are summed to obtain a total error. During training, the model's internal parameters are adjusted iteratively. In each iteration, the parameters are updated in the direction of reducing the total error, causing the total error to gradually decrease during training. When the change in the total error is lower than a preset threshold after several iterations, the model can be considered to have learned a relatively stable input-output correspondence.

[0075] 3-6 Normalization of electroplating layer state description quantities.

[0076] To facilitate the use of state description quantities with unified dimensions in subsequent steps, after the training phase, the range of target values ​​for electroplating state description quantities in the training workpiece set is statistically analyzed. For the overall thickness deviation from the target value of the description quantity for all training workpieces, the minimum and maximum values ​​are found. The minimum value is considered the lower bound of the overall thickness deviation, and the maximum value is considered the upper bound of the overall thickness deviation. These two values ​​form a closed interval, corresponding to the range of overall deviation values ​​that occurred during the training phase. For the target values ​​of local risk description quantities for all training workpieces, the minimum and maximum values ​​are found in the same way, forming a closed interval for local risk values.

[0077] During online operation, after the electroplating layer state estimation model outputs the overall thickness deviation estimate and the local risk estimate for the entire workpiece, the output is linearly normalized according to the upper and lower bounds obtained during the training phase. For example, for the overall thickness deviation estimate, the lower bound of the overall thickness deviation is first subtracted, and then the difference is compared with the difference between the upper and lower bounds of the overall thickness deviation to obtain a dimensionless value between zero and one. If the upper and lower bounds of the overall thickness deviation are equal during the training phase, it indicates that the overall deviation has not changed in the training data, and the normalized value of the overall thickness deviation can be directly set to zero. The local risk component is normalized in the same way. After normalization, in the dimensionless vector of the binary electroplating layer state description, the values ​​of both components are between zero and one. The value near zero indicates that the corresponding state is close to the most ideal state in the training data, and the value near one indicates that the corresponding state is close to the situation with the most severe deviation or the highest risk in the training data.

[0078] 3-7 Calculation logic for online electroplating layer state description quantities.

[0079] During the online operation phase, for each workpiece processed in step S2, its state input vector is obtained. This vector is then fed into the pre-trained electroplating layer state estimation model to obtain two descriptive components from the model output. These components are then converted into dimensionless overall thickness deviation descriptive quantity and dimensionless local risk descriptive quantity using the normalization method described in step S3.6. For online monitoring, the dimensionless overall thickness deviation descriptive quantity directly reflects the level of uniformity of the overall thickness of the current workpiece, while the dimensionless local risk descriptive quantity directly reflects the risk level of significant deviations at a few locations on the current workpiece. These two values ​​will serve as important inputs for constructing the quantitative results of appearance deviation in step S4.

[0080] 3-8 Calculation logic for process feature sensitivity information.

[0081] The state input vector contains two types of components: image features and process features. Sensitivity analysis focuses on the influence of process features on the descriptive quantity of the electroplating layer state. Therefore, firstly, according to the feature concatenation order in step S2, a set of all component indices belonging to process features is listed. For each process feature component index in the set, during the training phase, statistics are performed on the training workpiece set to find the minimum and maximum values ​​of this component on the training workpiece. The dynamic range of this process feature component during the training phase is obtained by subtracting the minimum value from the maximum value. Then, a perturbation scaling factor is set for this process feature component. The perturbation scaling factor is a decimal between zero and one, selected by process engineers through experiments during the parameter tuning phase. Usually, a value much smaller than one is chosen so that the perturbation occupies only a small portion of the dynamic range. Multiplying the dynamic range by the perturbation scaling factor yields the perturbation amplitude of this process feature component.

[0082] For the workpiece whose sensitivity needs to be calculated, two perturbation vectors are constructed around the current state input vector. One is a positive perturbation vector, which, based on the original state input vector, increases the perturbation amplitude only for the current process feature component, while keeping the other components unchanged. The other is a negative perturbation vector, which, based on the original state input vector, decreases the perturbation amplitude only for the current process feature component, while keeping the other components unchanged. The positive perturbation vector is fed into the electroplating layer state estimation model to obtain a set of dimensionless electroplating layer state descriptor outputs, and the negative perturbation vector is fed into the electroplating layer state estimation model to obtain another set of dimensionless electroplating layer state descriptor outputs. For the overall thickness deviation component, the difference between the positive and negative perturbation output values ​​is calculated, and the ratio of this difference to twice the perturbation amplitude is taken as the local rate of change around the current process feature component; for the local risk component, the local rate of change is calculated in the same way. The absolute value of the local rate of change reflects the response speed of the electroplating layer state descriptor component to changes in the process feature component.

[0083] To synthesize the influence of two state components on the same process feature component into a single scalar sensitivity, the larger of the absolute values ​​of the two local rates of change is taken as the initial sensitivity value of the current workpiece to the current process feature component. The larger this initial sensitivity value, the more significant the small changes in the process feature component near the current operating point will be in terms of either overall thickness deviation or local risk. By repeating the above operation for all process feature components, a set of initial sensitivity values ​​corresponding one-to-one with each process feature component can be obtained.

[0084] 3-9 Normalization and output format of sensitivity information.

[0085] During the training phase or on online running data over a period of time, the raw sensitivity values ​​corresponding to each process feature component are statistically analyzed, and the minimum and maximum values ​​of the raw sensitivity values ​​for each process feature component are calculated. For any raw sensitivity value, if the maximum value is greater than the minimum value, it is converted into a dimensionless sensitivity value between zero and one through a linear mapping: first, the minimum value is subtracted from the raw sensitivity value, and then the difference is divided by the difference between the maximum and minimum values; the resulting ratio is the normalized sensitivity. If the minimum and maximum raw sensitivity values ​​of a certain process feature component in the statistical data are equal, it indicates that within the statistical data range, this process feature component does not bring about an observable change to the electroplating layer state description; in this case, the corresponding normalized sensitivity can be directly set to zero.

[0086] After normalization, a set of sensitivity values ​​between zero and one is obtained for each workpiece. Each sensitivity value corresponds one-to-one with a process feature component in the state input vector. A sensitivity value close to one indicates that the corresponding process feature component has a significant impact on the electroplating layer state description within the training statistical range, while a sensitivity value close to zero indicates that the corresponding process feature component has a limited impact on state changes within the current process range. In step S4, when constructing the process adjustment reliability quantification results, this sensitivity information will be used in conjunction with historical process adjustment records to distinguish the intensity of different process features.

[0087] Through the processing in step S3, offline thickness measurement or defect annotation information is introduced at the state input vector level to construct an overall thickness deviation descriptor and a local risk descriptor. The electroplating layer state estimation model provides a dimensionless output with a unified scale. At the same time, process feature sensitivity information is calculated for each process feature component, enabling the electroplating production line to obtain a dual-channel state characterization that can distinguish between overall thickness uniformity and local abnormal risks. It also obtains quantitative results on the degree of influence of process features on state changes. This solves the problem in the background technology that electroplating quality is only represented by acceptance criteria or the number of single defects, which cannot reflect the thickness distribution pattern and cannot be traced back to specific process features. It reduces experience-based trial-and-error adjustment behavior and establishes a stable mapping relationship between online detection results and subsequent process parameter adjustments. This provides a direct input basis for constructing quantitative results of appearance deviation and quantitative results of the intensity of process adjustment in subsequent steps.

[0088] Step S3, based on the state input vector and combined with offline thickness measurement or defect annotation results, trains the electroplating layer state estimation model, outputs a dimensionless electroplating layer state description quantity, and calculates dimensionless process feature sensitivity information for each process feature component. Step S4, based on the above, simultaneously utilizes the current workpiece image and state description quantity to provide a quantitative result reflecting the current appearance deviation, and utilizes historical process adjustment records and sensitivity information to provide a quantitative result reflecting the reliability of each process feature adjustment. Then, based on these two quantitative results, a mapping relationship from the appearance deviation degree to the process parameter adjustment range is constructed. Adjustment coefficients for correcting subsequent workpiece adjustable process parameter settings are extracted from the mapping relationship, so that feedback control behavior no longer relies on empirical trial and error, but is based on the quantified deviation degree and the quantified adjustment reliability.

[0089] 4-1 Construction of the quantitative results of appearance deviation.

[0090] In step S3, each workpiece has obtained a pair of dimensionless electroplating state descriptor values. One component represents the normalized result of the overall thickness deviation descriptor, and the other component represents the normalized result of the local risk descriptor. The values ​​of both components fall between zero and one. A value close to zero indicates that the corresponding state is close to the ideal state in the training sample, and a value close to one indicates that the corresponding state is close to the severely deviated state in the training sample.

[0091] In the state input vector, step S2 has already generated several normalized image features, including a normalized feature of the texture difference between the central and edge regions, and a normalized feature of the ratio of bright to dark areas. These features are also compressed into a numerical range of zero and one. The texture difference feature between the central and edge regions is used to reflect the degree of imbalance in texture intensity between the central and edge regions of the workpiece, while the bright to dark area ratio feature is used to reflect the degree of difference between the areas of bright and dark regions in the image.

[0092] When constructing the quantitative results of appearance deviation, a morphological deviation index is first constructed within the electroplating layer state description quantity. The morphological deviation index is calculated by comparing the normalized value of the overall thickness deviation description quantity with the normalized value of the local risk description quantity, and selecting the larger value as the morphological deviation index. Thus, in terms of thickness distribution, the morphological deviation index will be reflected as long as either the overall thickness non-uniformity or the local extreme deviation reaches a severe level.

[0093] Subsequently, an image deviation index is constructed within the image features. The image deviation index is calculated by comparing the normalized feature of the texture difference between the center and edge regions and the normalized feature of the brightness-dark area ratio. The larger value is selected as the image deviation index. If the texture difference feature value is close to one, it indicates a very significant difference in texture intensity between the center and edge regions; if the brightness-dark area ratio feature value is close to one, it indicates a very prominent difference in the area ratio between bright and dark regions in the image.

[0094] After constructing the morphological deviation index and the image deviation index, a total appearance deviation index is introduced. The total appearance deviation index is calculated by comparing the morphological deviation index and the image deviation index, selecting the larger value as the total appearance deviation index. A value close to zero indicates that both thickness and image appearance are within the normal range; a value close to one indicates that at least one of the thickness or image appearance is approaching the severe deviation range recorded in the training data. In this way, thickness and image information are uniformly converted into a single quantitative result of appearance deviation with a value between zero and one.

[0095] 4-2 Historical process adjustment records sorting and event screening.

[0096] The quantification of process adjustment reliability needs to be based on real historical adjustment behavior. Therefore, a set of process feature indices is defined along the process feature dimension. Each index corresponds to a process feature component in the state input vector, such as electroplating current, electroplating time, solution temperature, etc. For each process feature index, within a set historical time window, all events involving adjustments to that process feature are extracted from production records, forming a set of process adjustment events.

[0097] Each process adjustment event includes one workpiece before adjustment and one immediately adjacent workpiece after adjustment. The workpiece before adjustment is the one that underwent electroplating before the process parameter modification was executed, and the immediately adjacent workpiece after adjustment is the first workpiece that underwent electroplating after the process parameter modification was executed. For each event record, the following information is saved: the workpiece number before adjustment, the workpiece number after adjustment, the numerical change of the process feature component between the two workpieces, and the total appearance deviation index for each of the two workpieces. The total appearance deviation index has been defined in one of steps S4 and can be directly calculated from the output and image features of step S3.

[0098] To filter out minor adjustments that have almost no physical impact from historical records, a process adjustment magnitude threshold is introduced. The threshold value lies between zero and the difference between the process feature's numerical range during training. The logic is as follows: within a historical time window, the absolute values ​​of all adjustments for the process feature are collected, sorted from smallest to largest, and the smallest samples at the top and the largest samples at the bottom are removed to eliminate extreme cases. Then, a representative value is selected from the remaining values ​​as the process adjustment magnitude threshold. This ensures that the process adjustment magnitude threshold lies within the common levels of historical adjustment behavior, unaffected by occasional extreme adjustments.

[0099] Meanwhile, to distinguish between natural fluctuations in appearance deviation and effective changes caused by adjustments, an appearance deviation change threshold is introduced. The appearance deviation change threshold is set between zero and one, and its acquisition logic is as follows: In a stable production data set without process adjustments, the absolute value of the difference between the total index of appearance deviations of adjacent workpieces is calculated, resulting in a set of natural fluctuation amplitude samples. These samples are then sorted, and a value near the middle position of the sorted sequence is selected as the appearance deviation change threshold, making the threshold slightly higher than the typical amplitude of natural fluctuations, so that only changes significantly exceeding the natural fluctuation range are retained in subsequent statistics.

[0100] After the aforementioned thresholds are prepared, for each process adjustment event, first calculate the absolute value of the change in the process characteristic value of the event, and then calculate the difference between the total index of the degree of appearance deviation of the workpiece before and after the adjustment. Then, take the absolute value of this difference. Only when the absolute value of the change in the process characteristic is not less than the process adjustment range threshold, and the absolute value of the difference in appearance deviation is not less than the appearance deviation change threshold, is the event considered a significant adjustment event. For significant adjustment events, if the difference in appearance deviation is negative and its absolute value is not less than the appearance deviation change threshold, it is recorded as a successful event, indicating a significant decrease in the degree of appearance deviation after adjustment; if the difference in appearance deviation is positive and its absolute value is not less than the appearance deviation change threshold, it is recorded as a failed event, indicating a significant increase in the degree of appearance deviation after adjustment; if the absolute value of the difference in appearance deviation is less than the appearance deviation change threshold, it indicates that the impact of this adjustment on the degree of appearance deviation falls within the natural fluctuation range, and it is recorded as a neutral event, not included in the success or failure statistics.

[0101] Through the above event filtering and classification process, each process feature index can obtain the number of successful events and the number of failed events within a specified historical window, which lays the foundation for the construction of reliability indicators for subsequent process adjustments.

[0102] 4-3 Construction of sensitivity threshold and reliability quantification results of process adjustment.

[0103] Step S3 has provided a dimensionless sensitivity value for each workpiece and each process feature component. The sensitivity value is between zero and one; the closer the value is to one, the more significant the impact of the process feature on the electroplating layer state description under the current operating conditions; the closer the value is to zero, the weaker the impact. The reliability quantification results for process adjustments should only include those adjustment events that, according to the model's determination, have a significant impact on the state description. Therefore, low-sensitivity events need to be filtered out using a sensitivity threshold.

[0104] The sensitivity threshold ranges between zero and one. The logic for obtaining it is as follows: From the set of all process feature sensitivity values ​​calculated during the training phase, all sensitivity values ​​are arranged in ascending order, and the value near the middle of the sequence is selected as the sensitivity threshold. This selection method ensures that the sensitivity threshold is roughly located near the median level of all sensitivity samples, thus prioritizing applications with moderate or higher sensitivity levels in the statistics.

[0105] For a given process feature index, a set of significant adjustment events has already been obtained in section 4.2, and each event has been labeled as successful, neutral, or failed. Now, a sensitivity threshold is introduced. For each event, the dimensionless sensitivity value of the corresponding process feature component of the workpiece before adjustment is re-examined. Only events with a sensitivity not lower than the sensitivity threshold before adjustment are retained for success and failure counting. Events with a sensitivity lower than the threshold are considered "insufficiently clear influence channels" and are removed from the success and failure statistics.

[0106] Among the retained events, the number of successful events and the number of failed events are counted separately. The number of successful events is divided by the sum of the number of successful and failed events. The result, assuming the denominator is not zero, is a ratio between zero and one. This ratio is defined as the initial reliability index for process adjustment. When the sum of the number of successful and failed events is zero, it indicates that there are no usable events under the constraints, and the initial reliability index for process adjustment is directly set to zero. When the initial reliability index is close to one, it indicates that in highly sensitive and significantly variable events, successful events far outnumber failed events; when the value is close to zero, it indicates that failed events dominate.

[0107] Considering that when the number of statistical events is very limited, the initial reliability index of process adjustments is prone to overly optimistic or pessimistic judgments due to individual events, a smoothing coefficient is introduced for correction. To this end, the total number of events is obtained by first summing the number of successful events and the number of failed events. Then, a positive constant is preset as the smoothing control benchmark. The total number of events is divided by the sum of the total number of events and this constant, resulting in a smoothing coefficient with a value between zero and one. When the total number of events is much smaller than this constant, the smoothing coefficient is close to zero, indicating insufficient statistical information; when the total number of events is significantly larger than this constant, the smoothing coefficient gradually approaches one, indicating relatively sufficient statistical information.

[0108] The final process adjustment reliability index is obtained by multiplying the smoothing coefficient by the initial process adjustment reliability index. This constructed index remains between zero and one: when a large number of highly sensitive and significant adjustment events accumulate with a high success rate, the index value approaches one; when the number of events is extremely small or failures predominate, the index value approaches zero. This index reflects the degree of success of a given process characteristic in historical adjustment behavior and will play a role in shaping the shape of the deviation from the adjustment magnitude mapping curve.

[0109] 4-4 Mapping and adjustment coefficient generation between appearance deviation and process parameter adjustment range.

[0110] After obtaining the overall index of appearance deviation and the reliability index of process adjustment, it is necessary to construct a mapping relationship from the overall index of appearance deviation to the proportion of process parameter adjustment. The mapping relationship consists of two levels: the first is a basic mapping that only considers the degree of appearance deviation, and the second is to apply shape adjustment controlled by the reliability of process adjustment on the basic mapping to obtain the adjustment coefficient.

[0111] The basic mapping principle expresses the fundamental principle that the more severe the appearance deviation, the closer the parameter adjustment range should be to the upper limit. The specific method of basic mapping is as follows: the value of the total appearance deviation index is squared to obtain a new dimensionless value, denoted as the basic adjustment intensity. The value of the basic adjustment intensity remains between zero and one. When the total appearance deviation index is close to zero, the basic adjustment intensity is close to zero, indicating that significant adjustments to process parameters are not recommended; when the total appearance deviation index is close to one, the basic adjustment intensity is close to one, indicating that parameter adjustments close to the maximum permissible range should be adopted. Due to the use of squaring, the basic adjustment intensity is still significantly lower than one for moderate deviations, which better meets the engineering requirement of gradual adjustment for minor deviations and stricter adjustment for severe deviations.

[0112] The reliability index for process adjustment is used to control the curvature of the mapping curve. To this end, a shape adjustment coefficient is introduced. This coefficient is selected by process engineers during the trial operation phase based on the production line response, and its value is between zero and one. The closer the shape adjustment coefficient is to zero, the closer it is to a linear mapping; the closer the shape adjustment coefficient is to one, the more significant the adjustment effect on the curve shape.

[0113] Based on the shape adjustment coefficient and the process adjustment reliability index, a power-law parameter is constructed for each process feature. The calculation logic of the power-law parameter is as follows: take 1 as the baseline value, and add it to the product of the shape adjustment coefficient and 1 minus the difference between the process adjustment reliability index and the baseline value. Thus, when the process adjustment reliability index is close to 1, the power-law parameter is close to 1, and the mapping relationship is close to linear; when the process adjustment reliability index value is low, the power-law parameter is significantly greater than 1, and the mapping curve compresses the adjustment range when the appearance deviation is small, and still maintains an output value close to 1 when the appearance deviation is close to 1.

[0114] When calculating the adjustment coefficient, the following approach is used for each workpiece and each process feature: the basic adjustment strength is used as the base of the exponentiation, and the exponent parameter corresponding to the process feature is used as the exponent. The basic adjustment strength is then exponentially calculated to obtain the adjustment coefficient. The adjustment coefficient remains numerically between zero and one. When the process adjustment reliability index is close to one, the exponent is close to one, and the adjustment coefficient is close to the value of the basic adjustment strength. When the process adjustment reliability index is close to zero, the exponent is close to one. Adding the shape adjustment coefficient, when the appearance deviation is in the low to medium range, the adjustment coefficient after exponentiation is significantly lower than the basic adjustment strength. When the appearance deviation is close to one, the adjustment coefficient remains close to one. In this way, when the process adjustment reliability is low, a more conservative adjustment strategy is adopted for small deviations, while sufficient adjustment range is reserved for large deviations, achieving adaptive adjustment based on historical performance.

[0115] After obtaining the adjustment coefficient, it needs to be mapped to the actual process parameter adjustment amount. For each process feature, the maximum allowable range of a single parameter adjustment is given in advance during the process design stage. This maximum allowable range is jointly determined by equipment capacity, quality risk, and safety standards, and is used as a constant throughout the implementation process. During the online control stage, for each workpiece and each process feature, the adjustment coefficient is multiplied by the maximum allowable adjustment range applied to that process feature, and the product is used as the absolute range of this parameter adjustment. The adjustment direction of the process parameters is determined by the electroplating process rules and the deviation direction of the electroplating layer state description quantity. For example, when the overall thickness deviates from the description quantity and the local risk description quantity indicates that the thickness is too large and accompanied by the risk of overheating, the adjustment direction is reduced for current-type process features; when the description quantity indicates that the thickness is too small, the adjustment direction is increased for current-type process features. The adjustment coefficient mainly controls the magnitude of the adjustment range, while the adjustment direction is jointly determined by the state deviation direction and process knowledge.

[0116] Through the processing in step S4, the electroplating production line no longer relies solely on experience thresholds or single defect counts to manually determine how to adjust process parameters. Instead, it first uses dimensionless overall thickness deviation descriptors, dimensionless local risk descriptors, and texture difference features and light-dark area ratio features between the center and edge regions to construct a total index of appearance deviation. Then, it combines historical process adjustment records and dimensionless process feature sensitivity information to construct a process adjustment reliability index. A square-form basic mapping is used in conjunction with a power exponent controlled by the reliability of process adjustment to generate an adjustment coefficient that varies within the range of zero to one. This guides each process feature parameter to be adjusted in a controllable and directional manner within the preset maximum adjustment range. In this way, the successful experience of each actual adjustment is precipitated into a repeatable quantitative mapping, reducing blind trial and over-adjustment and improving the closed-loop consistency and convergence stability between online detection and feedback control.

[0117] Specifically, the above description is only a preferred embodiment of this application and is not intended to limit this application.

[0118] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0119] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for online detection and feedback control of electroplated layers based on machine learning, characterized in that, Including the following steps: S1. Acquire online observation data related to the electroplating layer formation process on the electroplating production line, and integrate the surface image of the workpiece after electroplating with the process operation record of the corresponding electroplating section according to the workpiece identification to obtain the original observation data. S2 inputs the original observation data into the feature filtering and reconstruction module, and filters out image features reflecting changes in the appearance of the electroplating layer and process features reflecting the strength of the electroplating process according to the preset discrimination criteria. The filtered features are then reorganized and numerically standardized to form a state input vector for electroplating state analysis. S3. Input the state input vector into the electroplating layer state estimation model. The model outputs an electroplating layer state description quantity that represents the overall thickness deviation trend and local anomaly risk, and outputs the sensitivity information of the degree of influence of each process feature on the electroplating layer state description quantity. S4. Based on the electroplating layer state description quantity and sensitivity information, construct two quantitative results from the current workpiece image and historical process adjustment records, respectively, reflecting the degree of appearance deviation and the reliability of process adjustment. Use the reliability of process adjustment to shape the mapping relationship between the degree of appearance deviation and the adjustment range of process parameters, and determine the adjustment coefficient for correcting the adjustable process parameter settings of subsequent workpieces based on the mapping relationship.

2. The online detection and feedback control method for electroplated layers based on machine learning according to claim 1, step S1 includes the following: The workpiece number, entry time, and exit time are recorded to form the time boundary of the electroplating interval. Process operation records within the electroplating interval are collected at fixed time intervals. The time alignment threshold interval is determined based on the distance between the exit position and the image acquisition position and the range of conveying speed, and the electroplated surface images are selected accordingly. The sharpness threshold interval is calculated based on the calibration image, and the electroplated surface images are selected accordingly. The workpiece number, electroplating interval time boundary, process operation records, and electroplated surface images selected by the time alignment threshold interval and sharpness threshold interval are integrated into the original observation dataset.

3. The online detection and feedback control method for electroplated layers based on machine learning according to claim 2, step S2 includes the following: The grayscale reference of the electroplated surface image in the original observation data is adjusted, and three image candidate features are constructed: total texture, texture difference between the central and edge regions, and ratio of bright and dark areas. Three process candidate features are constructed for the electroplating current time series: current integral, high current duration, and current jump amplitude, thus forming a set of image candidate features and a set of process candidate features.

4. The online detection and feedback control method for electroplated layers based on machine learning according to claim 3, step S2 further includes the following: In the calibrated workpiece set, a monotonic consistency index is calculated for each candidate feature in the image candidate feature set and the process candidate feature set based on the reference appearance index. The formal features are selected by constructing a monotonic consistency threshold interval through sorting. The formal features are then subjected to linear numerical normalization based on the minimum and maximum values ​​of the calibrated samples. The normalized formal image features and formal process features are connected in a fixed order to form a state input vector, which is then input into the electroplating layer state estimation model.

5. The online detection and feedback control method for electroplated layers based on machine learning according to claim 4, step S3 includes the following: Thickness measurements are performed at multiple measurement locations on the surface of the workpiece in the training workpiece set. An overall thickness deviation descriptor is constructed based on the average value of the measured thickness. Local risk descriptors are constructed by sorting the thickness deviations and selecting extreme deviation locations according to the local risk proportion threshold. The overall thickness deviation descriptor and the local risk descriptor are combined to form the target value of the electroplating layer state descriptor.

6. The online detection and feedback control method for electroplated layers based on machine learning according to claim 5, step S3 further includes the following: The state input vector and the target value of the electroplating state descriptor are input into the electroplating state estimation model. The model parameters are trained by minimizing the sum of the absolute values ​​of the differences between the overall thickness deviation descriptor and the predicted and target values ​​of the local risk descriptor. The upper and lower bounds of the target value of the electroplating state descriptor are statistically analyzed in the training workpiece set, and the model output is linearly normalized to obtain the dimensionless electroplating state descriptor.

7. The online detection and feedback control method for electroplated layers based on machine learning according to claim 6, step S3 further includes the following: In the state input vector, the set of process feature component indices is determined in a predetermined order. The perturbation amplitude is set according to the value range of the process feature components in the training workpiece set and the perturbation ratio coefficient. Positive and negative perturbations are applied to a single process feature component around the state input vector to obtain two sets of dimensionless electroplating layer state description quantities. Finite difference change rates are constructed for the two sets of output values, and the maximum value of the absolute value of the change rate of the overall thickness deviation description quantity and the local risk description quantity is taken as the original sensitivity value. Then, linear normalization is performed on the upper and lower bounds of the original sensitivity value obtained from training statistics to generate process feature sensitivity information.

8. The online detection and feedback control method for electroplated layers based on machine learning according to claim 7, step S4 includes the following: The morphological deviation index is obtained by taking the maximum value of the dimensionless overall thickness deviation descriptor and the dimensionless local risk descriptor. The image deviation index is obtained by taking the maximum value of the normalized feature of the texture difference between the center and the edge regions and the normalized feature of the ratio of light and dark areas. Finally, the total appearance deviation index is obtained by taking the maximum value of the morphological deviation index and the image deviation index.

9. The online detection and feedback control method for electroplated layers based on machine learning according to claim 8, step S4 further includes the following: In terms of process feature index, pairing records of workpieces before and after adjustment are extracted from historical process adjustment records. Process adjustment amplitude threshold and appearance deviation change threshold are introduced, and significant adjustment events are screened by combining dimensionless process feature sensitivity threshold. Based on the number of successful events and the number of failed events in the significant adjustment events, an initial index of process adjustment reliability is constructed. Then, the smoothing coefficient calculated based on the number of events and the smoothing control constant is used to shrink the initial index of process adjustment reliability to obtain the process adjustment reliability index.

10. The online detection and feedback control method for electroplated layers based on machine learning according to claim 9, step S4 further includes the following: Using the overall index of appearance deviation as the independent variable, a basic adjustment intensity is constructed through squaring. A power exponent parameter is constructed using the process adjustment reliability index and the shape adjustment coefficient. The basic adjustment intensity is then exponentially operated on according to the power exponent parameter to generate an adjustment coefficient between zero and one. Finally, the adjustment coefficient is multiplied by the maximum adjustment range preset for each process feature to determine the adjustment range of the process parameters of the subsequent workpiece on the corresponding process feature.

Images