An adaptive compensation electroplating method for surface defects of superfine metal wire

Abstract

This invention relates to an adaptive compensation electroplating method for surface defects of ultrafine metal wires in the field of information technology. The method includes: acquiring a continuous wire image during a deformation process; extracting pixel differences in the continuous wire image using a convolutional neural network to obtain surface pit features and diameter fluctuation values; constructing a first physical state sequence corresponding to the deformation process based on the surface pit features and the diameter fluctuation values; processing the target feature vector, the initial deposition current, and the initial transmission speed using a long short-term memory network to output a target deposition current and a target transmission speed; calculating the expected deposition layer thickness for the deposition process based on the target deposition current and the target transmission speed; and executing the deposition process using the updated deposition current and the target transmission speed to obtain a target deposit.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to an adaptive compensation electroplating method for surface defects in ultrafine metal wires. Background Technology

[0002] Ultrafine metal wires play an indispensable and crucial role in high-end manufacturing fields such as precision electronics and medical devices. The quality of their surface coating directly determines the conductivity and lifespan of the final product.

[0003] Traditional production models typically separate wire stretching and thinning from surface electroplating into two independent processes. This physical segmentation not only increases the number of times the wire is repeatedly handled and wound within the workshop, easily leading to mechanical scratches and oxidation on the wire surface, but also severs the status perception channel between upstream and downstream processes. Existing electroplating production lines often use fixed operating parameters, relying entirely on manual experience for pre-setting, lacking the ability to respond to the real-time status of the wire.

[0004] Because of the information gap between the stretching and electroplating processes, the microscopic physical changes generated in the preceding process cannot be transmitted to the subsequent process, leading to serious quality control issues. During high-speed stretching, metal wire inevitably experiences minor diameter fluctuations or surface pits due to mechanical stress. When wires with these microscopic differences enter the electroplating stage, the fixed electroplating parameters cannot adapt to the dynamic changes in the wire. For example, if a section of wire has a smaller actual diameter due to fluctuations in tensile tension, or if minor scratches appear on the surface, maintaining the original fixed current intensity and wire speed in the electroplating tank will result in excessively thick plating buildup on that section of wire, or even plating peeling off at the defective area, ultimately rendering the entire spool of wire unusable. This severe mismatch between the fluctuations in the front-end physical state and the fixed processing parameters at the back end makes it difficult to improve the finished product yield.

[0005] Therefore, how to break down the physical and information barriers between the stretching and electroplating processes, so that the electroplating process can perceive the size and surface condition of the wire after stretching in real time, and dynamically adjust the electroplating parameters accordingly to compensate for the deviation at the front end, has become a key issue in achieving high-quality continuous production of ultra-fine metal wires. Summary of the Invention

[0006] This invention provides an adaptive compensation electroplating method for surface defects in ultrafine metal wires, mainly comprising:

[0007] A continuous wire image of the deformation process is acquired, and pixel differences in the continuous wire image are extracted using a convolutional neural network to obtain surface pit features and diameter fluctuation values. Based on the surface pit features and diameter fluctuation values, a first physical state sequence corresponding to the deformation process is constructed. If the diameter fluctuation value in the first physical state sequence is greater than a preset fluctuation threshold, the first physical state sequence is linearly interpolated and smoothed to obtain a second physical state sequence. A target feature vector is extracted from the second physical state sequence to obtain the initial operating settings for the deposition process, and the initial deposition current and initial transmission speed are extracted from the initial operating settings. The Long Short-Term Memory (LSTM) network processes the target feature vector, the initial deposition current, and the initial transmission velocity, and outputs the target deposition current and the target transmission velocity. Based on the target deposition current and the target transmission velocity, the expected deposition layer thickness for the deposition process is calculated. If the expected deposition layer thickness is less than a preset thickness threshold, the target deposition current is increased according to a preset compensation ratio to obtain an updated deposition current. If the expected deposition layer thickness is greater than or equal to the preset thickness threshold, the target deposition current is used as the updated deposition current. The deposition process is executed using the updated deposition current and the target transmission velocity to obtain the target deposit. Furthermore, the step of acquiring the continuous wire image of the deformation process, and extracting pixel differences in the continuous wire image using a convolutional neural network to obtain surface pit features and diameter fluctuation values, includes: acquiring a grayscale image sequence of the continuous wire and performing normalization processing to obtain a standardized image; dividing the standardized image into stress concentration regions and material uniformity regions according to the degree of deformation; using a convolutional neural network to extract features from the standardized image, outputting an image pixel gradient matrix; extracting surface texture features based on the locations of abrupt gradient changes in the image pixel gradient matrix to obtain the defect location coordinates of the surface pits; and then, based on the defect location coordinates... The radial distribution of pixel gradients in the corresponding image region is statistically analyzed to obtain the pit depth distribution value. Based on the defect location coordinates and the pit depth distribution value, an edge detection algorithm is used to extract the contour boundary point set of the continuous wire in the standardized image. The Euclidean distance between adjacent boundary points is calculated to obtain the cross-sectional diameter sequence of the substrate. The diameter variation amplitude is obtained by the difference between the maximum and minimum values ​​in the cross-sectional diameter sequence. The ratio of the diameter variation amplitude to the maximum deviation of the distance from each cross-sectional boundary point to the center is taken as the cross-sectional roundness deviation. Based on the ratio relationship between the diameter variation amplitude and the cross-sectional roundness deviation, the surface pit features and diameter fluctuation values ​​are obtained.Furthermore, the step of constructing a first physical state sequence corresponding to the deformation process based on the surface pit features and the diameter fluctuation values ​​includes: extracting the pit depth values ​​at each defect location from the surface pit features; counting the number of pits per unit length of substrate to obtain a defect density value; calculating the diameter change rate between adjacent measurement sections in conjunction with the diameter fluctuation values; matching the diameter change rate with the pit depth values ​​at corresponding positions according to the process position coordinates to obtain a cross-sectional deformation distribution sequence; extracting process position points of the substrate at fixed intervals along the deformation direction; collecting the surface temperature field distribution and the tension fluctuation amount applied by the deformation equipment at each process position point; and calculating the stress distribution of the material at the corresponding position based on the temperature field distribution and the tension fluctuation amount. The numerical values ​​are calculated as follows: For each measured section in the cross-sectional deformation distribution sequence, the stress distribution field value and temperature field distribution data at the corresponding location are extracted. The cross-sectional deformation, the stress distribution field value, the temperature field distribution data, and the tension fluctuation are concatenated into a physical parameter vector. The corresponding process location coordinates and the timestamp of the acquisition time are added to each physical parameter vector. All physical parameter vectors are arranged in chronological order according to the timestamps. The defect density value is inserted between adjacent physical parameter vectors as a transition marker. Based on the comparison result between the cross-sectional deformation and the preset deformation threshold, a quality level identifier is marked for each physical parameter vector. A first physical state sequence containing the physical parameter vector, the quality level identifier, and the timestamp is constructed. Furthermore, the step of performing linear interpolation smoothing on the first physical state sequence to obtain a second physical state sequence if the diameter fluctuation value in the first physical state sequence is greater than a preset fluctuation threshold includes: if the diameter fluctuation value between adjacent time nodes in the first physical state sequence exceeds the preset fluctuation threshold, marking the positions of abnormal nodes that exceed the fluctuation threshold, and determining whether the distribution density of the abnormal nodes meets the conditions for triggering smoothing; based on the distribution density determination result of the abnormal nodes, using a linear interpolation method on all abnormal nodes in the first physical state sequence, and taking the effective diameter measurement values ​​on both sides of the abnormal node as a reference, calculating the value at the abnormal node. The interpolated values ​​are used to replace the values ​​to obtain the corrected diameter values ​​of each node. If the difference between the corrected diameter value and the value of the adjacent node is still greater than the preset fluctuation threshold, then the corrected diameter value is calculated again by linear interpolation until the diameter fluctuation values ​​between all nodes do not exceed the preset fluctuation threshold, and the final diameter value of each node is determined. According to the final diameter values ​​of each node, they are rearranged in the original time order to obtain the second physical state sequence. For the diameter values ​​in the second physical state sequence, the overall fluctuation dispersion of the sequence is calculated. If the dispersion is lower than the preset dispersion threshold, then the second physical state sequence is determined to have passed the smoothing quality check and the second physical state sequence is obtained.Furthermore, the step of extracting target feature vectors from the second physical state sequence to obtain the initial operating settings of the deposition process, and extracting the initial deposition current and initial transmission speed from the initial operating settings, includes: aligning the second physical state sequence along the time axis to obtain a time-aligned state dataset; using a sliding window to extract fixed-length state segments based on the time-aligned state dataset, extracting statistical features from each state segment to obtain a target feature vector set; reading the initial operating setting record from the deposition process database; and if the process number in the initial operating setting record matches the process number corresponding to the target feature vector set, parsing the initial deposition current and initial transmission speed from the initial operating setting record to obtain the initial parameter pair. Based on the initial deposition current and initial transmission velocity in the initial parameter pair, a two-dimensional parameter space coordinate is constructed. Each feature quantity in the target feature vector set is projected onto the two-dimensional parameter space coordinate to obtain the feature projection distribution matrix. The feature projection distribution matrix is ​​fitted and trained using a support vector regression model. If the fitting error is lower than the preset current deviation threshold and velocity deviation threshold, the predicted deposition current value and the predicted transmission velocity value are output to obtain the operation parameter prediction result. Based on the operation parameter prediction result, the difference between the predicted deposition current value and the predicted transmission velocity value and the initial deposition current and the initial transmission velocity are calculated respectively. If the difference exceeds the preset parameter fluctuation threshold, the parameter correction command is triggered to extract the initial deposition current and the initial transmission velocity in the initial operation setting. Furthermore, the step of processing the target feature vector, the initial deposition current, and the initial transmission velocity through a Long Short-Term Memory (LSTM) network to output the target deposition current and the target transmission velocity includes: segmenting the target feature vector into frames using a sliding window method based on the temporal dependencies of each dimension, obtaining a temporal segment sequence with a fixed time step; feeding the temporal segment sequence into an LSM network, updating the hidden state at each time step through input gates, forget gates, and output gates, obtaining the hidden state sequence corresponding to each time step; and, based on the output of the last time step of the hidden state sequence, combining the initial deposition current and the initial transmission velocity, outputting the target deposition current and target transmission velocity through a fully connected... The output layer performs a linear transformation on the hidden layer state at the last time step to obtain the predicted value of the target deposition current and the predicted value of the target transmission velocity. If the difference between the predicted value of the target deposition current and the initial deposition current exceeds a preset threshold, it is determined that there is a deviation in the current deposition state, triggering dynamic adjustment and correction of the target deposition current. If the difference between the predicted value of the target transmission velocity and the initial transmission velocity does not exceed the preset threshold, the target transmission velocity is directly determined as the current output value. Based on the dynamically adjusted and corrected target deposition current and the determined target transmission velocity, the gradient of the parameters of each layer of the Long Short-Term Memory network is calculated through the error backpropagation algorithm, the network weight matrix and bias parameters are updated, and the target deposition current and target transmission velocity are output.Further, the step of calculating the expected deposition layer thickness of the deposition process based on the target deposition current and the target transmission speed; if the expected deposition layer thickness is less than a preset thickness threshold, then increasing the target deposition current according to a preset compensation ratio to obtain a new deposition current; if the expected deposition layer thickness is greater than or equal to the preset thickness threshold, then using the target deposition current as the new deposition current, includes: calculating the immersion time of the substrate in the treatment solution based on the target transmission speed and the length of the deposition tank passage section to obtain an immersion time value per unit substrate segment; multiplying the immersion time value by the target deposition current to obtain the electrolysis flux; and using the electrolysis flux... The process involves determining the metal deposition quality on the substrate; obtaining the substrate immersion surface area based on the metal deposition quality; dividing the metal deposition quality by the metal density and then by the substrate immersion surface area to obtain the expected deposition layer thickness for the deposition process; obtaining the thickness deviation value between the actual detected deposition layer thickness in the current deposition process and a preset thickness threshold; if the thickness deviation value is negative, multiplying the absolute value of the thickness deviation value by a preset compensation ratio to obtain the current compensation amplitude; adding the current compensation amplitude to the preset target deposition current to obtain the updated deposition current; if the thickness deviation value is non-negative, directly using the target deposition current as the updated deposition current.Furthermore, the step of performing the deposition process using the updated deposition current and the target transmission speed to obtain the target deposit includes: using a temperature sensor array to monitor the temperature distribution of the treatment liquid in the deposition tank in real time, obtaining the temperature values ​​of each monitoring point, calculating the average temperature value of the monitoring points as the current treatment liquid temperature, if the current treatment liquid temperature is lower than a preset lower limit, then activating a heating device to raise the treatment liquid temperature to a standard temperature range, if the current treatment liquid temperature is higher than a preset upper limit, then activating a cooling device to lower the treatment liquid temperature to a standard temperature range; establishing a resistivity calculation model based on the treatment liquid temperature within the standard temperature range and the updated deposition current, using the resistivity calculation model to calculate the resistivity value of the current treatment liquid, collecting the voltage difference between the substrate surface and the anode through a voltage sensor, and combining the resistivity value with the voltage... The pressure difference value is used to calculate the contact impedance of the substrate surface; the rate of change of the contact impedance is obtained; if the absolute value of the rate of change exceeds a preset rate of change threshold, it is determined that there is an uneven deposition layer on the substrate surface, and the target transmission speed is adjusted; if the absolute value of the rate of change does not exceed the preset rate of change threshold, the target transmission speed is kept constant; a laser rangefinder is used to continuously scan and measure the diameter of the deposited substrate to obtain the measurement sequence data of the substrate diameter, and the standard deviation of the diameter values ​​in the measurement sequence data is calculated. If the standard deviation is greater than a preset diameter deviation threshold, it is determined that there is an abnormal fluctuation in the substrate diameter, and the combination parameter of the update deposition current and the target transmission speed corresponding to the time when the abnormality occurs is recorded. At least one parameter of the update deposition current or the target transmission speed is adjusted according to the combination parameter to obtain the target deposit.

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

[0009] This invention addresses the problem of uneven electroplating thickness caused by pitting defects and diameter fluctuations on the surface of wires during the stretching process. It utilizes a convolutional neural network to extract surface pit features and diameter fluctuation values ​​from continuous wire images in real time, constructing a feature sequence reflecting the physical state of the wire. When the diameter fluctuation exceeds a threshold, linear interpolation smoothing is performed to eliminate abnormal interference. Then, a long short-term memory network is used to fuse the smoothed target feature vector with the initial current and wire speed parameters of the electroplating process, predicting and outputting a target electroplating current and target wire speed that match the actual state of the wire. Furthermore, the electroplating current is dynamically compensated based on whether the expected plating thickness meets the standard. Ultimately, adaptive compensation electroplating for wire surface defects is achieved, ensuring uniform plating thickness and product quality stability for the target metal wire. Attached Figure Description

[0010] Figure 1 This is a flowchart of an adaptive compensation electroplating method for surface defects in ultrafine metal wires according to the present invention. Detailed Implementation

[0011] To further understand the content of this invention, a detailed description of the invention is provided in conjunction with the accompanying drawings and embodiments. The specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention. It should also be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0012] like Figure 1 The adaptive compensation electroplating method for surface defects of ultrafine metal wires in this embodiment may specifically include:

[0013] S101. Obtain a continuous wire image of the stretching process, and extract the pixel differences in the continuous wire image through a convolutional neural network to obtain surface pit features and diameter fluctuation values.

[0014] Grayscale image sequences of continuous wire are acquired using an industrial camera on the stretching production line. Image regions with different light reflection intensities are normalized to obtain standardized images with uniform pixel value distribution. The grayscale variance of local regions in the standardized image is calculated. Regions with a variance exceeding 25 are marked as stress concentration areas, regions with a variance below 10 are marked as areas with uniform material stretching, and regions with a variance between 10 and 25 are marked as transition areas. A convolutional neural network is used to extract features from the standardized image. Using the standardized image as network input, a 3×3 Sobel operator convolution kernel slides across the image, calculating the horizontal gradient Gx and vertical gradient Gy at each pixel location. The gradient magnitude M is obtained as the square root of the sum of the squares of Gx and Gy, outputting the image pixel gradient matrix. Surface texture features are extracted from locations in the image pixel gradient matrix where the gradient magnitude exceeds a preset threshold of 30, indicating the presence of surface pits. The coordinates of the defect locations of these surface pits are then obtained. For the image region corresponding to the defect location coordinates, a radial profile is sampled every 0.1 mm along the pit boundary towards the center. The radial distribution of pixel gradients on each profile is statistically analyzed, and the cumulative change in gradient magnitude from the boundary to the center, ΣΔM, is calculated. Using a pre-calibrated depth conversion coefficient k equal to 0.015 mm per gradient unit, k is multiplied by ΔM to obtain the pit depth value d at that radial location, forming the pit depth distribution value. The Euclidean distance from the pit center to the boundary of the nearest stress concentration area is calculated. When the distance is less than 5 mm, it is determined to be a defect caused by stress concentration; a depth exceeding 0.08 mm is a severe defect; a depth between 0.03 and 0.08 mm is a moderate defect; and a depth less than 0.03 mm is a minor defect. Based on the defect location coordinates and the pit depth distribution value, the Canny edge detection algorithm is used to extract the contour boundary point set of the continuous wire in the standardized image. The Euclidean distance between adjacent boundary points is calculated to obtain the cross-sectional diameter sequence of the wire. The diameter change amplitude ΔD is obtained by the difference between the maximum value Dmax and the minimum value Dmin in the cross-sectional diameter sequence. For each cross-section, extract the set of boundary points, calculate the distance from each boundary point to the centroid of the cross-section, and take the maximum value rmax and the minimum value rmin among all distances. Calculate the roundness deviation RON according to the ISO 1101 standard, which is equal to rmax minus rmin, to obtain the cross-sectional roundness deviation. Calculate the diameter variation rate η, which is equal to ΔD divided by the average diameter Davg. Define the influence coefficient of elongation uniformity λ, which is equal to η divided by RON. When λ is less than 0.5, the material elongation uniformity has a small impact on geometric stability; λ between 0.5 and 1.2 has a moderate impact; and λ exceeding 1.2 has a significant impact.

[0015] In the quality inspection of the stretching process, the acquisition of continuous wire images relies on a high-speed industrial camera system arranged along the production line. The camera is installed at the exit of the stretching die, and a ring light source is used to achieve uniform illumination of the wire surface. The camera continuously captures images of the wire surface at a rate of thirty frames per second, forming a time-series image sequence. The light source uses a white LED array, with the illumination angle controlled at forty-five degrees to ensure that minute surface depressions produce clear shadow contrast. The image resolution is set to 2,000 x 1,000 pixels, with each pixel corresponding to an actual wire surface size of 0.02 millimeters. The processing of the continuous wire images by a convolutional neural network constitutes the core of this technical solution.

[0016] For pixel difference extraction in step S101, the convolutional neural network employs a three-layer convolutional structure to locate anomalies on the wire surface. The first layer uses 64 convolutional kernels, each 3×3 pixels in size, with a stride of 1 pixel, to extract basic texture features from the wire surface. During the sliding process, the convolutional kernels perform weighted summation on local regions; when a texture abrupt change or grayscale anomaly is detected, the feature response value increases significantly. The second layer increases the number of convolutional kernels to 128, maintaining the 3×3 pixel kernel size. After dimensionality reduction via max pooling, the feature map size is reduced to one-quarter of the original image. The third layer uses 256 convolutional kernels to extract high-level semantic features, and finally outputs the coordinate range of the defect area through a fully connected layer. During network training, a dataset containing 5000 normal wire images and 3000 defect images is used, with a learning rate of 0.001 and 50 iterations to ensure the network achieves an accuracy of over 96% in identifying minute surface defects.

[0017] The pixel difference extraction process relies on the sensitivity of the convolutional kernel to the gray values ​​of adjacent pixels. When there are pit defects on the surface of the wire, the gray value of the pixels in that area will be significantly lower than that of the surrounding normal surface, forming a sudden change in gray-level gradient. The convolutional kernel calculates the difference between the center pixel and the surrounding pixels, transforming this gradient change into an activation response in the feature map. For example, the weight matrix of an edge detection convolutional kernel may be set with positive values ​​at the center and negative values ​​at the surrounding positions, so that the edge of the pit produces a strong pair of positive and negative responses. The second convolutional layer of the network contains 128 convolutional kernels, further refining the feature representation and combining the edge information of the shallow layer into more complex shape patterns.

[0018] In one embodiment, the network introduces residual connections on top of the basic convolutional structure to preserve shallow feature information. Residual connections pass the input directly to deeper layers via skip connections, adding it to the convolutional output before feeding it into the next layer. This design solves the vanishing gradient problem in deep network training, allowing the network to stack to more than 50 layers. For pixel difference extraction of wire images, residual connections ensure that minute surface texture changes are not lost during deep propagation. The network's third and fourth convolutional layers contain 256 and 512 convolutional kernels, respectively, forming a feature hierarchy from local texture to global defect patterns. To avoid spatial accuracy loss due to pooling downsampling, the network uses convolutional layers with a stride of 2 instead of pooling operations. Stride convolutions halve the feature map size while performing feature extraction, reducing computational cost while preserving the positional information of defect edges. This operation makes the network robust to slight wire jitter and positional shifts, while ensuring pixel-level defect localization accuracy.

[0019] To ensure accurate localization of the pits, the network uses strided convolutions instead of traditional pooling layers for downsampling. Stride convolutions reduce the feature map size by increasing the stride of the convolution kernel, preserving complete spatial location information while reducing computational cost. Specifically, in the network layer requiring dimensionality reduction, the stride is set to 2, halving the feature map size while maintaining the kernel size at 3×3. This achieves both feature extraction and downsampling. Compared to pooling, which loses some pixel-level localization information, strided convolutions directly sample the original features using learnable convolution parameters, ensuring that the coordinates of the pit edges and center are accurately transmitted to subsequent network layers. The final output pit detection box coordinate error can be controlled within 2 pixels.

[0020] For example, pixel differences are quantified using a 3×3 convolutional kernel to calculate the gray-level difference between adjacent pixels. Regions with a difference exceeding 25 are marked as candidate regions. Network training uses 5000 images of wire materials with labeled pit locations. The convolutional kernel weights are adjusted using backpropagation, with an initial learning rate of 0.001. The loss function is cross-entropy loss L = -Σylog(p), where y is the ground truth label and p is the predicted probability, measuring the difference between the predicted pixel category and the ground truth label. After 200 training epochs, the convolutional kernel weights are fixed. The network outputs a pixel-level difference heatmap for newly acquired wire material images. Regions with heatmap values ​​greater than 0.8 are identified as suspected defect locations; this threshold is determined through validation set adjustments.

[0021] This system employs a fully convolutional neural network architecture. Each layer of the network uses convolutional operations to process features, with the last layer having only one output channel, corresponding to a single-class defect detection task. The fully convolutional structure preserves the spatial dimension of the feature map, outputting a segmentation mask of the same size as the input image. The value of each pixel in the mask represents the probability that the location belongs to a defect, ranging from zero to one. Based on test results from 500 samples, the detection accuracy is highest when the threshold is set between 0.45 and 0.55. Therefore, 0.5 is chosen as the binarization threshold to convert the probability mask into defect region labels. This pixel-level classification method learns defect features directly from the original image data without the need for pre-designed feature extraction rules, making it adaptable to wire inspection under different surface roughness and lighting conditions. Surface pit features are extracted based on the probability mask output by the network. The system includes a dedicated morphological analysis module for post-processing the mask, comprising a connected component labeling unit and a shape filtering unit. In an image, pits appear as dark circular or elliptical patches with pixel grayscale values ​​30 to 80 levels lower than the surrounding background, and their edges exhibit grayscale abrupt changes with a gradient magnitude greater than 20. Connected component labeling units perform region aggregation on the binarized mask, scanning the image using an eight-neighborhood connection criterion. When two pixels are adjacent in the horizontal, vertical, or diagonal direction and both have a label value of 1, they are grouped into the same connected component and assigned a unified label.

[0022] Preferably, morphological feature parameters are calculated for each candidate connected component, including area, perimeter, roundness, and aspect ratio. Roundness is defined as four times pi multiplied by the area divided by the square of the perimeter; a value close to 1 indicates that the region is nearly circular. The roundness of pit defects is typically between 0.7 and 0.95, while the roundness of linear defects such as stretch scratches is below 0.4. By setting a roundness threshold, pits can be distinguished from other types of defects. The aspect ratio is obtained by fitting an ellipse; the aspect ratio of pits is generally less than 1.5, while that of fine scratches can reach 5 or higher.

[0023] In one embodiment, the depth information of the pit feature is obtained through the gradient radial distribution method described in step S101. This method statistically analyzes pixel gradient changes along the pit boundary to establish a radial depth distribution curve from the edge to the center. The maximum decrease in the curve corresponds to the maximum depth value of the pit. The extracted pit features are output in the form of structured data, including center coordinates, area, roundness, and depth distribution values. The diameter fluctuation value is obtained through continuous measurement of the wire edge. After the convolutional neural network identifies the boundary position between the wire and the background, a measurement section is set every 0.5 mm along the longitudinal direction of the wire. At each section, the pixel coordinates of the left and right edges are located, the pixel distance between the two edges is calculated, and converted into the actual diameter value. The continuously measured diameter values ​​are arranged according to position to form a diameter sequence. The standard deviation of the diameter difference between adjacent measurement points in the sequence is calculated as the diameter fluctuation value. When the standard deviation exceeds 0.02 mm, it is determined to be an abnormal diameter fluctuation.

[0024] Specifically, the system continuously measures the diameter along the length of the wire at a density of 500 measurement points per meter. The diameter value at each measurement point is calculated by multiplying the pixel distance by the actual size of a single pixel. The measurement sequence forms time-series data, and the instantaneous rate of change of diameter fluctuation is obtained by calculating the diameter difference between adjacent measurement points. The fluctuation values ​​are quantified using standard deviation; a larger standard deviation indicates poorer diameter consistency.

[0025] In one possible implementation, the system calculates the standard deviation σ of the diameter sequence as an indicator of fluctuation stability; the smaller the standard deviation, the more stable the stretching process. Differential thresholds are set for different wire specifications: when the wire diameter D < 2 mm, a standard deviation σ < 0.03 mm is considered acceptable, and σ > 0.06 mm triggers an alarm; when 2 mm ≤ D ≤ 5 mm, a standard deviation σ < 0.05 mm is considered acceptable, and σ > 0.1 mm triggers an alarm; when D > 5 mm, a standard deviation σ < 0.08 mm is considered acceptable, and σ > 0.15 mm triggers an alarm. This tiered threshold setting is based on the processing accuracy requirements and equipment control capabilities of wires with different diameters. Smaller diameter wires are more sensitive to fluctuations and require stricter control, while larger diameter wires allow for a more relaxed fluctuation range to balance production efficiency.

[0026] It should be noted that the accuracy of diameter measurement is affected by surface reflection and vibration of the wire. To improve measurement stability, the system performs median filtering on the measurement results of three consecutive frames, taking the median value as the final output. This filtering method can effectively suppress accidental measurement errors while preserving the true diameter variation trend. The fluctuation values ​​and pit feature data together constitute the basis for wire quality assessment, providing feedback information for the dynamic adjustment of stretching process parameters.

[0027] S102. Based on the surface pit characteristics and the diameter fluctuation value, construct a first physical state sequence corresponding to the stretching process.

[0028] The pit depth values ​​at each defect location are extracted from the surface pit feature dataset. The number of pits within each meter of wire is counted to obtain the defect density value. The diameter difference between adjacent measurement sections is divided by the distance between adjacent sections to obtain the diameter change rate R = ΔD / ΔL, where ΔD is the diameter difference between adjacent sections and ΔL is the distance between adjacent sections. The diameter change rate and the corresponding pit depth value are matched according to the linear process position coordinates from the starting point to the ending point along the stretching direction. The weighted sum of the absolute value of the diameter change rate and the pit depth value is taken as the cross-sectional deformation Q = 0.6 × |R| + 0.4 × H, where H is the pit depth value. The cross-sectional deformation distribution sequence is obtained. Along the stretching direction, process location points of the wire are extracted at 0.5-second intervals. At each process location point, an infrared thermal imager is used to collect the surface temperature field distribution of the wire. A tension sensor is used to collect the tension fluctuation applied by the stretching equipment at a frequency of 100 Hz. Based on the spatial distribution characteristics of the temperature gradient in the temperature field distribution and the variation amplitude of the tension fluctuation at each process location point, finite element analysis software is used, inputting the temperature field data and tension data as boundary conditions. An elastoplastic constitutive model is used to calculate the stress distribution field values ​​of the material at the corresponding locations. For each measurement section in the cross-sectional deformation distribution sequence, the stress distribution field values ​​and temperature field distribution data at the corresponding locations are extracted. The cross-sectional deformation, stress distribution field values, temperature field distribution data, and tension fluctuation are arranged sequentially to form a 4-dimensional physical parameter vector. The corresponding process location coordinates and the timestamp of the acquisition time are added to each physical parameter vector. All physical parameter vectors are arranged in chronological order according to the timestamps. Each physical parameter vector is labeled with a quality grade identifier based on the comparison between the cross-sectional deformation and the preset deformation threshold. The preset deformation threshold is set to 3% of the standard diameter of the wire. When the cross-sectional deformation is less than the preset deformation threshold, it is labeled as qualified. When the cross-sectional deformation is greater than or equal to the preset deformation threshold but less than twice the preset deformation threshold, it is labeled as warning. When the cross-sectional deformation is greater than or equal to twice the preset deformation threshold, it is labeled as unqualified. A first physical state sequence containing physical parameter vectors, quality grade identifiers, process location coordinates, and timestamps is constructed.

[0029] The first physical state sequence is constructed based on the comprehensive processing of surface pit features, diameter fluctuation values, cross-sectional deformation, stress distribution field, and temperature field distribution. This sequence records the dynamic evolution of wire quality status during the stretching process, with time as the axis. The physical state sequence transforms discrete test results collected at fixed intervals into a time-continuous state representation using cubic spline interpolation. Each state point corresponds to a comprehensive evaluation of wire quality at a specific moment or location on the production line. A single state quantity in the sequence contains information in seven dimensions: surface defect severity, diameter consistency level, cross-sectional deformation degree, stress concentration index, temperature deviation, tension stability, and the time of state occurrence. The surface defect severity is quantified by the density and severity of pit features. Pit density is defined as the number of pits per unit length of wire surface, and severity is expressed as the ratio of pit depth to wire radius. The diameter consistency level is characterized by the standard deviation of the fluctuation values; a standard deviation less than 0.02 mm is considered high consistency, 0.02 to 0.05 mm is medium consistency, and greater than 0.05 mm is low consistency. The cross-sectional deformation is quantified by the maximum deviation between the measured cross-sectional profile and the ideal circular profile. The stress concentration index is the ratio of the peak stress to the average stress at the die exit. The temperature deviation is the absolute value of the difference between the measured temperature and the target temperature. Tension stability is measured by the coefficient of variation of tension fluctuations. The time information corresponds to the time the wire spends in the stretching die. This multi-dimensional state representation allows the physical state sequence to fully reflect the quality evolution trajectory of the stretching process. The transformation of pit features into state quantities requires the establishment of a mapping mechanism from spatial distribution features to time series states. The mapping rules include three steps: time calibration, feature aggregation, and state assignment. In the time calibration stage, the wire passes through the die at a constant speed v in the steady-state stretching region during the stretching process. The length of the acceleration / deceleration transition zone from the die inlet to the outlet is L0, and the transition zone is processed using a linear velocity model. Each pit feature data includes its position coordinate x on the wire surface. For pits with position coordinate x greater than L0, their formation time t is calculated using the formula t = L0 divided by v0 plus the difference between x and L0 divided by v, where v0 is 0.8 times the average velocity of the transition zone. For pits with location coordinates x less than or equal to L0, their formation time t is calculated using the formula t = x divided by v0. In the feature aggregation stage, a time window of 0.5 seconds is set. Features of all pits falling within the same time window are statistically aggregated, and the total number and average depth of pits within that window are calculated as input values ​​for the surface defect level at that moment. In the state assignment stage, the aggregated pit statistics are combined with simultaneously acquired data on diameter fluctuations, cross-sectional deformation, stress distribution, temperature field, and tension fluctuations to form a complete state vector for that moment. Arranging all state vectors in chronological order constitutes the first physical state sequence.

[0030] For example, if the wire stretching speed is 2 meters per second and a detected dent is located 5 meters from the starting point, then the time corresponding to this dent is 2.5 seconds. All dents detected within a given time period are arranged by time to form a time series of dent events. To quantify the degree of surface defect, the system divides the time axis into fixed-length time windows, each corresponding to a production duration of 0.5 seconds. The number of dents N and the average depth D within each time window are counted. The dent depth is extracted from the 3D contour data using an image processing algorithm, and the depth value typically ranges from 0.1 mm to 2 mm. The number of dents and the average depth are weighted and summed using weighted coefficients. The calculation formula is S = w1 multiplied by N plus w2 multiplied by D, where w1 is 0.6 and w2 is 0.4, thus obtaining the defect degree score S for that time window.

[0031] Specifically, the weighting factor for the number of dents is set to 0.6, and the weighting factor for the average area is set to 0.4, ensuring that the score reflects both the frequency of defect occurrence and the severity of individual defects. Let the normalized value for the number of dents be N, and the normalized value for the average area be A, then the dent status score S = (N × 0.6 + A × 0.4) × 10. The score value is normalized to the range of 0 to 10, where 0 represents a defect-free state and 10 represents a severe defect state. The state mapping of diameter fluctuation values ​​is achieved using a sliding statistical window. A laser diameter gauge is configured on the production line as the diameter measurement device, installed 1 meter downstream of the dent visual inspection camera, continuously outputting diameter values ​​at a density of 500 points per meter, forming a high-density measurement sequence. To align with the time window of the dent characteristics, the system performs segmented statistical processing on the diameter measurement sequence. Each 0.5-second time window corresponds to a 1-meter length of wire, containing 500 diameter measurement points. The standard deviation σ and range R of the diameter values ​​within this segment are calculated; the standard deviation reflects the overall dispersion of the fluctuation, and the range reflects the maximum fluctuation amplitude. The standard deviation and range are normalized by dividing by their respective thresholds. The standard deviation threshold is set to 0.05 mm, and the range threshold is set to 0.2 mm. The average of the normalized standard deviation and range is used to obtain the diameter consistency score for this time window: D = [(σ / 0.05 + R / 0.2) / 2] × 10. The score value is also mapped to the range of 0 to 10; the higher the value, the more severe the diameter fluctuation and the worse the consistency. Temporal and spatial alignment is a key step in constructing a unified physical state sequence. The pit visual inspection camera and the laser diameter gauge have millisecond-level time delays and centimeter-level spatial deviations. The system installs a photoelectric encoder on the production line as a synchronous trigger source. When the wire passes the encoder position, the encoder outputs a pulse signal, which is simultaneously sent to the camera controller and the diameter gauge data acquisition card through a hardware trigger circuit, ensuring that their time bases are consistent. Spatial deviations are eliminated through precise calibration of the sensor installation position. The system records a distance difference of 1 meter between the camera and the diameter gauge, and combines this with a linear speed of 2 meters per second to convert this difference into a 0.5-second time offset, which is then added to the measurement time.

[0032] For example, the camera is located 0.5 meters after the mold exit, and the diameter measuring device is located 1 meter after the exit, with a spatial interval of 0.5 meters, corresponding to a time difference of 0.25 seconds. When processing diameter data, the system advances all timestamps by 0.25 seconds to align them with the time reference of the pit features. After completing the temporal and spatial alignment, the system merges the defect severity score and diameter consistency score within the same time window into a single state vector.

[0033] In one embodiment, the state vector is represented in multi-dimensional vector form, containing five physical parameters: timestamp, cross-sectional deformation, stress distribution field, temperature field, and tension fluctuation. The timestamp records the moment corresponding to the state in seconds; the cross-sectional deformation records the degree of deformation of the steel strip cross-section; the stress distribution field records the stress distribution on the steel strip surface; the temperature field records the temperature distribution of the steel strip; and the tension fluctuation records the amplitude of tension fluctuation in the steel strip. Based on these physical parameters, a defect score and a fluctuation score are calculated as auxiliary evaluation indicators through weighted calculation. The defect score D = 0.3 × cross-sectional deformation + 0.4 × stress anomaly + 0.3 × temperature deviation, and the fluctuation score F = 0.6 × tension fluctuation + 0.4 × velocity fluctuation. Both scores range from 0 to 10 floating-point numbers. The state sequence consists of state vectors arranged in chronological order, and the sequence length depends on the monitored production duration. For 10 minutes of continuous production, a 0.5-second time window can generate 1200 state vectors, forming a physical state sequence containing 1200 time points.

[0034] Understandably, the encoding method of the physical state sequence affects the efficiency of subsequent state analysis and pattern recognition. The system uses discretization encoding to convert continuous score values ​​into state category labels. Defect scores and fluctuation scores are each divided into three levels, corresponding to minor, moderate, and severe states, respectively. A defect score less than three is classified as minor, three to seven as moderate, and greater than seven as severe. Fluctuation scores use the same threshold. The three-level divisions across the two dimensions combine to form nine possible state categories, each corresponding to a unique state code.

[0035] For example, minor defects and slight fluctuations are encoded as state one, while severe defects and severe fluctuations are encoded as state nine. State-encoded sequences are more suitable for state transition analysis and anomaly detection than the original numerical sequences, while also reducing the burden of data storage and transmission.

[0036] The first physical state sequence specifically refers to the quality status record during the stretching process. This sequence is constructed by real-time acquisition of surface integrity and diameter uniformity indicators from online inspection equipment. Surface integrity indicators include scratch depth, oxide spot area, and surface roughness value, while diameter uniformity indicators include cross-sectional diameter measurements and ellipticity deviation. These indicators are directly affected by the wear of the stretching die, lubricant coating uniformity, and stretching rate stability. The system generates the state sequence at a frequency of one set of data per meter of length, with each set containing the average and standard deviation of five measurement points. By constructing this sequence, the system can transform real-time inspection data into a structured archive containing timestamps, location identifiers, and quality parameters, providing a data foundation for process parameter optimization and defect tracing. When a surface scratch depth exceeding 0.02 mm or a diameter deviation exceeding 0.05 mm is detected, the system automatically marks the location and associates it with the then-current die usage time and lubrication status parameters.

[0037] S103. If the diameter fluctuation value in the first physical state sequence is greater than the preset fluctuation threshold, then the first physical state sequence is subjected to linear interpolation smoothing to obtain the second physical state sequence.

[0038] A sensor is used to collect diameter measurements of the target object at consecutive time points. These measurements are arranged chronologically to obtain a first physical state sequence, which is a one-dimensional numerical sequence composed of diameter measurements at each time point. The absolute value of the diameter difference between adjacent time points in the first physical state sequence is calculated as the fluctuation value. If the fluctuation value exceeds a preset fluctuation threshold of 2 mm, the time point is marked as an abnormal node. The ratio of the total number of abnormal nodes to the total number of nodes in the sequence is calculated as the distribution density. If the distribution density is greater than or equal to 0.15, a smoothing process is triggered. The marked abnormal nodes are corrected using a linear interpolation method. The diameter measurements of the two nearest valid nodes whose fluctuation values ​​do not exceed the threshold are used as reference points. The interpolated value at the abnormal node is calculated as Di = Da + t × Db - Da, where Da is the diameter value of the preceding valid node, Db is the diameter value of the following valid node, and t is the ratio of the time interval between the abnormal node and the preceding valid node to the time interval between the two valid nodes, resulting in the corrected diameter value Di. If the absolute value of the difference between the corrected diameter value and the values ​​of its adjacent nodes is still greater than the preset fluctuation threshold of 2 mm, the search range of the reference point is expanded to the second and third valid nodes before and after, and the interpolation calculation is repeated until the fluctuation values ​​between all nodes do not exceed the preset fluctuation threshold of 2 mm, thus determining the final diameter value of each node. Based on the final diameter values ​​of each node, they are rearranged in the original time order to obtain the second physical state sequence. The diameter fluctuation values ​​of all adjacent nodes in the second physical state sequence are within the preset fluctuation threshold range. For the diameter values ​​in the second physical state sequence, the ratio of the overall standard deviation to the mean of the sequence is calculated as the fluctuation dispersion σ / μ, where σ is the sequence standard deviation and μ is the sequence mean. If the dispersion is lower than the preset dispersion threshold of 0.08, the second physical state sequence is determined to have passed the smoothing quality check, and a qualified second physical state sequence is output.

[0039] The fluctuation threshold is set based on the quality specifications of the wire products and the processing tolerance requirements of downstream processes. For standard copper wire drawing production, the system presets the diameter fluctuation threshold to 0.12 mm, which corresponds to a 3% fluctuation rate of the nominal diameter of the wire. The system smooths the original diameter fluctuation data sequence. When the diameter fluctuation values ​​of any three consecutive sampling points in the sequence exceed 0.12 mm, the sequence segment is determined to have significant fluctuation characteristics, and the smoothing process needs to be initiated. The fluctuation determination adopts a sliding window scanning mechanism, with the window length set to 3 sampling points, corresponding to a production time of 1.5 seconds. The system advances the window point by point from the beginning of the sequence, moving one sampling point at a time, and calculates the arithmetic mean of the three fluctuation values ​​within the window. When the average value exceeds the threshold of 0.12 mm, the center point of the window is marked as the node to be processed, and the difference between the maximum and minimum fluctuation values ​​within the window is recorded as an auxiliary indicator of the fluctuation severity. After completing the full sequence scan, the system obtains the set of all node locations that need smoothing and the fluctuation severity values ​​of each node. The core of linear interpolation smoothing lies in reconstructing the numerical distribution of a region using stable sampling points on both sides of the fluctuation anomaly region. The system first identifies the nearest stable sampling points before and after each node to be processed. Stable sampling points are defined as those with a diameter fluctuation value less than 0.08 mm, a threshold 33% lower than the fluctuation threshold to ensure the reliability of the baseline point. Assuming a fluctuation anomaly is detected at time 5.5 seconds, the system searches forward to find a stable point at time 4.0 seconds with a diameter fluctuation value of 0.06 mm, and searches backward to find a stable point at time 7.0 seconds with a fluctuation value of 0.07 mm. A 3-second interpolation interval is formed between the two stable points, containing 6 sampling points whose fluctuation values ​​need to be recalculated. The interpolation calculation uses a piecewise linear function, taking the fluctuation value of the initial stable point as the interpolation start point and the fluctuation value of the final stable point as the interpolation end point, establishing a linear relationship between the two points. Let the initial stable point fluctuation value be V1, the final stable point fluctuation value be V2, the interpolation interval time span be Δt, and the time difference between the interpolation point and the initial point be t. Then, the smoothed fluctuation value of the interpolation point is V = V1 + (V2 - V1) × t / Δt. In this example, V1 = 0.06 mm, V2 = 0.07 mm, Δt = 3 seconds, and the time difference of the sampling point at time 5.5 seconds is t = 1.5 seconds. The smoothed fluctuation value is V = 0.06 + (0.07 - 0.06) × 1.5 / 3 = 0.065 mm. The smoothed fluctuation values ​​of all sampling points within the interpolation interval are calculated sequentially, and the abnormal values ​​at corresponding positions in the original sequence are replaced to form a locally smoothed numerical distribution. After smoothing, the system discretizes and encodes the processed fluctuation data sequence to generate a state category label for the first physical state sequence. When there are multiple discontinuous fluctuation abnormal regions in the sequence, the system performs interpolation processing independently for each region, and identifies the stable reference point of each region separately to avoid mutual interference between different abnormal regions.

[0040] During processing, if anomalies are encountered near the start or end of the sequence, and a stable reference point cannot be found before the anomaly, the fluctuation score of the first state point in the sequence is used as the starting reference. If a stable reference point cannot be found after the anomaly, the fluctuation score of the last state point in the sequence is used as the ending reference. The fluctuation score sequence after interpolation and smoothing maintains consistency with the surface pit feature scores and location information of the original sequence, updating only the diameter fluctuation score component in the triplet state vector. Since the fluctuation scores have been normalized from zero to ten in the previous steps, the interpolated values ​​automatically remain within this range. The system directly recombines the smoothed fluctuation scores with the original pit scores and location information into a triplet state vector. The newly generated state sequence is the second physical state sequence. This sequence retains the true measurement data of the stable regions in the original sequence while eliminating numerical abrupt changes in the fluctuation anomaly regions, thus enhancing the continuity and analyzability of the overall sequence. During the smoothing process, the system synchronously records a correction log for each corrected state point. The log contains three data items: the state point location index i, the original fluctuation score Voriginal, and the corrected fluctuation score Vcorrected. The difference ΔV = Vcorrected - Voriginal serves as the correction magnitude index for that location. In step S104, the correction log data is used as the observation sequence input for the Hidden Markov Model. Based on the distribution characteristics of the correction magnitude ΔV, the system identifies the accumulation area of ​​process anomalies. When the absolute value of the correction magnitude of five or more consecutive state points exceeds 1.5, it is determined that there is a persistent process deviation in that section, triggering the anomaly tracing analysis process to locate the corresponding production batch and equipment parameter records.

[0041] S104. Extract the target feature vector from the second physical state sequence, obtain the initial operation settings of the electroplating process, and extract the initial electroplating current and initial trace speed from the initial operation settings.

[0042] Step 1: For the acquisition cycle of the electroplating process, acquire the second physical state sequence from the sensor acquisition end. The second physical state sequence includes the sensing values ​​of wire tension, plating thickness, and current fluctuation records. Align the second physical state sequence along the time axis to obtain a time-aligned state dataset. Step 2: Based on the time-aligned state dataset, use a sliding window to extract fixed-length state segments. Extract statistical features from each segment, including mean, standard deviation, kurtosis coefficient, and skewness coefficient. The kurtosis coefficient characterizes the rudimentation of the data distribution, and the skewness coefficient characterizes the symmetry of the data distribution, resulting in a target feature vector set. Step 3: Read the initial operation setting record from the electroplating process database. If the process number in the initial operation setting record matches the process number corresponding to the target feature vector set, parse the initial electroplating current and initial wire speed from the initial operation setting record to obtain the initial parameter pair. Step 4: Based on the initial electroplating current and initial trace speed in the initial parameter pair, construct a two-dimensional parameter space coordinate system. Establish a coordinate system with the initial electroplating current as the horizontal axis and the initial trace speed as the vertical axis. Use the mean trace tension of each state segment in the target feature vector set as the horizontal axis correction value and the standard deviation of the coating thickness sensing value as the vertical axis correction value. Project each feature quantity to the two-dimensional parameter space through coordinate mapping relationship to obtain the feature projection distribution matrix. Each row of this matrix corresponds to the projection coordinate of a time window. Step 5: The feature projection distribution matrix is ​​fitted and trained using a support vector regression model. The coordinates of the feature projection distribution matrix are used as input, and the actual electroplating current and trace speed at the corresponding time are used as output labels. The support vector regression model uses a radial basis function kernel, with the penalty parameter C set to 100 and the kernel function parameter γ set to 0.01. It is optimized by a grid search method. After training, the root mean square error (RMSE) between the predicted and actual values ​​is calculated, where RMSE = √(Σ(y predicted - y actual)² / n), and n is the number of samples. If the RMSE of the current is less than 2 amperes and the RMSE of the speed is less than 0.5 meters per minute, the predicted values ​​of the electroplating current and trace speed are output to obtain the predicted results of the operating parameters. Step 6: Based on the predicted operating parameters, calculate the differences between the predicted electroplating current and the predicted trace speed and the initial electroplating current and initial trace speed, respectively. If the absolute value of the current difference exceeds 5 amperes or the absolute value of the speed difference exceeds 1 meter per minute, a parameter correction command is triggered. A weighted fusion method is used for correction. The corrected electroplating current = initial electroplating current × 0.3 + predicted electroplating current × 0.7, and the corrected trace speed = initial trace speed × 0.3 + predicted trace speed × 0.7, thus obtaining the corrected electroplating operating parameter set. Step 7: Based on the corrected electroplating operating parameter set, sequentially issue the electroplating current control command and the trace speed control command to the electroplating execution unit according to the process cycle, obtaining the final operating control command set for the current process cycle.

[0043] The feature vector extraction module receives a second physical state sequence as the input data source. This sequence contains three-dimensional state parameters such as the wire diameter measurement, surface roughness quantitative score, and coating thickness deviation at the sampling time. The system first performs a Fourier frequency domain transform on the time-series data in the sequence, converting the state fluctuations in the time domain into an energy distribution spectrum in the frequency domain. By extracting the top 5 dominant frequency components with the highest energy proportions in the spectrum, a frequency domain feature subset is constructed. These 5 dominant frequency components correspond to 0.5 Hz, 1.2 Hz, 2.8 Hz, 4.1 Hz, and 5.6 Hz, respectively, with energy proportions of 32%, 24%, 18%, 13%, and 9%, respectively. The remaining 4% of the energy is distributed among other frequency components. The system then calculates the statistical characteristics of the sequence within a time window, using a 10-second sliding window that moves from the start to the end of the sequence. Each 5-second slide creates a 50% overlap. Four statistical measures are calculated for the data within the window: mean, standard deviation, kurtosis, and skewness. In a typical window, the mean diameter measurement is 3.98 mm, the standard deviation is 0.04 mm, the kurtosis is 2.7, and the skewness is -0.15. The system further extracts extreme point features from the sequence by detecting the location and values ​​of local maxima and minima, recording three parameters: the average time interval between adjacent extreme points, the average amplitude change, and the total number of extreme points. Within a 60-second analysis window, 7 maxima and 8 minima were identified, with an average time interval of 4.3 seconds and an average amplitude change of 0.18 mm between adjacent extreme points. The five frequency domain features, four statistical features, and three extreme value features extracted above are weighted and summed according to preset weight coefficients to form a target feature vector with a dimension of 12. The first to fifth dimensions correspond to the energy values ​​of the five main frequency components, and the weight coefficients are 0.065, 0.061, 0.058, 0.056, and 0.050, respectively. The sixth to ninth dimensions correspond to the mean, standard deviation, kurtosis coefficient, and skewness coefficient, and the weight coefficients are 0.092, 0.088, 0.085, and 0.085, respectively. The tenth to twelfth dimensions correspond to the average time interval, average amplitude change, and total number of extreme points, and the weight coefficients are 0.085, 0.085, and 0.080, respectively. The numerical range of each component in this vector is normalized to the interval between 0 and 1 through minimum and maximum value standardization. The initial operation settings for the electroplating process are stored in the process parameter database of the production management system. The database uses product model and wire specification as index keys. The system queries the corresponding process parameter record based on the product identifier of the current processing task and reads the values ​​of the initial electroplating current field and the initial wire speed field from the record.For copper core wire with a diameter of 4 mm, the database records show that the initial electroplating current is set to 250 amperes. This current value is calculated based on Faraday's law of electrolysis, where I represents the electroplating current, v represents the wire speed set to 1.2 meters per minute, S represents the surface area per unit length of the wire as 5.02 square decimeters per meter, δ represents the target plating thickness as 8 micrometers, ρ represents the copper density as 8.96 grams per cubic centimeter, M represents the copper molar mass as 63.5 grams per mole, n represents the copper ion valence state as 2, F represents the Faraday constant as 96485 coulombs per mole, and η represents the current efficiency as 0.92. According to the formula I = v × S × δ × ρ × n × F ÷ M ÷ η, the calculated current is 250 amperes.

[0044] The initial wire speed field is set to 12 meters per minute. This speed parameter is related to the length of the electroplating tank and the desired coating uniformity. With an effective tank length of 6 meters, the wire spends 30 seconds in the tank at a speed of 12 meters per minute, ensuring sufficient electrochemical reaction. The speed setting must match the preceding stretching process, with the speed difference between the two processes controlled within 0.5 meters per minute to avoid sudden tension changes exceeding 20 Newtons at the junction. The system uses the initial electroplating current of 250 amperes and the initial wire speed of 12 meters per minute as the baseline input for subsequent parameter optimization modules. These two parameters, along with the extracted target feature vector, constitute the complete input dataset for the electroplating quality prediction model. The model analyzes the mapping relationship between the wire's physical state and the initial process parameters reflected in the feature vector to identify parameter configuration patterns that may lead to coating quality deviations.

[0045] S105. Process the target feature vector, the initial electroplating current, and the initial trace speed through a long short-term memory network, and output the target electroplating current and the target trace speed.

[0046] Step 1: Acquire the raw time-series signals collected during the electroplating process, including the electroplating tank potential, plating solution concentration sensor readings, and wiring tension values. Normalize these multi-source signals to obtain the target feature vector. Step 2: Based on the temporal dependencies of each dimension in the target feature vector, use a sliding window method to segment the target feature vector into frames, obtaining a time-series segment sequence with a fixed time step length, which serves as the input to the Long Short-Term Memory (LSTM) network. Step 3: Feed the time-series segment sequence into the LSM network. Update the hidden state at each time step using input gates, forget gates, and output gates. The memory unit retains long-range dependency information related to the electroplating state changes at each time step, obtaining the hidden state sequence corresponding to each time step. Step 4: Based on the output of the last time step of the hidden state sequence, combined with the initial electroplating current and initial wiring speed, perform a linear transformation on the hidden state at the last time step through a fully connected output layer to obtain the predicted values ​​of the target electroplating current and target wiring speed. Step 5: If the absolute value of the difference between the predicted target plating current and the initial plating current exceeds 15% of the initial plating current, it is determined that there is a deviation in the current plating state. A proportional-integral-derivative (PID) control algorithm is used to dynamically adjust and correct the target plating current, with the proportional coefficient set to 0.8, the integral coefficient to 0.05, and the derivative coefficient to 0.02. The corrected current value is taken as the final target plating current. If the absolute value of the difference between the predicted target trace speed and the initial trace speed does not exceed 10% of the initial trace speed, the predicted target trace speed is directly taken as the final target trace speed. Step 6: Based on the dynamically adjusted and corrected target plating current and the determined target trace speed, the gradient of each layer parameter of the Long Short-Term Memory (LSTM) network is calculated using the backpropagation algorithm. The network weight matrix and bias parameters are updated to obtain the final output values ​​of the converged target plating current and target trace speed.

[0047] The input layer of the Long Short-Term Memory (LSTM) network receives a 14-dimensional composite vector, which is formed by concatenating a 12-dimensional target feature vector with 2-dimensional initial process parameters. The 12-dimensional target feature vector includes four electroplating tank potential signals, four plating solution concentration sensor readings, and four trace tension values. These are processed using a minimum-maximum normalization method, with the normalization formula x' = (x - xmin) / (xmax - xmin), where x is the original value, and xmin and xmax are the historical minimum and maximum values ​​of the feature, respectively. The historical range of the electroplating current is 200 to 400 amperes, with the initial value of 250 amperes normalized to 0.625. The historical range of the trace speed is 8 to 25 meters per minute, with the initial value of 12 meters per minute normalized to 0.48. The network hidden layer is configured with three stacked Long Short-Term Memory (LSTM) unit layers, each containing 64 neurons. The initial bias of the forget gate is set to 1.5 to enhance long-term dependent memory capabilities. The input gate uses a hyperbolic tangent activation function to constrain the unit state within the range of -1 to +1. An attention mechanism module is inserted after the second LSM layer. This module calculates attention weights using a learnable query matrix Q, key matrix K, and value matrix V. The weight matrix has a dimension of 20 x 20, corresponding to the correlation between 20 time steps. After training, the weight coefficient of time steps strongly correlated with coating thickness fluctuations reaches 0.78. The system feeds the input vector sequence of 20 consecutive time steps into the network. Each time step corresponds to a 3-second process state sampling period. The first LSM layer processes the original input sequence and outputs an intermediate hidden state tensor with a dimension of 20 x 64. The second layer receives this tensor and filters key temporal dependencies through a gating mechanism. The third layer, Long Short-Term Memory (LSTM), performs deep abstraction of prior features. The number of neurons in the output layer is set to 2, corresponding to the two optimization parameters: target electroplating current and target trace speed. The network uses mean squared error as the loss function and iteratively updates the weight parameters through backpropagation. The learning rate is set to 0.001, and the batch size during training is 32 samples. After 500 iterations, the loss value converges to 0.0047. During the inference phase, the network output layer yielded normalized values ​​of 0.84 and 0.313. Based on the inverse normalization formula x = x' × (xmax - xmin) + xmin, the target electroplating current was calculated to be 0.84 × 200 + 200, equaling 268 amperes, a 7.2% increase from the initial value. The target trace speed was 0.313 × 17 + 8, equaling 13.321 meters per minute. After safety margin correction by the dynamic control module, this was reduced to 10.5 meters per minute, a 12.5% ​​decrease from the initial setting. This correction value is the final output result after correction via backpropagation and dynamic control in step five above. Attention mechanism analysis showed that the cross-correlation coefficient between plating thickness deviation and trace speed fluctuation was -0.63. Reducing the speed helps extend the immersion time of the wire in the electrolyte from 30 seconds to 34.3 seconds.

[0048] S106. Calculate the expected coating thickness of the electroplating process based on the target electroplating current and the target trace speed.

[0049] Based on the target trace speed and the effective length of the electroplating tank, the immersion time of the trace in the plating solution is calculated by dividing the effective length of the electroplating tank by the target trace speed, yielding the immersion time value t per unit trace segment, in seconds. The immersion time value t is multiplied by the target electroplating current I to obtain the electrolytic flux Q = I × t, in coulombs. The electrolytic flux represents the total charge received by the trace during the electroplating process. Based on the electrolytic flux, the metal deposition mass m = Q × M ÷ F ÷ n is calculated according to Faraday's law of electrolysis, where F is the Faraday constant (96500 coulombs per mole), M is the molar mass of the metal, and n is the valence state number of the metal ions. For copper plating, M is 63.5 g / mol, n is 2, and the metal density ρ is 8.96 g / cm³. The immersion surface area S of the trace is obtained, calculated based on the trace width and the effective length of the electroplating tank. Divide the metal deposition mass m by the metal density ρ and then by the immersion plating surface area S of the trace to obtain the expected plating thickness d for the electroplating process: d = m ÷ ρ ÷ S.

[0050] In one embodiment, the calculation of the expected coating thickness is based on a quantitative relationship model established by Faraday's law of electrochemical deposition. This model converts the target electroplating current into the amount of metal ion reduction per unit time, and determines the effective duration of the wire in the electrolytic cell by combining the wire speed.

[0051] The core parameters of the coating thickness calculation model include current throughput, electrochemical equivalent of metal ions, current efficiency coefficient, and wire surface area. The electrochemical equivalent represents the mass of metal deposited per coulomb of charge; for copper coatings, this value is 0.329 mg / coulomb, and for nickel coatings, it is 0.304 mg / coulomb. In the actual plating process, the total charge Q is equal to the product of current I and time t, i.e., Q = I × t. Therefore, given the current and plating time, the mass of deposited metal can be calculated as m = electrochemical equivalent × Q × current efficiency coefficient. The current efficiency coefficient reflects the proportion of current actually involved in metal deposition. It is affected by factors such as electrolyte temperature, additive concentration, and cathode polarization. For every 10 degrees Celsius increase in temperature, the current efficiency increases by 3% to 5%. However, the formation of an adsorption layer on the cathode surface by brightener molecules in the additive reduces efficiency by approximately 8%.

[0052] For example, after the system acquires the target electroplating current of 268 amperes, it first calculates the current density distribution on the surface of the wire. The wire is a cylindrical conductor with a diameter of 1.5 mm, and its circumferential surface area, calculated in length units, is 47.12 square centimeters per meter, or 0.4712 square decimeters per meter. Assuming the length of the electroplating section of the wire is 3.5 meters, the total surface area is 1.649 square decimeters. When the current is uniformly distributed, the current density reaches 162.5 amperes per square decimeter. However, in actual production, the distribution of the current on the cylindrical surface exhibits non-uniform characteristics. The current density is higher on the anode side, reaching 186.9 amperes per square decimeter, while it drops to 145.6 amperes per square decimeter on the opposite side. This difference causes a 15% fluctuation in the circumferential thickness of the plating layer. To correct this phenomenon, the system introduces a geometric shape factor for weighted average calculation of current density. This factor obtains the electric field distribution of the wire under a specific anode configuration through finite element simulation. The circumference is divided into sixteen sector regions and assigned weight coefficients to each. The weight of the four regions on the facing side is 1.15, the weight of the four regions on each side is 1.00, and the weight of the four regions on the facing side is 0.90. The weighted equivalent current density ρ_eff = (4 × 186.9 × 1.15 + 8 × 162.5 × 1.00 + 4 × 145.6 × 0.90) ÷ 16 = 164.8 amperes per square decimeter, where ρ_eff is the equivalent current density.

[0053] The wire travel speed directly determines the residence time of the wire in the electrolyte wetting zone, which is the time window for metal ions to complete the discharge reaction on the cathode surface. The effective working length of the electrolytic cell is designed to be 3.5 meters. When the target wire travel speed is set to 10.5 meters per minute, the time for the wire to pass through this zone is calculated to be 20 seconds. However, considering the lag time required for the establishment of the electrolyte turbulent diffusion layer, the actual effective deposition time needs to deduct the influence of the transition zone between the inlet and outlet sections. Within the 0.4-meter inlet section, electrolyte convection is not yet fully established, and the ion mass transfer rate only reaches 60% of the steady-state value. This rate directly determines the deposition efficiency of 60%. Within the 0.3-meter outlet section, the current distribution is affected by boundary effects, and the deposition efficiency is 75%. The middle section of 2.8 meters is the fully effective deposition zone. The weighted effective time was calculated using a piecewise integration method. The travel velocity was v = 0.175 m / s. The passage time t1 of the inlet section was 0.4 ÷ v = 2.29 seconds, with an effective contribution of t1 × 0.6 = 1.37 seconds. The passage time of the fully effective section was t2 = 2.8 ÷ v = 16 seconds, with an effective contribution of 16 seconds. The passage time of the outlet section was t3 = 0.3 ÷ v = 1.71 seconds, with an effective contribution of t3 × 0.75 = 1.29 seconds. The cumulative weighted effective deposition time was 18.66 seconds.

[0054] In one possible implementation, the coating quality is calculated using a modified Faraday formula, substituting the target plating current of 268 amperes, the weighted effective deposition time of 20.0 seconds, the electrochemical equivalent of 0.000091 g / ampere-second, and the current efficiency coefficient of 0.85 into the calculation process. The current throughput Q is 268 × 20.0 = 5360 ampere-seconds, and the theoretical deposition mass mtheoretical = Q × K = 5360 × 0.000091 ≈ 0.488 g. Considering the current efficiency correction, the actual deposition mass mactual = 0.488 × 0.85 ≈ 0.415 g. The copper of this mass is uniformly distributed on the surface of the wire. Based on the density of copper (8.9 grams per cubic centimeter), the volume V = 0.415 ÷ 8.9 ≈ 0.0466 cubic centimeters. Distributing this to the product of the circumferential surface area of ​​4.71 square centimeters per meter of wire and the length of the wire passing through the electrolytic cell (3.5 meters), the total surface area S = 4.71 × 3.5 = 16.49 square centimeters. Therefore, the average coating thickness d = 0.0466 ÷ 16.49 ≈ 0.00283 centimeters, or 28.3 micrometers.

[0055] The current efficiency coefficient is dynamically calculated based on real-time operating condition monitoring data. The system acquires an electrolyte temperature sensor reading of 55 degrees Celsius. This sensor is a PT100 platinum resistance thermometer with an accuracy of ±0.1 degrees Celsius, installed 20 cm below the cathode plate in the middle of the electrolyte bath, and calibrated quarterly. The cathode potential deviates by -0.45 volts relative to the saturated calomel reference electrode. The additive consumption rate is monitored by the replenishment pump flow meter as 0.2 liters per hour. Combining this with the total electrolyte volume of 500 liters and the initial addition amount, the current brightener concentration is calculated to be 1.2 grams per liter. The temperature correction factor F_T is calculated based on a reference temperature of 45 degrees Celsius. The current temperature of 55 degrees Celsius corresponds to a temperature difference ΔT of 10 degrees Celsius. For every 1 degree Celsius increase in temperature, the efficiency improves by 0.003. The correction factor F_T = 1 + 0.003 × 10 = 1.03. Here, the lower limit of the temperature efficiency improvement is used for a conservative estimate. The increased negative shift in cathode potential reflects intensified polarization of the cathode reduction reaction, leading to an increased proportion of side reactions such as hydrogen evolution. The baseline efficiency of 0.85 needs to be multiplied by a reduction factor F_E = 0.98. The relationship between brightener concentration and the correction factor is set in segments: the correction factor F_A is 0.92 when the concentration is above 1.5 g / L; the correction factor linearly increases when the concentration is between 1.0 and 1.5 g / L; the current concentration of 1.2 g / L corresponds to F_A = 0.96. After comprehensive correction, the actual current efficiency coefficient η = 0.85 × 1.03 × 0.98 × 0.96 = 0.822. This value is used for real-time calculation of subsequent coating thickness and power consumption.

[0056] Preferably, the system introduces an edge thickness compensation algorithm to handle the plating buildup at the wire ends. The current density in the end region experiences a tip discharge effect due to geometric abrupt changes, resulting in a thickness 20% to 30% greater than the center section. The compensation strategy adjusts the anode plate spacing of the corresponding electrolytic cell section at the end, increasing the standard spacing from 8 cm to 11 cm, thereby reducing the electric field strength in this area. Considering the edge boundary electric field distortion effect, the effective spacing is 0.85 times the physical spacing; the standard effective spacing is 6.8 cm, and the adjusted effective spacing at the end is 9.35 cm. Based on the inverse relationship between current density j and effective spacing d, j = σE = σU / d, where σ is the electrolyte conductivity, U is the electrolytic cell voltage, and E is the electric field strength, the end current density is reduced from 7.1 amperes per square decimeter to 5.2 amperes per square decimeter, a reduction of 26.8%, with the thickness difference controlled within 8%. Simultaneously, the system monitors the impact of wire tension fluctuations on coating uniformity. When the tension fluctuates from the standard value of 120 Newtons to more than ±15 Newtons, the immersion depth of the wire in the electrolyte shifts by millimeters, resulting in changes in the effective working area. The system uses a servo tension controller to limit the fluctuation amplitude to within ±5 Newtons to ensure immersion depth stability.

[0057] For example, in a certain batch of production, the actual measured coating thickness ranged from 2.75 micrometers to 2.91 micrometers, with a deviation from the expected calculated value of 2.83 micrometers within ±3%, verifying the accuracy of the calculation model. The sources of deviation were analyzed and quantified into three factors: surface roughness differences in the wire accounted for 55% of the contribution, electrolyte composition drift accounted for 40%, and the surface condition of the anode plate accounted for 5%. Surface roughness was characterized by the Ra value; when Ra increased from 0.8 micrometers to 1.2 micrometers, the local current density increased by 8% to 12%, resulting in a corresponding increase in coating thickness. To address the main sources of deviation, the system monitored the copper ion concentration and sulfuric acid concentration in the electrolyte online. When the copper ion concentration deviated from the standard value of 45 grams per liter by more than 2 grams per liter, compensation was achieved by correcting the conductivity coefficient κ in the model. The correction amount for κ was the 0.85th power of the ratio of the measured concentration to the standard concentration, ensuring that the accuracy of the expected thickness calculation met the process control requirements.

[0058] S107. If the expected coating thickness is less than a preset thickness threshold, the target electroplating current is increased according to a preset compensation ratio to obtain an updated electroplating current. If the expected coating thickness is greater than or equal to the preset thickness threshold, the target electroplating current is used as the updated electroplating current.

[0059] The thickness deviation value between the actual detected coating thickness in the current electroplating process and the preset thickness threshold is obtained. If the thickness deviation value is negative, the absolute value of the thickness deviation value is multiplied by the preset compensation ratio to obtain the thickness compensation amount. The thickness compensation amount is multiplied by the thickness current conversion coefficient to obtain the current compensation amplitude. The thickness current conversion coefficient is pre-calibrated according to the electroplating process parameters, and the typical value is 0.08 amperes per micrometer to 0.15 amperes per micrometer. The preset target electroplating current is added to the current compensation amplitude to obtain the updated electroplating current. If the thickness deviation value is zero or positive, the target electroplating current is directly used as the updated electroplating current. A current sensor is used to collect the updated electroplating current in real time. The collection period is set to 3 seconds and the sampling frequency is set to 10 Hz. 30 current values ​​are recorded within the collection period. The difference between the maximum and minimum values ​​of these current values ​​is calculated to obtain the current fluctuation amplitude. The fluctuation threshold is calculated to be equal to the target electroplating current multiplied by 0.05. The current fluctuation amplitude is compared with the fluctuation threshold. If the current fluctuation amplitude is less than the fluctuation threshold, the updated electroplating current is determined to be in a stable state. The updated electroplating current is recorded as the standard current value of the current process. According to the plating type and thickness requirements of subsequent processes, the corresponding process adaptation coefficient is retrieved from the process parameter database. The typical value range of the process adaptation coefficient is 0.8 to 1.2. The standard current value is multiplied by the process adaptation coefficient to obtain the adjusted current setting value. The adjusted current setting value is output to the current control module of the subsequent electroplating process as the basis for setting the target electroplating current of the next process.

[0060] In one embodiment, the thickness threshold is set based on a multi-level judgment standard established according to the technical specifications and process stability requirements of GB / T3953-2009 for coated wires for communication. This threshold is not a fixed value but a range that is dynamically adjusted according to the wire specifications, the type of coating metal, and the application scenario. For copper coatings used in communication cables, the thickness threshold is usually set between 3.0 micrometers and 3.5 micrometers. The lower limit of this range ensures that the coating resistivity does not exceed 1.724 × 10^-8 ohm-meters, thus meeting the national standard requirements for conductivity. The upper limit controls material costs and avoids the risk of microcracks caused by coating internal stress exceeding 150 MPa. The system extracts the target coating thickness specification of the current batch of wires from the product database. When the specification value is 3.2 micrometers, the process control threshold is taken as 93% to 97% of the specification value. The calculated values ​​are T_min = 3.2 × 0.93 = 2.976 micrometers and T_max = 3.2 × 0.97 = 3.104 micrometers, where T_min is the lower limit of the threshold and T_max is the upper limit of the threshold. The lower threshold was set to account for fluctuations in coating uniformity. Statistical process control data from 12 consecutive months showed that the circumferential thickness difference of the wire fluctuated between 8% and 12% in actual production. The lower threshold was set to ensure that even the thinnest part met the minimum performance requirements. The upper threshold, on the other hand, reserved space for thickness increases in subsequent processes. During the subsequent cleaning and drying process, copper atoms on the coating surface undergo a trace oxidation reaction with residual water molecules to generate a thin layer of cuprous oxide. This oxide layer has a lattice constant greater than that of metallic copper, resulting in an apparent thickness increase of 0.05 micrometers. Therefore, the final finished product thickness control upper limit was set to T_max - 0.05 = 3.104 - 0.05 = 3.054 micrometers, while the lower limit remained at 2.976 micrometers, ensuring that the finished product thickness fell within the range of 2.976 to 3.054 micrometers and did not exceed the ±5% allowable deviation range of the original specification of 3.2 micrometers.

[0061] Specifically, the system uses a current density and time integration method to calculate the expected coating thickness. Input parameters include a current density of 0.5 amperes per square decimeter, an electroplating time of 45 minutes, a current efficiency of 92%, an electrochemical equivalent of nickel of 0.304 milligrams per coulomb, and a nickel density of 8.9 grams per cubic centimeter. The calculated expected coating thickness is 2.976 micrometers. This value is compared with the lower limit of the thickness threshold of 2.976 micrometers, and it is determined that the expected thickness is exactly equal to the lower limit of the threshold, which is at the critical state of the coating quality requirements. The system further calculates that the deviation between this thickness and the target thickness of 3.2 micrometers is -0.224 micrometers, with a deviation rate of 7%. The system's built-in deviation level classification rules define a deviation of 3% to 5% as slightly insufficient, 5% to 10% as moderately insufficient, and more than 10% as severely insufficient. The current deviation of 7% belongs to the moderately insufficient level, and the electroplating parameters need to be adjusted to improve the coating thickness.

[0062] The determination of the compensation ratio relies on the correlation model between the missing amount of coating thickness and the electroplating current response characteristic, and this model establishes the mapping relationship between the current increment and the thickness increment through regression analysis of historical production data. The system collects 2300 batches of production records of the same wire specification in the past 3 months, and the samples cover two seasons of spring and summer, including working conditions with an ambient temperature of 18 to 32 degrees Celsius and an electrolyte concentration fluctuation of ±5%. 1400 batches of data in which the coating thickness is more than 0.05 micrometers lower than the design value and the thickness after current compensation reaches the range of ±0.03 micrometers of the design value are screened out as training samples, and the remaining 900 batches include 600 batches of cases with no insufficient thickness and 300 batches of cases that are still unqualified after compensation, which are used for model boundary verification. Each sample contains four parameters: initial current value I0, thickness missing amount Δd, compensation current increment ΔI, and actual thickness increment Δd' after compensation. A multiple linear regression method is used to establish the model ΔI=α×Δd+β×I0+γ, where α is the reference compensation coefficient, β is the initial current influence coefficient, and γ is a constant term. Regression analysis gives α=32% / micrometer, β=0.008, γ=0.15 amperes, the model coefficient of determination R²=0.92, and the residual standard deviation is 0.18 amperes. The reference compensation coefficient α=32% / micrometer means that for every 1 micrometer of missing thickness, the current needs to be increased by 32%, that is, a missing thickness of 0.1 micrometer corresponds to a current increase of 3.2%. Considering the nonlinear saturation effect under a large missing amount, a correction coefficient η is introduced. When the missing rate r=Δd / design thickness exceeds 5%, η is calculated according to a piecewise linear function: when r≤5%, η=1.0; when 5%<r≤10%, η=1.0-3×(r-5%); when r>10%, η=0.85. When the current thickness missing amount is 0.146 micrometers and the design thickness is 3.2 micrometers, the missing rate r=0.146 / 3.2=4.56%, corresponding to the correction coefficient η=1.0. The reference compensation ratio is 0.146×32%=4.67%, and the final compensation ratio adjusted by the correction coefficient is 4.67%×1.0=4.67%, which is rounded to 4.7% according to the system accuracy requirement. The mean prediction error of the model on the verification set is 0.21 amperes, and the deviation between the actual thickness increment and the predicted value of 90% of the samples is within ±0.05 micrometers.

[0063] In one possible implementation, the current increase is calculated by multiplying the target plating current of 268 amps by a compensation coefficient of 0.0458 to obtain a current increment of 12.27 amps. The updated plating current is the original current plus the increment, i.e., 280.27 amps. The compensation coefficient k is defined as k equal to the thickness compensation rate η divided by the current efficiency coefficient ε, where the current efficiency coefficient ε is 0.95, representing the ratio of the actual deposition amount to the theoretical deposition amount. The thickness compensation rate η is equal to 0.0435, therefore k is equal to 0.0458, a dimensionless coefficient. The application of the current increment follows a gradual adjustment principle. The system breaks down the increment into three step-like increases to avoid concentration polarization that could lead to coating burn-out. The number of steps is selected based on the Nernst diffusion layer response characteristics. When the amplitude of a single current jump exceeds 50% of the total increment, the time it takes for the diffusion layer thickness δ to compress from the initial 0.05 mm to 0.02 mm exceeds 15 seconds, leading to ion depletion. Three steps can control the single jump within 40% of the total increment, ensuring that the diffusion layer recovers within 8 seconds. The first step increases the total increment by 30%, or 3.68 amperes. After maintaining this current for 5 seconds, the coating growth rate is monitored. According to Faraday's law, the coating thickness growth rate v is proportional to the current density J, i.e., v equals M multiplied by J divided by n multiplied by F multiplied by ρ, where M is the molar mass of the metal, n is the charge number, F is the Faraday constant, and ρ is the density. The first step increases the current by 30% of the total increment, but it acts on a local high current density area, causing the current density in that area to increase by 60%, resulting in a 60% increase in the rate. If the rate increase meets the target, the second step increases the current by 40%, or 4.91 amperes. The third step makes up the remaining 30%, or 3.68 amperes. The step interval time is set according to the diffusion layer recovery time τ, which is equal to the square of δ divided by twice the diffusion coefficient D. The diffusion coefficient D is 5 times 10 to the power of negative 6 square centimeters per second. It takes τ equal to 6.25 seconds for the diffusion layer thickness δ to recover from the compressed state of 0.02 mm to the stable state of 0.05 mm. The actual interval is set to 5 to 8 seconds and adjusted according to the change of diffusion coefficient in the electrolyte temperature range of 45 degrees Celsius to 65 degrees Celsius. For every 10 degrees Celsius increase in temperature, the diffusion coefficient increases by 1.3 times and the interval is shortened to 5 seconds.

[0064] It should be noted that when the expected plating thickness is greater than or equal to the thickness threshold, the system determines that the current electroplating parameters meet the process requirements and do not need adjustment. The target electroplating current is then directly output as the updated electroplating current to the power control module. In this case, the system will still record the correspondence between the current current value and the actual plating thickness, storing this data in the historical database for subsequent optimization and updates of the compensation ratio model.

[0065] Preferably, after updating the electroplating current generation, the system initiates a closed-loop verification mechanism, re-executing the coating thickness calculation process to predict the compensated coating thickness. The verification calculation uses the same Faraday's law model as the initial calculation, substituting the updated electroplating current of 280.27 amperes into the model, while keeping other parameters such as trace speed, current efficiency coefficient, and effective deposition time unchanged. The calculated expected thickness after compensation is 2.96 micrometers. The system then compares this expected thickness with the lower threshold of 2.976 micrometers. If it is still less than the threshold, the compensation process is executed again, but this time the compensation ratio needs to be multiplied by an attenuation factor of 0.7 to prevent over-compensation. The current increment for the second compensation is 70% of the initial increment, and the upper limit for the cumulative current increment of multiple iterations of compensation is set to 15% of the initial current. This upper limit is determined based on three constraints: the maximum output capacity of the power supply is 120% of the rated current; the maximum current carrying capacity of the electrolyte at a temperature of 45 degrees Celsius corresponds to a current density not exceeding 3.5 amperes per square decimeter; and the uniformity requirement that the thickness difference between the coating edge and the center region will exceed 5% when the current increase exceeds 15%. When the cumulative increment reaches the upper limit but the thickness still does not meet the standard, the system triggers an anomaly warning and automatically suspends the current batch production, controls the current parameters, and records the process data at the moment of the anomaly. Upon receiving the warning, the operator follows the standard inspection procedure to sequentially measure whether the copper ion concentration in the electrolyte is within the range of 45 to 55 g / L, whether the sulfuric acid concentration is within the range of 180 to 220 g / L, check for the presence of a passivation layer on the anode plate surface, and confirm that the distance between the cathode trace and the anode plate meets the standard value of 25 mm. After the inspection, targeted adjustments are made based on the deviations: if the copper ion concentration is below 45 g / L, copper sulfate solution is added to the target concentration; if the anode plate is passivated, a new anode plate is replaced; if the distance deviation exceeds 2 mm, the fixture position is recalibrated. After adjustment, the system enters the recovery process, first running at 80% of the initial current for 10 minutes to conduct a stability test. Thickness is measured on a sample taken from the test section. If the thickness reaches 2.976 micrometers or more and the uniformity deviation is less than 5%, the production lock is released, the electroplating current is recalculated, and normal production resumes. The closed-loop verification mechanism typically iterates no more than three times. Statistical data shows that 92% of cases with insufficient thickness can meet the standard after the first compensation, 7% of cases require a second compensation, and only 1% require a third compensation or trigger an anomaly warning.

[0066] In a certain batch of production, the initial target electroplating current was 250 amps. Based on the linear relationship model between current and thickness, the expected plating thickness was calculated to be 2.65 micrometers. The product specifications for this batch required a target thickness of 2.9 micrometers, with an allowable deviation range of 2.7 to 3.1 micrometers. The system detected that the expected thickness of 2.65 micrometers was below the lower limit of 2.7 micrometers, a shortfall of 0.25 micrometers. According to the compensation model, the missing thickness ΔT = 0.25 micrometers, and the baseline thickness increment ΔT0 = 0.1 micrometers corresponds to a baseline current increase of 3.2%. The calculated compensation ratio is ΔT / ΔT0 × 3.2% = 0.25 / 0.1 × 3.2% = 8%. The system calculated the current increment as 250 × 8% = 20 amps, and updated the electroplating current to 270 amps. Recalculating according to the linear relationship, the current increase of 8% corresponds to a thickness increment of 0.25 micrometers. After the update, the expected thickness is 2.65 + 0.25 = 2.9 micrometers, which is within the threshold range of 2.7 to 3.1 micrometers, and the verification is successful. The updated current was officially applied to production control. Due to uneven current distribution and slight fluctuations in plating solution concentration during the electroplating process, the measured plating thickness at different locations in this batch ranged from 2.88 to 2.96 micrometers, all within the allowable deviation range of 2.7 to 3.1 micrometers, and all met the product specifications.

[0067] S108. Perform the electroplating process using the updated electroplating current and the target trace speed to obtain the target metal line.

[0068] Five temperature sensors are evenly arranged along the length of the bottom of the electroplating tank to form a temperature sensor array, with a spacing of one-quarter of the length of the electroplating tank. The temperature distribution of the electrolyte is monitored in real time, and the temperature value of each monitoring point is obtained. The average value of the temperature values ​​of the monitoring points is calculated as the current electrolyte temperature. If the current electrolyte temperature is lower than the preset lower limit of 40 degrees Celsius, the heating device is activated to raise the electrolyte temperature to the standard temperature range of 45 to 50 degrees Celsius. If the current electrolyte temperature is higher than the preset upper limit of 50 degrees Celsius, the cooling device is activated to lower the electrolyte temperature to the standard temperature range. A resistivity calculation model is established based on the electrolyte temperature T within the standard temperature range and the updated electroplating current I. The resistivity is calculated as ρ = ρ0 + α × (T - T0) + β × I, where ρ0 is the reference resistivity, T0 is the reference temperature, α is the temperature coefficient, and β is the current coefficient. The resistivity of the current electrolyte is calculated using the resistivity calculation model. The voltage difference U between the substrate surface and the anode is collected by a voltage sensor. The contact impedance R = U / I of the substrate surface is calculated based on Ohm's law, combined with the resistivity value and the voltage difference. The contact impedance value is acquired every second. The ratio of the difference in contact impedance between adjacent moments to the time interval is calculated as the rate of change. If the absolute value of the rate of change exceeds a preset rate of change threshold of 0.5 ohms per second, it is determined that there is uneven plating on the substrate surface. The adjustment range is selected according to the positive or negative attribute of the rate of change: when the rate of change is positive, the target routing speed is reduced by 15%; when the rate of change is negative, the target routing speed is reduced by 10%. If the absolute value of the rate of change does not exceed the preset rate of change threshold, the target routing speed remains unchanged. A laser diameter gauge is used to continuously scan and measure the diameter of the electroplated metal wire. The laser diameter gauge includes a transmitter and a receiver. The wire diameter is calculated by measuring the width of the laser beam blocked by the metal wire. The diameter data is acquired 10 times per second to obtain the measurement sequence data of the wire diameter. The standard deviation of the diameter values ​​in the measurement sequence data is statistically analyzed. If the standard deviation is greater than a preset diameter deviation threshold of 0.02 mm, it is determined that there is an abnormal fluctuation in the wire diameter. The combination parameters of the updated electroplating current and the target routing speed corresponding to the time of the abnormality are recorded. The ratio of current value to trace speed in the combined parameters is extracted as the current speed ratio K=I / V. This ratio reflects the amount of charge received per unit length of metal wire. The difference between the current speed ratio and the preset standard current speed ratio K0 is calculated to obtain the current speed ratio deviation ΔK=K-K0. When ΔK is greater than zero, the target trace speed is increased by 5% or the renewal plating current is decreased by 5%. When ΔK is less than zero, the target trace speed is decreased by 5% or the renewal plating current is increased by 5%, so that the absolute value of the current speed ratio deviation is reduced to within the allowable range of 0.1.A spectrometer is used to analyze the composition of the metal coating on the surface of the electroplated metal wire, obtaining the mass percentage data of the main metal elements in the coating. This mass percentage data is then compared with a preset standard component ratio to form a vector. The Euclidean distance between the two vectors is calculated as a matching index. If the matching index is greater than a preset matching threshold of 3%, a component deviation is determined in the metal deposition process, and a component deviation warning signal is output to the electrolyte circulation system. Based on the component deviation warning signal, the electrolyte circulation system is triggered to replenish and adjust the concentration of metal ions in the electrolyte. An ion concentration detection device is used to measure the concentration C of the target metal ions in the adjusted electrolyte. This concentration value is fed back to the current control module, a microprocessor unit integrated into the electroplating control system. The current control module calculates the appropriate electroplating current parameter I' = k × C × V × F / M according to Faraday's law of electrolysis, where k is a correction coefficient, V is the line speed, F is the Faraday constant, and M is the molar mass of the metal. This completes the closed-loop electroplating control of the target metal wire and obtains the target metal wire.

[0069] The system determines the required compensation current value for this batch based on the previous coating thickness deviation analysis results. This compensation value is calculated using a multiple linear regression model based on the measured thickness and process parameters of the most recent 200 batches in the historical database. The model input consists of three variables: target thickness, trace speed, and bath temperature. The output is the recommended operating current. The model output value for this batch is 280.27 amperes. This value, along with the target trace speed of 60 meters per minute, is input as execution parameters to the electroplating process control system. The system synchronously transmits instructions to the power control module and the trace drive module via industrial Ethernet. The time synchronization error of the two signals is controlled within 0.5 milliseconds. This accuracy is based on the fact that the copper ion diffusion time constant of the electroplating reaction is approximately 2 to 5 milliseconds. The synchronization error must be less than one-tenth of the diffusion time to avoid abrupt changes in coating thickness along the trace direction. After receiving the update current command, the power module, following the pre-set stepped loading strategy, gradually increases the operating current from an initial value of 50 amps to the target value of 280.27 amps, in steps of 30 amps each, with each step lasting 5 seconds. This process requires 8 steps and approximately 40 seconds to complete. This strategy is designed to prevent sudden current surges that could cause instantaneous breakdown of the anode surface passivation layer, leading to localized burning. During the loading process, the inter-electrode voltage U and operating current I are sampled every 2 seconds. The real-time equivalent resistance of the electrolyte is calculated using the formula R = U divided by I. This equivalent resistance includes the solution bulk resistance, anode-cathode contact resistance, and concentration polarization equivalent resistance. If the sum of the resistances deviates from the reference value of 0.045 ohms by more than 5%, i.e., exceeds the range of 0.043 to 0.047 ohms, voltage compensation is triggered. The compensation algorithm adopts incremental PI control. When the resistance rises, the cell voltage is increased to maintain a constant current. The voltage adjustment amount delta_U is equal to Kp multiplied by the resistance deviation plus Ki multiplied by the cumulative deviation value. The proportional coefficient Kp is 15 volts per ohm, and the integral coefficient Ki is 8 volts per ohm second. The cell voltage is dynamically adjusted within the range of 12 volts to 14.5 volts. The lower limit of this range is determined by the copper ion reduction potential plus the overpotential, and the upper limit is limited by the anode oxygen evolution side reaction potential. The cable drive module calculates the take-up reel motor speed to be 1200 rpm based on the speed command of 60 meters per minute. This speed is derived from the take-up reel diameter of 159 mm and the cable speed using the formula v = π multiplied by D multiplied by n. The inverter accelerates the cable from rest to the target speed at a ramp rate of 0.3 Hz per second. This acceleration rate is calculated based on the cable's unit length mass of 0.8 grams per meter and the maximum allowable tension of 1.5 Newtons. During acceleration, the tension sensor monitors the cable tension in real time, controlling it within the range of 0.8 Newtons to 1.2 Newtons. The lower limit of this range ensures that the cable does not slack or slip, while the upper limit prevents the cable from stretching and deforming, which would reduce the cross-sectional area. When the cable exceeds the range, the tension compensator adjusts the braking torque at the take-up end. The torque increment is proportional to the tension deviation, with a proportionality coefficient of 0.05 Nm per Newton.The copper sulfate solution in the electroplating tank, with a copper ion concentration of 80 g / L, is maintained at 52 degrees Celsius by an independent temperature control system. The temperature fluctuation range does not exceed ±0.5 degrees Celsius. According to the Arrhenius equation, the copper ion deposition rate increases by approximately 1.8% for every 1 degree Celsius increase in temperature. The temperature control algorithm dynamically adjusts the heating power between 3 kW and 5 kW through a PID controller. The three coefficients of the PID controller are a proportional coefficient of 200 W / °C, an integral coefficient of 50 W / °C / s, and a derivative coefficient of 80 W / s / °C. The power range is determined by the tank's heat capacity of 18 kJ / °C and the ambient heat dissipation power of 1.2 kW. During the formal electroplating process, the system continuously reads the wire's travel position L and cumulative charge Q with a sampling period of 500 milliseconds. Based on Faraday's law, m equals Q multiplied by M divided by n multiplied by F to calculate the copper deposition mass m, where M is the molar mass of copper (64 g / mol), n is the electron transfer number 2, and F is the Faraday constant (96485 coulombs / mol). Then, based on the copper density of 8.96 g / cm³ and the wire surface area, the mass is converted into the plating thickness d. When the deviation between the real-time calculated thickness and the target thickness of 3.2 μm exceeds 2%, i.e., exceeds the range of 3.14 to 3.26 μm, the system initiates current fine-tuning. The adjustment amount delta_I equals Kp1 multiplied by the thickness deviation plus Ki1 multiplied by the cumulative deviation value, where the proportional coefficient Kp1 is 0.6 amperes per micrometer and the integral coefficient Ki1 is 0.12 amperes per microsecond. The current correction range is limited to ±8 amperes to prevent over-adjustment. After production is completed, the system writes the actual average line speed (59.8 meters per minute), average operating current (280.1 amperes), average tank temperature (52.1 degrees Celsius), and sampled measured coating thickness (3.08 to 3.15 micrometers) of this batch into the production history database. Every 500 new records accumulated in the database trigger an update of the compensation model. The update uses the rolling window least squares method to refit the regression coefficients. When calculating the new coefficients, the most recent 100 batches of data are assigned a weight of 1.0, the 100 to 300 batches of data are assigned a weight of 0.6, and the 300 to 500 batches of data are assigned a weight of 0.3, ensuring that the model remains highly sensitive to recent changes in process conditions.

[0070] It should be noted that the above examples are merely some specific embodiments of the present invention. Obviously, the present invention is not limited to the above embodiments and many variations are possible. All variations that can be directly derived or conceived by those skilled in the art from the content disclosed in this invention should be considered within the scope of protection of this invention.

Claims

1. An adaptive compensation electroplating method for surface defects in ultrafine metal wires, characterized in that, include: A continuous wire image of the deformation process is acquired. Pixel differences in the continuous wire image are extracted using a convolutional neural network to obtain surface pit features and diameter fluctuation values. Based on the surface pit features and diameter fluctuation values, a first physical state sequence corresponding to the deformation process is constructed. If the diameter fluctuation value in the first physical state sequence is greater than a preset fluctuation threshold, the first physical state sequence is linearly interpolated and smoothed to obtain a second physical state sequence. A target feature vector is extracted from the second physical state sequence to obtain the initial operating settings of the deposition process. The initial deposition current and initial transmission speed are extracted from the initial operating settings. The target feature vector, the initial deposition current, and the initial transmission speed are processed using a long short-term memory network to output the target deposition current and target transmission speed. Based on the target deposition current and target transmission speed, the expected deposition layer thickness of the deposition process is calculated. If the expected deposition layer thickness is less than a preset thickness threshold, the target deposition current is increased according to a preset compensation ratio to obtain an updated deposition current. If the expected deposition layer thickness is greater than or equal to the preset thickness threshold, the target deposition current is used as the updated deposition current. The deposition process is performed using the updated deposition current and the target transport speed to obtain the target deposit.

2. The adaptive compensation electroplating method for surface defects of ultrafine metal wires as described in claim 1, characterized in that, The process of obtaining a continuous wire image during the deformation process, and extracting pixel differences from the continuous wire image using a convolutional neural network to obtain surface pit features and diameter fluctuation values, includes: acquiring a grayscale image sequence of the continuous wire and performing normalization processing to obtain a standardized image; dividing the standardized image into stress concentration regions and material uniformity regions according to the degree of deformation; using a convolutional neural network to extract features from the standardized image, outputting an image pixel gradient matrix; extracting surface texture features based on the locations of abrupt gradient changes in the image pixel gradient matrix to obtain the defect location coordinates of the surface pits; and targeting the defect location coordinates... The radial distribution of pixel gradients in the image region is statistically analyzed to obtain the pit depth distribution value. Based on the defect location coordinates and the pit depth distribution value, an edge detection algorithm is used to extract the contour boundary point set of the continuous wire in the standardized image. The Euclidean distance between adjacent boundary points is calculated to obtain the cross-sectional diameter sequence of the substrate. The diameter variation amplitude is obtained by the difference between the maximum and minimum values ​​in the cross-sectional diameter sequence. The ratio of the diameter variation amplitude to the maximum deviation of the distance from each cross-sectional boundary point to the center is taken as the cross-sectional roundness deviation. Based on the ratio relationship between the diameter variation amplitude and the cross-sectional roundness deviation, the surface pit features and diameter fluctuation values ​​are obtained.

3. The adaptive compensation electroplating method for surface defects of ultrafine metal wires as described in claim 1, characterized in that, The step of constructing a first physical state sequence corresponding to the deformation process based on the surface pit features and the diameter fluctuation values ​​includes: extracting the pit depth values ​​at each defect location from the surface pit features; counting the number of pits per unit length of substrate to obtain the defect density value; calculating the diameter change rate between adjacent measurement sections in conjunction with the diameter fluctuation values; matching the diameter change rate with the pit depth values ​​at corresponding positions according to the process position coordinates to obtain a cross-sectional deformation distribution sequence; extracting process position points of the substrate at fixed intervals along the deformation direction; collecting the surface temperature field distribution and the tension fluctuation amount applied by the deformation equipment at each process position point; and calculating the stress distribution field of the material at the corresponding position based on the temperature field distribution and the tension fluctuation amount. Numerical values; for each measured section in the cross-sectional deformation distribution sequence, extract the stress distribution field value and temperature field distribution data at the corresponding location, and concatenate the cross-sectional deformation, the stress distribution field value, the temperature field distribution data, and the tension fluctuation amount into a physical parameter vector. Add the corresponding process location point coordinates and the timestamp of the acquisition time to each physical parameter vector; arrange all physical parameter vectors in chronological order according to the timestamps, insert the defect density value as a transition mark between adjacent physical parameter vectors, and label each physical parameter vector with a quality level identifier based on the comparison result of the cross-sectional deformation and the preset deformation threshold, and construct a first physical state sequence containing the physical parameter vector, the quality level identifier, and the timestamp.

4. The adaptive compensation electroplating method for surface defects of ultrafine metal wires as described in claim 1, characterized in that, If the diameter fluctuation value in the first physical state sequence is greater than a preset fluctuation threshold, then the first physical state sequence is subjected to linear interpolation smoothing to obtain a second physical state sequence. This includes: if the diameter fluctuation value between adjacent time nodes in the first physical state sequence exceeds the preset fluctuation threshold, then marking the abnormal node positions that exceed the fluctuation threshold, and determining whether the distribution density of the abnormal nodes meets the conditions for triggering smoothing; based on the distribution density determination result of the abnormal nodes, applying a linear interpolation method to all abnormal nodes in the first physical state sequence, using the effective diameter measurement values ​​on both sides of the abnormal node as a benchmark, calculating the interpolation replacement value at the abnormal node to obtain the corrected diameter value of each node; if the difference between the corrected diameter value and the value of the adjacent node is still greater than the preset fluctuation threshold, then performing linear interpolation calculation again on the corrected diameter value until the diameter fluctuation value between all nodes does not exceed the preset fluctuation threshold, and determining the final diameter value of each node; rearranging according to the original time order based on the final diameter value of each node to obtain a second physical state sequence; calculating the overall fluctuation dispersion of the sequence for the diameter values ​​in the second physical state sequence, and if the dispersion is lower than a preset dispersion threshold, then determining that the second physical state sequence passes the smoothing quality check and obtaining the second physical state sequence.

5. The adaptive compensation electroplating method for surface defects of ultrafine metal wires as described in claim 1, characterized in that, The step of extracting target feature vectors from the second physical state sequence to obtain the initial operating settings of the deposition process, and extracting the initial deposition current and initial transmission speed from the initial operating settings, includes: aligning the second physical state sequence along the time axis to obtain a time-aligned state dataset; using a sliding window to extract fixed-length state segments based on the time-aligned state dataset, extracting statistical features from each state segment to obtain a target feature vector set; reading the initial operating setting record from the deposition process database; if the process number in the initial operating setting record matches the process number corresponding to the target feature vector set, parsing the initial deposition current and initial transmission speed from the initial operating setting record to obtain an initial parameter pair; and so on. Based on the initial deposition current and initial transmission velocity in the initial parameter pair, a two-dimensional parameter space coordinate is constructed. Each feature quantity in the target feature vector set is projected onto the two-dimensional parameter space coordinate to obtain the feature projection distribution matrix. The feature projection distribution matrix is ​​fitted and trained using a support vector regression model. If the fitting error is lower than the preset current deviation threshold and velocity deviation threshold, the predicted deposition current value and the predicted transmission velocity value are output to obtain the operation parameter prediction result. Based on the operation parameter prediction result, the difference between the predicted deposition current value and the predicted transmission velocity value and the initial deposition current and the initial transmission velocity are calculated respectively. If the difference exceeds the preset parameter fluctuation threshold, the parameter correction command is triggered to extract the initial deposition current and the initial transmission velocity in the initial operation setting.

6. The adaptive compensation electroplating method for surface defects of ultrafine metal wires as described in claim 1, characterized in that, The process of processing the target feature vector, the initial deposition current, and the initial transmission velocity through a Long Short-Term Memory (LSTM) network to output the target deposition current and the target transmission velocity includes: dividing the target feature vector into frames using a sliding window method based on the temporal dependencies of each dimension, obtaining a temporal segment sequence with a fixed time step; feeding the temporal segment sequence into the LSM network, updating the hidden state at each time step through input gates, forget gates, and output gates, obtaining the hidden state sequence corresponding to each time step; and outputting the target deposition current and the target transmission velocity through a fully connected layer based on the output of the last time step of the hidden state sequence. The layer performs a linear transformation on the hidden layer state at the last time step to obtain the predicted value of the target deposition current and the predicted value of the target transmission velocity. If the difference between the predicted value of the target deposition current and the initial deposition current exceeds a preset threshold, it is determined that there is a deviation in the current deposition state, and dynamic adjustment and correction of the target deposition current is triggered. If the difference between the predicted value of the target transmission velocity and the initial transmission velocity does not exceed the preset threshold, the target transmission velocity is directly determined as the current output value. Based on the dynamically adjusted and corrected target deposition current and the determined target transmission velocity, the gradient of the parameters of each layer of the Long Short-Term Memory Network is calculated through the error backpropagation algorithm, the network weight matrix and bias parameters are updated, and the target deposition current and target transmission velocity are output.

7. The adaptive compensation electroplating method for surface defects of ultrafine metal wires as described in claim 1, characterized in that, The process involves calculating the expected deposition layer thickness based on the target deposition current and the target transmission speed. If the expected deposition layer thickness is less than a preset thickness threshold, the target deposition current is increased according to a preset compensation ratio to obtain a refresh deposition current. If the expected deposition layer thickness is greater than or equal to the preset thickness threshold, the target deposition current is used as the refresh deposition current. This includes: calculating the immersion time of the substrate in the treatment solution based on the target transmission speed and the length of the deposition tank section to obtain the immersion time value per unit substrate section; multiplying the immersion time value by the target deposition current to obtain the electrolysis flux; and determining the electrolysis flux based on the electrolysis flux. The metal deposition quality on the substrate; based on the metal deposition quality, the immersion surface area of ​​the substrate is obtained, and the metal deposition quality is divided by the metal density and then divided by the immersion surface area of ​​the substrate to obtain the expected deposition layer thickness of the deposition process; the thickness deviation value between the actual detected deposition layer thickness in the current deposition process and the preset thickness threshold is obtained; if the thickness deviation value is negative, the absolute value of the thickness deviation value is multiplied by the preset compensation ratio to obtain the current compensation amplitude, and the preset target deposition current is added to the current compensation amplitude to obtain the updated deposition current; if the thickness deviation value is non-negative, the target deposition current is directly used as the updated deposition current.

8. The adaptive compensation electroplating method for surface defects of ultrafine metal wires as described in claim 1, characterized in that, The process of performing the deposition step using the updated deposition current and the target transmission speed to obtain the target deposit includes: using a temperature sensor array to monitor the temperature distribution of the treatment liquid in the deposition tank in real time, acquiring the temperature values ​​of each monitoring point, calculating the average value of the temperature values ​​of the monitoring points as the current treatment liquid temperature, and if the current treatment liquid temperature is lower than a preset lower limit, activating a heating device to raise the treatment liquid temperature to a standard temperature range; if the current treatment liquid temperature is higher than a preset upper limit, activating a cooling device to lower the treatment liquid temperature to the standard temperature range; establishing a resistivity calculation model based on the treatment liquid temperature within the standard temperature range and the updated deposition current, using the resistivity calculation model to calculate the resistivity value of the current treatment liquid, acquiring the voltage difference between the substrate surface and the anode using a voltage sensor, and combining the resistivity value with the voltage difference... The contact impedance of the substrate surface is calculated; the rate of change of the contact impedance is obtained; if the absolute value of the rate of change exceeds a preset rate of change threshold, it is determined that there is an uneven deposition layer on the substrate surface, and the target transmission speed is adjusted; if the absolute value of the rate of change does not exceed the preset rate of change threshold, the target transmission speed is kept constant; the diameter of the deposited substrate is continuously scanned and measured using a laser rangefinder to obtain measurement sequence data of the substrate diameter, and the standard deviation of the diameter values ​​in the measurement sequence data is calculated. If the standard deviation is greater than a preset diameter deviation threshold, it is determined that there is an abnormal fluctuation in the substrate diameter, and the combination parameter of the update deposition current and the target transmission speed corresponding to the time when the abnormality occurs is recorded. At least one parameter of the update deposition current or the target transmission speed is adjusted according to the combination parameter to obtain the target deposit.

Images