A purity real-time detection method and system for a copper powder extraction process

Abstract

The application provides a copper powder extraction process purity real-time detection method and system, and particularly relates to the technical field of intelligent sensor detection, and comprises the following steps: capturing a surface initial image sequence and an electromagnetic response original sequence of copper powder; extracting a color deviation index of the copper powder in the surface initial image sequence; extracting a magnetic permeability fluctuation amplitude sequence and an electrical conductivity decline slope sequence from the electromagnetic response original sequence to obtain a fusion feature vector; determining a pollution state level; analyzing the electromagnetic response original sequence to determine an internal impurity distribution proportion; comparing the internal impurity distribution proportion with a surface oxide layer thickness distribution unevenness degree and a hue offset angle in the color deviation index to obtain a comprehensive purity evaluation score; and generating a regulation parameter adjustment instruction according to the time variation of the comprehensive purity evaluation score to obtain an optimized production parameter set. The application effectively solves the problem that the surface oxidation and internal impurities are difficult to be synchronously and accurately detected and regulated in the copper powder production, and improves the purity consistency.

Description

A method and system for real-time detection of purity during copper powder extraction. Technical Field

[0001] This invention relates to the field of intelligent sensor technology, and in particular to a method and system for real-time detection of the purity of copper powder during extraction. Background Technology

[0002] Copper powder, as a core raw material in high-end manufacturing fields such as electronics, new energy batteries, and powder metallurgy, directly determines the conductivity, mechanical strength, and service life of downstream products. In high-precision applications, it has become a crucial indicator for industrial upgrading. As copper powder extraction processes accelerate towards continuous and large-scale production, the need for real-time control of copper powder purity during production is becoming increasingly urgent. Only by promptly understanding purity changes can process parameters be effectively adjusted to avoid the generation of batches of substandard products.

[0003] Currently, industrial sites generally rely on the traditional method of offline sampling followed by laboratory testing. While this method yields accurate data, the time from sampling to results often takes several hours or even longer. This causes production adjustments to lag significantly behind the actual material conditions, frequently resulting in situations where large quantities of substandard copper powder have already been produced before the problem is discovered, leading to raw material waste and equipment idleness. Furthermore, online detection methods are mostly limited to measuring single physical quantities, making it difficult to address the complex reality that copper powder is simultaneously affected by multiple impurities and oxidation during extraction. This results in a significant discrepancy between the test results and the true purity.

[0004] The unique aspect of copper powder extraction lies in the fact that the material is constantly in a dynamic and highly active state. Purity is influenced by multiple factors, including the degree of oxidation, the presence of ferromagnetic impurities, and the addition of other non-copper metals. Among these, the thickness of the oxide layer and changes in surface color are the most intuitive yet most difficult to quantify, while magnetic and electrical properties reflect the deeper distribution of internal impurities. These two characteristics are closely related: the more severe the surface oxidation, the darker and redder the color; simultaneously, the presence of magnetic substances such as iron significantly alters the overall electromagnetic response. Conversely, a decrease in conductivity is often associated with the combined accumulation of oxidation and impurities. Fluctuations in a single indicator cannot clearly pinpoint a specific source of contamination, easily leading to misjudgments.

[0005] Therefore, the key issue in achieving real-time purity detection during the extraction process is how to simultaneously capture the surface color characteristics and electromagnetic property changes of copper powder on a high-speed copper powder production line, and effectively integrate these two seemingly independent types of information to accurately distinguish pure copper powder from copper powder in various typical impurities and oxidized states. Summary of the Invention

[0006] To address the technical problem of poor real-time performance and accuracy of purity detection during copper powder extraction in existing technologies, this invention proposes a method and system for real-time purity detection during copper powder extraction.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A method for real-time detection of purity during copper powder extraction includes the following steps:

[0009] S1: The surface image data of copper powder flow is captured by the production line image acquisition device to generate the initial surface image sequence, and the electromagnetic signal of copper powder is collected by the electromagnetic sensor to generate the original electromagnetic response sequence.

[0010] S2: A convolutional neural network is used to extract the color deviation index of copper powder in the initial image sequence of the surface, including the degree of oxide layer inhomogeneity, color ratio and the clarity of the darkened boundary of the particle edge;

[0011] S3: Extract the permeability fluctuation amplitude sequence and conductivity decrease slope sequence from the original electromagnetic response sequence, and then fuse them with the color deviation index to obtain the fused feature vector;

[0012] S4: Use a support vector machine to classify the fused feature vectors, determine the impurity type and oxidation level, and identify the contamination state level;

[0013] S5: Determine whether the contamination state level exceeds the preset contamination threshold. If not, output the contamination state level of the copper powder. If so, process the original electromagnetic response sequence obtained in S1 to obtain a refined electromagnetic feature sequence.

[0014] S6: The decision tree ensemble method is used to analyze the correlation pattern between the peak intensity of the eddy current loss signal and the reduction ratio of the hysteresis loop area in the refining electromagnetic characteristic sequence, thereby determining the distribution ratio of internal impurities.

[0015] S7: Cross-compare the distribution ratio of internal impurities with the unevenness of surface oxide layer thickness distribution and hue shift angle in the color deviation index to obtain a comprehensive purity evaluation score;

[0016] S8: Generate adjustment instructions for control parameters based on the time change of the comprehensive purity evaluation score, and obtain the optimized set of production parameters.

[0017] Preferably, step S1 includes the following steps:

[0018] S1-1: The surface image data of copper powder during its flow is captured by the production line image acquisition device, and the electromagnetic signal of the copper powder is collected by the electromagnetic sensor.

[0019] S1-2: The surface image data is processed using denoising technology to generate an initial surface image sequence, including surface image data and corresponding image timestamps;

[0020] S1-3: Remove low-frequency drift and high-frequency noise from the electromagnetic signal to obtain the original sequence of electromagnetic response, including permeability, conductivity and corresponding signal timestamps.

[0021] Preferably, in step S1-2, for the initial surface image sequence, the following step is further included:

[0022] The existing Sobel operator edge detection algorithm is used to extract the edges of each surface image in the initial image sequence to obtain the copper powder particle contour. Then, the Otsu adaptive thresholding method is used to segment all the copper powder particle contours to obtain the copper powder particle region. Then, the gray mean and standard deviation of the copper powder particle region are calculated and used as the surface texture feature parameters.

[0023] Preferably, step S2 includes the following steps:

[0024] S2-1: Normalize the initial image sequence of the surface to obtain the image matrix;

[0025] S2-2: The trained convolutional neural network is used to extract the oxide layer feature map of the image matrix, and the standard deviation of the oxide layer thickness in the oxide layer feature map is calculated as the degree of oxide layer non-uniformity.

[0026] S2-3: Extract the oxide layer region from the image matrix and calculate the ratio of the oxide layer region to the total pixels to obtain the color ratio;

[0027] S2-4: The existing Sobel operator edge detection algorithm is used to extract the boundaries of copper powder particles in the image matrix to obtain the boundary clarity;

[0028] S2-5: Weighted fusion of oxide layer unevenness, color ratio, and boundary clarity yields a color deviation index.

[0029] Color deviation index = 0.4 × unevenness of oxide layer + 0.3 × color ratio + 0.3 × (1 - normalized value of boundary sharpness).

[0030] Preferably, step S3 includes the following steps:

[0031] S3-1: Extract the peak value of permeability in the original sequence of electromagnetic response, and calculate the fluctuation amplitude sequence based on the peak value, that is, the difference between the highest and lowest values ​​of the signal;

[0032] S3-2: Perform least-squares linear fitting on the conductivity in the original sequence of electromagnetic response to obtain a descending slope sequence;

[0033] S3-3: The permeability fluctuation amplitude sequence, the conductivity decrease slope sequence, and the color deviation index are fused to obtain the fused feature vector.

[0034] Preferably, step S4 includes the following steps:

[0035] S4-1: First, the support vector machine is trained using a classification training set;

[0036] S4-2: Input the fused feature vectors into the trained support vector machine to output the classification result;

[0037] S4-3: Determine the pollution status level based on the preset pollution rules and classification results.

[0038] Preferably, step S6 includes the following steps:

[0039] S6-1: First, extract the peak intensity of the eddy current loss signal and the reduction ratio of the hysteresis loop area from the refined electromagnetic feature sequence;

[0040] S6-2: Using the random forest algorithm as the decision tree ensemble model, correlation analysis was performed on the peak intensity of the eddy current loss signal and the reduction ratio of the hysteresis loop area to obtain the correlation analysis results.

[0041] S6-3: Based on the correlation analysis results, divide the first intensity interval and the second intensity interval; when it is in the first intensity interval, use the median value of the hysteresis loop area reduction ratio to determine the first impurity ratio; when it is in the second intensity interval, use the average of the partial maximum value and the median value of the hysteresis loop area reduction ratio to determine the second impurity ratio.

[0042] The internal impurity distribution ratio = (first impurity ratio + second impurity ratio) / sample mass × 100%.

[0043] Preferably, step S7 includes the following steps:

[0044] S7-1: Collect multi-point spectral reflectance data of copper powder to obtain the hue shift angle;

[0045] S7-2: Calculate the overall purity assessment score based on the unevenness of the oxide layer, the proportion of internal impurities, and the hue shift angle.

[0046]

[0047] in, This indicates the overall purity assessment score; , , These represent the weighting coefficients for hue shift angle, oxide layer unevenness, and internal impurity distribution ratio, respectively. This represents the average value of the hue shift angle; Indicates the degree of unevenness of the oxide layer; This indicates the proportion of internal impurities.

[0048] Preferably, step S8 includes the following steps:

[0049] S8-1: Obtain the comprehensive purity assessment score of copper powder within a time period to obtain the time series of the comprehensive purity assessment score;

[0050] S8-2: Then, the linear regression algorithm is used to fit the time series of the comprehensive purity assessment score to obtain the comprehensive purity assessment score curve, and the slope value is calculated.

[0051] S8-3: Generate adjustment instructions for control parameters based on the slope value to obtain an optimized set of production parameters.

[0052] This invention also provides a real-time purity detection system for copper powder extraction, comprising a data acquisition module, a first feature extraction module, a second feature extraction module, a contamination state level determination module, a third feature extraction module, an impurity ratio determination module, a purity assessment score calculation module, and a control module; wherein,

[0053] The data acquisition module is used to acquire and capture surface image data during the flow of copper powder to generate an initial surface image sequence, and at the same time acquire the electromagnetic signals of copper powder to generate an original electromagnetic response sequence.

[0054] The first feature extraction module is used to extract the color deviation index of copper powder in the initial image sequence of the surface using a convolutional neural network, including the degree of oxide layer unevenness, color ratio and the clarity of the darkened boundary of the particle edge.

[0055] The second feature extraction module is used to extract the permeability fluctuation amplitude sequence and the conductivity decrease slope sequence from the original electromagnetic response sequence, and then fuse them with the color deviation index to obtain a fused feature vector;

[0056] The pollution state level determination module is used to determine the pollution state level based on the fused feature vector;

[0057] The third feature extraction module is used to process the acquired original electromagnetic response sequence to obtain a refined electromagnetic feature sequence.

[0058] The impurity ratio determination module is used to determine the internal impurity distribution ratio based on the refining electromagnetic characteristic sequence.

[0059] The purity assessment score calculation module is used to calculate the comprehensive purity assessment score based on the distribution ratio of internal impurities, the unevenness of the oxide layer, and the hue shift angle.

[0060] The control module is used to adjust the production parameters according to the generated control parameter adjustment instructions, thereby obtaining the optimized production parameter set.

[0061] In summary, by adopting the above technical solution, the present invention has at least the following beneficial effects compared with the prior art:

[0062] This invention captures images of the flowing surface of copper powder using a production line image acquisition device and simultaneously acquires magnetic and conductivity signals from an electromagnetic sensor to obtain an initial surface image sequence and an original electromagnetic response sequence. For the surface image sequence, a convolutional neural network is used to extract the unevenness of oxide layer thickness distribution, color proportion, and clarity of darkened particle edges, forming a color deviation index. This index is then fused with the instantaneous fluctuation amplitude of magnetic permeability and the slope of conductivity decrease in the electromagnetic sequence to obtain a comprehensive feature vector. A support vector machine is used to classify the fused features, determining the impurity type, oxidation degree, and contamination level. When the contamination level exceeds a threshold, an adaptive bandpass filter is applied to the original electromagnetic sequence to remove periodic noise, resulting in a refined electromagnetic feature sequence. A decision tree ensemble analysis is then used to analyze the correlation between the peak value of eddy current loss and the reduction ratio of the hysteresis loop area to determine the internal impurity distribution ratio. The internal impurity ratio is cross-compared with the surface color deviation index to generate a comprehensive purity evaluation score. Based on the time series change of the score, control commands are generated and sent to the extraction process control system to optimize production parameters. The impact of optimized parameters on the images and electromagnetic sequences is continuously monitored, and the purity index is updated, forming a closed-loop real-time control system.

[0063] This invention effectively solves the problem of simultaneous and accurate detection and control of surface oxidation and internal impurities in copper powder production, significantly improving the consistency of copper powder purity and production stability. Attached Figure Description

[0064] Figure 1 is a schematic diagram of a method for real-time detection of purity in a copper powder extraction process according to an exemplary embodiment 1 of the present invention.

[0065] Figure 2 is a schematic diagram of the data acquisition and processing flow according to an exemplary embodiment 1 of the present invention.

[0066] Figure 3 is a schematic diagram of the color deviation index extraction process according to an exemplary embodiment 1 of the present invention.

[0067] Figure 4 is a schematic diagram of the fusion feature vector extraction process according to exemplary embodiment 1 of the present invention.

[0068] Figure 5 is a schematic diagram of the pollution state level determination process according to exemplary embodiment 1 of the present invention.

[0069] Figure 6 is a schematic diagram of the correlation analysis between the peak intensity of the eddy current loss signal and the reduction ratio of the hysteresis loop area according to exemplary embodiment 1 of the present invention.

[0070] Figure 7 is a schematic diagram of the comprehensive purity evaluation score curve according to exemplary embodiment 1 of the present invention.

[0071] Figure 8 is a schematic diagram of a real-time purity detection system for copper powder extraction process according to an exemplary embodiment 2 of the present invention. Detailed Implementation

[0072] The present invention will be further described in detail below with reference to embodiments and specific implementation methods. However, this should not be construed as limiting the scope of the above-described subject matter of the present invention to the following embodiments; all technologies implemented based on the content of the present invention fall within the scope of the present invention.

[0073] In the description of this invention, it should be understood that the terms "longitudinal", "lateral", "up", "down", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0074] Example 1

[0075] As shown in Figure 1, this invention provides a method for real-time detection of purity during copper powder extraction, specifically including the following steps:

[0076] S1: The surface image data of the copper powder during its flow is captured by the production line image acquisition device to generate an initial surface image sequence. At the same time, the electromagnetic signal of the copper powder is collected by the electromagnetic sensor to generate an original electromagnetic response sequence.

[0077] Specifically, as shown in Figure 2, S1 includes the following steps:

[0078] S1-1: The surface image data of copper powder during its flow is captured by the production line image acquisition device, and the electromagnetic signal of the copper powder is collected by the electromagnetic sensor.

[0079] In this embodiment, both the production line image acquisition device and the electromagnetic sensor are intelligent sensors. For example, the production line image acquisition device can be a high-speed industrial camera (such as Dalsa, FLIR, etc.) and installed on the side of the copper powder flow channel; the electromagnetic sensor can be a dual-coil structure probe and installed above the copper powder flow channel.

[0080] In this embodiment, a high-speed industrial camera captures surface image data of copper powder flowing in a transparent flow channel with an inclination angle of 30 degrees at a acquisition rate of 2000 frames per second and a resolution of 2048×1536 pixels (using high-resolution imaging technology to record dynamic changes); the camera lens uses a 50mm focal length and is equipped with an 850nm near-infrared filter to reduce ambient light interference.

[0081] In this embodiment, the electromagnetic sensor uses a dual-coil structure probe with a working frequency of 500kHz. The excitation coil is supplied with a sinusoidal current with an amplitude of 0.8A. The output voltage of the detection coil is converted into an electromagnetic signal (including permeability and conductivity) by a 24-bit synchronous sampling rate of 10MSPS ADC. Each frame consists of 1024 sampling points, and each frame corresponds to an image frame.

[0082] In this embodiment, the trigger signal of the production line image acquisition device is synchronized with the trigger signal of the electromagnetic sensor, thereby achieving timestamp alignment to obtain continuous surface image data.

[0083] S1-2: The surface image data is processed using denoising technology to generate an initial surface image sequence, including surface image data and corresponding image timestamps.

[0084] In this embodiment, median filtering denoising technology can be used to remove salt-and-pepper noise from the surface image data, thereby generating an initial surface image sequence.

[0085] In this embodiment, for the initial surface image sequence, the following steps are also included:

[0086] The existing Sobel operator edge detection algorithm is used to extract the edges of each surface image in the initial image sequence to obtain the copper powder particle contour. Then, the Otsu adaptive thresholding method is used to segment all the copper powder particle contours (separating particles and background) to obtain the copper powder particle region. Then, the gray mean and standard deviation of the copper powder particle region are calculated and used as surface texture feature parameters.

[0087] The formula for calculating the average gray value of the copper powder particle region is as follows:

[0088]

[0089] In formula (1), This represents the average grayscale value of the copper powder particle region; N represents the total number of pixels in the copper powder particle region. This represents the grayscale value of the i-th pixel in the copper powder particle region. The pixel grayscale value of the copper powder particle region can be extracted using a binary mask obtained through the Otsu adaptive thresholding method.

[0090] The formula for calculating the standard deviation of grayscale in the copper powder particle region is as follows:

[0091]

[0092] In formula (2), This represents the standard deviation of grayscale values ​​in the copper powder particle region. This represents the average grayscale value of the copper powder particle region; N represents the total number of pixels in the copper powder particle region. This represents the grayscale value of the i-th pixel in the copper powder particle region.

[0093] S1-3: Remove low-frequency drift and high-frequency noise from the electromagnetic signal to obtain the original sequence of electromagnetic response, including permeability, conductivity and corresponding signal timestamps.

[0094] In this embodiment, a digital bandpass filter can be used to remove low-frequency drift and high-frequency noise, retaining the signal in the 480kHz to 520kHz frequency band, thereby obtaining the original electromagnetic response sequence.

[0095] S2: A convolutional neural network is used to extract the color deviation index of copper powder in the initial image sequence of the surface, including the degree of oxide layer inhomogeneity, color ratio and the clarity of the darkened boundary of the particle edge.

[0096] Specifically, as shown in Figure 3, S2 includes the following steps:

[0097] S2-1: Normalize the initial image sequence of the surface to obtain the image matrix.

[0098] In this embodiment, each image in the initial surface image sequence has a resolution of 1920x1080. After normalization, the pixel values ​​of the image matrix range from 1 to 1.

[0099] S2-2: The trained convolutional neural network is used to extract the oxide layer feature map of the image matrix, and the standard deviation of the oxide layer thickness in the oxide layer feature map is calculated as the degree of oxide layer non-uniformity.

[0100] In this embodiment, the convolutional neural network can adopt the existing VGG-16 architecture CNN model, which contains 16 layers, where the convolutional layers use 3x3 convolutional kernels with a stride of 1, and the pooling layers use 2x2 max pooling with a stride of 2.

[0101] The training set (including images with and without oxide layers labeled) is input into the convolutional neural network for optimization, and the output is a feature map containing oxide layers. The loss function is mean squared error (MSE), the optimizer is Adam, the learning rate is 0.001, and the model converges after 30 epochs, thus completing the training of the convolutional neural network.

[0102] In this embodiment, after the image matrix is ​​input into the trained convolutional neural network, it will output a feature map containing the oxide layer, and the oxide layer thickness will be labeled. Then, the standard deviation of the oxide layer thickness is calculated as the degree of oxide layer non-uniformity.

[0103]

[0104] In formula (3), The standard deviation of the oxide layer thickness indicates the degree of oxide layer non-uniformity. This represents the average thickness of the oxide layer; This indicates the total number of oxide layer images; This represents the thickness of the oxide layer in the feature map of the j-th oxide layer.

[0105] In this embodiment, a standard deviation of 0.15 is assumed to indicate a high degree of unevenness in the oxide layer thickness distribution. The larger the standard deviation, the more obvious the unevenness.

[0106] S2-3: Extract the oxide layer region from the image matrix and calculate the ratio of the oxide layer region to the total pixels to obtain the color ratio.

[0107] In this embodiment, the extraction range of the oxide layer region can be defined as: R channel value greater than 180, G channel value between 80 and 120, and B channel value less than 100.

[0108] Assuming the ratio of the oxide layer area to the total pixels is 0.35, this means the color ratio (mainly reddish-brown) is 35%.

[0109] S2-4: The existing Sobel operator edge detection algorithm is used to extract the boundaries of copper powder particles in the image matrix and obtain the boundary gradient magnitude, i.e. boundary sharpness (this is an existing technology and will not be elaborated here).

[0110] Assuming the boundary sharpness is 12.5 and the threshold is set to 10, if the boundary sharpness is greater than or equal to the boundary threshold, the boundary is considered sharp; if the boundary sharpness is less than the boundary threshold, the boundary is considered blurry.

[0111] S2-5: Weighted fusion of oxide layer unevenness, color ratio, and boundary clarity yields a color deviation index.

[0112] Color deviation index = 0.4 × unevenness of oxide layer + 0.3 × color ratio + 0.3 × (1 - normalized value of boundary sharpness).

[0113] Assuming a calculation result of 0.28, it indicates a moderate color deviation, and parameters need to be adjusted in subsequent processes to optimize surface quality.

[0114] In this embodiment, if the color deviation index exceeds the preset deviation index range, the image matrix is ​​enhanced (contrast enhancement, noise reduction, sharpening, etc.), and the color deviation index is recalculated.

[0115] Through the above methods, logical connections are formed between the various indicators. The degree of unevenness and color ratio affect the overall deviation, while the clarity of the boundary serves as a supplementary factor to ensure comprehensive analysis and meet the needs of industrial testing.

[0116] S3: Extract the permeability fluctuation amplitude sequence and the conductivity decrease slope sequence from the original electromagnetic response sequence, and then fuse them with the color deviation index to obtain the fused feature vector.

[0117] Specifically, as shown in Figure 4, S3 includes the following steps:

[0118] S3-1: Extract the peak value of the permeability in the original sequence of the electromagnetic response, and calculate the fluctuation amplitude sequence based on the peak value, that is, the difference between the highest and lowest values ​​of the signal.

[0119] For example, the permeability in the original sequence of the electromagnetic response is divided into multiple windows according to time sequence, each window having a corresponding fluctuation amplitude, thus forming a fluctuation amplitude sequence R(t). The fluctuation amplitude of the permeability within the first 500ms window is calculated, yielding a fluctuation amplitude of [0.0214, 0.0387] H / m; the fluctuation amplitude of the permeability within the second 500ms window is calculated, yielding a fluctuation amplitude of [0.0126, 0.0453] H / m.

[0120] S3-2: Perform least-squares linear fitting on the conductivity in the original sequence of electromagnetic response to obtain a descending slope sequence.

[0121] In this embodiment, the conductivity data in the original sequence of electromagnetic response is smoothed by a Savitzky-Golay filter (window length 31, polynomial order 3). Then, 150ms data segments are taken before and after each time point for least squares linear fitting to obtain the slope sequence S(t), where S(1000) = -0.187S / m·ms.

[0122] S3-3: The permeability fluctuation amplitude sequence, the conductivity decrease slope sequence, and the color deviation index are fused to obtain the fused feature vector.

[0123] In this embodiment, the color deviation index is first normalized to map its value range to the interval between 0 and 1. For example, the original color deviation ΔE has a maximum of 8.7 and a minimum of 0.3. After min-max normalization, the color deviation feature sequence C(t) is obtained, and the value at time t=1000 is 0.624.

[0124] The permeability fluctuation amplitude sequence, conductivity decrease slope sequence, and color deviation feature sequence are concatenated into vectors at the same time point to form a fused feature vector F(t)=[C(t), R(t), S(t)].

[0125] Taking t=1000 as an example, the fused feature vector F(1000)=[0.624,0.0321,-0.187] is obtained. This vector retains the visual perception information of color deviation and integrates the transient fluctuation intensity and trend change characteristics of electromagnetic response, forming a high-dimensional input with temporal correlation and multi-physics complementarity, which can be directly used for defect classification or quantitative prediction in subsequent machine learning models. S4: The fused feature vector is classified using a support vector machine to determine the impurity type and oxidation degree and to determine the pollution state level.

[0126] Specifically, as shown in Figure 5, S4 includes the following steps: S4-1: In this embodiment, the support vector machine is first trained using a classification training set.

[0127] The classification training set includes 1000 samples, each labeled with impurity type (metallic and non-metallic, labeled based on the permeability fluctuation range sequence and conductivity decrease slope sequence) and oxidation degree (slight, moderate, and severe, labeled based on color deviation index). The training process uses a radial basis function kernel, with a penalty parameter C of 1.0 and a kernel parameter γ of 0.5.

[0128] S4-2: Input the fused feature vectors into the trained support vector machine to output the classification results.

[0129] In this embodiment, principal component analysis algorithm can be used to reduce the dimensionality of the fused feature vector to form a two-dimensional vector. This process is achieved through matrix decomposition, which reduces data redundancy and retains 95% of the information.

[0130] For example, the fused feature vector is F(1000)=[0.624,0.0321,-0.187], the weight of principal component 1 is 0.6, the weight of principal component 2 is 0.4, and the two-dimensional vector is [6.3, 3.2]. This two-dimensional vector is input into the trained support vector machine, and the classification result is classified as a metal impurity with a moderate degree of oxidation, and the classification confidence is 0.85.

[0131] S4-3: Determine the pollution status level based on the preset pollution rules and classification results.

[0132] For example, if the impurity is metallic and the degree of oxidation is moderate, the pollution level is set at level 3 (out of 5); if the impurity is non-metallic and the degree of oxidation is slight, the pollution level is level 1. See Table 1 for details.

[0133] Table 1 Pollution Status Levels

[0134]

[0135] S4-4: Automatically generate a pollution status report, including impurity type, oxidation degree and pollution status level, and store the results in the database. At the same time, trigger the equipment maintenance reminder interface to ensure the continuity of subsequent business processes.

[0136] For example, when the pollution level exceeds level 2, the maintenance scheduling algorithm is automatically invoked to optimize equipment maintenance time and reduce downtime losses.

[0137] The above process is automated through algorithms and data-driven approaches. Logically, it forms a closed loop from feature extraction to classification and then to state evaluation, ensuring the accuracy of the judgment and its relevance to business operations.

[0138] S5: Determine whether the contamination level exceeds (is greater than or equal to) the preset contamination threshold. If not (is less than the preset contamination threshold), output the purity detection result of the copper powder, i.e., the contamination level. If it is (is greater than or equal to the preset contamination threshold), process the original electromagnetic response sequence obtained in S1 to obtain the refined electromagnetic feature sequence.

[0139] In this embodiment, when the pollution state level exceeds (is greater than or equal to) a preset pollution threshold, it is determined that the original electromagnetic response sequence obtained in S1 contains periodic noise interference. Therefore, in this embodiment, for the original electromagnetic response sequence containing periodic noise interference, the present invention applies an adaptive bandpass filter to the original electromagnetic response sequence, specifically using a Butterworth filter, setting the passband frequency range to 10Hz to 50Hz, in order to remove periodic noise interference and obtain a refined electromagnetic feature sequence.

[0140] For example, if a strong noise component with a frequency of 20Hz is detected in the original electromagnetic response sequence, the amplitude of this component will decrease from 5.2 to 0.3 after filtering. The filtering algorithm decomposes the signal into frequency domain components through fast Fourier transform, and then attenuates the data outside the target frequency band to ensure that the main signal characteristics are preserved.

[0141] In this embodiment, without verifying the processing effect, the present invention calculates the root mean square value of the refined electromagnetic feature sequence. The result decreases from the original 3.8 to 1.2, indicating that noise interference has been significantly reduced. At the same time, by comparing with historical data, it is confirmed that the stability of the feature sequence has been improved by about 30%, thereby providing a more reliable data basis for subsequent electromagnetic signal classification or anomaly detection.

[0142] The above processes are all completed automatically by the system, achieving a logical closed loop through preset algorithms and parameters to ensure the integrity from contamination detection to feature extraction. Simultaneously, it is linked to business requirements; for example, the refining sequence can be directly input into the anomaly detection model to improve detection accuracy. S6: For the refining electromagnetic feature sequence, a decision tree ensemble method is used to analyze the correlation pattern between the peak intensity of the eddy current loss signal and the reduction ratio of the hysteresis loop area, thereby determining the internal impurity distribution ratio. S6-1: First, the peak intensity of the eddy current loss signal and the reduction ratio of the hysteresis loop area are extracted from the refining electromagnetic feature sequence.

[0143] In this embodiment, the magnetic flux density sequence B is obtained by refining the permeability and magnetic field strength in the electromagnetic characteristic sequence, and then the dB / dt is calculated by numerical differentiation; in the dB / dt sequence, the maximum value point is found, which corresponds to the peak intensity of the eddy current loss signal.

[0144] In this embodiment, data points of magnetic field strength and magnetic induction intensity for a complete cycle can be obtained from the refined electromagnetic feature sequence. Then, the hysteresis loop area (including dynamic hysteresis loop area and static hysteresis loop area) can be calculated using existing numerical integration methods (see reference: Calculation of memristor hysteresis loop area: This code calculates memristor hysteresis loop area - MATLAB development). Finally, the hysteresis loop area reduction ratio is obtained: Hysteresis loop area reduction ratio = (static hysteresis loop area - dynamic hysteresis loop area) / static hysteresis loop area × 100%.

[0145] S6-2: Using the random forest algorithm as the decision tree ensemble model, the correlation analysis between the peak intensity of the eddy current loss signal and the reduction ratio of the hysteresis loop area is performed to obtain the correlation analysis results.

[0146] In this embodiment, the peak intensity of the eddy current loss signal and the reduction ratio of the hysteresis loop area are divided into a training set (80%) and a test set (20%). The training set is then input into the decision tree ensemble model for training (setting the number of decision trees in the forest n_estimators=200, the maximum tree depth max_depth=8, the minimum number of samples required for internal node splits min_samples_split=5, and using 5-fold cross-validation to optimize the parameters). The six-dimensional features of the peak intensity P of the eddy current loss signal—mean, variance, maximum, minimum, kurtosis, and skewness (which can be obtained through existing calculation methods)—are used as the input to the decision tree ensemble method, and the reduction ratio of the hysteresis loop area is used as the output. Finally, the test set is input into the trained decision tree ensemble model, and the correlation analysis results (i.e., the hysteresis loop slope) are output, as shown in Figure 6.

[0147] As shown in Figure 6, the curve exhibits a rapid upward convex growth, indicating that the increase in eddy current loss is nonlinear as the short circuit intensifies (the loop area shrinks). The initial growth is relatively gradual, while the increase is sharp in the middle and later stages. This typically conforms to the fitting pattern of exponential or power functions.

[0148] In this embodiment, the coefficient of determination (R²) can be used to reflect the proportion of the variation of the target variable explained by the model. The closer it is to 1, the stronger the ability of the statistical characteristics of the eddy current loss signal to explain the reduction ratio of the hysteresis loop area, and the stronger the correlation between the two.

[0149] The decision tree ensemble model achieved an R² of 0.892 on the validation set. During the analysis, the importance ranking of features from the random forest was extracted, revealing that the mean importance of peak intensity was 0.41, kurtosis was 0.23, variance was 0.18, and the rest were relatively low.

[0150] S6-3: Based on the correlation analysis results, the interval is divided into a first intensity interval and a second intensity interval, and the distribution values ​​of the reduction ratio of the hysteresis loop area in the first intensity interval and the second intensity interval are statistically analyzed; when it is the first intensity interval, the median value of the reduction ratio of the hysteresis loop area is used to determine the first impurity ratio; when it is the second intensity interval, the larger value of the reduction ratio of the hysteresis loop area (exceeding the median value, for example, the average of the maximum value and the median value can be selected) is used to determine the second impurity ratio;

[0151] The internal impurity distribution ratio = (first impurity ratio + second impurity ratio) / sample mass × 100%.

[0152] For example, the percentage of the total content of major impurity elements (Fe, Cr, Ni, etc.) in the sample mass, P_imp = 0.023%.

[0153] In this embodiment, when the slope of the hysteresis loop is less than a preset slope threshold, that is, the area where the reduction ratio of the area does not change significantly with the peak intensity, it is divided into the first intensity interval; when the slope of the hysteresis loop is greater than or equal to the preset slope threshold, it may have a linear relationship, that is, the reduction ratio of the area increases significantly with the peak intensity, and it is divided into the second intensity interval.

[0154] Further analysis of individual samples using SHAP values ​​revealed that a sudden change in the peak intensity sequence of a certain sample resulted in a SHAP contribution of +0.094, directly increasing the impurity distribution ratio to 0.36. This indicates a strong correlation between local high-amplitude peak clusters and large-sized inclusions within the material. Finally, the impurity distribution ratio was quantitatively mapped to 0.34±0.03 through the output of the integrated model, achieving non-destructive inference from electromagnetic features to internal defect distribution.

[0155] S7: Cross-compare the distribution ratio of internal impurities with the surface oxide layer unevenness and hue shift angle in the color deviation index to obtain a comprehensive purity evaluation score.

[0156] S7-1: Collect multi-point spectral reflectance data of copper powder to obtain the hue shift angle.

[0157] Specifically, after collecting multi-point spectral reflectance data of copper powder, the reflectance curve of each measurement point was recorded at 5nm intervals in the 400-700nm wavelength range. Then, the CIE 1976 standard was used to analyze the data. Color space conversion formula calculates the hue angle at each point and color saturation ,in, This represents the red-green axis (-128 to +127). The yellow-blue axis is represented as (-128 to +127). Then, using the center point of the surface as a reference, the hue offset angle Δh (in degrees) between the center point and the remaining points is calculated, and the maximum and average values ​​of Δh for all measurement points are statistically analyzed. And standard deviation, for example, the measured maximum value of Δh is 8.7°, the average value is 3.2°, and the standard deviation is 2.1°.

[0158] S7-2: Calculate the overall purity assessment score based on the degree of unevenness of the oxide layer, the distribution ratio of internal impurities, and the hue shift angle.

[0159] In this embodiment, the formula for calculating the comprehensive purity evaluation score is:

[0160]

[0161] In formula (4), This indicates the overall purity assessment score; , , These are weighting coefficients representing the hue shift angle, the degree of unevenness of the oxide layer, and the proportion of internal impurities, respectively, for example, 0.42, 0.35, and 0.23; This represents the average value of the hue shift angle; Indicates the degree of unevenness of the oxide layer; This indicates the proportion of internal impurities.

[0162] In this embodiment, purity levels can be classified according to the S value: S≥85 is excellent, 65≤S<85 is good, and S<65 is in need of improvement. For example, in this embodiment, the comprehensive purity evaluation score of copper powder is 37.86, which belongs to the need of improvement level. This indicates that the uneven distribution of the surface oxide layer thickness and the large hue shift are the main factors causing the low purity. At the same time, the internal impurity content also exceeds the target threshold, requiring targeted optimization of the growth process and subsequent surface treatment. In this embodiment, the final comprehensive purity evaluation result is archived in the database using data storage technology, and the latest evaluation record is obtained through an automatic update mechanism. S8: Generate adjustment instructions for control parameters based on the time change of the comprehensive purity evaluation score, and send them to the extraction process control system to obtain the optimized production parameter set. S8-1: Obtain the comprehensive purity evaluation score of copper powder within a time period to obtain the comprehensive purity evaluation score time series.

[0163] It can collect 12 hours of copper powder data, and after processing, obtain the comprehensive purity evaluation score time series: [92.3, 92.5, 92.1, 91.8, 91.4, 91.9, 92.7, 93.2, 93.8, 94.1, 94.0, 93.6];

[0164] In this embodiment, the time series of the comprehensive purity evaluation score is also smoothed to obtain the smoothed sequence [92.18,92.22,92.12,91.92,92.18,92.92,93.36,93.76,94.06,93.90].

[0165] S8-2: Then, the linear regression algorithm is used to fit the time series of the comprehensive purity assessment score to obtain the comprehensive purity assessment score curve, as shown in Figure 7, and the slope value is calculated.

[0166] In this embodiment, as shown in Figure 7, the slope value is 0.214 (indicating an upward trend in purity), and the standard deviation of the sequence over the past 7 days is 0.86, indicating that the volatility is at a moderate level.

[0167] S8-3: Generate adjustment instructions for control parameters based on the slope value to obtain an optimized set of production parameters.

[0168] In this embodiment, when the slope value is less than 0.15 and the overall purity evaluation score is less than 93.8, the key parameters need to be adjusted in a positive direction. Therefore, the adjustment direction is calculated by combining the historical parameter-purity response surface model (using quadratic polynomial regression R²=0.972), such as increasing the temperature and decreasing the rotation speed.

[0169] For example, the instructions for adjusting the control parameters include increasing the proportion of ethanol in the solvent ratio by 0.8 percentage points (from the current 38.2% to 39.0%), increasing the extraction temperature by 1.2℃ (from 48.5℃ to 49.7℃), and decreasing the stirring speed by 12 rpm (from 385 rpm to 373 rpm).

[0170] The control parameter adjustment command is encapsulated in JSON format as {"timestamp":"2026-01-20T00:45:00","adjustments":[{"param":"ethanol_ratio","value":39.0,"delta":0.8},{"param":"extract_temp","value":49.7,"delta":1.2},{"param":"stir_speed","value":373,"delta":-12}]}, and sent to the subscription topic / extract / control / cmd of the extraction process control system via MQTT protocol at QoS=1 level. After receiving the command, the control system automatically updates the corresponding PID parameters and setpoints, completing the closed-loop optimization adjustment. The entire process is completed by the edge computing node within 2.3 seconds without manual intervention. S9: Continuously monitor the impact of the optimized production parameter set on the initial surface image sequence and the original electromagnetic response sequence, thereby updating the color deviation index and the internal impurity distribution ratio and forming a closed-loop real-time purity control process. Specifically, firstly, an initial image sequence of the strip surface under optimized production parameters is acquired using an industrial camera at a frequency of 30 frames per second, with each frame having a resolution of 2048×1536 pixels. Then, an improved U-Net network is used to perform semantic segmentation on each frame of the image, segmenting normal regions, color abnormal regions, and suspected impurity regions. The network outputs the category probability of each pixel and performs binarization processing with 0.75 as a confidence threshold to obtain a color deviation binary mask.

[0171] Meanwhile, the electromagnetic eddy current flaw detection equipment collects the original impedance sequence at a sampling frequency of 10kHz. Every 2000 sampling points is used as an analysis window. The sequence within the window is subjected to a fast Fourier transform, and the sum of the characteristic peak values ​​in the amplitude spectrum from 10kHz to 50kHz is extracted as the electromagnetic response intensity index, and its value is denoted as Z_em.

[0172] Next, the color deviation index CDI is calculated, which is the proportion of the area of ​​color-abnormal pixels in the whole image multiplied by the average color Euclidean distance deviation. The color deviation is calculated using the Lab color space with a ΔE value. Pixels with a ΔE greater than 3.2 are included in the deviation area, and the CDI range is usually between 0.00 and 0.18.

[0173] Then, the electromagnetic response intensity Z_em is compared with the baseline value Z_base of historical pure samples. If Z_em / Z_base is greater than 1.12, it is determined that there is a risk of internal impurities. Combined with the spatial distribution of the surface color deviation mask, the suspected impurity points are divided into 3 categories by K-means clustering and the proportion of each category is calculated to form the internal impurity distribution ratio vector P=[p1,p2,p3], for example [0.62,0.31,0.07]; where p1 represents the distribution ratio of the first type of impurity (e.g., Fe), p2 represents the distribution ratio of the second type of impurity (e.g., Cr), and p3 represents the distribution ratio of the third type of impurity (e.g., Ni).

[0174] Finally, the CDI and P vectors are input into a pre-trained BP neural network purity prediction model. This model has 12 hidden neurons and uses the tanh activation function to output a real-time purity estimate Q. The deviation of Q from the target purity of 0.995 is e = Q_target - Q. If e is greater than 0.003, a closed-loop adjustment command is triggered. The adjustment amount is mapped to a 2.4% reduction in rolling speed, a 7.8℃ increase in annealing temperature, and an 11.2% increase in cooling water flow rate via the Modbus protocol, completing the adaptive parameter update for the next cycle. This achieves a closed-loop purity control driven by the fusion of surface and internal features.

[0175] Example 2

[0176] Based on the real-time purity detection method for copper powder extraction provided in Example 1, the present invention also provides a real-time purity detection system for copper powder extraction, as shown in Figure 8, which includes a data acquisition module, a first feature extraction module, a second feature extraction module, a contamination state level determination module, a third feature extraction module, an impurity ratio determination module, a purity evaluation score calculation module, and a control module.

[0177] Specifically, the first output terminal of the data acquisition module is connected to the input terminal of the first feature extraction module, and the output terminal of the first feature extraction module is connected to the input terminal of the second feature extraction module; the output terminal of the second feature extraction module is connected to the input terminal of the contamination state level determination module; the second output terminal of the data acquisition module is connected to the input terminal of the third feature extraction module, the output terminal of the third feature extraction module is connected to the input terminal of the impurity ratio determination module, the output terminal of the impurity ratio determination module is connected to the input terminal of the purity assessment score calculation module, and the output terminal of the purity assessment score calculation module is connected to the input terminal of the control module.

[0178] The data acquisition module includes a production line image acquisition device and an electromagnetic sensor, which are used to acquire and capture surface image data when copper powder flows to generate an initial surface image sequence, and at the same time acquire the electromagnetic signals of copper powder to generate an original electromagnetic response sequence.

[0179] The first feature extraction module is used to extract the color deviation index of copper powder in the initial image sequence of the surface using a convolutional neural network, including the degree of oxide layer unevenness, color ratio and the clarity of the darkened boundary of the particle edge.

[0180] The second feature extraction module is used to extract the permeability fluctuation amplitude sequence and the conductivity decrease slope sequence from the original electromagnetic response sequence, and then fuse them with the color deviation index to obtain a fused feature vector;

[0181] The pollution state level determination module is used to determine the pollution state level based on the fused feature vector;

[0182] The third feature extraction module is used to process the acquired original electromagnetic response sequence to obtain a refined electromagnetic feature sequence.

[0183] The impurity ratio determination module is used to determine the internal impurity distribution ratio based on the refining electromagnetic characteristic sequence.

[0184] The purity assessment score calculation module is used to calculate the comprehensive purity assessment score based on the distribution ratio of internal impurities, the unevenness of the oxide layer, and the hue shift angle.

[0185] The control module is used to adjust the production parameters according to the generated control parameter adjustment instructions, thereby obtaining the optimized production parameter set.

[0186] Those skilled in the art will understand that the above embodiments are specific examples of implementing the present invention, and in practical applications, various changes in form and detail may be made without departing from the spirit and scope of the present invention.

Claims

1. A method for real-time detection of purity during copper powder extraction, characterized in that, Includes the following steps: S1: Surface image data of copper powder flow is captured by the production line image acquisition device to generate an initial surface image sequence. Simultaneously, electromagnetic signals of the copper powder are collected by an electromagnetic sensor to generate a raw electromagnetic response sequence. S2: A convolutional neural network is used to extract color deviation indicators of the copper powder from the initial surface image sequence, including oxide layer inhomogeneity, color ratio, and particle edge darkening clarity. S3: The permeability fluctuation amplitude sequence and conductivity decrease slope sequence are extracted from the raw electromagnetic response sequence and then fused with the color deviation indicators to obtain a fused feature vector. S4: A support vector machine is used to classify the fused feature vector to determine... S5: Determine if the contamination level exceeds the preset contamination threshold; otherwise, output the contamination level of the copper powder. If so, process the original electromagnetic response sequence obtained in S1 to obtain the refining electromagnetic feature sequence. S6: Use the decision tree ensemble method to analyze the correlation pattern between the peak intensity of the eddy current loss signal and the reduction ratio of the hysteresis loop area in the refining electromagnetic feature sequence to determine the internal impurity distribution ratio. S7: Cross-compare the internal impurity distribution ratio with the surface oxide layer non-uniformity and hue shift angle in the color deviation index to obtain a comprehensive purity evaluation score. S8: Generate adjustment instructions for control parameters based on the time change of the comprehensive purity assessment score, and obtain the optimized set of production parameters; S2 includes the following steps: S2-1: Normalize the initial surface image sequence to obtain an image matrix; S2-2: Use a trained convolutional neural network to extract the oxide layer feature map of the image matrix and calculate the standard deviation of the oxide layer thickness in the oxide layer feature map as the degree of oxide layer non-uniformity; S2-3: Extract the oxide layer region in the image matrix and calculate the ratio of the oxide layer region to the total pixels to obtain the color ratio; S2-4: Use the existing Sobel operator edge detection algorithm to extract the copper powder particle boundaries in the image matrix to obtain the boundary sharpness; S2-5: Perform weighted fusion of oxide layer non-uniformity, color ratio, and boundary sharpness to obtain a color deviation index; S6 includes the following steps: S6-1: First, extract the peak intensity of the eddy current loss signal and the reduction ratio of the hysteresis loop area from the refined electromagnetic feature sequence; S6-2: ... Using the random forest algorithm as the decision tree ensemble model, a correlation analysis is performed on the peak intensity of the eddy current loss signal and the reduction ratio of the hysteresis loop area to obtain the correlation analysis results; S6-3: Divide the first intensity interval and the second intensity interval according to the correlation analysis results; when the correlation analysis result is in the first intensity interval, the median value of the reduction ratio of the hysteresis loop area is used to determine the first impurity ratio; when the correlation analysis result is in the second intensity interval, the average of the partial maximum value and the median value of the reduction ratio of the hysteresis loop area is used to determine the second impurity ratio; then determine the internal impurity distribution ratio according to the first impurity ratio and the second impurity ratio; S7 includes the following steps: S7-1: Collect multi-point spectral reflectance data of copper powder to obtain the hue shift angle; S7-2: Calculate the comprehensive purity evaluation score according to the degree of oxide layer non-uniformity, the internal impurity distribution ratio and the hue shift angle.

2. The method for real-time purity detection in the copper powder extraction process as described in claim 1, characterized in that, S1 includes the following steps: S1-1: Capturing surface image data of copper powder flow using a production line image acquisition device, and simultaneously acquiring electromagnetic signals of copper powder using an electromagnetic sensor; S1-2: Processing the surface image data using denoising technology to generate an initial surface image sequence, including surface image data and corresponding image timestamps; S1-3: Removing low-frequency drift and high-frequency noise from the electromagnetic signals to obtain the original electromagnetic response sequence, including permeability, conductivity, and corresponding signal timestamps.

3. The method for real-time purity detection in the copper powder extraction process as described in claim 2, characterized in that, In step S1-2, for the initial surface image sequence, the following steps are also included: using the existing Sobel operator edge detection algorithm to extract the edges of each surface image in the initial surface image sequence to obtain the copper powder particle contours, and then using the Otsu adaptive thresholding method to segment all the copper powder particle contours to obtain the copper powder particle regions; then calculating the gray mean and standard deviation of the copper powder particle regions, and using the gray mean and standard deviation of the copper powder particle regions as surface texture feature parameters.

4. The method for real-time purity detection in the copper powder extraction process as described in claim 1, characterized in that, S3 includes the following steps: S3-1: Extract the peak value of the permeability in the original electromagnetic response sequence and calculate the fluctuation amplitude sequence based on the peak value; S3-2: Perform least squares linear fitting on the conductivity in the original electromagnetic response sequence to obtain the descending slope sequence; S3-3: Fuse the permeability fluctuation amplitude sequence, the conductivity descending slope sequence, and the color deviation index to obtain the fused feature vector.

5. The method for real-time purity detection in the copper powder extraction process as described in claim 1, characterized in that, The S4 includes the following steps: S4-1: First, the support vector machine is trained using a classification training set; S4-2: The fused feature vector is input into the trained support vector machine to output the classification result; S4-3: The pollution state level is determined according to the preset pollution rules and the classification result.

6. The method for real-time purity detection in the copper powder extraction process as described in claim 1, characterized in that, S8 includes the following steps: S8-1: Obtain the comprehensive purity evaluation score of copper powder within a time period to obtain a comprehensive purity evaluation score time series; S8-2: Then, use a linear regression algorithm to fit the comprehensive purity evaluation score time series to obtain a comprehensive purity evaluation score curve and calculate the slope value; S8-3: Generate adjustment instructions for control parameters based on the slope value to obtain an optimized set of production parameters.

7. A real-time purity detection system for copper powder extraction process based on the method of any one of claims 1-6, characterized in that, The system includes a data acquisition module, a first feature extraction module, a second feature extraction module, a contamination state level determination module, a third feature extraction module, an impurity ratio determination module, a purity assessment score calculation module, and a control module. The data acquisition module collects and captures surface image data of the flowing copper powder to generate an initial surface image sequence, and simultaneously collects the electromagnetic signals of the copper powder to generate a raw electromagnetic response sequence. The first feature extraction module uses a convolutional neural network to extract color deviation indicators of the copper powder from the initial surface image sequence, including oxide layer inhomogeneity, color ratio, and particle edge darkening and boundary clarity. The second feature extraction module extracts a permeability fluctuation amplitude sequence from the raw electromagnetic response sequence. The system first extracts the conductivity decrease slope sequence and then fuses it with the color deviation index to obtain a fused feature vector. A contamination state level determination module determines the contamination state level based on the fused feature vector. A third feature extraction module processes the acquired original electromagnetic response sequence to obtain a refining electromagnetic feature sequence. An impurity ratio determination module determines the internal impurity distribution ratio based on the refining electromagnetic feature sequence. A purity assessment score calculation module calculates the comprehensive purity assessment score based on the internal impurity distribution ratio, oxide layer inhomogeneity, and hue shift angle. A control module generates control parameter adjustment instructions based on the time change of the comprehensive purity assessment score, thereby obtaining an optimized set of production parameters.

Images