A method for controlling the thickness of a plating layer of an IC lead frame continuous plating strip

Abstract

The application relates to an IC lead frame continuous electroplating strip material coating thickness regulation method in the information technology field, which comprises the following steps: obtaining a target processing piece surface image from an image acquisition device, adopting a neural network model to perform edge feature extraction on the target processing piece surface image to obtain a surface texture feature matrix; adopting a time sequence network model to perform time sequence feature analysis on the local current density sequence to obtain a multi-region collaborative compensation current value; generating a control instruction for a channel covered by the local distortion region coordinates through the multi-region collaborative compensation current value to obtain a multi-channel current regulation instruction set; and adjusting output current corresponding to the local distortion region coordinates according to the multi-channel current regulation instruction set, reacquiring secondary scanning data from the thickness detection device to obtain updated thickness profile data.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to a method for controlling the thickness of the strip plating layer in continuous electroplating of IC lead frames. Background Technology

[0002] In the field of semiconductor packaging material manufacturing, the roll-to-roll continuous electroplating process for integrated circuit lead frames is a core step that determines the chip's soldering and corrosion resistance. The uniformity of the plating thickness directly affects the yield and reliability of the final product, and is a key technological bottleneck that urgently needs to be overcome in this field.

[0003] Current production lines largely rely on offline sampling and destructive measurements at the end of the production line to assess quality. This approach has an extremely long feedback cycle and cannot respond to real-time thickness changes caused by anode consumption, fluctuations in chemical composition, or slight strip misalignment during production. Furthermore, traditional online thickness measurement equipment often uses fixed-position local scanning, making it difficult to quickly obtain the true thickness distribution of different functional areas on the complex pattern of the lead frame.

[0004] In a high-speed electroplating bath, the dense pins and large pads on the lead frame create complex patterns that induce strong edge and shielding effects, resulting in highly uneven electric field distribution across the strip width. This spatial electric field imbalance directly causes significant differences in local plating thickness. For example, when the pin area at the strip edge has an excessively thick plating due to electric field concentration, the central pad area may still be too thin. Traditional control methods often only offer crude adjustments to the total current or overall strip speed to address this local thickness imbalance. Increasing the total current to increase the thickness of the central pads further exacerbates over-plating of the edge pins, potentially leading to serious defects such as burning. This global adjustment approach is completely inadequate to address the thickness unevenness caused by local electric field distortion, easily resulting in a control dilemma where addressing one aspect at the expense of others.

[0005] Therefore, how to perceive the thickness profile of different functional areas in the width direction of the strip in real time during high-speed continuous production, and to implement multi-regional fine-grained collaborative compensation for local electric field distortion, has become a key issue in overcoming the bottleneck of roll-to-roll electroplating thickness uniformity. Summary of the Invention

[0006] This invention provides a method for controlling the thickness of the strip plating in continuous electroplating of IC lead frames, mainly including:

[0007] A surface image of the target workpiece is acquired from an image acquisition device. An edge feature extraction is performed on the surface image using a neural network model to obtain a surface texture feature matrix. Based on the marked positions in the surface texture feature matrix, initial electric field acquisition data is acquired at corresponding positions within an electrochemical processing device. This initial electric field acquisition data is then matrix-mapped to obtain a spatial electric field distribution matrix. Scanning sampling points are located using the field strength distribution of each grid in the spatial electric field distribution matrix. Initial thickness contour data for these scanning sampling points is acquired from a thickness detection device. This initial thickness contour data is then denoised to obtain a denoised coating thickness contour sequence. For the coating thickness contour sequence, if the difference between the maximum and minimum thickness values ​​in the sequence exceeds a preset threshold... If the difference value is found, the continuous sampling point segment where the difference is located is determined to be a local distortion region, and the coordinates of the local distortion region are obtained. Based on the coordinates of the local distortion region, real-time sensing current acquisition data is obtained from the current detection device, and the real-time sensing current acquisition data is normalized to obtain a local current density sequence with a time dimension. A time-series network model is used to perform time-series feature analysis on the local current density sequence to obtain multi-region collaborative compensation current values. Based on the multi-region collaborative compensation current values, control commands are generated for the channels covered by the coordinates of the local distortion region to obtain a multi-channel current adjustment command set. Based on the multi-channel current adjustment command set, the output current corresponding to the coordinates of the local distortion region is adjusted, and secondary scan data is obtained again from the thickness detection device to obtain updated thickness contour data. Furthermore, the step of extracting edge features from the surface image of the target workpiece using a neural network model to obtain a surface texture feature matrix includes: performing pixel-level conversion processing on the surface image of the target workpiece using a grayscale conversion algorithm to obtain a grayscale image of the target workpiece; performing multi-scale edge feature extraction on the grayscale image of the target workpiece using the neural network model to obtain an edge feature map of the target workpiece; calculating the pixel gradient value of each pixel based on the edge feature map of the target workpiece, and if the pixel gradient value is greater than a preset gradient threshold, determining that the region where the pixel corresponding to the pixel gradient value is located is a surface defect region; extracting local texture pixels for the surface defect region, and constructing a local texture matrix based on the local texture pixels according to a preset arrangement rule; and performing matrix dimensionality reduction and feature preservation processing on the local texture matrix using a singular value decomposition algorithm to obtain the surface texture feature matrix.Furthermore, the step of performing matrix mapping on the initial electric field acquisition data to obtain a spatial electric field distribution matrix includes: dividing the initial electric field acquisition data into a two-dimensional plane grid coordinate system according to preset array coordinates and width coordinates to obtain a discrete electric field coordinate point set; identifying missing nodes in the discrete electric field coordinate point set that do not contain acquisition data, and using a bilinear interpolation algorithm to perform numerical filling and matrix mapping processing on the missing nodes to obtain an initial spatial electric field distribution matrix; obtaining a preset electrode spacing parameter, and calculating the potential difference between adjacent grid nodes based on the initial spatial electric field distribution matrix and the electrode spacing parameter to determine the electric field gradient matrix; if the edge node values ​​in the electric field gradient matrix are greater than a preset gradient threshold, then using a preset compensation coefficient to perform numerical correction on the edge nodes to complete edge distortion compensation, thereby obtaining the spatial electric field distribution matrix. Furthermore, the step of denoising and filtering the initial thickness contour data to obtain a denoised coating thickness contour sequence includes: if there are edge distortion data in the initial thickness contour data that exceed a preset distortion threshold, then the edge distortion data is removed using a data cleaning algorithm to obtain first thickness contour data; a sliding window denoising and filtering algorithm is used on the first thickness contour data, and high-frequency noise interference in the first thickness contour data is eliminated by setting the step size and window size of the sliding window to obtain denoised second thickness contour data; the second thickness contour data is serialized and spliced ​​according to the position order of the scanning sampling points to construct continuous thickness distribution features and determine the denoised coating thickness contour sequence. Furthermore, if the difference between the maximum and minimum thickness values ​​in the coating thickness contour sequence is greater than a preset threshold, then determining that the continuous sampling point segment where the difference is located is a local distortion region and obtaining the coordinates of the local distortion region includes: classifying the coating thickness contour sequence using a support vector machine to identify abnormal contour sequences that deviate from the normal thickness range; extracting the maximum and minimum thickness values ​​in the abnormal contour sequences; if the difference between the maximum and minimum thickness values ​​in the abnormal contour sequences is greater than a preset thickness difference threshold, then determining that the continuous sampling point segment between the sampling point corresponding to the maximum thickness value and the sampling point corresponding to the minimum thickness value is a local distortion region; and mapping the local distortion region to obtain the coordinates of the local distortion region.Furthermore, the normalization processing of the real-time sensed current acquisition data to obtain a local current density sequence with a time dimension includes: determining the corresponding channel in the array of the electrochemical processing equipment based on the coordinates of the local distortion region; acquiring the real-time sensed current acquisition data and the corresponding timestamp through the current detection device of the corresponding channel; if the real-time sensed current acquisition data exceeds a preset current threshold, then normalizing the real-time sensed current acquisition data in combination with the channel identifier of the corresponding channel to eliminate the dimensional differences between different channels and obtain the local current density; and generating the local current density sequence with a time dimension by matching and combining the local current density with the timestamp using a time series splicing method. Furthermore, the step of using a time-series network model to perform time-series feature analysis on the local current density sequence to obtain multi-region collaborative compensation current values ​​includes: acquiring local current density sequences of multiple monitoring points within the electrochemical treatment device; using the time-series network model to extract features from the local current density sequences to obtain time-series feature vectors; evaluating the degree of electrode corrosion between adjacent channels based on the current decay trend in the time-series feature vectors; determining the impedance change rate between adjacent channels based on the degree of electrode corrosion; quantifying the impedance deviation gradient between adjacent channels using the impedance change rate; constructing a coupling relationship matrix between adjacent channels based on the impedance deviation gradient; determining multi-region compensation benchmark values ​​based on the coupling relationship matrix; allocating the compensation benchmark values ​​to regions based on the regional location information of the electrochemical treatment device; adjusting the output of the power supply circuit based on the results of the region allocation to obtain the multi-region collaborative compensation current values. Furthermore, determining the compensation benchmark value for multiple regions based on the coupling relationship matrix includes: if an element in the coupling relationship matrix is ​​greater than a preset matrix element threshold, then the element greater than the preset matrix element threshold is determined as a target element; the corresponding channel weight is determined based on the position coordinates of the target element in the coupling relationship matrix, where the channel weight characterizes the degree of influence of different channels on the overall current distribution; a weighted calculation is performed by combining the channel weight with the timing feature vector to fuse the timing feature information and spatial coupling information of multiple channels; global compensation features are extracted based on the weighted calculation result, and the global compensation features are numerically transformed through a preset benchmark mapping function; the compensation benchmark value for multiple regions is determined based on the numerical transformation result.Furthermore, the step of generating control commands for the channels covered by the coordinates of the local distorted regions using the multi-region collaborative compensation current values ​​to obtain a multi-channel current adjustment command set includes: determining the spacing parameter and conductivity ratio based on the coordinates of the local distorted regions; obtaining the target density by dividing the multi-region collaborative compensation current values ​​by the spacing parameter and multiplying by the conductivity ratio; processing the target density and the coordinates of the local distorted regions using a preset channel mapping model to determine the corresponding target channel; obtaining the deposition thickness corresponding to the target channel, and if the deposition thickness is less than a preset thickness threshold, obtaining a preset edge compensation value; obtaining the target current density by adding the edge compensation value to the target density; determining the voltage amplitude and period duration based on the target current density, and obtaining the duty cycle value by dividing the voltage amplitude by the period duration; generating corresponding control commands based on the duty cycle value and the target channel, traversing all coordinates of the local distorted regions to obtain multiple control commands, and combining multiple control commands to obtain the multi-channel current adjustment command set. Furthermore, the step of adjusting the output current corresponding to the coordinates of the local distortion region according to the multi-channel current adjustment instruction set, and obtaining the updated thickness contour data from the thickness detection device again, includes: parsing the set of distortion coordinates corresponding to the coordinates of the local distortion region, obtaining the output current at each coordinate in the set of distortion coordinates, and constructing an initial current distribution matrix; performing state estimation on the initial current distribution matrix using a Kalman filter algorithm to obtain a filtered current distribution matrix; calculating the deposition rate at each coordinate in the set of distortion coordinates by combining a preset electrode spacing and ion concentration; calculating the current adjustment amount based on the deposition rate, and adjusting the output current using the current adjustment amount; obtaining the secondary scan data of the surface of the target workpiece from the thickness detection device, and generating the updated thickness contour data based on the secondary scan data; calculating the thickness deviation between the updated thickness contour data and the preset target thickness; if the thickness deviation exceeds a preset deviation threshold, extracting edge effect features from the updated thickness contour data, calculating a thickness compensation coefficient based on the edge effect features and the target thickness, and determining the final coating thickness based on the thickness compensation coefficient.

[0008] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects:

[0009] This invention discloses an intelligent control method for strip coating thickness in a continuous electroplating production line. Addressing the challenge of localized distortion in strip coating thickness caused by uneven electric field distribution and anode current fluctuations during continuous electroplating production, making real-time and accurate compensation difficult, this invention first acquires images of the strip surface from an industrial camera and extracts an edge texture feature matrix using a convolutional neural network. Based on the strip location segments marked in the matrix, it obtains initial electric field data within the electroplating tank and maps this data to a spatial electric field distribution matrix using the grid coordinates of the anode array and the strip width. This allows for the location of scanning sampling points along the strip width direction. The method then uses laser measurement... The thickness sensor acquires the thickness profile and uses a sliding window for noise reduction. The coordinates of the local distortion area are determined based on the maximum and minimum thickness difference. Then, the real-time current data of the Hall sensor of the corresponding anode channel is retrieved and normalized into a local current density sequence with a time dimension. The timing characteristics are analyzed by combining the long short-term memory network with the current coupling relationship of adjacent channels. Multi-region collaborative compensation current values ​​and multi-channel anode control commands are generated to accurately adjust the anode output current in the distortion area and perform secondary scanning verification. This achieves closed-loop linkage control of electric field, current and thickness, which significantly improves the uniformity of coating thickness and production line stability. Attached Figure Description

[0010] Figure 1 This is a flowchart of a method for controlling the thickness of the strip plating in continuous electroplating of IC leadframes according to the present invention. Detailed Implementation

[0011] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. 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.

[0012] like Figure 1 This embodiment of a method for controlling the thickness of the strip plating in continuous electroplating of IC leadframes may specifically include:

[0013] Step S101: Obtain a strip surface image from an industrial camera on the continuous electroplating production line. First, perform grayscale conversion on the strip surface image to obtain a strip grayscale image. Then, use a convolutional neural network to extract edge features from the strip grayscale image to obtain a strip surface texture feature matrix M. The grayscale conversion uses a weighted method, combining three-channel pixel values ​​into a single-channel grayscale value using R*0.299+G*0.587+B*0.114 to reduce color interference caused by coating reflection. The convolutional neural network contains three convolutional layers and two pooling layers, with a 3*3 kernel size and a stride of 1. The number of output channels is 16, 32, and 64 respectively. The last layer of feature map is flattened and normalized to obtain a 64*64 texture feature matrix M. This matrix M serves as the input to the defect area identification and localization module in subsequent step S102, and is used to compare it with a pre-stored standard strip texture template to determine the type and location of surface defects such as scratches and pits.

[0014] Based on the strip grayscale image obtained after preprocessing, edge features are extracted using a convolutional neural network to obtain a strip edge feature map. On this feature map, the Sobel operator is used to calculate the gradient components Gx and Gy of each pixel along the horizontal and vertical directions, respectively, and the pixel gradient value is synthesized according to G=sqrt(Gx*Gx+Gy*Gy), where G is the synthesized gradient magnitude. When G is greater than a preset threshold T, the region where the pixel is located is determined as a defect region. T is adaptively determined based on the overall gradient mean and standard deviation of the strip surface, generally ranging from 40 to 80; in this embodiment, it is set to 60.

[0015] For each defective region, based on the minimum bounding rectangle of its connected components, pixel grayscale values ​​are extracted block by block using a 16*16 sliding window as local texture pixels. The pixels within each window are expanded row by row and stacked sequentially to construct a local texture matrix of size N*256, where N is the number of windows. Singular value decomposition is performed on this matrix, and the top k components are retained after sorting the singular values ​​from largest to smallest. k is taken as the minimum value that makes the cumulative energy proportion not less than 95%. In this embodiment, k is 8, finally resulting in an N*8 strip surface texture feature matrix.

[0016] The surface texture feature matrix of the strip will be used as input for subsequent defect classification and grade assessment steps, and will be compared with the preset defect template library to achieve defect type identification and severity determination.

[0017] In one implementation, an image of the strip surface is first acquired from an industrial camera on a continuous electroplating production line.

[0018] Specifically, during high-speed operation of a continuous electroplating production line, the surface of the strip may develop minute defects due to fluctuations in electroplating process parameters or mechanical friction. To capture these surface conditions in real time, a high-resolution industrial camera is installed along the transport direction of the strip. This industrial camera continuously captures images of the passing electroplated strip at a preset sampling frequency, thereby acquiring surface images containing information about the strip's surface morphology. These images accurately reflect the uniformity of the electroplated layer and the presence of any abnormal textures. Furthermore, after acquiring the strip surface images, to reduce the interference of changes in ambient lighting on image quality, basic grayscale processing and filtering denoising can be performed on the strip surface images to highlight structural information in the images, providing a clearer data foundation for subsequent feature extraction.

[0019] For example, a convolutional neural network is used to extract edge features from the strip surface image.

[0020] It should be noted that a convolutional neural network (CNN) is a deep learning model specifically designed for processing data with a grid-like structure. Internally, it contains multiple alternating convolutional and pooling layers. In this embodiment, edge feature extraction refers to using the convolutional kernels in the network to perform sliding window calculations on the image, capturing locations where pixel grayscale values ​​change drastically. In the scenario of continuous electroplating of strip materials, a normal electroplated surface is usually smooth and has uniform grayscale, exhibiting a high degree of consistency. However, if the electroplating solution concentration is abnormal or the conductive rollers have poor contact, defects such as scratches, pinholes, incomplete plating, or plating peeling will appear on the strip surface. The edge regions of these defects will exhibit obvious grayscale steps under optical imaging, i.e., abrupt changes in pixel values ​​from bright to dark or from dark to bright. CNNs, through pre-trained weight parameters on a large number of strip samples, can adaptively identify and extract these edge features representing defects or special textures, without requiring manual design of complex feature extraction operators.

[0021] In one possible implementation, the convolutional layers of a convolutional neural network contain multiple sets of convolutional kernels of different sizes.

[0022] For example, a smaller convolutional kernel is used to capture the edges of tiny pits on the strip surface, while a larger kernel is used to extract longer scratches or watermarks. During the convolution operation, the convolutional kernel slides across the pixel matrix of the strip surface image with a set stride. The weights of the convolutional kernel are multiplied and added to the pixel values ​​at the corresponding positions in the image, and then processed by a non-linear activation function to generate a feature map containing preliminary edge information.

[0023] Understandably, to reduce the dimensionality of the data and preserve the most significant edge features, convolutional neural networks typically connect pooling layers after convolutional layers. Pooling layers, by sampling local maxima or averages of the feature maps, not only reduce the number of parameters required for subsequent computations but also enhance the translation invariance of edge features to minor offsets or deformations of the strip on the production line.

[0024] Specifically, for the specific business processes of continuous electroplating production lines, the extraction of surface texture features of strips needs to overcome interference factors such as the high reflectivity of metal surfaces and imaging blurring caused by production line vibrations. In the deep network structure, convolutional neural networks perform complex combinations and abstractions of the simple edge features extracted from the shallow layers.

[0025] For example, shallow networks might only extract isolated short line segments, corner points, or tiny patch edges, while deep networks, with their larger receptive field, can combine these local line segments into texture patterns with specific orientations, distribution densities, and periodicity. Suppose that during electroplating, the adhesion of anode mud causes continuous stripe-like defects on the strip surface; a deep convolutional network can recognize this complex texture structure composed of numerous interwoven parallel edges. Alternatively, when the electroplating solution temperature is too high, causing the coating to crystallize coarsely and form an orange-peel texture, the network can capture this irregular but densely distributed set of tiny edges and encode it as a high-dimensional feature representation. This layer-by-layer abstraction and fusion process, from simple local edges to complex global textures, ensures a comprehensive and in-depth analysis of the true physical state of the strip surface. Based on the above steps, after multi-layer processing by the convolutional neural network, the final strip surface texture feature matrix is ​​obtained. This strip surface texture feature matrix is ​​a multi-dimensional numerical matrix, where each element corresponds to the texture intensity and edge direction information of a local area on the strip surface. This matrix not only condenses the key morphological features in the original strip surface image and filters out redundant background information, but also transforms it into a structured data form that computers can process efficiently.

[0026] Preferably, by using the above-mentioned method of obtaining the surface texture feature matrix of the strip through a convolutional neural network, an objective and quantitative expression of the surface state of the continuously electroplated strip can be achieved, avoiding the problems of subjective differences and low efficiency caused by traditional manual visual inspection, and providing reliable data support for subsequent quality assessment or defect classification.

[0027] Step S102: Based on the strip position segment marked in the strip surface texture feature matrix, obtain the initial electric field acquisition data of the corresponding position in the electroplating tank, perform matrix mapping on the initial electric field acquisition data according to the grid coordinates of the anode array and the strip width, calculate the electric field gradient matrix on this basis, further generate the steady-state electric field matrix, and obtain the spatial electric field distribution matrix.

[0028] Based on the strip location segments marked in the strip surface texture feature matrix, initial electric field acquisition data is read from the corresponding positions within the electroplating tank. Using a preset anode array coordinate as the horizontal axis and the strip width coordinate as the vertical axis, the acquired data is projected onto a two-dimensional plane and divided into a grid, forming a discrete electric field coordinate point set. For missing nodes not covered by the sensor in the point set, a bilinear interpolation algorithm is used to fill in the missing nodes with the electric field values ​​of the four surrounding valid nodes. The filled results are then mapped to the corresponding positions in the matrix according to the grid row and column indices, yielding the spatial electric field distribution matrix E.

[0029] Obtain the preset electrode spacing parameter d, which is 25 mm. Calculate the potential difference by dividing the difference in electric field values ​​of adjacent grid nodes by d, i.e., G = (E(i,j) - E(i+1,j)) / d, where E(i,j) represents the electric field value of the node in the i-th row and j-th column of the spatial electric field distribution matrix, and G is the electric field gradient at the corresponding position. Assemble the G values ​​of all adjacent nodes into an electric field gradient matrix according to the original grid arrangement.

[0030] The nodes located in the first and last columns of the grid along the width of the strip are marked as edge nodes. The preset gradient threshold is 0.8 volts per millimeter, and the preset compensation coefficient k is 0.85. When the G value of the edge node is greater than 0.8, the node value is updated to G multiplied by k using a multiplicative correction method to offset the distortion caused by the divergence of electric field lines at the edge of the tank. If it is less than or equal to the threshold, the original value is retained. The corrected matrix is ​​the steady-state electric field matrix.

[0031] In one implementation, in order to achieve precise perception and control of the electric field distribution during the electroplating process, it is first necessary to determine the specific spatial position of the strip in the electroplating tank.

[0032] Specifically, the system extracts marked strip location segments from the strip surface texture feature matrix constructed in the previous steps. These strip location segments reflect the real-time physical coordinate range of specific textured regions within the electroplating bath during continuous operation, thus providing an accurate physical reference for subsequent electric field data spatial matching. Based on the determined strip location segments, the system further acquires initial electric field data at the corresponding locations within the electroplating bath.

[0033] For example, multiple electric field sensor arrays, each encapsulated in a PTFE shell and containing a titanium-based platinum-plated electrode, are pre-deployed on a non-conductive support on the outer wall of the electroplating tank and the back side of the anode plate. The sensor probes penetrate the electrolyte area through ceramic seals, exposing only a 2mm diameter sensing surface. The shells are resistant to temperatures up to 80°C and to strong acid and alkali environments with pH values ​​from 0 to 14. The insulation resistance is no less than 1000 megohms, and the rated withstand voltage is no less than 1500 volts. The sensors can operate stably under anode current densities of 5 to 15 amps per square decimeter. The sensor arrangement follows the principle of maintaining a distance of at least 30mm from the edge of the anode plate and a 50mm spacing between array units to minimize disturbance to the original electric field distribution. The distortion rate of the local electric field by a single probe is controlled to within 2%. Each sensor undergoes zero-point and gain calibration using a standard electric field source before being put into use, and drift compensation is performed at 10-second intervals during operation. When a specific section of the strip passes through the sensor coverage area, the strip position encoder synchronously triggers a data acquisition command to record the original physical quantities such as electric field strength and potential difference in that area at the current moment using a sampling rate of 1 kHz. This set of values, fed back in real time by discrete sensor nodes, is transmitted to the host computer via fiber optic isolation, thus constituting the initial electric field acquisition data.

[0034] It should be noted that due to the fixed installation location of the sensors and the dynamic movement of the strip, these initial data are often irregular and sparse in spatial distribution, and cannot be directly used for subsequent refined process analysis. In order to transform the irregular initial electric field acquisition data into a structured spatial data model, a unified coordinate reference system needs to be introduced.

[0035] It is understandable that the electric field distribution in the electroplating tank is mainly affected by the arrangement of the anode plates and the geometric dimensions of the strip itself.

[0036] In one possible implementation, the system constructs a grid coordinate system based on the anode array and the strip width. Specifically, using the physical boundary of the electroplating tank as a reference, the space within the tank is defined as longitudinal along the strip's running direction and transverse along the direction perpendicular to the strip's running direction, and then meshed separately. The granularity of the transverse division is strictly set quantitatively with reference to the strip width W, with the width of a single transverse grid being W / 20, meaning 20 equidistant grids are evenly divided within the strip's coverage area. Simultaneously, a range of W*0.1 extends from each side of the strip as a stray current region, which is also divided with a granularity of W / 20, with two additional grids added to each side, resulting in a final total of 24 transverse grids. The longitudinal division is based on the physical arrangement of the anode array, mapping the centerline of each independent anode plate to a longitudinal node, and mapping the centerline of the gap between adjacent anode plates to another longitudinal node, thus ensuring that the number of longitudinal grid nodes equals the number of anode plates plus the number of gaps. Through this intersecting gridding process, a two-dimensional grid coordinate system is established that accurately reflects the relative positions of the core conductive components within the electroplating tank. The row indices of the matrix correspond to the grid along the horizontal direction (strip width), and the column indices correspond to the grid along the vertical direction (strip running direction). Furthermore, the initial electric field acquisition data is matrix-mapped according to the aforementioned grid coordinates. Since the initial electric field acquisition data originates from discrete sensor nodes, their physical coordinates often cannot perfectly coincide with the nodes of the grid coordinates. Therefore, the core of matrix mapping lies in spatial interpolation and coordinate alignment.

[0037] For example, the system first extracts the spatial coordinates of each initial electric field acquisition data point and projects them into a pre-established grid coordinate system. Then, it uses an inverse distance weighted interpolation algorithm to calculate the theoretical electric field value at each grid node.

[0038] Specifically, for any given grid node, the system searches for multiple initial electric field data points within its spatial neighborhood. Weights are assigned to these data points based on their physical distance to the grid node, with closer points receiving larger weights. This weighted average is then used to calculate the electric field intensity value for that grid node. This calculation is performed by traversing all grid nodes, mapping and filling the originally disordered discrete data points into regular grid cells. After this matrix mapping process, a spatial electric field distribution matrix is ​​finally obtained. The rows and columns of this matrix correspond to the vertical and horizontal grid indices in the grid coordinate system, respectively, and each element represents the quantized electric field intensity within the corresponding spatial grid cell. This structured matrix not only intuitively presents the electric field gradient changes between the anode array and the strip within the electroplating tank but also effectively eliminates data blind spots caused by uneven sensor distribution.

[0039] Preferably, in a real-world continuous electroplating production line scenario for metal strips, considering the potential slight vibrations or deviations that may occur during high-speed operation of the strip, the marking of the strip position segments and the acquisition of initial electric field data need to maintain a high-frequency synchronous refresh rate. By continuously updating the spatial electric field distribution matrix, the system can capture in real time the local electric field distortion caused by strip position deviation, thereby ensuring the uniformity and quality stability of the final electroplated layer thickness.

[0040] Step S103: Locate the scanning sampling point in the width direction of the strip by the field strength distribution of each grid in the spatial electric field distribution matrix, obtain the initial thickness contour data of the sampling point from the laser thickness sensor, and perform sliding window noise reduction filtering on the initial thickness contour data to obtain the noise-reduced coating thickness contour sequence.

[0041] A spatial electric field distribution matrix is ​​obtained. Based on the field strength distribution of each grid in the spatial electric field distribution matrix, scanning sampling points along the width direction of the strip are located. Initial thickness contour data of the scanning sampling points is obtained from a laser thickness sensor. If edge distortion data exceeding a preset distortion threshold exists in the initial thickness contour data, the edge distortion data is removed to obtain first thickness contour data. A moving average algorithm is used to perform sliding window noise reduction filtering on the first thickness contour data to obtain denoised second thickness contour data. The second thickness contour data is serialized and stitched together according to the position order of the scanning sampling points to determine the denoised coating thickness contour sequence.

[0042] For example, in a continuous hot-dip galvanizing or electroplating production line for strip, the strip passes through a specific detection area during high-speed operation. The spatial electric field distribution matrix is ​​a dataset characterizing the spatial distribution of electric field intensity within this detection area. This matrix divides the physical space of the strip's movement into multiple two-dimensional virtual grids, each corresponding to a specific spatial coordinate and recording the real-time electric field intensity value at that coordinate. By acquiring and parsing this matrix, the system can non-contactly sense the strip's real-time position and lateral attitude changes in space, thereby providing a precise physical position reference for subsequent thickness measurements.

[0043] Specifically, locating the scanning sampling points along the width of the strip by analyzing the field strength distribution of each grid in the spatial electric field distribution matrix is ​​a crucial step in achieving high-precision measurement. In actual strip coating scenarios, the sensor node array deployed above and below the production line serves both as the excitation and emission function of the high-frequency electric field and as the acquisition function of distorted signals. When the strip, as a conductor, enters this space, it causes distortion of the original uniform electric field distribution. The spatial electric field distribution matrix is ​​constructed based on the normalized initial electric field data acquired by the aforementioned sensor nodes, with the rows and columns corresponding to the horizontal and vertical coordinates of the physical space. Due to the distribution characteristics of charge on the conductor surface, the edge region of the strip will produce a significant edge effect, causing a sharp nonlinear change in the field strength value of the grid at the edge, while the field strength distribution in the central region of the strip is relatively gentle. The system first iterates through the field strength values ​​of each grid along the row direction of the matrix, i.e., the width direction of the strip, and calculates the spatial gradient of the field strength between adjacent grids, G=(E2-E1) / d, where E1 and E2 are the field strength values ​​of adjacent grids, and d is the physical spacing between the grids. When the absolute value of the gradient of three or more consecutive grids exceeds eight times the average gradient of the central stable region, and the gradient sign changes from positive to negative or from negative to positive within this consecutive interval, the physical coordinates corresponding to these grids can be accurately determined as the left and right edge positions of the strip. In this embodiment, the edge distortion threshold is set to 0.15 volts per millimeter. After determining the positions of the two edges, the system calculates the actual effective width of the strip in the current operating state through the coordinate difference. Subsequently, within the effective width range, the system divides the entire width direction into multiple equally spaced measurement intervals according to the preset measurement resolution requirements. To avoid the influence of severe edge fluctuations on the measurement, the system also offsets inward by a safety margin of 2% of the bandwidth and not less than 5 millimeters. Finally, the center coordinates of each measurement interval are mapped to specific scanning sampling points. For the adaptive density adjustment strategy, the variance of the electric field strength σ² within the sliding window is used as the criterion. When σ² is less than 0.02, it is determined to be a region with small fluctuations, and sampling is performed at the base density. When σ² is greater than 0.08, it is determined to be a region with significant increases, and the sampling density is increased to 2 to 3 times the base value. This non-contact positioning method based on the distribution of electric field strength can effectively overcome the positioning errors caused by lateral deviation, serpentine movement, or edge damage of the strip during high-speed operation, ensuring that each sampling point can always accurately fall within the effective coating area of ​​the strip.

[0044] In one possible implementation, to further improve the measurement adaptability under complex working conditions, the positioning of scanning sampling points can also adopt an adaptive density adjustment strategy.

[0045] It should be noted that the flatness of the strip surface or the uniformity of the coating thickness is often not absolutely consistent, and some areas may have local thickness anomalies. Since the spatial electric field distribution within the electroplating tank is directly related to the charge deposition rate on the strip surface, and the deposition rate determines the local coating growth state, there is a one-to-one coordinate mapping relationship between the projection area of ​​the electric field distribution matrix along the strip width direction and the downstream laser thickness measurement detection area. The system aligns the data in each column of the electric field matrix along the strip width direction with the thickness measurement scanning coordinates using a pre-calibrated geometric transformation matrix, and then calculates the variance of the field strength fluctuation within the effective width region in real time. If the variance of a certain width segment is less than 0.05, the deposition process in that segment is considered stable, and a lower density of scanning sampling points is assigned, with a sampling interval of 5 mm. If the variance exceeds 0.2, it is determined that the corresponding strip position in that segment may experience warping or abnormal coating accumulation, and the sampling interval is automatically shortened to 1 mm. If the variance is between these two values, the interval is determined by linear interpolation. This dynamic adjustment mechanism, based on the causal mapping from electric field to thickness, accurately captures minute local thickness variations while ensuring overall measurement efficiency. Based on the scanning sampling point coordinates determined in the above steps, the system controls the scanning probe of the laser thickness sensor to move to the corresponding position, or directly aligns it with these coordinates using a multi-probe array, acquiring the initial thickness profile data of the sampling points from the laser thickness sensor. The laser thickness sensor emits a laser beam and receives reflected light, using the triangulation principle to measure the distance between the sensor probe and the upper and lower surfaces of the strip, thereby calculating the total thickness value including the substrate and coating. These thickness values, arranged sequentially along the width direction, constitute the initial thickness profile data reflecting the transverse cross-sectional morphology of the strip. Furthermore, due to the mechanical vibration of the rolling mill, interference from zinc dust floating in the air, and the dark current noise of the laser sensor itself, the acquired initial thickness profile data often contains a large number of high-frequency spikes and abnormal abrupt changes. Therefore, it is necessary to perform sliding window weighted median noise reduction filtering on the initial thickness profile data. This filtering is a smoothing technique based on local data features, the core of which lies in using the spatial correlation between adjacent sampling points to correct abnormal data. In practice, the system first sets a fixed-length one-dimensional data window based on the strip's running speed and sampling resolution, such as an odd-length window containing seven consecutive sampling points. This window starts at the beginning of the initial thickness profile data and slides towards the end of the data sequence in fixed steps of one sampling point. At each sliding stop position of the window, the system extracts all thickness values ​​within the window's coverage area and performs weighted median filtering. The weighted median filtering process involves first assigning weights based on the physical distance of each sampling point within the window from the center point. The weights are calculated using w = 1 / (1+d), where d is the index distance from the sampling point to the center point; values ​​closer to the center point have higher weights.Then, these weighted thickness values ​​are arranged in ascending order, and the value whose cumulative weight reaches half of the total weight is selected as the new, corrected thickness value for the center sampling point of the window. Compared to simple arithmetic mean filtering, this weighted median filtering strategy can not only effectively remove isolated maximum noise points caused by dust momentarily obstructing the laser path, or minimum noise points caused by tiny pits on the strip surface, but also perfectly preserve the contour edge features of the strip edge region or the actual step change in coating thickness. This processing mechanism completely avoids the problems of contour edge blurring and thickness feature distortion that are easily caused by conventional linear filtering algorithms, and restores the true physical morphology of the strip surface to the greatest extent.

[0046] Understandably, as the sliding window completes its traversal of the sampling point data across the entire width direction, the system reassembles all the filtered and replaced thickness values ​​according to their original spatial coordinate order, ultimately obtaining a denoised coating thickness profile sequence. This sequence not only eliminates random interference introduced by the environment and equipment but also accurately reproduces the smooth distribution of coating thickness along the strip width, providing a reliable data source with a high signal-to-noise ratio for subsequent coating quality assessment and closed-loop control of coating process parameters.

[0047] Step S104: For the coating thickness profile sequence, if the difference between the maximum thickness value and the minimum thickness value in the coating thickness profile sequence is greater than a preset threshold, then the continuous sampling point segment where the difference is located is determined as a local distortion region, and the coordinates of the local distortion region are obtained.

[0048] A coating thickness contour sequence is obtained. The mean, standard deviation, range, skewness, kurtosis, and the maximum value of the thickness difference between adjacent sampling points for each contour sequence are used as feature vectors and input into a pre-trained support vector machine (SVM) classifier for classification. The SVM uses historically collected coating thickness contour sequences as training samples and manually labeled normal and abnormal contours as binary classification labels. It is trained using a radial basis function kernel function. When the output label is 1, it is determined to be an abnormal contour sequence; when the label is 0, it is determined to be a normal contour sequence. For the determined abnormal contour sequence, the thickness values ​​in the sequence are traversed. If the difference between the maximum and minimum thickness values ​​in the abnormal contour sequence is greater than a preset threshold of 5 micrometers, then the continuous sampling point segment between the sampling point corresponding to the maximum thickness value and the sampling point corresponding to the minimum thickness value is determined as a local distortion region. The starting and ending sampling point numbers of this continuous sampling point segment are recorded as the coordinates of the local distortion region for subsequent process correction.

[0049] For example, the coating thickness profile sequence refers to a set of thickness values ​​obtained by sampling at equal intervals along a specific path on the surface of the workpiece to be measured. Each element in the sequence corresponds to a specific physical sampling point on the workpiece surface, and its value represents the actual coating thickness at that point.

[0050] In one embodiment, for the coating thickness profile sequence, the system traverses the sequence to extract the maximum and minimum thickness values ​​and calculates the difference between them.

[0051] Understandably, this difference directly reflects the maximum fluctuation range of the coating thickness along the contour path. Subsequently, this difference is compared with a preset threshold. The preset threshold is typically set based on the process tolerance requirements of a specific workpiece or historical quality data, representing the maximum allowable normal thickness fluctuation range. If the calculated difference is greater than the preset threshold, it indicates an abnormal thickness abrupt change or non-uniformity in the contour sequence. In this case, it is necessary to further locate the continuous sampling point segment where the difference occurs.

[0052] Specifically, the system uses the sampling points with the maximum and minimum thickness values ​​as benchmarks to extend the search in both directions, calculating the thickness change gradient between adjacent sampling points one by one. The thickness change gradient is defined as the absolute value of the thickness difference between two adjacent sampling points divided by the distance between the two points, denoted as g = |t2 - t1| / d, where t1 and t2 are the thickness values ​​of adjacent sampling points, and d is the distance between the two sampling points. When the g value of an adjacent sampling point is lower than a preset smoothing threshold G0, the thickness change is considered to be stable, and the search continues to extend in that direction. The smoothing threshold G0 is determined according to the statistical characteristics of the thickness change gradient of the entire contour, taking 0.3 times the sum of the mean μ and standard deviation σ of the gradient sequence of all sampling points, i.e., G0 = 0.3 * (μ + σ). In conventional coating detection scenarios, the typical value range of G0 is 0.05 μm to 0.5 μm. For example, for precision coatings with small thickness fluctuations, 0.1 μm can be used. Therefore, the continuous sampling point interval containing the maximum and minimum values ​​and exhibiting drastic thickness variations is identified as a local distortion region. Further, after identifying the local distortion region, its coordinates are obtained. These coordinates can be represented as the start and end indices of the continuous sampling point segment within the entire contour sequence, or they can be mapped to the actual physical spatial coordinates of the workpiece surface through coordinate system transformation. Through these steps, the specific location of coating thickness anomalies can be accurately pinpointed, enabling rapid spatial localization of surface quality defects.

[0053] Step S105: Continuing from the electric field simulation comparison in step S104, the coordinates of the local distortion region are provided as a two-dimensional set of points on the anode array plane. Each point corresponds to a distortion center and the set of anode channel numbers it covers. Based on these coordinates, the corresponding channel is located, and real-time sensing current acquisition data is read from the Hall sensor connected to that channel at a sampling frequency of 1000 Hz. The real-time sensing current acquisition data is normalized according to the anode channel number and timestamp. The current value of each channel is divided by the rated full-scale current I_max of that channel, and aligned to a unified time reference t to obtain a local current density sequence J(c,t) with a time dimension, where c is the anode channel number and t is the timestamp.

[0054] Based on the coordinates of the local distortion region, the corresponding channel is determined in the anode array. Real-time sensing current data and its corresponding timestamp are acquired through the Hall sensor of the corresponding channel. If the real-time sensing current data exceeds a preset threshold, it is normalized by combining the channel identifier of the corresponding channel to obtain the local current density. A time-series splicing method is used to generate a local current density sequence with a time dimension based on the local current density and the timestamp.

[0055] In one implementation, the process of acquiring real-time sensing current data from the Hall sensors of the corresponding channels of the anode array based on the coordinates of the local distortion region first requires establishing a mapping relationship between a virtual spatial coordinate system and the physical anode distribution of the electrolytic cell. Multiple sets of anode carbon blocks arranged in a matrix are typically suspended above the electrolytic cell, with each set of anode carbon blocks and its guide rod forming an independent anode channel. The system uses a preset two-dimensional grid mapping table to resolve the input coordinates of the local distortion region into specific anode channel numbers, thereby accurately locating the physical region where the anomaly has occurred.

[0056] Specifically, after identifying the target anode channel, the system sends a data acquisition command to the Hall sensor mounted on the guide rod of that channel. The Hall sensor utilizes the magnetic field effect to non-contactly measure the strong direct current flowing through the anode guide rod. To ensure the real-time nature and accuracy of the data, the sensor continuously reads the current values ​​at a millisecond-level sampling frequency and converts these raw analog signals with precise timestamps into digital signals, forming real-time sensing current acquisition data. This data not only includes the absolute amplitude of the current but also records the exact moment the data was generated.

[0057] It should be noted that due to the different service lives of the various anode channels within the electrolytic cell, there are significant differences in the length, volume, and contact resistance of the old and new anode carbon blocks, resulting in an uneven distribution of the base current under normal operating conditions. Directly comparing raw absolute current data can easily lead to misjudgments. Therefore, the collected data must be normalized.

[0058] For example, the real-time sensing current acquisition data is normalized according to the anode channel and timestamp, and the final output is uniformly referred to as the "local current density sequence". Its generation process involves multiple dimensions of business calculations. The first step is the spatial dimension current density conversion. The system retrieves the current equipment ledger information of the corresponding anode channel, obtains the effective conductive base area S of the anode carbon block, divides the real-time absolute current value I collected by the Hall sensor by S, and obtains the intermediate absolute current density j, i.e., j=I / S. Then, the channel benchmark is standardized. The system extracts the historical average current density j0 of the specific anode channel in the past stable operating cycle as the benchmark value, and calculates the dimensionless intermediate quantity r according to r=(j-j0) / j0. This step eliminates the inherent differences caused by the different ages of different channels, so that all channel data are on the same evaluation benchmark. In the time dimension, since there may be slight sampling clock offsets between Hall sensors of different channels, the system sets a uniform time slice in milliseconds, for example, 100 milliseconds as a slice node. Based on the timestamp attached to each data point, a linear interpolation method is used to align the r values ​​of each channel to the same time slice node. After completing the above three steps of spatial transformation, channel standardization and time alignment, the output r value sequence organized by channel and by time slice is uniformly named "local current density sequence" and used as the input for the subsequent step S105. Other terms such as "relative deviation ratio" or "standardized current density state value" will no longer be used throughout the text.

[0059] In one possible implementation, timestamp processing also involves smoothing and filtering transient disturbances. During operation, the electrolytic cell may experience brief current spikes due to feeding actions or fluctuations in the bus magnetic field. The system uses timestamps to set a sliding time window and performs a weighted average calculation on the normalized data within the window, thereby eliminating high-frequency noise and retaining low-frequency trend data that accurately reflects the anode's operating state. Furthermore, after the aforementioned spatial and temporal normalization processing, the system concatenates the discrete data points in chronological order to obtain a local current density sequence with a time dimension. This sequence is essentially a one-dimensional data set containing multiple time nodes and their corresponding normalized current density state values, fully recording the dynamic trajectory of the current distribution evolution over time within the local distortion region.

[0060] Understandably, obtaining such a time-varying local current density sequence can provide crucial data support for the refined management of electrolytic cells. By observing the fluctuation amplitude and trend of this sequence, technicians can intuitively determine whether there is a long-term risk of overload heating at the local anode, or whether there is a sudden drop in current due to anode detachment.

[0061] For example, in the actual operation of a large-scale aluminum electrolysis plant, the system detected thermal field distortion in an area with coordinates of three rows and five columns. The system immediately locked onto the corresponding anode channel No. 5 and retrieved real-time sensing current acquisition data from its Hall sensor over the past ten minutes. Combining this with the area parameters of anode No. 5, which is currently in the middle of its service life, the system converted the absolute current value into current density and compared it with the normal baseline value of the channel from the previous day to calculate the deviation rate. Subsequently, the system aligned these deviation rates to one data point per second based on the timestamp, ultimately generating a local current density sequence containing 600 nodes. This sequence clearly shows that the current density of the channel exhibits a step-like upward trend over ten minutes, indicating to the operators that a shortening of the electrode spacing due to a long anode package may have occurred at this location, requiring timely intervention.

[0062] Step S106: Use a long short-term memory network to perform time-series feature analysis on the local current density sequence, and combine the current coupling relationship between adjacent channels to obtain the multi-region collaborative compensation current value.

[0063] Monitoring points were evenly spaced on the anode plate array of the electrolytic cell, and local current density sequences were collected at each monitoring point. The sampling period was 1 second, and each input segment had a length of 60 time steps. Each point had a single channel, forming an input tensor of shape 60 x 1. A two-layer stacked Long Short-Term Memory (LSTM) network was used to extract features from this sequence, with 64 hidden units. The output layer used the hidden state of the last time step as the temporal feature vector, with a dimension of 64. Training samples were approximately 20,000 segments from the historical operating data of the electrolytic cell, labeled with current decay rates and corresponding measured impedance values. These were divided into training and validation sets at an 8:2 ratio. The Adam optimizer was used with a learning rate of 0.001 and a loss function of mean squared error. The iterations were performed for 100 rounds until the validation loss converged.

[0064] Linear regression fitting is performed on the time-series feature vector to obtain the current decay slope k. The larger the absolute value of the slope, the more severe the corrosion of the anode plate. The corrosion degree C is defined as the normalized absolute value of k, with a value range of 0 to 1. The impedance change rate R between adjacent channels is obtained by multiplying the mean corrosion degree of the two channels by an empirical coefficient α, where α is set to 0.85, reflecting the increase in contact resistance caused by the expansion of the corrosion area.

[0065] Further calculate the absolute value of the difference in impedance change rate between adjacent channels to obtain the impedance deviation gradient G. Arrange all adjacent channel pairs of G according to the channel number to form an N-by-N symmetric coupling relationship matrix M, where N is the total number of channels, the diagonal elements are set to zero, and non-adjacent positions are diffused and filled with an attenuation factor of 0.5.

[0066] Elements in the coupling matrix M are compared with a preset threshold of 0.6. Elements greater than 0.6 are considered strongly coupled target elements, corresponding to channels with significant erosion correlations. The value of each target element is divided by the sum of all target elements in that row to obtain the normalized channel weight w. Channel weights below the threshold are reset to zero. w is then multiplied element-wise by the temporal feature vector for each channel and summed to obtain the compensation baseline value for each region.

[0067] Based on the regional location information of the electrolytic cell, the channel is mapped according to physical partitions such as upper, middle, lower or left, middle, right, etc. The compensation reference value of the corresponding region is summarized as the target compensation current value of the region. Then, the output current of each partition is adjusted in real time according to the target value by the controllable rectifier module in the power supply circuit, and finally the multi-region collaborative compensation current value is obtained.

[0068] In one implementation, to achieve dynamic equilibrium of the anode current distribution inside the electrolytic cell, it is necessary to perform in-depth analysis on the collected local current density sequences of each anode carbon block and calculate the corresponding compensation current. The specific process is as follows: Local current densities of each anode carbon block are collected at a sampling period of 1 second, forming a sliding window sequence of length 60; the sequence is denoised and averaged to obtain the average current density j_avg; then the deviation Δj_i = j_i - j_avg between the current current density j_i and j_avg of each anode carbon block is calculated; when |Δj_i| exceeds a preset threshold of 0.05 A / cm², the anode is adjusted according to the compensation current I_c = k * Δj_i * S, where S is the effective conductive area of ​​the anode carbon block, and k is the adjustment coefficient, ranging from 0.8 to 1.2, thereby balancing the anode currents.

[0069] Specifically, a long short-term memory network is used to perform time-series feature analysis on the local current density sequence, and combined with the current coupling relationship between adjacent channels, the multi-region collaborative compensation current value is obtained.

[0070] For example, Long Short-Term Memory (LSTM) networks are deep learning models specifically designed for processing time-series data, effectively capturing the long-term dependencies of local current density changes over time. In practice, the preprocessed local current density sequence is used as input to the LTM network. This network comprises three core components: a forget gate, an input gate, and an output gate. The forget gate determines which information from the previous cell state needs to be discarded, such as abnormal current spikes caused by instantaneous load fluctuations in history. The input gate determines which information from the new input data at the current moment needs to be updated to the cell state, such as the current continuous downward trend in current density. The output gate calculates the hidden state at the current moment based on the updated cell state. Through this gating mechanism, the LTM network can process the local current density sequence step by step, extracting a temporal feature vector reflecting the dynamic evolution of current density. This temporal feature vector not only contains the current current state but also incorporates historical current change trend information, providing a solid data foundation for subsequent compensation calculations.

[0071] It should be noted that the various channels within a fuel cell system do not exist in isolation; there are complex physical and electrical interactions between adjacent channels.

[0072] Specifically, the current coupling relationship between adjacent channels refers to the cascading changes in the current distribution of adjacent channels caused by a change in the current in one channel. To quantify this coupling relationship, a current coupling coefficient matrix is ​​constructed. The elements of this matrix represent the degree of mutual influence between different channels. In actual operation, the determination of the coupling coefficient depends on physical parameters such as the bipolar plate structure of the fuel cell, the flow field distribution, and the conductivity of the membrane electrode assembly.

[0073] For example, when a channel experiences insufficient reactive gas supply due to localized flooding, its local current density will significantly decrease. At this time, due to the conductivity of the bipolar plates, adjacent channels will share some of the current load, leading to an increase in the current density of adjacent channels. To quantify this coupling relationship, the single cell is discretized into N channel units along the flow field direction. Each unit forms an equivalent circuit network using parallel current sources, film resistance Rm, contact resistance Rc, and lateral conduction resistance Rt between it and adjacent channels. Node voltages are provided by the plate potential sampling points, and node currents are measured by segmented current acquisition chips. Based on historical steady-state operating data, the least squares method is used to identify the parameters Rm, Rc, and Rt. Typical values ​​are Rm between 20 and 80 milliohms, Rc between 5 and 15 milliohms, and Rt between 50 and 200 milliohms. The static coupling weight wij is defined as the ratio of the change in current in channel j caused by a unit change in current in channel i. It is obtained by inverting the node admittance matrix from the identified resistor network and then normalizing it. Diagonal elements are fixed at 1, while off-diagonal elements generally fall between 0.05 and 0.35. Adjacent channels have higher weights than non-adjacent channels. The dynamic adjustment stage uses real-time current, temperature, inlet humidity, and voltage drop as inputs. When the channel voltage drop deviates from the baseline by more than 15% or the temperature deviation exceeds 3 degrees Celsius, the corresponding row and column weights are corrected. The correction coefficient k = 1 + α*ΔP / P0 + β*ΔT / T0, where α is 0.4, β is 0.2, ΔP and ΔT are the measured deviations, and P0 and T0 are the baseline values. After correction, it is normalized again by row to form a current coupling coefficient matrix reflecting the current system state. Based on the above steps, the time-series feature vector output by the Long Short-Term Memory network is further fused with the constructed current coupling coefficient matrix.

[0074] Feature fusion is achieved through information transfer using a single-layer graph convolution, with the entire process completed via a single matrix multiplication. The temporal feature vectors of each region are stacked sequentially to form a feature matrix H, with dimensions N*T, where N is the number of regions in the power grid and T is the temporal feature dimension (e.g., N = 32, T = 64). The current coupling coefficient matrix A has dimensions N*N, where element A[i,j] represents the current influence of region j on region i, ranging from 0 to 1, with diagonal lines set to 1 to preserve their own characteristics. The fused feature matrix H' = A * H, with an output dimension still N*T, where the i-th row represents the fused feature of region i after absorbing the current state of adjacent regions. To avoid numerical drift caused by differences in node degrees, A is normalized row-wise, ensuring the sum of each row's elements is 1 before participating in the multiplication operation. The fused feature matrix is ​​then fed into the downstream collaborative state discrimination module to characterize the evolution of current coupling between multiple regions.

[0075] Understandably, obtaining multi-region collaborative compensation current values ​​is the ultimate goal of achieving system equilibrium control. Based on the fused feature vector obtained above, a fully connected neural network layer is used for mapping calculation. This fully connected layer converts the high-dimensional fused feature vector into specific current value outputs. During the calculation process, the system first needs to determine the target equilibrium current density value, which is determined by three parts: First, the reference current density is obtained based on the ratio of the system's rated capacity to the effective operating area of ​​each region, usually falling within the range of 0.8 to 1.2 amperes per square centimeter; second, the current density of each region during steady-state operation is weighted and averaged based on historical operating data from the past 30 days, with the weights decreasing daily according to a time decay factor of 0.95, resulting in an empirical correction term; third, the target value is dynamically adjusted based on real-time collected ambient temperature, load rate, and battery state of charge through a preset operating condition correction coefficient table, for example, the target value is reduced by 0.05 amperes per square centimeter for every 10 degrees Celsius increase in temperature. The three factors are weighted and fused in proportions of 0.5, 0.3, and 0.2 to obtain the target equilibrium current density value J_ref at the current moment. The fully connected layer calculates the required compensation current magnitude and direction for each region based on the deviation between the current density of each region and J_ref in the fused feature vector, as well as the coupling influence factor between regions. The output results are used to drive the subsequent equalization control execution unit.

[0076] For example, in regions where the current density is lower than the target value and is significantly affected by negative coupling from adjacent channels, the system calculates a positive compensation current value to improve the response rate of that region; conversely, in regions where the current density is too high, a negative compensation current value is calculated.

[0077] Preferably, by employing the analysis method described above, which combines long short-term memory networks with current coupling relationships, accurate calculation of multi-region collaborative compensation current values ​​can be achieved. This method fully considers the time lag and spatial correlation of current changes, effectively avoiding system oscillations caused by traditional independent channel compensation strategies, and improving the operational stability of fuel cell systems under complex operating conditions.

[0078] Step S107: Using the multi-region collaborative compensation current value, generate an anode control command for the anode channel covered by the coordinates of the local distortion region to obtain a multi-channel current adjustment command set.

[0079] The compensation current and distortion coordinates are obtained. Based on the distortion coordinates, the spacing parameter d and conductivity ratio k are determined by looking up a table in the process database. Here, d represents the spacing between adjacent electrodes in millimeters, ranging from 5 to 50, and k is the relative conductivity coefficient of the conductive material in that region, dimensionless, ranging from 0.6 to 1.2. The target density J is calculated according to J = I / d*k, where I is the compensation current in amperes, and J is in amperes per millimeter, characterizing the current distribution intensity per unit spacing. The target density J and distortion coordinates are input into a preset channel mapping model, which is a nearest neighbor matching lookup table built based on historical calibration data. Using the coordinate range and density range as keys and the channel number as the value, the corresponding target channel is output. The deposition thickness corresponding to the target channel is obtained. If the deposition thickness is less than a preset thickness threshold of 8 micrometers, a preset edge compensation value is obtained. The dimension of this compensation value is consistent with the target density, in amperes per millimeter, ranging from 0.05 to 0.30. The target current density is obtained by adding the edge compensation value to the target density. The pulse high-level duration t and period duration T are determined by looking up a table based on the target current density, both in microseconds. The duty cycle is calculated using t / T, resulting in a dimensionless ratio ranging from 0.1 to 0.9. Based on the duty cycle and the target channel, corresponding adjustment commands are generated. Multiple adjustment commands are obtained by traversing all distorted coordinates, and these commands are combined to obtain a multi-channel current adjustment command set.

[0080] In one implementation, after obtaining the multi-region collaborative compensation current value, the system needs to convert it into a control signal that can be executed by the underlying device.

[0081] Specifically, the first step is to establish a spatial mapping relationship between the coordinates of the local distortion region and the physical anode channels. Since electrolytic machining equipment typically has multiple independent anode channels arranged in an array, each channel has a fixed range of physical coordinates within the machining area. The system uses a coordinate matching algorithm to compare the identified coordinates of the local distortion region with a preset anode channel coordinate matrix, thereby accurately locating one or more target anode channels covering the distortion region. Furthermore, for the located target anode channels, the system generates specific anode control commands based on multi-region collaborative compensation current values.

[0082] For example, when a local distortion region spans three adjacent anode channels, the system does not simply distribute the total compensation current evenly. Instead, it performs a weighted allocation based on the coverage area ratio of each channel within the distortion region and the electric field distribution characteristics. After calculating the independent compensation current value required for each target channel, the system encapsulates it into a standardized anode control command. This command includes not only the hardware address identifier of the target channel and the calculated absolute current setpoint, but also dynamic parameters for current regulation, such as the current rise slope, pulse duty cycle, and regulation response time. This refined command generation process ensures that each channel can cooperate with each other when outputting compensation current, avoiding new electric field distortions caused by sudden current changes in a single channel.

[0083] Understandably, after generating instructions for all target channels, the system aligns the timing of these dispersed anode control instructions and packages the data to obtain a multi-channel current regulation instruction set. This instruction set, as a complete control data package, is synchronously sent to the underlying power control module. After parsing the instruction set, the underlying module drives the corresponding solid-state relays or thyristor components to achieve synchronous and precise regulation of the multi-channel current. In this way, local electric field distortion can be effectively eliminated, ensuring the uniformity and stability of the overall processing.

[0084] In step S108, after the current adjustment is completed in S107, wait 5 to 10 seconds for the anode current to stabilize and for the physical response of the local coating thickness to be completed. Then, trigger the laser thickness sensor to perform a second scan of the coordinate range of the local distortion area on the strip surface. The scanning resolution is consistent with that in S101, and the sampling step is 2 mm. Map the second scan data to the corresponding position in the thickness profile matrix T according to the original coordinate system, and replace the thickness value before adjustment to obtain the updated thickness profile data T'. At the same time, calculate the thickness difference Δd = T' - T at the same coordinate point before and after adjustment, and statistically analyze the mean and standard deviation of Δd in the distortion area for closed-loop convergence determination in S109.

[0085] Based on the multi-channel current adjustment command set issued by S107, the set of distortion coordinates corresponding to the local distortion region is extracted. The measured anode current values ​​at each coordinate after being set by S107 are read to construct an initial current distribution matrix I0, which is used to perform closed-loop verification of the execution effect of the S107 command. The Kalman filter algorithm is used to estimate the state of I0, suppress sampling noise and transient fluctuations, and the filtered current distribution matrix I1 is obtained. Combining the preset electrode spacing d values ​​of 8 mm to 15 mm and ion concentration c values ​​of 60 g / L to 90 g / L, the deposition rate v at each coordinate is calculated according to the Faraday electrolytic deposition relationship, where v = k * I1 * c / d, and k is an electrochemical constant determined by the metal ion valence state and current efficiency. In this embodiment, it is taken as 2.3 multiplied by 10 to the power of -4. The difference between v and the desired deposition rate v0 of S107 is used to obtain the current fine-tuning amount ΔI=α*v0-v / k*d / c, where α is a convergence factor ranging from 0.6 to 0.8. This is then superimposed on the anode current to complete the secondary fine-tuning, thereby achieving a fine correction of the coarse-tuning result of S107.

[0086] After fine-tuning, secondary laser scanning data of the strip surface is acquired from the laser thickness sensor to generate updated contour data T_x, where x is the coordinate of the strip width direction. The thickness deviation ΔT_x between T_x and the target thickness T0 is calculated. If the absolute value of ΔT_x at any point exceeds a preset threshold of 0.5 micrometers, edge effect features are extracted from the contour data, including the thickness drop amplitude h1, drop width w1, and edge thickness gradient g1 within a 15 mm area on both sides of the strip edge. Based on the above features, the thickness compensation coefficient β is calculated according to β=1+h1 / T0*w1 / W*g1, where W is the total width of the strip, and the value of β is limited to between 1.02 and 1.15. If the value exceeds the limit, the boundary value is taken. The final coating thickness T_final=β*T_x, and T_final is sent back to S107 for the next cycle current reference update.

[0087] In one possible implementation, to address the thickness uniformity control requirements in a continuous strip electroplating production line, the system, after generating a multi-channel current adjustment instruction set, needs to translate it into actual physical control actions. The multi-channel current adjustment instruction set contains information on current increases and decreases at different spatial locations on the strip surface. This information is calculated based on thickness deviations identified in previous steps and aims to eliminate localized thickness anomalies on the strip surface through precise electrochemical reaction rate control.

[0088] Specifically, adjusting the anode output current corresponding to the coordinates of the locally distorted region is a complex process involving spatial mapping and multi-channel power supply coordinated control. In actual electroplating tanks, an anode array consisting of multiple independently controllable anode plates is typically arranged, with each anode plate corresponding to an independent current output channel. The system first extracts the coordinates of the locally distorted region, which includes the lateral position in the strip width direction and the longitudinal position in the running direction. Subsequently, the control module uses a preset coordinate mapping matrix to convert the coordinates of the locally distorted region into a specific channel number in the anode array.

[0089] For example, if the coordinates of a localized distortion region indicate a thinner distortion on the left edge of the strip, the mapping matrix will locate several anode channels corresponding to that left edge. Next, the system parses the multi-channel current regulation instruction set, extracts the current regulation amplitude and regulation time parameters for these channels, and sends analog control signals to the corresponding rectifiers via a programmable logic controller, thereby precisely changing the output current of these specific anode channels. This point-to-point current regulation mechanism can directly change the electric field distribution corresponding to the localized distortion region, thereby adjusting the metal ion deposition rate in that region.

[0090] It is understandable that changes in the anode output current will cause changes in the local current density in the electroplating solution. After the set electroplating reaction delay time, the thickness of the metal coating in the locally distorted areas on the strip surface will undergo corresponding physical changes. Areas that were originally thinner will have accelerated deposition due to increased current, while areas that were originally thicker will have slower deposition due to decreased current, thus achieving dynamic thickness compensation.

[0091] For example, to verify the actual effect of current regulation and form closed-loop control, the system needs to acquire secondary scan data of the strip surface from the laser thickness sensor. The laser thickness sensor is typically installed at the inspection station after the electroplating tank outlet, using the high-frequency laser triangulation principle to perform a transverse reciprocating scan of the running strip surface. When the strip passes through the current regulation area and reaches the inspection station, the sensor emits a laser beam to illuminate the strip surface and receives the reflected light signal. By calculating the displacement of the light spot, the real-time height information of the strip surface is obtained. This continuous set of height information constitutes the secondary scan data. Further, obtaining the updated thickness profile data requires a series of data processing steps on the secondary scan data. Due to factors such as strip vibration and moisture interference in the industrial environment, the original secondary scan data usually contains high-frequency noise. The system first uses a moving average filtering algorithm to smooth the secondary scan data, eliminating abnormal jump points. Subsequently, the system timestamps the filtered data with the strip's running speed and tension parameters to ensure that the secondary scan data completely coincides with the initial coordinates of the locally distorted area in spatial coordinates. After coordinate alignment and reference plane compensation, the system fits the discrete scan point data into a continuous surface, ultimately generating updated thickness profile data. This data accurately reflects the latest state of the strip surface thickness distribution after multi-channel current adjustment intervention.

[0092] It should be noted that obtaining the updated thickness profile data through the above steps provides an accurate data foundation for the next round of iterative control of the system. This closed-loop mechanism based on precise multi-channel current adjustment and real-time feedback from laser thickness measurement effectively overcomes the shortcomings of traditional global current control in addressing local thickness unevenness, enabling high-precision control of the overall thickness consistency of the strip.

[0093] 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 controlling the thickness of the strip plating layer in continuous electroplating of IC lead frames, characterized in that, include: The surface image of the target workpiece is acquired from the image acquisition device, and the edge features of the surface image of the target workpiece are extracted using a neural network model to obtain a surface texture feature matrix. Based on the marked position segments in the surface texture feature matrix, initial electric field acquisition data at corresponding positions within the electrochemical processing equipment is obtained. Matrix mapping is performed on the initial electric field acquisition data to obtain a spatial electric field distribution matrix. Using the field strength distribution of each grid in the spatial electric field distribution matrix, scanning sampling points are located. Initial thickness contour data of the scanning sampling points is obtained from the thickness detection device. Noise reduction filtering is applied to the initial thickness contour data to obtain a denoised coating thickness contour sequence. For the coating thickness contour sequence, if the difference between the maximum and minimum thickness values ​​is greater than a preset threshold, the continuous sampling point segment containing the difference is determined to be a local distortion region, and local distortion is obtained. The coordinates of the local distorted region are varied. Real-time sensing current data is acquired from the current detection device based on the coordinates of the local distorted region. This real-time sensing current data is normalized to obtain a local current density sequence with a time dimension. A time-series network model is used to perform time-series feature analysis on the local current density sequence to obtain multi-region collaborative compensation current values. Control commands are generated for the channels covered by the coordinates of the local distorted region based on these multi-region collaborative compensation current values, resulting in a multi-channel current adjustment command set. The output current corresponding to the coordinates of the local distorted region is adjusted according to the multi-channel current adjustment command set, and secondary scan data is acquired again from the thickness detection device to obtain updated thickness contour data.

2. The method for controlling the thickness of the strip plating in continuous electroplating of IC lead frames as described in claim 1, characterized in that, The step of extracting edge features from the surface image of the target workpiece using a neural network model to obtain a surface texture feature matrix includes: performing pixel-level conversion processing on the surface image of the target workpiece using a grayscale conversion algorithm to obtain a grayscale image of the target workpiece; performing multi-scale edge feature extraction on the grayscale image of the target workpiece using the neural network model to obtain an edge feature map of the target workpiece; calculating the pixel gradient value of each pixel based on the edge feature map of the target workpiece, and if the pixel gradient value is greater than a preset gradient threshold, determining that the region where the pixel corresponding to the pixel gradient value is located is a surface defect region; extracting local texture pixels for the surface defect region, and constructing a local texture matrix based on the local texture pixels according to a preset arrangement rule; and performing matrix dimensionality reduction and feature preservation processing on the local texture matrix using a singular value decomposition algorithm to obtain the surface texture feature matrix.

3. The method for controlling the thickness of the strip plating in continuous electroplating of IC lead frames as described in claim 1, characterized in that, The step of performing matrix mapping on the initial electric field acquisition data to obtain a spatial electric field distribution matrix includes: dividing the initial electric field acquisition data into two-dimensional plane grid coordinates according to preset array coordinates and width coordinates to obtain a discrete electric field coordinate point set; identifying missing nodes in the discrete electric field coordinate point set that do not contain acquisition data, and using a bilinear interpolation algorithm to perform numerical filling and matrix mapping processing on the missing nodes to obtain an initial spatial electric field distribution matrix; obtaining a preset electrode spacing parameter, and calculating the potential difference between adjacent grid nodes based on the initial spatial electric field distribution matrix and the electrode spacing parameter to determine the electric field gradient matrix; if the edge node values ​​in the electric field gradient matrix are greater than a preset gradient threshold, then using a preset compensation coefficient to perform numerical correction on the edge nodes to complete edge distortion compensation, thereby obtaining the spatial electric field distribution matrix.

4. The method for controlling the thickness of the strip plating in continuous electroplating of IC lead frames as described in claim 1, characterized in that, The step of denoising and filtering the initial thickness contour data to obtain a denoised coating thickness contour sequence includes: if there are edge distortion data in the initial thickness contour data that exceed a preset distortion threshold, then the edge distortion data is removed by a data cleaning algorithm to obtain first thickness contour data; a sliding window denoising and filtering algorithm is used to apply a sliding average algorithm to the first thickness contour data, and high-frequency noise interference in the first thickness contour data is eliminated by setting the step size and window size of the sliding window to obtain denoised second thickness contour data; the second thickness contour data is serialized and spliced ​​according to the position order of the scanning sampling points to construct continuous thickness distribution features and determine the denoised coating thickness contour sequence.

5. The method for controlling the thickness of the strip plating in continuous electroplating of IC lead frames as described in claim 1, characterized in that, If the difference between the maximum and minimum thickness values ​​in the coating thickness contour sequence is greater than a preset threshold, then the continuous sampling point segment where the difference is located is determined to be a local distortion region, and the coordinates of the local distortion region are obtained. This includes: classifying the coating thickness contour sequence using a support vector machine to identify abnormal contour sequences that deviate from the normal thickness range; extracting the maximum and minimum thickness values ​​in the abnormal contour sequences; if the difference between the maximum and minimum thickness values ​​in the abnormal contour sequences is greater than a preset thickness difference threshold, then determining the continuous sampling point segment between the sampling point corresponding to the maximum thickness value and the sampling point corresponding to the minimum thickness value as a local distortion region; and mapping the local distortion region to obtain its coordinates.

6. The method for controlling the thickness of the strip plating in continuous electroplating of IC lead frames as described in claim 1, characterized in that, The normalization process for the real-time sensing current acquisition data to obtain a local current density sequence with a time dimension includes: determining the corresponding channel in the array of the electrochemical processing equipment based on the coordinates of the local distortion region; acquiring the real-time sensing current acquisition data and the corresponding timestamp through the current detection device of the corresponding channel; if the real-time sensing current acquisition data exceeds a preset current threshold, normalizing the real-time sensing current acquisition data in combination with the channel identifier of the corresponding channel to eliminate the dimensional differences between different channels and obtain the local current density; and generating the local current density sequence with a time dimension by matching and combining the local current density with the timestamp using a time series splicing method.

7. The method for controlling the thickness of the strip plating in continuous electroplating of IC lead frames as described in claim 1, characterized in that, The step of using a time-series network model to perform time-series feature analysis on the local current density sequence to obtain multi-region collaborative compensation current values ​​includes: acquiring local current density sequences at multiple monitoring points within the electrochemical treatment device; extracting features from the local current density sequences using the time-series network model to obtain a time-series feature vector; assessing the degree of electrode corrosion between adjacent channels based on the current decay trend in the time-series feature vector; determining the impedance change rate between adjacent channels based on the degree of electrode corrosion; quantifying the impedance deviation gradient between adjacent channels using the impedance change rate; constructing a coupling relationship matrix between adjacent channels based on the impedance deviation gradient; determining multi-region compensation benchmark values ​​based on the coupling relationship matrix; allocating the compensation benchmark values ​​to regions based on the regional location information of the electrochemical treatment device; adjusting the output of the power supply circuit based on the region allocation result to obtain the multi-region collaborative compensation current values.

8. The method for controlling the thickness of the strip plating in continuous electroplating of IC lead frames as described in claim 7, characterized in that, The step of determining the compensation benchmark value for multiple regions based on the coupling relationship matrix includes: if an element in the coupling relationship matrix is ​​greater than a preset matrix element threshold, then the element greater than the preset matrix element threshold is determined as a target element; the corresponding channel weight is determined based on the position coordinates of the target element in the coupling relationship matrix, wherein the channel weight characterizes the degree of influence of different channels on the overall current distribution; the channel weight is combined with the timing feature vector for weighted calculation, and the timing feature information and spatial coupling information of multiple channels are fused; global compensation features are extracted based on the weighted calculation result, and the global compensation features are numerically transformed through a preset benchmark mapping function, and the compensation benchmark value for multiple regions is determined based on the numerical transformation result.

9. The method for controlling the thickness of the strip plating in continuous electroplating of IC lead frames as described in claim 1, characterized in that, The step of generating control commands for the channels covered by the coordinates of the local distorted regions using the multi-region collaborative compensation current values ​​to obtain a multi-channel current adjustment command set includes: determining a spacing parameter and conductivity ratio based on the coordinates of the local distorted regions; obtaining a target density by dividing the multi-region collaborative compensation current values ​​by the spacing parameter and multiplying by the conductivity ratio; processing the target density and the coordinates of the local distorted regions using a preset channel mapping model to determine the corresponding target channel; obtaining the deposition thickness corresponding to the target channel, and if the deposition thickness is less than a preset thickness threshold, obtaining a preset edge compensation value; obtaining a target current density by adding the edge compensation value to the target density; determining the voltage amplitude and period duration based on the target current density, and obtaining a duty cycle value by dividing the voltage amplitude by the period duration; generating corresponding control commands based on the duty cycle value and the target channel, traversing all coordinates of the local distorted regions to obtain multiple control commands, and combining multiple control commands to obtain the multi-channel current adjustment command set.

10. The method for controlling the thickness of the strip plating in continuous electroplating of IC leadframes as described in claim 1, characterized in that, The step of adjusting the output current corresponding to the coordinates of the local distortion region according to the multi-channel current adjustment instruction set, and obtaining updated thickness contour data from the thickness detection device by acquiring secondary scan data again, includes: parsing the set of distortion coordinates corresponding to the coordinates of the local distortion region, obtaining the output current at each coordinate in the set of distortion coordinates, and constructing an initial current distribution matrix; performing state estimation on the initial current distribution matrix using a Kalman filter algorithm to obtain a filtered current distribution matrix; calculating the deposition rate at each coordinate in the set of distortion coordinates by combining a preset electrode spacing and ion concentration; calculating the current adjustment amount based on the deposition rate, and adjusting the output current using the current adjustment amount; acquiring secondary scan data of the surface of the target workpiece from the thickness detection device, and generating the updated thickness contour data based on the secondary scan data; calculating the thickness deviation between the updated thickness contour data and the preset target thickness; if the thickness deviation exceeds a preset deviation threshold, extracting edge effect features from the updated thickness contour data, calculating a thickness compensation coefficient based on the edge effect features and the target thickness, and determining the final coating thickness based on the thickness compensation coefficient.

Images