Electroplating quality prediction method and system based on machine learning

Abstract

The present application relates to the technical field of industrial quality intelligent detection and prediction, in particular to a plating quality prediction method and system based on machine learning, which divides workpiece grid points and converts signals to generate current density space sequences, divides sections by comparing space-time data to generate deposition feature distribution sets, constructs plating layer thickness space distribution data sets in combination with thickness differences before and after, generates thickness change interval sequences by analyzing continuation states, forms associated records by embedding power supply current and plating solution temperature and fills in data, and generates plating layer quality distribution prediction results.The present application constructs a space-time association structure by grid points, associates current density with actual thickness change point by point to express stage characteristics continuously, identifies continuation and turning relationship of thickness evolution by hierarchical organization and forms evolution tracks, fills in time gaps in combination with power supply state and plating solution temperature, improves quality distribution integrity, and provides a basis for process state tracing and local quality trend judgment.

Description

Technical Field

[0001] This invention relates to the field of intelligent detection and prediction technology for industrial quality, and in particular to a method and system for predicting electroplating quality based on machine learning. Background Technology

[0002] The field of intelligent industrial quality detection and prediction technology encompasses product quality monitoring, defect identification, and process optimization during the manufacturing process. Its core content involves the collection, analysis, and modeling of data generated during manufacturing. This includes using sensors to collect data on workpiece surface characteristics, process parameters, and environmental conditions; cleaning and feature extraction of this data; and then using statistical analysis or predictive models to assess product quality and predict trends. The entire technology system comprehensively covers the complete chain from data acquisition and processing to predictive analysis, encompassing production process control, process parameter optimization, and anomaly monitoring, providing intelligent quality management solutions for industrial manufacturing.

[0003] Among them, the electroplating quality prediction method and system based on machine learning refers to the use of machine learning algorithms to model and analyze various process parameters and monitoring data generated during the electroplating process. The main technical aspects covered include the collection and processing of influencing factors such as electroplating solution concentration, current density, temperature, flow rate, and workpiece surface condition. Using time-series data and multi-dimensional features as input, a mapping relationship between electroplating quality and process parameters is established through regression or classification models. Specifically, data preprocessing, feature selection, and model training are employed to complete data input processing and model building, thereby forming a predictive system that can be used to predict the thickness, uniformity, and surface defect probability of the electroplated layer.

[0004] In existing technologies, when the mapping relationship between process parameters and quality results is used as the basis for analysis, the data organization mainly revolves around the feature set. The quality change information formed at different times is prone to a discrete distribution. When there are continuous changes or stage transitions in local areas on the workpiece surface, it is difficult to accurately reflect the continuous evolution law in the quality change process. For example, when the same area gradually thickens or changes from thickening to slowing down at different stages, the relevant change features are easily recorded in a scattered manner, resulting in an unclear quality change path, which in turn affects the tracking of process status, the judgment of quality trends, and the early identification of quality risks in local areas. Summary of the Invention

[0005] To address the technical problems existing in the prior art, embodiments of the present invention provide a method and system for predicting electroplating quality based on machine learning. The technical solution is as follows: A machine learning-based method for predicting electroplating quality includes the following steps: S1: Obtain the geometric dimensions and surface boundary of the surface to be measured, delineate the acquisition area and divide the grid points according to the surface morphology, record the spatial position and point identification, call the probe deployment information to establish the acquisition correspondence, continuously read the current signal and convert and organize it to generate a current density spatial sequence. S2: Based on the current density spatial sequence, establish the connection relationship between adjacent grid points, compare the current density level and spatial change direction at the sampling time, mark the interval category by comparing the interval boundary, determine the change direction and category crossing along the time sequence, divide the continuous segment or new segment, and generate a set of sedimentary feature distributions. S3: Based on the set of deposition feature distributions, determine the point interval category and continuous segment, call the initial thickness and end thickness detection data, compare the actual thickness change results according to the point identifier, and associate and aggregate them with the category segment to generate a coating thickness spatial distribution dataset. S4: Extract the actual thickness variation results of the spatial distribution dataset of the coating thickness, arrange them according to the location and time sequence of the segment, compare the thickness direction and interval category of adjacent segments, merge the continuous intervals or create new intervals, record the start and end segments and the range of thickness variation, and generate a thickness variation interval sequence. S5: Based on the thickness variation interval sequence, extract continuous segments and thickness variation ranges, align grid points according to sampling time and spatial point order, read rectified power supply current and electroplating solution temperature and embed points, match actual thickness variation results and fill in missing time positions to generate coating quality distribution prediction results.

[0006] A machine learning-based electroplating quality prediction system, the system comprising: The grid signal acquisition module acquires the geometric dimensions and boundaries of the surface to be measured, delineates the acquisition area and divides grid points according to the surface morphology, records the spatial location and point identification, calls the probe deployment information to establish the acquisition correspondence, continuously reads the current signal and converts and organizes it to generate a current density spatial sequence. The spatiotemporal segmentation module establishes the connection relationship between adjacent grid points based on the current density spatial sequence, compares the current density levels and spatial change directions at the sampling time, marks the interval categories by comparing the interval boundaries, determines the change direction and category crossing along the time sequence, divides continuous segments or new segments, and generates a set of sedimentary feature distributions. The data difference association module, based on the set of deposition feature distributions, determines the point interval category and continuous segment, calls the initial thickness and end thickness detection data, compares them according to the point identifier to obtain the actual thickness change result, and associates and aggregates them with the category segment to generate a coating thickness spatial distribution dataset. The evolution trend analysis module extracts the actual thickness change results of the spatial distribution dataset of coating thickness, arranges them by point location and segment time sequence, compares the thickness direction and interval category of adjacent segments, merges continuous intervals or creates new intervals, records the start and end segments and the range of thickness change, and generates a thickness change interval sequence. The working condition data fusion module extracts continuous segments and thickness change ranges based on the thickness change interval sequence, aligns grid points according to sampling time and spatial point order, reads rectifier power supply current and electroplating solution temperature and embeds them into the points, matches the actual thickness change results and fills in the missing time positions, and generates coating quality distribution prediction results.

[0007] The beneficial effects of the technical solutions provided by the embodiments of the present invention include at least the following: In this invention, a spatiotemporal correlation data structure covering the workpiece surface is constructed by grid points. The current density change status, interval category information and actual thickness change results are correlated point by point, so that the stage change characteristics of the deposition process can be continuously expressed. Through the hierarchical organization of continuous segments and thickness change intervals, the continuity and transition relationships in the thickness evolution process can be identified, forming a thickness evolution trajectory with temporal correlation characteristics. At the same time, the multi-source correlation is completed by combining the rectifier power supply output status and electroplating solution temperature information, filling the time gap in the thickness change record, improving the completeness and coherence of the quality distribution results, enhancing the ability to identify the quality status of complex surface areas, and providing a basis for process status traceability and local quality trend judgment. Attached Figure Description

[0008] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a flowchart illustrating the acquisition process of S1 in this invention; Figure 3 This is a flowchart illustrating the acquisition process of S2 in this invention; Figure 4 This is a flowchart illustrating the acquisition process of S3 in this invention; Figure 5 This is a flowchart illustrating the acquisition process of S4 in this invention; Figure 6 This is a flowchart of the acquisition process for S5 of the present invention. Detailed Implementation

[0009] The following detailed explanation illustrates the specific implementation methods: The basic implementation examples are as follows: Figure 1 As shown: A machine learning-based method for predicting electroplating quality includes the following steps: S1: Obtain the geometric dimensions and surface boundaries of the electroplated workpiece surface to be measured, determine the area to be collected based on the surface morphology, divide it into multiple grid points, record the spatial position and point identification of each grid point, call the layout information of the preset current density probe, establish the acquisition correspondence based on the positional correspondence between the grid points and the current density probes, read the current signal of the corresponding current density probe according to the acquisition correspondence during continuous sampling, convert it into current density data of each grid point, and associate and organize the current density data according to the sampling time and spatial point order to form a record that has both temporal continuity and spatial correspondence, and generate a current density spatial sequence. Please see Figure 2 Specifically, S1 includes: S11: Obtain the geometric dimensions and surface boundaries of the electroplated workpiece's surface to be measured. Determine the area to be collected based on the surface morphology. Divide the area into grid points, determine whether each grid point falls within the surface boundary range, and write the grid points that fall within the boundary range into their spatial location and point identifier to generate a grid point index; details are as follows: To obtain the geometric dimensions and surface boundaries of the electroplated workpiece, optical mapping of the workpiece surface is performed using an industrial high-precision 3D laser scanner, or geometric parameters are directly extracted from the digitally designed CAD model file to obtain the length. mm and width mm, combined with the coordinates of the four corner vertices of the surface boundary (0,0), (400,0), (400,250), (0,250), a two-dimensional closed region is defined in a plane. This two-dimensional closed region is used as the coverage area of ​​the area to be collected, and the grid division step size is set. mm, the number of grid points in the length direction was calculated to be 81 and the number of grid points in the width direction was 51 through equal-interval division. The axial intersection points were determined as the coordinates of the grid points to be verified. subscript The index is an incrementing integer from 0 to 80 for each grid point along the length direction. The grid points in the width direction are indexed incrementally, with values ​​ranging from 0 to 50. Coordinates are extracted, and a ray-crossing polygon inclusion algorithm is used to check if they lie within the actual workpiece contour to adapt to irregularly shaped workpieces. For the standard rectangular boundary in this example, the x-coordinates of the coordinate points are extracted. It compares its value with the boundary minimum of 0 and the maximum of 400, and extracts the ordinate. And compare it with the boundary minimum value 0 and the maximum value 250, when it satisfies and Furthermore, when determining the location using the polygon interior algorithm, it is confirmed that the grid point is inside the surface boundary. If the determination condition is not met, it means that the grid point falls into the redundant edge blank area outside the workpiece and is removed without being stored. The valid point coordinates that pass the determination are written into the non-volatile storage register of the space. A unique point identifier code is assigned to each spatial position. For example, the point with coordinates (10,15) is assigned the identifier code P_002_003 for physical positioning. All the identifier codes and the corresponding coordinate values ​​are stored in the static memory in the form of key-value pairs to form a structured spatial topology index table and generate the grid point index.

[0010] S12: Retrieve the grid point index and preset current density probe deployment information. Based on the correspondence between the spatial location of the grid points and the deployment location of the current density probes, determine the matching status between the grid point identifier and the current density probe number, write the matching status into the point acquisition sequence, and establish a point-probe correspondence table; details are as follows: Based on the generated grid point index, the spatial coordinates (10, 15) corresponding to the identifier P_002_003 are extracted. Simultaneously, the preset current density probe deployment information is retrieved from the factory-installed configuration file. The current density probe is a microelectrode array sensor, specifically... The array consists of 100 square, 99.99% pure platinum microelectrodes arranged in a matrix, each with a side length of 1 mm and a spacing of 0.2 mm, embedded in a 0.1 mm thick flexible polyimide printed circuit board substrate. Each probe is pre-tested and inspected before leaving the factory. Two-point calibration was performed in a copper sulfate reference plating solution to ensure response consistency. The probe features an array of 4131 independent miniature beryllium copper spring probes on its back, with a silicone magnetic adsorption layer integrated at the substrate edge to ensure conformal fit with complex workpiece surfaces and maintain a fixed normal gap of 0.5 mm. The effective stimulated surface area of ​​a single probe is [missing information]. Standardized as The total area of ​​the 100 platinum electrodes directly exposed to the electroplating solution is strictly defined. The flexible circuit board is rigidly attached to the electroplating fixture via positioning holes machined by a high-precision CNC machine tool. Combined with magnetic adsorption from the substrate, this ensures precise alignment of the entire deployment position with the grid points, allowing for the retrieval of the probe's physical coordinates. And the probe hardware number S_01, set the matching tolerance radius threshold. mm, this threshold is set based on the center position of the probe sensing end face combined with the machining tolerance range of the automated fixing fixture, and the Euclidean geometric two-dimensional spacing between the grid points and the probe placement position is calculated. By substituting actual production environment data for calculation, if the grid point coordinates are (10.0, 15.0) and the probe's physical coordinates are (10.5, 14.8), the calculation yields... mm, comparison spacing Matching tolerance threshold ,determination If the connection is established, mark the matching status of grid point P_002_003 and probe S_01 as a valid connection code. If the spacing... If the probe is found to have mechanically deviated, an abnormal error code is written to the status bit and an alarm is triggered. Probes that meet the valid connection conditions are sorted according to the rule of increasing the horizontal coordinate first and then the vertical coordinate. The sorted hardware addresses are mapped to the acquisition task buffer, and the mapping relationship entries are recorded in the storage matrix to establish a point probe correspondence table.

[0011] S13: Based on the probe mapping table, read the current signal from the corresponding current density probe during continuous sampling, perform current density data conversion on the current signal, associate and organize the current density data according to the sampling time and spatial point order, determine whether the sampling time corresponds to the point identifier, and generate a current density spatial sequence; as detailed below. Based on the point probe mapping table, and utilizing the peripheral circuitry of a multi-channel, high-precision 24-bit analog-to-digital converter chip with integrated high-speed multiplexing scanning switch, the system reads the weak milliampere-level current signal transmitted from the corresponding current density probe during continuous sampling according to a preset time-division multiplexing scanning strategy. The sampling clock period is set to... ms, the time of probe S_01 is obtained through the hardware physical interface of the analog-to-digital converter. The original current value mA, call current density conversion factor ,in The effective stimulated surface area of ​​the probe's sensing electrode is given by a value of . Calculate the current current density value And uniformly converted into standard dimensions. The converted current density data, along with the sampling timestamp 20260603_100001 generated by the high-precision hardware clock source and the corresponding point identifier P_002_003, are encapsulated into a structure and stored in a temporary data frame structure. The data frames are arranged linearly in chronological order. At the same time, all point identifiers are traversed at each sampling node, and the preset point reference value is called. Compare the data with the actual number of points in the current data frame. If the number equals 4131, the data is considered complete. If the number does not equal 4131, it indicates that there is partial packet loss in the hardware transmission. This triggers an abnormal alarm and uses the valid data from the previous moment to perform zero-order hold-and-fill. The point identifiers in the data frame are extracted and the physical coordinates are retrieved. The current density values ​​are constructed into a spatiotemporal two-dimensional dynamic matrix according to the time and spatial dimensions. The consistency of the timestamp of each element is verified, and mismatched redundant entries are removed to generate a spatial sequence of current density.

[0012] S2: Based on the current density data in the current density spatial sequence, establish the connection relationship between adjacent grid points according to the spatial point order, compare the current density data of adjacent grid points at the same sampling time, record the current density high-low relationship and spatial change direction, compare the current density data of each grid point with the preset current density interval boundary, mark the interval category at the corresponding sampling time, continuously compare the current density data and interval category of the same grid point along the sampling time sequence, divide the time period with the same current density change direction and the interval category not crossed into continuous segments, divide the time period with the current density change direction changed or the interval category crossed into new segments, organize the interval category and segment results according to the time sequence and spatial point order, and generate a set of sedimentary feature distributions; Please see Figure 3 Specifically, S2 includes: S21: Based on the current density data in the current density spatial sequence, establish the connection relationship between adjacent grid points according to the spatial point order. Compare the current density data of adjacent grid points at the same sampling time, determine the current density sorting status and spatial change direction based on the sequential relationship of the point readings, write the judgment entries into the point connection order, and generate the adjacent direction code; as detailed below: Based on the current density data in the current density spatial sequence, the corresponding sampling time is extracted from the spatiotemporal two-dimensional dynamic matrix. The grid point coordinates (10, 15) with identifier P_002_003 and its measured current density value are as follows. Synchronously retrieve the current density measurements of spatially adjacent identifiers P_003_003 at the same timestamp. The adjacency relationship is based on the grid step size. For mm determination, compare the indexes of two points. If the absolute value of the difference between the horizontal indices is 1 and the absolute value of the difference between the vertical indices is 0, then they are determined to be horizontally first-order adjacency. Calculate the difference. ; Set the current density variation sensitivity threshold This threshold is used to filter electromagnetic interference baseline noise and is combined with the inherent resolution of the probe hardware. The three-fold redundancy ensures that only the true electric field gradient change is captured. The difference of 0.07 is compared with the threshold of 0.03. And with the direction being positive, assign an incrementing status identifier code DIR_01 to the spatial change direction from point P_002_003 to P_003_003. If the absolute value of the difference is... This indicates that the local electric field is extremely stable and is assigned a stability identifier code DIR_00. If the difference is negative and the absolute value is... This indicates a significant attenuation and assigns a decreasing identifier code DIR_02. The code is calculated and compared for all 4131 grid points in the entire domain. The increasing, stable, or decreasing identifier codes are written into the one-dimensional state vector in the order of the points. The comparison results are associated with the corresponding point connection topology table through bit operations to generate the adjacency direction code.

[0013] S22: Call the adjacent direction code and the preset current density interval boundary, compare the current density data of each grid point with the interval boundary arrangement, determine the interval category based on the sampling time, point identifier, and boundary assignment status, and write the category entry into the corresponding sampling time in point order to obtain the interval category code; specifically as follows: By invoking the adjacent direction code and the preset current density interval boundary, the experimentally determined boundary values ​​of the three-level electrochemical-physical interval are extracted from the parameter storage area. These are the low-current scorching boundary values. Medium current standard deposition boundary value And the boundary value for high current hydrogen evolution and coking: These boundary values ​​were obtained through standard Hull cell calibration. The specific standardization steps were as follows: using a 267 ml standard Hull cell... A standard brass sample, its surface polished with 800-grit sandpaper and degreased by cathodic electrolysis, was subjected to continuous constant current electrodeposition for 5 minutes with a total current of 2A applied by a constant DC power supply. The sample was then observed under a 200x industrial optical microscope and a 3D white light interferometer. The distance from the high-current end to the physical boundary point where the coating coverage was less than 90% was measured. Substitute into Haring's empirical formula ( The total current is 2A. The mapping is calculated for a distance of cm, and the boundary value of low current plating leakage is objectively extracted. Simultaneously, an interferometer is used to measure the surface roughness. The region boundary distance is accompanied by a microscopic dark brown dendritic crystal hydrogen evolution morphology. Measure and substitute the values ​​into the above formula to extract the boundary value for high-current hydrogen evolution and coking. ; The region between the two corresponds to the current standard deposition boundary value. Scope of coverage, invocation time Current density at point P_002_003 Perform nested conditional judgments to determine if 1.55 falls within the closed interval. And it falls into the semi-open interval If the condition is true, the interval is classified as a second-class standard electrodeposition region and assigned the identifier CAT_02. If it falls into... Then assign the weak current defect area identifier CAT_01. If it falls into Then, the high current oversaturation region identifier CAT_03 is assigned, and the current density of the adjacent point P_003_003 is extracted simultaneously. The duplicate determination belongs to CAT_02. For all grid points at each time, the category identifiers are concatenated in ascending order of the points to form a string sequence, which is then embedded with the current timestamp 20260603_100001. The result is a composite mapping matrix generated by stacking the data vertically along the time axis to obtain the interval category code.

[0014] S23: Based on the interval category code and adjacent direction code, the current density data and interval category of the same grid point are continuously compared along the sampling time sequence to determine the continuity state of the change direction and the interval crossing state. The continuous time period is assigned to the continuous segment, and the crossing time period is assigned to a new segment. The data are then organized according to the time sequence and spatial point order to generate a set of sedimentary feature distributions; as detailed below: Based on the interval category code and the adjacent direction code, the grid point P_002_003 in the continuous sampling sequence is extracted. to The timing state parameters in the data, retrieve the time. The interval category code CAT_02 and the adjacent direction code DIR_01, and the subsequent time... The interval category code CAT_02 and the adjacent direction code DIR_01 are compared to determine the category attribute of the grid point at two adjacent time points. and Since both category codes are CAT_02 and equal, the category interval is determined to be uncrossed. Next, the direction codes are determined to be consistent. Because both change directions are DIR_01, the thickness change direction is determined to be in a valid continuation state. and Record the data within the same continuous segment and initialize the segment duration timer. If fluctuations in current density data cause the segment category code to change to CAT_03, or the adjacent direction code to change to DIR_02, it indicates that a local electrodeposition non-steady-state shift has occurred, breaking the original steady state. If the category code change or direction code change is valid, trigger the segment truncation command. to Store it as the first continuous segment SEG_01 and record the start time 10:00:01 and end time 10:00:02, while also recording the state abrupt change. The starting anchor point is marked as the new segment SEG_02. The entire sampling time is traversed in this manner. An independent temporal segmented linked list is established for each grid point. The linked list nodes are mapped with spatial coordinates, categories, and directions in a high dimension to form a spatiotemporal distribution grid and generate a set of sedimentary feature distributions.

[0015] S3: Based on the interval categories and segment results in the deposition feature distribution set, determine the interval category and continuous segment to which each grid point belongs at the corresponding sampling time according to the time order and spatial point order. Call the initial thickness detection data and electroplating end thickness detection data of each grid point, and establish the correspondence between the thickness detection data and the grid point according to the point identifier. Compare the initial thickness and electroplating end thickness of the same grid point to obtain the actual thickness change result. Associate and organize the actual thickness change result with the interval category and continuous segment of the corresponding grid point. Collect the actual thickness change results of each grid point according to the spatial point order to generate a coating thickness spatial distribution dataset. Please see Figure 4 Specifically, S3 includes: S31: Based on the interval categories and segment results in the sedimentary feature distribution set, determine the interval category and continuous segment to which each grid point belongs at the corresponding sampling time according to the temporal and spatial point order. Arrange the point identifier, sampling time, interval category, and continuous segment accordingly, determine the segment continuity status of the point identifier in adjacent sampling times, and write the determination entry into the sample feature items to generate point segment identifiers; the details are as follows: Based on the interval categories and segment results in the sedimentary feature distribution set, the identifier P_002_003 is retrieved from the sampling sequence. to Full feature data in the data, extraction time The corresponding interval category code CAT_02 and the continuous segment number SEG_01 are used. Simultaneously, the interval category code CAT_02 and the segment number SEG_01 of adjacent time points are extracted. The segment numbers at two time points are compared. Since SEG_01 equals SEG_01, the segment is determined to be validly continued and a status code ST_1 is assigned. If time... If the segment number is changed to SEG_02, it means that the previous physical segment has ended and is determined to be a breakpoint and is assigned the status code ST_0. The point identifier, time, category, continuous segment and status code are arranged in a matrix column vector alignment. Establish a feature vector for each sampling node Where P_ID is the point identifier code, For timestamp, CAT is the interval category code, SEG is the continuous segment number, ST is the status code, and ST is the preset feature item integrity check value. Count the dimensions of the constructed feature vectors and match them with the dimensions of the feature vectors. The comparison is performed. If the result is true, it is written to the sample feature register. If it is not equal to 5, it is determined that there is a missing feature and it is removed as dirty data. The entire domain of 4131 points is traversed and alignment and judgment are performed at each time step. The generated feature vectors are arranged in time major order and point minor order to form the underlying data architecture and generate point segment identifiers.

[0016] S32: Call the point segment identifier, and call the initial thickness detection data and electroplating end thickness detection data of each grid point. Establish the correspondence between the thickness detection data and the grid points according to the point identifier. Arrange the initial thickness detection points and the end thickness detection points accordingly. Determine if the point identifier is consistent in the two sets of thickness detection data. Compare the initial thickness and the end thickness of the same grid point to obtain the actual thickness change result; as detailed below: The system retrieves the point segment identifiers and extracts the initial thickness detection data for each grid point from the online database of the high-precision industrial X-ray fluorescence thickness gauge. Electroplating end thickness detection data Using the identifier P_002_003 as the index key, the initial measurement value of 10.2 μm and the final measurement value of 25.5 μm were extracted. The initial and final thickness detection point lists were retrieved, and the point identifiers in the two lists were confirmed to be consistent through string matching. The actual thickness change was then calculated. ,in This represents the actual change. To conclude the measured thickness, The initial measured thickness is used as the basis for calculation. μm, setting a closed interval for the physical rationality of thickness variation μm, this range is estimated based on the product of the maximum allowable electroplating time and the maximum electrochemical deposition rate of the plating solution. The determination is made based on whether the actual thickness change of 15.3 falls within this range. Once a valid thickness result is established and confirmed, if it exceeds this range, the thickness gauge is deemed malfunctioning or there is a mechanical scratch defect on the workpiece surface. An error placeholder code of -999 is assigned, and manual re-inspection is triggered. The above operation is repeated for all grid points. The values ​​are stored in a floating-point array indexed by the point identifier to ensure that the physical dimension increment is accurately quantified throughout the process, thus obtaining the actual thickness change result.

[0017] S33: Based on the actual thickness variation results and the point segment identifiers, the actual thickness variation results are associated and organized with the corresponding grid point interval categories and continuous segments. The actual thickness variation results for each grid point are collected according to the spatial point order. The continuity status of thickness variation entries in the point order is determined, and continuous entries are written into the sample label field to generate a coating thickness spatial distribution dataset; specifically as follows: Based on the actual thickness change results and the point segment identifiers, the actual thickness change result of 15.3 μm corresponding to point identifier P_002_003 is extracted, and its full feature information in the point segment identifier is retrieved, including the interval category sequence {CAT_02,CAT_02,CAT_03…} and the continuous segment sequence {SEG_01,SEG_01,SEG_02…}. 15.3 μm is used as the target value and appended to the end of all time series feature vectors for the corresponding point to establish a mapping relationship. The total vector, The input vector is the time-series feature. To label the thickness variation, the entire data record is aggregated in ascending order of point index. Spatial continuity checks are performed on the thickness variation entries to verify the physical continuity of the electric field distribution. The thickness variation results of 15.3 μm and 15.8 μm corresponding to spatially adjacent points P_002_003 and P_003_003 are extracted, and the absolute value of the difference is calculated. μm, call the preset spatial thickness gradient threshold μm, this threshold is set based on the continuous distribution characteristics of edge effects, for determination If the difference is established and the current entry is determined to be a continuous entry with normal thickness distribution, then the absolute value of the difference is... If a physical mutation or coating peeling defect occurs, an abnormal code ERR_1 is added. The data items that pass the verification are written into the sample label field and the matrix is ​​reorganized according to the machine learning tensor format. The grid coordinates, time series, current features, segment attributes and thickness labels are encapsulated into a unified data storage block to generate a coating thickness spatial distribution dataset.

[0018] S4: Extract the actual thickness variation results from the spatial distribution dataset of coating thickness. Taking each grid point as the processing object, arrange the thickness variation correlation results corresponding to each grid point according to the time sequence of continuous segments. Compare the thickness variation direction and interval category corresponding to adjacent continuous segments, and record the continuity status of thickness variation between adjacent continuous segments. When the thickness variation direction is consistent and the interval category has not been crossed, the next continuous segment is merged into the current thickness variation interval. When any of the following occurs, such as the change in thickness variation direction or the crossing of interval category, the current thickness variation interval is ended, and a new thickness variation interval is established with the next continuous segment as the starting point. Record the starting and ending continuous segments and the thickness variation range of the current thickness variation interval and the new thickness variation interval respectively, and generate a thickness variation interval sequence. Please see Figure 5 Specifically, S4 includes: S41: Extract the actual thickness variation results from the spatial distribution dataset of coating thickness. Taking each grid point as the processing object, arrange the thickness variation correlation results corresponding to each grid point according to the time sequence of continuous segments. Perform sequential verification based on point identifier, continuous segment, thickness change direction, and interval category to determine the continuity status of thickness change between adjacent continuous segments and generate segment continuity identifiers; as detailed below: The actual thickness variation results in the spatial distribution dataset of coating thickness are extracted, and the total thickness increment of 15.3 μm corresponding to point P_002_003 is obtained. The associated continuous segment index list is called, and the first segment SEG_01 and its start and end times 10:00:01 and 10:00:10 are extracted by temporal sorting. The second segment SEG_02 and its start and end times 10:00:11 and 10:00:20 are extracted. The orientation status bit DIR_01 and category attribute code CAT_02 recorded in SEG_01 are retrieved, and the orientation status bit DIR_01 and category attribute code CAT_02 recorded in SEG_02 are retrieved simultaneously. The absolute value of the difference between the two orientation status bits is calculated. The difference of 0 is determined to be equal to the preset consistency logic threshold. And if true, then the absolute value of the difference between the two category attribute codes is calculated. The class difference is determined again to be equal to If true, perform a logical AND operation on the two judgment results. If and only if the absolute values ​​of the differences between the two physical attributes are both 0, mark the continuation state of the adjacent segment as a binary valid continuation code 1 in the status register. If either difference is not 0, it indicates a reversal of direction or a jump in category, and is marked as a binary interrupt code 0, representing a change in electrochemical steady state. Bind the generated binary code to the adjacent segment combination {SEG_01, SEG_02} and insert a continuation judgment entry into the time sequence log table to generate a segment continuation identifier.

[0019] S42: Based on the segment continuation identifier, compare the thickness change direction and interval category of adjacent consecutive segments to determine whether the thickness change direction is consistent or the interval category is crossed. When the thickness change direction is consistent and the interval category is not crossed, merge the next consecutive segment into the current thickness change interval. When either state changes, the thickness change interval is obtained; as follows: Based on the segment continuation identifier, the thickness change direction and interval category of adjacent consecutive segments are compared. The continuation relationship binary code of point P_002_003 is retrieved from the segment relationship mapping table and the value 1 is extracted. It is determined that the current state is a valid continuation state with the same thickness change direction and no interval category crossing. The termination time of the next segment SEG_02, 10:00:20, is taken as the termination time of the current merged interval, and the start time of SEG_01, 10:00:01, is determined as the starting boundary of the thickness change interval. If the binary code of the next set of adjacent segments {SEG_02, SEG_03} is 0, the truncation logic is triggered. It is determined that at least one of the following is true: either the direction changes or the category crosses. The category code CAT_03 of SEG_03 is extracted and compared with the category code CAT_02 of SEG_02 to determine the continuation relationship. If true, immediately terminate the interval extension and define 10:00:01 to 10:00:20 as a complete thickness change interval INT_01. The standard for determining the consistency of the thickness change direction is that the absolute value of the direction code difference falls into the closed interval [0,0]. The standard for determining the interval category crossing is that the absolute value of the category code difference falls into the closed interval [1,3]. The time-series segmentation of all 4131 points is cyclically traversed and merged for determination. When the continuity relationship is 0, it represents the physical boundary and blocks excessive merging. The continuous time period that meets the conditions is encapsulated into a uniquely numbered spatial-temporal block to obtain the thickness change interval.

[0020] S43: Based on the thickness variation interval and segment continuation identifier, organize the current thickness variation interval and the newly created thickness variation interval in sequence, record the starting continuous segment, the ending continuous segment, and the thickness variation range, determine the arrangement status of adjacent thickness variation intervals, write the arrangement entries into the sample interval label, and generate a thickness variation interval sequence; as detailed below: Based on the thickness variation interval and segment continuation identifier, the current thickness variation interval and the newly created thickness variation interval are sequentially organized, and the starting continuous segment, ending continuous segment, and thickness variation range are recorded. The interval sequence generated by point P_002_003 is retrieved, and the first interval INT_01 is recorded as starting with SEG_01 and ending with SEG_02. The current density of all sampling points within the interval is extracted. The thickness variation range is obtained by calling the thickness distribution formula. The formula is: ,in The physical increment of thickness within the desired interval. The total thickness increment over the entire process is 15.3 μm, numerator term For the measured values ​​of the time series iteration variables within the interval denominator For the measured values ​​corresponding to the full-duration time series iterative variables , The step size is 1 second. Let be the total number of sampling points. The fractional term represents the percentage of the integral of the electrical quantity within the interval. Substituting the data into the calculation, if the percentage of the integral of the current INT_01 is 0.4, then the thickness increment of that interval is calculated as follows: μm, record the starting value of the thickness range of this interval as 0 and the ending value as 6.12. For the next interval INT_02, record the starting value as 6.12 and the ending value as 15.3. Retrieve the adjacent interval identifier codes to verify the timing continuity, calculate the difference between the end time of INT_01 and the start time of INT_02, and set the arrangement benchmark difference. s, compare the time difference with ,like s is determined to be a seamless, continuous arrangement and written to the tag LB_01. The 's' indicates that there is a stop breakpoint, which is determined to be a non-continuous arrangement and written to the label LB_00. The interval features and labels of all grid points are assembled to generate a thickness variation interval sequence.

[0021] S5: Based on the current thickness change interval and the newly created thickness change interval in the thickness change interval sequence, extract the corresponding continuous segments and thickness change ranges, align them with the grid points according to the sampling time and spatial point order, read the rectifier power supply output current and electroplating solution temperature during continuous sampling, organize them into rectifier power supply current records and electroplating solution temperature records, embed them with the grid points under the corresponding sampling time, form a quality prediction association record, call the actual thickness change results and match them with the quality prediction association record according to the continuous segments and thickness change ranges, fill in the missing sampling time positions for the matched thickness change results according to the sampling time order, and generate the coating quality distribution prediction results.

[0022] Please see Figure 6 Specifically, in S5, this includes: S51: Based on the current thickness change interval and the newly created thickness change interval in the thickness change interval sequence, extract the corresponding continuous segments and thickness change ranges. Align them with grid points according to the sampling time and spatial point order. Perform sequential judgment on continuous segments, thickness change ranges, and point identifiers, and write the alignment entries into the sample time series field to generate a segment point index; the details are as follows: Based on the current and newly created thickness change intervals in the thickness change interval sequence, extract the corresponding continuous segments and thickness change ranges. Align them with the grid points according to the sampling time and spatial point order, and extract the first interval INT_01 and the second interval INT_02 corresponding to the identifier code P_002_003. Retrieve the start and end segments SEG_01 and SEG_02 associated with INT_01 and extract the thickness range. μm, retrieve the clock recording nodes from 10:00:01 to 10:00:20 to align the sampling time and point markers, retrieve the current coordinates (10,15) and the thickness increment step value, and set the single-cycle step increment. The process is to divide the total range of 6.12 μm by the total number of samplings of 20 to obtain the result. μm / s, concatenate the cumulative thickness value per second with the corresponding segment number SEG_01 string, verify the point order, extract the current processing point number and the subsequent point number P_002_004 to calculate the spatial index difference, determine whether the difference is equal to the adjacent reference value 1, if the determination is true, it is confirmed as a continuous spatial sequence, if it is not equal to 1, it is determined as a topological break and triggers the discontinuity flag to jump to the starting point of the next region, integrate the spatial position, timestamp, segment attributes and stage thickness increment of each time into the waiting queue, determine whether the time sequence field is increasing, calculate the difference between adjacent times and determine whether it is always equal to the period 1s, if it is true, write it into the buffer, traverse all nodes from P_000_000 to P_080_050 of the entire workpiece, align the feature structure of the nodes under different segments to form a point index matrix, and generate the segment point index.

[0023] S52: Based on the segment point index, the rectifier power supply output current and electroplating solution temperature are read during continuous sampling, and organized into rectifier power supply current records and electroplating solution temperature records. Correspondence is determined based on sampling time and grid point, and the current and temperature records are embedded into the corresponding grid points at the sampling time to form a quality prediction correlation record; specifically as follows: Based on the segment point index, the rectifier power supply output current and electroplating solution temperature are read during continuous sampling. The DC output measurement value at time 10:00:05 is retrieved through the rectifier cabinet bus digital interface. A. The measured values ​​of the electroplating solution temperature at the same moment were retrieved using Pt100 platinum resistance thermometers around the inner wall of the tank and their programmable controller. ℃, extract the record rows corresponding to 10:00:05 and point P_002_003 from the segment point index and verify the sampling time alignment, determine the relationship between the sensor's native acquisition frequency of 10Hz and the refresh frequency of 1Hz, downsample the 10Hz raw data, take one value every 9 points to extract valid points that conform to the whole second of 1Hz, and set the temperature reference value. ℃ and fluctuation tolerance threshold ℃, this threshold is set based on the solubility of the solute and the diffusion isothermal tolerance, and the temperature deviation value is calculated. ℃, judge If the physical state is determined to be within the normal isothermal range, and the absolute value of the deviation is... The temperature of ℃ indicates a heating failure or thermal shock, and is classified as a fluctuating range, with an additional disturbance penalty factor applied. The rated reference value of the rectified current is then extracted simultaneously. A; And calculate the instantaneous relative volatility. Compare volatility with safety and stability coefficient determination Established, the current measurement value, temperature measurement value, and corresponding instantaneous current density are combined. The section identifier SEG_01 and spatial coordinate features are fused and spliced ​​to form a comprehensive feature vector, which is then embedded into the corresponding grid point storage space in a linear order of sampling time, cyclically covering the entire time period and points to form a quality prediction association record.

[0024] S53: Retrieve the quality prediction association records and actual thickness change results, match them according to continuous segments and thickness change ranges, judge the matching thickness change results and sampling time order, add missing sampling time positions to the thickness change entries of adjacent segments, organize the prediction entries according to spatial point order, and generate coating quality distribution prediction results; details are as follows: The quality prediction association records and actual thickness change results are retrieved and matched according to continuous segments and thickness change ranges. The actual thickness change label value of 15.3μm corresponding to grid point P_002_003 is extracted and determined as the target output term of the final value of supervised learning regression fitting. This label value is matched with the high-dimensional quality prediction association records in the time series, and the matching thickness range of 6.12μm in SEG_01 is extracted. Continuity logic is performed on the sampling time nodes of the prediction sequence to determine whether there are missing time points due to network jitter or sensor packet loss. If the 10:00:08 data item is missing between 10:00:07 and 10:00:09, mathematical interpolation prediction is calculated, and the cumulative thickness value at 10:00:07 is extracted. μm and cumulative value at 10:00:09 μm, call the mean linear compensation formula ,in For a moment Missing thickness padding value, and The cumulative value of adjacent valid sampling times before and after the missing data is calculated. μm, the interpolation prediction results are correlated with the missing positions to fill in the missing parts to form a complete prediction sequence, the prediction records from P_000_000 to P_080_050 are arranged in topological order, and the physical threshold for the qualified coating thickness is set. μm, this threshold is set with reference to the minimum uniform deposition thickness of the corrosion protection design life, comparing the predicted thickness with... When the final predicted thickness is 15.3 μm When the coating thickness reaches μm, the coating quality is considered to be within the acceptable range. If the thickness is μm, local thinning or missing plating defects are marked as unqualified areas and their coordinates are exported to the automatic re-plating production line. The global results are transformed into a high spatiotemporal resolution color mapping matrix to generate coating quality distribution prediction results.

[0025] A machine learning-based electroplating quality prediction system, comprising: The grid signal acquisition module acquires the geometric dimensions and boundaries of the surface to be measured, delineates the acquisition area and divides grid points according to the surface morphology, records the spatial location and point identification, calls the probe deployment information to establish the acquisition correspondence, continuously reads the current signal and converts and organizes it to generate a current density spatial sequence. The spatiotemporal segmentation module establishes the connection relationship between adjacent grid points based on the spatial sequence of current density, compares the current density levels and spatial change directions at the sampling time, marks the interval categories by comparing the interval boundaries, determines the change direction and category crossing along the time sequence, divides continuous segments or new segments, and generates a set of sedimentary feature distributions. The data difference correlation module, based on the set of deposition feature distribution, determines the category and continuous segment of the point interval, calls the initial thickness and end thickness detection data, compares them according to the point identifier to obtain the actual thickness change result, and associates and aggregates them with the category segment to generate a spatial distribution dataset of coating thickness. The evolution trend analysis module extracts the actual thickness change results of the spatial distribution dataset of coating thickness, arranges them by point location and segment time sequence, compares the thickness direction and interval category of adjacent segments, merges continuous intervals or creates new intervals, records the start and end segments and the range of thickness change, and generates a thickness change interval sequence. The working condition data fusion module extracts continuous segments and thickness change ranges based on the thickness change interval sequence, aligns grid points according to sampling time and spatial point order, reads rectifier power supply current and electroplating solution temperature and embeds them into the points, matches the actual thickness change results and fills in the missing time positions, and generates coating quality distribution prediction results.

[0026] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A machine learning-based method for predicting electroplating quality, characterized in that, Includes the following steps: S1: Obtain the geometric dimensions and surface boundary of the surface to be measured, delineate the acquisition area and divide the grid points according to the surface morphology, record the spatial position and point identification, call the probe deployment information to establish the acquisition correspondence, continuously read the current signal and convert and organize it to generate a current density spatial sequence. S2: Based on the current density spatial sequence, establish the connection relationship between adjacent grid points, compare the current density level and spatial change direction at the sampling time, mark the interval category by comparing the interval boundary, determine the change direction and category crossing along the time sequence, divide the continuous segment or new segment, and generate a set of sedimentary feature distributions. S3: Based on the set of deposition feature distributions, determine the point interval category and continuous segment, call the initial thickness and end thickness detection data, compare the actual thickness change results according to the point identifier, and associate and aggregate them with the category segment to generate a coating thickness spatial distribution dataset. S4: Extract the actual thickness variation results of the spatial distribution dataset of the coating thickness, arrange them according to the location and time sequence of the segment, compare the thickness direction and interval category of adjacent segments, merge the continuous intervals or create new intervals, record the start and end segments and the range of thickness variation, and generate a thickness variation interval sequence.

2. The electroplating quality prediction method based on machine learning according to claim 1, characterized in that: The current density spatial sequence includes the spatial location of each grid point, the sampling time sequence, and the current density value. The deposition feature distribution set includes the interval category of the grid point, the continuous segment identifier, and the current density change direction. The coating thickness spatial distribution dataset includes the initial thickness, the ending thickness, and the actual thickness change of each grid point. The thickness change interval sequence includes the start and end continuous segments of the thickness change interval, the thickness change range, and the continuity status of the thickness change direction.

3. The electroplating quality prediction method based on machine learning according to claim 1, characterized in that: The steps for obtaining S1 are as follows: S11: Obtain the geometric dimensions and surface boundary of the electroplated workpiece surface to be measured, determine the area to be collected based on the surface morphology, divide the area to be collected into grid points, determine whether the grid points fall within the surface boundary range, write the grid points that fall within the boundary range into the spatial location and point identifier, and generate the grid point index. S12: Call the grid point index and the preset current density probe deployment information, determine the matching status between the grid point identifier and the current density probe number according to the correspondence between the spatial position of the grid point and the deployment position of the current density probe, write the matching status into the point acquisition sequence, and establish a point probe correspondence table. S13: Based on the point probe correspondence table, read the current signal of the corresponding current density probe during continuous sampling, perform current density data conversion on the current signal, associate and organize the current density data according to the sampling time and spatial point order, determine whether the sampling time and point identifier correspond, and generate a current density spatial sequence.

4. The electroplating quality prediction method based on machine learning according to claim 1, characterized in that: The steps for obtaining S2 are as follows: S21: Based on the current density data in the current density spatial sequence, establish the connection relationship between adjacent grid points according to the spatial point order, compare the current density data of adjacent grid points at the same sampling time, determine the current density sorting status and spatial change direction based on the order of point readings, write the judgment entry into the point connection order, and generate the adjacent direction code. S22: Call the adjacent direction code and the preset current density interval boundary, compare the current density data of each grid point with the interval boundary arrangement, determine the interval category based on the sampling time, point identifier and boundary belonging status, and write the category entry into the corresponding sampling time along the point order to obtain the interval category code; S23: Based on the interval category code and the adjacent direction code, continuously compare the current density data and interval category of the same grid point along the sampling time sequence, determine the continuity state of the change direction and the interval crossing state, classify the continuous time period into the continuous segment, and list the crossing time period into a new segment. Organize according to the time sequence and spatial point order to generate a set of sedimentary feature distributions.

5. The electroplating quality prediction method based on machine learning according to claim 1, characterized in that: The steps for obtaining S3 are as follows: S31: Based on the interval category and segment results in the sedimentation feature distribution set, determine the interval category and continuous segment to which each grid point belongs at the corresponding sampling time according to the time order and spatial point order. Arrange the point identifier, sampling time, interval category, and continuous segment accordingly, determine the segment continuity status of the point identifier in adjacent sampling times, write the judgment entry into the sample feature item, and generate the point segment identifier. S32: Call the point segment identifier, and call the initial thickness detection data and electroplating end thickness detection data of each grid point. Establish the correspondence between the thickness detection data and the grid points according to the point identifier. Arrange the initial thickness detection points and the end thickness detection points accordingly. Determine whether the point identifier is consistent in the two thickness detection data. Compare the initial thickness and the end thickness of the same grid point to obtain the actual thickness change result. S33: Based on the actual thickness change results and the point segment identifiers, associate and organize the actual thickness change results with the corresponding grid point interval categories and continuous segments, collect the actual thickness change results of each grid point according to the spatial point order, determine the continuous status of the thickness change entries in the point order, write the continuous entries into the sample label field, and generate a coating thickness spatial distribution dataset.

6. The electroplating quality prediction method based on machine learning according to claim 1, characterized in that: The steps for obtaining S4 are as follows: S41: Extract the actual thickness change results from the spatial distribution dataset of the coating thickness, take each grid point as the processing object, arrange the thickness change association results of each grid point in the order of continuous segment time, check the point identifier, continuous segment, thickness change direction and interval category in sequence, determine the thickness change continuity status between adjacent continuous segments, and generate segment continuity identifier. S42: Based on the segment continuation identifier, compare the thickness change direction and interval category of adjacent continuous segments to determine the state of consistent thickness change direction and the state of interval category crossing. When the thickness change direction is consistent and the interval category is not crossed, merge the next continuous segment into the current thickness change interval. When any state changes, the thickness change interval is obtained. S43: Based on the thickness change interval and the segment continuation identifier, sort the current thickness change interval and the newly created thickness change interval in order, record the starting continuous segment, the ending continuous segment and the thickness change range, determine the arrangement status of adjacent thickness change intervals, write the arrangement entries into the sample interval label, and generate a thickness change interval sequence.

7. The electroplating quality prediction method based on machine learning according to claim 1, characterized in that: The steps for obtaining the result are as follows: The method further includes: S5: Based on the thickness change interval sequence, extract continuous segments and thickness change ranges, align grid points according to sampling time and spatial point order, read rectified power supply current and electroplating solution temperature and embed points, match actual thickness change results and fill in missing time positions to generate coating quality distribution prediction results. The coating quality distribution prediction results include quality prediction association records, thickness change results after missing time completion, and the influence of electroplating solution temperature and current.

8. The electroplating quality prediction method based on machine learning according to claim 7, characterized in that: The steps for obtaining S5 are as follows: S51: Based on the current thickness change interval and the newly created thickness change interval in the thickness change interval sequence, extract the corresponding continuous segment and thickness change range, align them with the grid points according to the sampling time and spatial point order, perform sequential judgment on the continuous segment, thickness change range and point identifier, write the alignment entries into the sample time sequence field, and generate the segment point index. S52: Based on the segment point index, the output current of the rectifier power supply and the temperature of the electroplating solution are read during continuous sampling, and organized into rectifier power supply current record and electroplating solution temperature record. Corresponding judgment is made for the sampling time and grid point, and the current record and temperature record are embedded into the grid point under the corresponding sampling time to form a quality prediction association record. S53: Call the quality prediction association record and the actual thickness change result, match them according to continuous segments and thickness change range, judge the matching thickness change result and sampling time order, add the missing sampling time position to the thickness change entry of the adjacent segment, organize the prediction entries according to the spatial point order, and generate the coating quality distribution prediction result.

9. A machine learning-based electroplating quality prediction system, characterized in that, The system is used in the machine learning-based electroplating quality prediction method according to any one of claims 1-8, the system comprising: The grid signal acquisition module acquires the geometric dimensions and boundaries of the surface to be measured, delineates the acquisition area and divides grid points according to the surface morphology, records the spatial location and point identification, calls the probe deployment information to establish the acquisition correspondence, continuously reads the current signal and converts and organizes it to generate a current density spatial sequence. The spatiotemporal segmentation module establishes the connection relationship between adjacent grid points based on the current density spatial sequence, compares the current density levels and spatial change directions at the sampling time, marks the interval categories by comparing the interval boundaries, determines the change direction and category crossing along the time sequence, divides continuous segments or new segments, and generates a set of sedimentary feature distributions. The data difference association module, based on the set of deposition feature distributions, determines the point interval category and continuous segment, calls the initial thickness and end thickness detection data, compares them according to the point identifier to obtain the actual thickness change result, and associates and aggregates them with the category segment to generate a coating thickness spatial distribution dataset. The evolution trend analysis module extracts the actual thickness change results of the spatial distribution dataset of coating thickness, arranges them by point location and segment time sequence, compares the thickness direction and interval category of adjacent segments, merges continuous intervals or creates new intervals, records the start and end segments and the range of thickness change, and generates a thickness change interval sequence. The working condition data fusion module extracts continuous segments and thickness change ranges based on the thickness change interval sequence, aligns grid points according to sampling time and spatial point order, reads rectifier power supply current and electroplating solution temperature and embeds them into the points, matches the actual thickness change results and fills in the missing time positions, and generates coating quality distribution prediction results.

Images