Conductive deposition evaluation method and system based on copper sulfur analysis

Abstract

The present application relates to the field of material analysis and electrical equipment fault diagnosis, and discloses a conductive deposition evaluation system based on copper-sulfur analysis, comprising: a basic detection module, which performs conductive deposition testing on a PCB sample and collects resistance time series data and CDF time series data related to the deposition process in real time; an element analysis module, which obtains concentration-related data of copper and sulfur elements and collects dendrite morphology SEM images and sulfur-oxygen element distribution maps of the PCB surface; a multi-modal data preprocessing module, which preprocesses the concentration-related data of copper and sulfur elements and the dendrite morphology SEM images; a deep learning fusion evaluation module, which extracts time series pattern features and image features, fuses the normalized concentration-related data of copper and sulfur elements, the time series pattern features and the image features, and inputs them into a trained machine learning model to obtain a risk evaluation result of conductive deposition. This improves the accuracy, objectivity and adaptability of conductive deposition evaluation.

Description

Technical Field

[0001] This invention relates to the fields of materials analysis and electrical equipment fault diagnosis, and in particular to a conductive deposition evaluation method and system based on copper-sulfur analysis. Background Technology

[0002] The formation of conductive deposits is closely related to the interaction between copper and sulfur, and their evaluation results are crucial for lubricant formulation optimization and transmission system fault tracing. Existing evaluation methods mainly rely on manual analysis of ICP (copper-sulfur concentration) data, SEM (microscopic morphology) data, and electrical resistance data, which have significant limitations:

[0003] The low correlation between data makes it difficult to establish a deep mapping relationship between microscopic elemental characteristics (such as sulfur and oxygen distribution, and copper dissolution morphology) and macroscopic depositional faults. Traditional methods fail to effectively integrate the temporal evolution patterns and spatial features of multimodal data, making it impossible to quantify the impact of dynamic elemental changes on deposition formation.

[0004] SEM image enhancement is ineffective. Existing enhancement techniques often employ a single modality or fixed model parameters, failing to consider the complex characteristics of image degradation. This results in insufficient visual clarity regarding dendrite morphology and sulfur and oxygen element distribution, making it difficult to support accurate feature extraction. Furthermore, the lack of flexible guidance mechanisms and parameter optimization strategies prevents dynamic adjustments to the enhancement process based on the actual image degradation.

[0005] The ability to extract time-series features is insufficient. Traditional time-series analysis methods struggle to capture key patterns in resistance / CDF data and are prone to feature redundancy, failing to effectively reflect the performance changes during conductive deposition.

[0006] Image feature extraction does not make full use of historical data, relying only on single-frame SEM images for feature analysis, ignoring the temporal correlation of the deposition process, resulting in a lack of dynamic evolution information in feature representation, which affects the generalization ability and accuracy of the evaluation model.

[0007] The evaluation results are highly subjective, relying on the experience and judgment of operators, resulting in poor consistency. Furthermore, the evaluation model cannot fully utilize historical elemental analysis data to optimize it, leading to poor adaptability to complex samples and insufficient evaluation accuracy, making it difficult to meet the actual needs of high-precision R&D and quality control.

[0008] Therefore, there is an urgent need for a conductive deposition analysis technology that can deeply integrate multimodal data, automatically and accurately extract features, and achieve objective and accurate evaluation to overcome the above-mentioned shortcomings. Summary of the Invention

[0009] The present invention aims to address the aforementioned deficiencies in existing conductive deposition evaluation methods and to provide a conductive deposition evaluation method and system based on copper-sulfur analysis.

[0010] In a first aspect, embodiments of the present invention provide a conductive deposition evaluation system based on copper-sulfur analysis, comprising:

[0011] The basic testing module is used to perform conductive deposition tests on PCB board samples and to acquire resistance timing data and cumulative distribution function (CDF) timing data related to the deposition process in real time.

[0012] The elemental analysis module is used to analyze the PCB board samples and lubricant samples after the basic detection module has been tested, to obtain the concentration-related data of copper and sulfur elements, and to collect SEM images of dendrite morphology and sulfur and oxygen element distribution maps on the PCB board surface.

[0013] The multimodal data preprocessing module is used to normalize the concentration-related data of copper and sulfur elements and enhance the dendritic morphology SEM image.

[0014] The deep learning fusion assessment module is used to extract key temporal pattern features from preprocessed resistance time series data and cumulative distribution function (CDF) time series data; extract dynamic evolution image features from enhanced current and historical dendrite morphology SEM images; and fuse normalized copper and sulfur concentration correlation data, temporal pattern features, and image features into a trained machine learning model to obtain the risk assessment results of conductive deposition.

[0015] Furthermore, the multimodal data preprocessing module employs a diffusion model with a condition-guided adaptation mechanism and an improved Koschmieder light scattering model to enhance the dendrite morphology SEM image, thereby improving the clarity of the dendrite morphology and elemental distribution.

[0016] Furthermore, the conditional guidance adaptation mechanism includes:

[0017] Four guidance schemes were constructed based on the combination of "guiding variable-guiding domain", where the guiding variable was selected from the intermediate noise image in the diffusion process. Or a clear image predicted by it. The guiding domain is selected from the natural image domain or the SEM image domain;

[0018] By comparing the enhancement effects and stability of the four schemes, the optimal guidance scheme was selected.

[0019] In the optimal guidance scheme, pseudo-label images are generated by processing the original SEM image using an image enhancement algorithm. ; Calculate the sharp image predicted by the diffusion model. With the pseudo-label image Fusion loss;

[0020] Apply the fusion loss to the predicted sharp image The negative gradient is used as a guiding signal to correct the mean of the reverse process of the diffusion model, thereby guiding the generation of the enhanced SEM image.

[0021] Furthermore, the multimodal data preprocessing module also employs the Koschmieder light scattering model to directly perform physical restoration and image enhancement on the original SEM image through its reverse restoration process. The parameter optimization method of the improved Koschmieder light scattering model includes:

[0022] The global background light A is estimated from the original SEM image using the Gaussian blur method, including: performing Gaussian blur on the original SEM image. After Gaussian kernel blurring, select the brightness level. The average value of the pixels is used as the initial value of the global background light A;

[0023] Initialize the medium transmission map T based on the generalized dark channel prior algorithm;

[0024] The original SEM image y is compared with the output image of the improved Koschmieder degradation model. The mean square error is used as the loss, and the gradient descent method is used to dynamically iteratively optimize the global background light A and the medium transmission map T.

[0025] Substituting the optimized global background light A and the medium transmission map T into the restoration formula of the improved Koschmieder light scattering model , where y is the original SEM image and J is the restored clear image, and the enhanced clear image J is calculated.

[0026] Furthermore, the deep learning fusion evaluation module includes:

[0027] The timing feature extraction unit uses the Shapelet Bottleneck model to extract features from the resistance timing data and the cumulative distribution function (CDF) timing data, and outputs a timing feature vector.

[0028] The image feature extraction unit employs a self-attention interaction mechanism based on search frames and memory frames to extract features from the enhanced current SEM image and historical SEM images from historical data, and outputs an image feature vector.

[0029] The conductive deposition risk assessment unit is used to concatenate the concentration data of copper and sulfur elements, time-series feature vectors, and image feature vectors, and input them into a multilayer perceptron model. The model adaptively learns the contribution weight of each feature to the deposition state and outputs a quantified conductive deposition risk score. Combining the risk score with a preset judgment threshold, a comprehensive assessment result is output.

[0030] Furthermore, the time-series feature extraction unit uses the Shapelet Bottleneck model to extract features from the resistance time-series data and the cumulative distribution function (CDF) time-series data, outputting a time-series feature vector, including:

[0031] Define a multi-length shapelet set H, and use Gaussian radial basis functions to convert time series data into logical predicates to capture key patterns;

[0032] The Shapelet diversity loss function is introduced into model training to avoid learning redundant features. Its formula is as follows:

[0033]

[0034] in, : Shapelet diversity loss function; K is the number of shapeslets; M is the number of time-series variables; |H| is the set of shapelet lengths. The number of length types included. and They represent the m-th variable of length l. and the One Shapelet;

[0035] The output includes a 3D time-series feature vector containing the average rate of change of resistance, CDF peak value, and fault occurrence time.

[0036] Furthermore, the self-attention interaction mechanism based on search frames and memory frames specifically includes:

[0037] The current SEM image to be analyzed is used as the search frame, and the associated SEM image selected from historical data is used as the memory frame to form a dual-frame input structure.

[0038] The search frame and memory frame are feature-encoded separately using a convolutional neural network with shared weights to obtain the search frame feature map. and memory frame feature map ;

[0039] Self-attention enhancement is applied to the search frame feature map, and the correlation weights between pixels within the search frame are calculated to generate a self-attention enhanced search frame feature map. This enhances the characteristic expression of the dendritic core region and the sulfur-oxygen aggregation region;

[0040] The association weights between the search frame and memory frame features are calculated, and the memory frame feature maps are weighted and fused based on the association weights to generate a memory feature map related to the search frame that contains spatiotemporal evolution information. ;

[0041] Feature map of the search frame after interaction and memory feature map Channel attention processing and spatial attention processing are performed sequentially, and then the images are stitched together to form the final image feature vector. Channel attention is used to select key feature channels related to sedimentary evolution, while spatial attention is used to enhance the target region and suppress background interference.

[0042] Furthermore, the step of calculating the correlation weight between the features of the search frame and the memory frame, and performing weighted fusion of the memory frame feature maps based on the correlation weight, specifically includes:

[0043] Calculate the similarity between each pixel in the search frame feature map and each pixel in the memory frame feature map to form an association weight matrix. ;

[0044]

[0045] in, Search frame number The pixel and the memory frame The similarity weight of each pixel is used to quantify the correlation strength between the features of two pixels. Search frame number The feature vector of each pixel contains features related to dendrite morphology and sulfur and oxygen element distribution at that location. : Memory Frame The feature vector of each pixel records the deposition-related features at that location at a historical moment; The transpose operation is used to convert a feature vector into a dimension suitable for dot product operations. Vector dot product operation; Feature channel number, which is the dimension of the feature vector of each pixel, represents the richness of the features; : Memory frame pixel index, value range and Consistent, used to iterate through all memory frame pixels to calculate the normalized denominator; : Normalized summation term, ensuring that all memory frame pixels correspond to the search frame's first pixel. The sum of the weights of each pixel is 1;

[0046] The memory frame feature map is weighted and averaged using the aforementioned correlation weight matrix to obtain a memory feature map related to the search frame that contains spatiotemporal evolution information. : .

[0047] Furthermore, the comprehensive evaluation results output by the conductive deposition risk assessment unit include:

[0048] The deposition risk level is determined by combining the conductive deposition risk score with traditional criteria that include at least one of the resistance threshold and CDF distribution characteristics.

[0049] By analyzing the weight ratio of each input feature in the multilayer perceptron model, the main causes of conductive deposition are identified.

[0050] In a second aspect, embodiments of the present invention also provide a conductive deposition evaluation method based on copper-sulfur analysis, applied to the system described in the first aspect, the method comprising:

[0051] Conductive deposition testing is performed using a basic detection module to obtain resistance timing data and CDF timing data.

[0052] The sample was analyzed by the elemental analysis module to obtain the concentration data of copper and sulfur elements, and SEM images of dendrite morphology and sulfur and oxygen element distribution spectrum were collected on the PCB board surface.

[0053] The concentration data of copper and sulfur elements were normalized, and the SEM image of the dendrite morphology was enhanced.

[0054] Key temporal pattern features are extracted from preprocessed resistance time-series data and cumulative distribution function (CDF) time-series data; dynamic evolution image features are extracted from enhanced current and historical dendrite morphology SEM images; and the concentration correlation data of copper and sulfur elements after normalization, temporal pattern features, and image features are fused and input into a trained machine learning model to obtain the risk assessment results of conductive deposition.

[0055] Compared with the prior art, the present invention achieves the following beneficial effects:

[0056] 1. Based on the conditional guided adaptation mechanism, this invention designs four SEM image enhancement guidance methods with combinations of "guided variables-guided domains" and selects the optimal guidance scheme. By correcting the diffusion model generation process through gradient guidance, the flexibility and accuracy of SEM image enhancement are improved, laying the foundation for subsequent feature extraction.

[0057] 2. This invention is based on an improved Koschmieder light scattering model parameter optimization method. By dynamically estimating the global background light and medium transmission map parameters, it solves the problem of poor enhancement effect caused by fixed parameters in traditional models, effectively reduces SEM image blur and color deviation, and clearly presents dendrite morphology and sulfur and oxygen element distribution.

[0058] 3. This invention constructs a Shapelet Bottleneck model for feature extraction from resistance / CDF time series data. It designs a Shapelet diversity loss function to avoid feature redundancy, accurately captures key patterns in time series data (such as resistance change rate and CDF peak position), and improves the discriminative ability of time series features.

[0059] 4. This invention extracts SEM image features based on the self-attention interaction mechanism of search frames and memory frames, making full use of the depositional evolution information of historical detection images, strengthening the feature expression of the target region, and making up for the lack of dynamic correlation of single-frame image features.

[0060] 5. This invention designs a search frame-memory frame interaction strategy to generate a weight matrix, and realizes the extraction of cross-frame spatiotemporal correlation features through pixel-level similarity calculation. It integrates the self-attention features of the search frame and the historical memory features to enrich the dimension and information content of image features.

[0061] Ultimately, by fusing multimodal data, a precise "elemental characteristics-deposition state" correlation model is constructed, which improves the accuracy, objectivity, and adaptability of conductive deposition assessment, and provides scientific and reliable technical support for lubricant formulation optimization and transmission system fault tracing. Attached Figure Description

[0062] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings. The drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0063] Figure 1 This is a schematic diagram of the module structure of a conductive deposition evaluation system based on copper-sulfur analysis provided in an embodiment of the present invention;

[0064] Figure 2 This is a schematic diagram of the system architecture of a conductive deposition evaluation system based on copper-sulfur analysis provided in an embodiment of the present invention;

[0065] Figure 3 This is a schematic diagram of the image enhancement processing flow according to an embodiment of the present invention;

[0066] Figure 4 This is a schematic diagram of the deep learning fusion evaluation structure according to an embodiment of the present invention;

[0067] Figure 5 This is a schematic diagram of the structure of an electronic device provided in Embodiment 3 of the present invention. Detailed Implementation

[0068] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the present invention, and not all of the structures.

[0069] Before discussing the exemplary embodiments in more detail, it should be mentioned that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe the operations (or steps) as sequential processes, many of the operations (or steps) can be performed in parallel, concurrently, or simultaneously. Furthermore, the order of the operations can be rearranged. The process can be terminated when its operation is completed, but it may also have additional steps not included in the figures. The process may correspond to a method, function, procedure, subroutine, subroutine, etc.

[0070] The acquisition, storage, use, and processing of data in this application comply with relevant national laws and regulations. It should be noted that existing industry solutions such as software, components, or models may be mentioned in the embodiments of this application. These should be considered exemplary and intended only to illustrate the feasibility of implementing the technical solution of this application, but do not imply that the applicant has already used or necessarily used such a solution.

[0071] Figure 1 This is a schematic diagram of the module structure of a conductive deposition evaluation system based on copper-sulfur analysis provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the system architecture of a conductive deposition evaluation system based on copper-sulfur analysis provided in an embodiment of the present invention. Figure 1 and Figure 2 As shown, a conductive deposition evaluation system 100 based on copper-sulfur analysis includes: a basic detection module 110, an elemental analysis module 120, a multimodal data preprocessing module 130, and a deep learning fusion evaluation module 140.

[0072] 1.1: Basic Detection Module 110

[0073] The basic testing module 110 includes a sample holding unit 111, a temperature monitoring unit 112, a resistance / CDF testing unit 113, and a circuit control unit 114. It is used to conduct conductive deposition tests and record key performance data. Furthermore, it is used to conduct conductive deposition tests on PCB board samples and to collect resistance timing data and cumulative distribution function (CDF) timing data related to the deposition process in real time.

[0074] Specifically, the basic testing process for conductive deposition is as follows: In this embodiment of the invention, a conductive deposition test is conducted for 500 hours, and resistance data and CDF (cumulative distribution function) data are recorded in real time; after the test, PCB board samples containing conductive deposits and lubricant samples are collected for subsequent elemental analysis.

[0075] 1.2: Elemental Analysis Module 120

[0076] The elemental analysis module 120 is used to analyze the PCB board samples and lubricant samples after the basic detection module 110 tests, obtain the concentration-related data of copper and sulfur elements, and acquire SEM images of dendrite morphology and sulfur and oxygen element distribution maps on the PCB board surface.

[0077] Specifically, the elemental analysis module 120 is equipped with an ICP analyzer and a scanning electron microscope (SEM). The ICP analyzer is used to acquire concentration-related data of copper and sulfur, including four core data types: soluble copper concentration (ppm), particulate copper mass (g), and sulfur concentration variation data. The SEM acquires two types of image data: dendrite morphology images of the PCB board surface (256×256 pixels resolution) and sulfur and oxygen element distribution maps, with the distribution map outputting the percentage of sulfur and oxygen content on the sample surface.

[0078] 1.3: Multimodal Data Preprocessing Module 130

[0079] The multimodal data preprocessing module 130 is used to preprocess the time-series data and image data output by the elemental analysis module 120. Specifically, it normalizes the concentration correlation data of copper and sulfur elements and enhances the dendrite morphology SEM images. This includes the following two steps:

[0080] Step 1.3.1: Time Series Data Preprocessing

[0081] The copper and sulfur concentration data obtained from ICP were normalized to eliminate the influence of dimensional differences on model training. Specifically, the min-max normalization method was used to map the data to the [0,1] interval.

[0082] Step 1.3.2: Image Data Preprocessing

[0083] Step 1.3.2 uses a conditional guided adaptation mechanism and an improved Koschmieder light scattering model to enhance the dendrite morphology SEM image. The core of this step is to use a guiding signal generated from the original SEM image to control and correct the image generation direction of the diffusion model, so as to output a clearer dendrite and element distribution image, thereby improving the clarity of dendrite morphology and element distribution. Figure 3 This is a schematic diagram of the image enhancement processing flow according to an embodiment of the present invention, as shown below. Figure 3 As shown, the specific steps are as follows:

[0084] Step 1.3.2.1: Construction of Condition-Guided Adaptation Mechanism

[0085] In a preferred embodiment, the conditional guided adaptation mechanism is implemented in the following way:

[0086] In this embodiment of the invention, the conditional guidance adaptation mechanism specifically involves constructing four guidance schemes based on the combination of "guiding variable-guiding domain," wherein the guiding variable is selected from intermediate noise images during the diffusion process. Or a clear image predicted by it. The guidance domain is selected from either the natural image domain or the SEM image domain; by comparing the enhancement effects and stability of four schemes, the optimal guidance scheme is selected; in the optimal guidance scheme, a pseudo-label image is generated by processing the original SEM image using an image enhancement algorithm. ; Calculate the clear image predicted by the diffusion model With pseudo-label images The fusion loss; applying the fusion loss to the predicted sharp image. The negative gradient is used as a guiding signal to correct the mean of the reverse process of the diffusion model, thereby guiding the generation of the enhanced SEM image.

[0087] More specifically, drawing inspiration from classifier guidance, the guiding signal for SEM image enhancement tasks is transformed into gradients to correct the mean of the reverse process in the diffusion model. First, SEM image enhancement is modeled as based on degraded input. A clear image of the original SEM image. The conditional probability formula for the reverse process of the posterior probability sampling process is:

[0088]

[0089] In the reverse diffusion model process, given the original SEM image... and the intermediate noise image at the current time. Under the condition of generating the intermediate noise image of the previous time step. The conditional probability. : Normalization constant, used to ensure the normalization of probability distribution. The fundamental transition probability of the reverse process in the diffusion model follows a Gaussian distribution, representing the transition from the initial state to the final state in the absence of a guiding signal. generate The probability of. The original SEM image (degraded input image) contains information on the dendrite morphology and sulfur and oxygen distribution to be enhanced. During the diffusion process The intermediate noise image at time step is a transitional state image during the image enhancement process. Original SEM image With intermediate noise image The matching degree is used to quantify the strength of the pilot signal; a higher matching degree indicates... The better the consistency with the features of the original image. This is achieved through... Constructing the matching loss function by taking the negative logarithm Further calculate the gradient The gradient is then incorporated into the mean correction of the reverse process to achieve the control of image generation by the guiding signal. :based on The constructed matching loss function, through the... Obtained by taking the negative logarithm, used for measurement and The differences. Matching loss function pairs The negative gradient is used as a guiding signal to correct the mean of the reverse process in the diffusion model. This guides the image generation process towards optimizing the enhancement target.

[0090] In this embodiment of the invention, the Diffusion Model serves as the primary image enhancement engine. Its task is to start from the original, blurry SEM image and, through a progressive denoising "reverse process," generate a clear, high-quality image. Its workflow involves a series of steps to gradually remove noise and blur from the image, ultimately outputting the enhanced image. In this process, a guiding signal is needed to ensure that the generated result is a correct, clear SEM image, and not some other irrelevant image. Preferably, the present invention uses the Koschmieder degradation model to provide this guiding signal for the generation process of the diffusion model.

[0091] Step 1.3.2.2: Design and Parameter Calculation of Four Guiding Methods

[0092] Based on the combined dimension of "guided variable - guided domain", four SEM image enhancement guided schemes are designed:

[0093] Step 1.3.2.2.1: Selection of guiding variables

[0094] Guiding variables include intermediate noise images during the diffusion process. and by Predicted clear image ,in The noise prediction function is calculated using the diffusion model, and the formula is as follows:

[0095]

[0096]

[0097]

[0098] : From the diffusion process Intermediate noise image at time step The predicted clear image is a target state image enhanced by SEM image, which includes optimized dendrite morphology and sulfur and oxygen element distribution characteristics. During the diffusion process The intermediate noise image at time step is a transitional state image during the image enhancement process. The noise predicted by the diffusion model is learned by the model from... Intermediate noise image at time step Extracted from [the image], used to reverse engineer a clear image. . : Pre-diffusion process step The cumulative product is used to quantify the degree of noise addition, and the formula is: . During the diffusion process The noise attenuation coefficient at time step, and the variance scheduling satisfy This is used to control the intensity of noise addition at each step. The fixed variance scheduling of the diffusion process is a preset constant sequence that determines the variance of the noise added to the original image at each step.

[0099] Step 1.3.2.2.2: Boot Domain Selection

[0100] The guidance domain includes the natural image domain and the SEM image domain. The natural image domain generates pseudo-label images by inputting the original SEM image into any existing image enhancement algorithm (such as multi-scale fusion algorithm). To achieve this, the SEM image domain will diffuse sample the image ( or Generate SEM-style images from an improved Koschmieder degradation model. The degradation model formula is as follows:

[0101]

[0102] SEM-style images generated using an improved Koschmieder degradation model are used for matching with the original SEM image during SEM image domain guidance. The clear image predicted by the diffusion model serves as the base image for the input of the degradation model. A: Medium transmission map, representing the proportion of radiation reaching the detector in the SEM imaging scene, affecting image sharpness and detail retention. B: Global background light, representing the ambient background brightness of the SEM imaging, used to correct overall image brightness deviations. Pseudo-labeled images in the natural image domain are generated by processing the original SEM image using existing image enhancement algorithms (such as multi-scale fusion algorithms) and serve as a guiding benchmark.

[0103] The four combination schemes are: Scheme 1 ( Natural Image Domain), Scheme 2 ( Natural Image Domain), Scheme ( SEM image domain) Option 4 SEM image domain).

[0104] Step 1.3.2.3: Selection and Implementation of the Optimal Guidance Scheme

[0105] By comparing the enhancement effects of the four schemes (indicated by dendrite morphological clarity and sulfur and oxygen element distribution recognition) and model stability (indicated by loss function convergence speed), the scheme was determined to be " The optimal solution is "variable + natural image domain". In practice, pseudo-labels are first generated using existing enhancement algorithms. ,calculate and Fusion loss:

[0106]

[0107] Pseudo-label images With predicted clear images The fusion loss function is used to measure the degree of difference between the two. Pseudo-labeled images from the natural image domain serve as a guiding benchmark for enhancement effects. The clear image predicted by the diffusion model is the target state image enhanced by SEM image. Mean absolute error loss ( The weighting coefficient of ) is set to 0.6, which is used to adjust the contribution ratio of this loss in the fusion loss. Multi-scale structural similarity loss The weighting coefficient of ) is set to 0.4, which is used to adjust the contribution ratio of this loss in the fusion loss. Mean absolute error loss, used for quantization and The absolute difference is at the pixel level. Multi-scale structural similarity loss is used to measure... and The multi-scale structural similarity is consistent with the visual characteristics of the human eye.

[0108] The fusion loss will affect the predicted sharp image. The negative gradient is used as a guiding signal to correct the mean of the reverse process of the diffusion model, thereby guiding the generation of the enhanced SEM image. Specifically, through the gradient... Mean of the reverse process in the corrected diffusion model (The corrected mean is) (where is the gradient scaling factor), guiding the inverse denoising process to generate a high-quality enhanced image. : Fusion loss function pair The negative gradient is used as a guiding signal to correct the mean of the reverse process in the diffusion model. . : The original mean of the reverse process of the diffusion model, used to generate the intermediate image of the previous time step. : Gradient scaling factor, with a value of 0.8, is used to adjust the strength of the guiding gradient and avoid excessively large gradients that could lead to model instability. :Right now fusion loss function pair The negative gradient.

[0109] In summary, the image data preprocessing (image enhancement) of this invention includes two stages: the first stage is dual-path image enhancement, and the second stage is result fusion and output. The first stage further includes two technical paths. Steps 1.3.2.1 to 1.3.2.3 are technical path 1 of this embodiment: enhancement based on a conditionally guided adaptive diffusion model, aiming to utilize the powerful generative capabilities of the diffusion model to generate a clear image rich in detail. The process is as follows:

[0110] (1) Input the original SEM image y into the diffusion model.

[0111] (2) The model generates a clear image step by step through a reverse denoising process. During this process, a "guide signal" is needed to ensure that the output is a correct SEM image.

[0112] (3) The core of the guidance mechanism: Four combination schemes of "guiding variable-guiding domain" were designed. Among them, the SEM image domain guidance scheme (scheme 3 and scheme 4) uses the forward process of the Koschmieder model.

[0113] (4) The role of the Koschmieder forward process: At a certain step in the diffusion process, there is an intermediate result (such as a clear prediction map). Substitute it into the positive degradation formula: A simulated degradation SEM image was obtained. .calculate The difference (loss) between the generated image and the original, blurred image y is the difference between the generated image and the original input image. The gradient generated by this loss is fed back to the diffusion model, telling it: "The image you generate, after physical degradation, should be as similar as possible to the original image." This constrains the direction of the diffusion model's generation, making it conform to physical laws.

[0114] (5) Finally, the diffusion model outputs an enhanced image.

[0115] Step 1.3.2.4: Parameter Optimization of the Improved Koschmieder Light Scattering Model

[0116] The multimodal data preprocessing module 130 also employs an improved Koschmieder light scattering model to directly perform physical restoration and image enhancement on the original SEM image through its reverse restoration process. The parameter optimization method of this improved Koschmieder light scattering model includes: estimating the initial value of the global background light A from the original SEM image using Gaussian blurring; and further includes: processing the original SEM image... After Gaussian kernel blurring, select the brightness level. The average value of pixels as Initial values; the medium transmission map T is initialized based on the generalized dark channel prior algorithm; the mean square error between the original SEM image and the model output image is used as the loss, and the gradient descent method is used to dynamically iteratively optimize the medium transmission map T; the global background light A and the optimized medium transmission map T are substituted into the restoration formula of the light scattering model to calculate the enhanced clear image.

[0117] In summary, steps 1, 3, 2, and 4 constitute the second technical path of the first stage of this invention: image enhancement based on the improved Koschmieder physical restoration model. The goal is to directly invert and calculate a clear image using a light scattering physical model. This process includes:

[0118] The global background light A is estimated using Gaussian blurring, and the original SEM image is then processed. After Gaussian kernel blurring, select the brightness level. The average value of the pixels is used as the value of the global background light A.

[0119] The medium transmission map T is initialized based on the GDCP (Generalized Dark Channel Prior) algorithm. After calculating the dark channel of the image (the image composed of the minimum values ​​of each pixel in the RGB three channels), the transmission map is then processed. The mean value of the dark channel within the sliding window is calculated as the initial T.

[0120] Parameter optimization: Gaussian blur, GDCP, and gradient descent are used to dynamically optimize the global background light A and the medium transmission map T. Preferably, gradient descent is used to dynamically update the global background light A and the medium transmission map T based on the original SEM image. Output of the degradation model The mean squared error is the loss function, and the iteration is performed 100 times (learning rate). Finally, the parameters of the stable global background light A and the medium transmission map T are obtained.

[0121] Image restoration: Substitute the optimized A and T, along with the original SEM image y, into the reverse restoration formula: Where y represents the original SEM image and J represents the restored clear image, thus completing the SEM image enhancement. This formula directly performs mathematical calculations to cancel out the blurring effect caused by scattering, outputting a physically restored clear image, effectively reducing image blur and color deviation problems, and improving the visual clarity of dendrite morphology and sulfur and oxygen element distribution.

[0122] Phase Two: Result Fusion and Output. In this invention, the output of Path 2 (Koschmieder Restoration Model), i.e., the physically restored image, is used as a pseudo-label guided by the natural image domain in Path 1 (Diffusion Model). In other words, the sharpened image J obtained by enhancing the original SEM image through the inverse restoration process of the Koschmieder light scattering model is used as the pseudo-label image generated in the conditional guided adaptation mechanism. The foundation.

[0123] In summary, this invention first obtains a basic enhancement result through a physical model, and then uses this result to guide a more powerful generative model, enabling it to generate higher-quality images. This is a clever fusion. Ultimately, the system selects the diffusion model as the optimal guiding scheme (…). The image generated under the variable + natural image domain is used as the final enhancement result.

[0124] 1.4: Deep Learning Fusion Evaluation Module 140

[0125] The deep learning fusion evaluation module 140 is used to extract key temporal pattern features from preprocessed resistance time series data and cumulative distribution function (CDF) time series data; extract dynamic evolution image features from enhanced current and historical dendrite morphology SEM images; fuse normalized copper and sulfur concentration data, temporal pattern features and image features, and input them into a trained machine learning model to obtain the risk assessment results of conductive deposition.

[0126] Furthermore, the deep learning fusion evaluation module 140 is used to realize deep learning feature extraction and fusion. It has a built-in temporal feature extraction unit 141, a multimodal data fusion model 142 and an evaluation result output unit 143. The image feature extraction unit 142 includes a convolutional neural network (CNN), and the evaluation result output unit 143 includes a joint architecture of multilayer perceptron (MLP). The CNN is used for SEM image feature extraction, and the MLP is used for multi-source data fusion analysis.

[0127] Figure 4 This is a schematic diagram of the deep learning fusion evaluation structure according to an embodiment of the present invention, as shown below. Figure 4 As shown, the deep learning fusion evaluation module 140 includes: a temporal feature extraction unit 141, an image feature extraction unit 142, and a conductive deposition risk assessment unit 143.

[0128] The timing feature extraction unit 141 uses the Shapelet Bottleneck model to extract features from the resistance timing data and the cumulative distribution function (CDF) timing data, and outputs a timing feature vector.

[0129] Furthermore, the timing feature extraction unit 141 uses the Shapelet Bottleneck model to extract features from the resistance timing data and the cumulative distribution function (CDF) timing data, and outputs a timing feature vector, including: a 3-dimensional timing feature vector containing the average rate of change of resistance, the peak value of CDF, and the fault occurrence time.

[0130] 1.4.1: Temporal Feature Extraction

[0131] The specific process of temporal feature extraction unit 141 performing temporal feature extraction is as follows:

[0132] 1.4.1.1: Define a collection of multi-length Shapelets

[0133] Time series data is converted into logical predicates using Gaussian radial basis functions to capture key patterns in the data (such as resistance change rate and CDF peak location).

[0134] 1.4.1.2: Introducing the Shapelet Diversity Loss Function

[0135] The Shapelet diversity loss function is introduced to avoid learning redundant features. The loss function formula is as follows:

[0136]

[0137] in, Shapelet diversity loss function is used to avoid the model learning redundant shapelet features and improve the ability to distinguish temporal features. : Number of Shapelets, i.e., the total number of feature subsequences preset to capture temporal patterns. : Number of timing data variables, which is 2 here (corresponding to resistance data and CDF data respectively). Shapelet length collection The number of length types included. For a preset set of different length values ​​(such as ). : Index of time series data variable, with a value range of 1 to This corresponds to different types of timing data (1 is resistance data, 2 is CDF data). Shapelet length, values ​​are taken from the Shapelet length set. , which represents the time step of a single Shapelet. Shapelet index, with a value range of 1 to... This is used to iterate through all Shapelets. Shapelet index, with a value range of 1 to... and Used for traversing except Other Shapelets besides these. : No. Among the time series variables, the one with a length of The A Shapelet is a feature subsequence that captures a specific temporal pattern of the variable. : No. Among the time series variables, the one with a length of The A Shapelet, with These are different Shapelets with the same variable and the same length. and The 2-norm (Euclidean distance) is used to measure the similarity between two Shapelets. : An exponential decay term based on Euclidean distance. The smaller the distance, the larger the value of this term, which represents the degree of redundancy between Shapelets.

[0138] The final output is a 3D time-series feature vector, including the average rate of change of resistance, CDF peak value, and fault occurrence time.

[0139] 1.4.2: Image Feature Extraction

[0140] The image feature extraction unit 142 uses a self-attention interaction mechanism based on search frames and memory frames to extract features from the enhanced current SEM image and historical SEM images from historical data, and outputs an image feature vector.

[0141] The image feature extraction unit 142 of this embodiment extracts SEM image features based on a self-attention interaction mechanism between search frames and memory frames. Further, the self-attention interaction mechanism based on search frames and memory frames specifically includes: using the current SEM image to be analyzed as the search frame, and combining it with a related SEM image selected from historical data as the memory frame, forming a dual-frame input structure; and using a convolutional neural network with shared weights to encode features in the search frame and memory frame respectively, obtaining a search frame feature map. and memory frame feature map The search frame feature map is subjected to self-attention enhancement, and the correlation weights between pixels within the search frame are calculated to generate a self-attention enhanced search frame feature map. This process strengthens the feature representation of the dendritic core region and the sulfur-oxygen aggregation region; calculates the correlation weight between the features of the search frame and the memory frame, and performs weighted fusion of the memory frame feature map based on the correlation weight to generate a memory feature map related to the search frame that contains spatiotemporal evolution information. Feature maps of search frames after interaction and memory feature map Channel attention processing and spatial attention processing are performed sequentially, and then the images are stitched together to form the final image feature vector. Channel attention is used to select key feature channels related to sedimentary evolution, while spatial attention is used to enhance the target region and suppress background interference.

[0142] The specific steps for extracting SEM image features based on the self-attention interaction mechanism between search frames and memory frames include:

[0143] 1.4.2.1: Constructing a dual-frame input structure

[0144] The current SEM image is defined as the search frame (the image to be analyzed, containing dendrite morphology and sulfur and oxygen distribution). Previous detection images of the same batch of samples are selected from the historical SEM image library as memory frames (reference images that record the sedimentary evolution process). Both are preprocessed and then input into the CNN.

[0145] 1.4.2.2: Feature Encoding and Self-Attention Enhancement

[0146] The search frame and memory frame are feature-encoded separately using CNN branches with shared weights to generate a search frame feature map. (Query features) and memory frame feature maps (Memory Features). Self-attention enhancement is applied to the search frame feature map, and the correlation weights between pixels within the search frame are calculated to generate the self-attention-enhanced search frame features. This enhances the feature representation of the dendrite core region and the sulfur-oxygen aggregation region.

[0147] 1.4.2.3: Search Frame - Memory Frame Interaction

[0148] Calculate the pixel-level similarity between the memory frame and the search frame to generate a weight matrix. The weight calculation formula is as follows:

[0149]

[0150] Search frame number The pixel and the memory frame The similarity weights of each pixel are used to quantify the correlation strength between the features of two pixels. Search frame number The feature vector of each pixel contains features related to the dendrite shape and sulfur and oxygen distribution at that location. : Memory Frame The feature vector of each pixel records the deposition-related features at that location at a historical moment. : Transpose operation, used to convert feature vectors to dimensions suitable for dot product operations. The vector dot product operation is used to calculate the similarity between the feature vectors of two pixels. Feature channel number, which is the dimension of the feature vector of each pixel, represents the richness of the features. The square root of the number of feature channels is used to scale the dot product result to avoid excessive gradients. : Exponential function, used to strengthen the weighting of highly similar pixel pairs. : Memory frame pixel index, value range and Consistent, used to iterate through all memory frame pixels to calculate the normalized denominator. : Normalized summation term, ensuring that all memory frame pixels correspond to the search frame's first pixel. The sum of the weights of each pixel is 1.

[0151] Based on the weight matrix A weighted average operation is performed on the memory frame feature map to obtain a memory feature map that contains spatiotemporal evolution information and is related to the search frame. This enables the extraction of spatiotemporal correlation features across frames.

[0152]

[0153] in, : Memory frame feature map, which contains the set of feature vectors of all memory frame pixels. The memory feature map associated with the search frame is obtained by weighted averaging of the memory frame feature map and the weight matrix, and the cross-frame spatiotemporal correlation features are fused.

[0154] 1.4.2.4: Enhanced Channel and Spatial Attention

[0155] For the feature map after interaction (including and The mean, maximum and standard deviation of each channel are calculated by the channel attention module to screen out the key channels related to dendrite growth and sulfur-oxygen accumulation; then the spatial attention module generates a spatial weight matrix to suppress the interference of background noise on the target area.

[0156] 1.4.2.5: Feature Integration and Output

[0157] Concatenating search frame features after attention enhancement Inter-frame interaction features Output the final interaction features After multi-scale feature fusion, a 256-dimensional micro-feature vector is generated, which covers dynamic evolution features such as dendrite length change rate (growth relative to memory frame), density gradient, sulfur and oxygen accumulation area expansion rate, and element content gradient difference.

[0158] The final interaction feature map is a fusion of search frame self-attention features and cross-frame interaction features. Feature concatenation operation merges two feature maps along the channel dimension. The self-attention-enhanced search frame feature map strengthens the feature representation of the target region (dendrite, sulfur-oxygen aggregation region) within the search frame. : Memory feature map associated with the search frame, containing related features of historical sedimentary evolution.

[0159] 1.4.3: Multimodal Data Fusion and Risk Assessment

[0160] The conductive deposition risk assessment unit 143 is used to concatenate the concentration data of copper and sulfur elements, time-series feature vectors and image feature vectors, and input them into the multilayer perceptron model. The model adaptively learns the contribution weight of each feature to the deposition state and outputs a quantified conductive deposition risk score. Combining the risk score with a preset judgment threshold, a comprehensive assessment result is output.

[0161] The conductive deposition risk assessment unit 143 is used to achieve multimodal data fusion. The specific process is as follows:

[0162] 1.4.3.1: Feature splicing

[0163] The ICP data (4-dimensional: soluble copper concentration, particle copper mass, initial sulfur concentration, and post-test sulfur concentration), time-series feature vector (3-dimensional), and image micro-feature vector (256-dimensional) are concatenated to form a 263-dimensional multimodal fusion feature.

[0164] 1.4.3.2: Weight Learning

[0165] The fused features are input into the MLP model, and the model parameters are optimized through training. This enables the MLP to adaptively learn the influence weights of each feature on the conductive deposition state, establishing a nonlinear mapping relationship between "elemental features and deposition state".

[0166] 1.4.3.3: Risk Score

[0167] MLP outputs a conductive deposition risk score, ranging from 0 to 10 points, with ≥6 points indicating a high risk level and <6 points indicating a low to medium risk level.

[0168] 1.4.3.4: Evaluation of Conductive Deposition Results

[0169] Furthermore, the comprehensive assessment results output by the conductive deposition risk assessment unit 143 include:

[0170] Combining the MLP risk score with traditional criteria (resistance threshold, CDF distribution characteristics), a comprehensive evaluation result is output. For example, when the risk score is ≥6 and the resistance exceeds the standard / CDF peak is abnormal, it is judged as high-risk deposition. By analyzing the weight ratio of each input feature in the multilayer perceptron model, the main causes of conductive deposition are determined. For example, if the weight ratio of soluble copper concentration is the highest, it is judged as copper dissolution-dominated; if the weight ratio of sulfur concentration change and sulfur-oxygen distribution characteristics is high, it is judged as sulfur-oxygen reaction-dominated.

[0171] This invention provides a conductive deposition assessment system based on copper-sulfur analysis. By employing a conditionally guided diffusion model and an improved Koschmieder physics model to synergistically enhance SEM images, the system significantly improves the clarity of dendrite and elemental distribution. Utilizing a Shapelet Bottleneck model and a search-memory frame interaction mechanism, it accurately extracts key features and dynamic evolution information from time-series data and image sequences, respectively. Finally, through multimodal feature fusion and an MLP model, it adaptively learns the weights of each feature to output an objective and quantitative risk assessment result. This system effectively overcomes the shortcomings of traditional methods, such as reliance on manual labor, one-sided feature extraction, and strong subjectivity in assessment. It achieves high-precision and highly objective intelligent assessment and tracing of the causes of conductive deposition, providing reliable technical support for lubricant formulation optimization and transmission system fault prevention.

[0172] Example 2

[0173] This invention also provides a conductive deposition evaluation method 200 based on copper-sulfur analysis, applied to the aforementioned conductive deposition evaluation system 100 based on copper-sulfur analysis. The conductive deposition evaluation method 200 based on copper-sulfur analysis includes:

[0174] S1: Conduct conductive deposition test through basic detection module 110 to obtain resistance timing data and CDF timing data;

[0175] S2: The elemental analysis module 120 analyzes the tested sample to obtain the concentration data of copper and sulfur elements, and collects SEM images of dendrite morphology and sulfur and oxygen element distribution maps on the PCB board surface.

[0176] S3: Normalize the concentration data of copper and sulfur elements, and enhance the SEM image of dendrite morphology.

[0177] S4: Extract key temporal pattern features from preprocessed resistance time series data and cumulative distribution function (CDF) time series data; extract dynamic evolution image features from enhanced current and historical dendrite morphology SEM images; fuse normalized copper and sulfur concentration data, temporal pattern features and image features, and input them into a trained machine learning model to obtain the risk assessment results of conductive deposition.

[0178] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the steps described in the method can be referred to the corresponding module working process in the foregoing system embodiments, and will not be repeated here.

[0179] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, because according to this disclosure, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily essential to this application.

[0180] Example 3

[0181] Figure 5 This is a schematic diagram of the structure of an electronic device provided in Embodiment 3 of the present invention. The electronic device 10 is intended to represent various forms of digital computers, and may also represent various forms of mobile devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the invention described and / or claimed herein.

[0182] like Figure 5 As shown, the electronic device 10 includes at least one processor 11 and a memory, such as a read-only memory (ROM) 12 or a random access memory (RAM) 13, communicatively connected to the at least one processor 11. The memory stores computer programs executable by the at least one processor. The processor 11 can perform various appropriate actions and processes based on the computer program stored in the ROM 12 or loaded into the RAM 13 from storage unit 18. The RAM 13 can also store various programs and data required for the operation of the electronic device 10. The processor 11, ROM 12, and RAM 13 are interconnected via a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0183] Multiple components in electronic device 10 are connected to I / O interface 15, including: input unit 16, such as keyboard, mouse, etc.; output unit 17, such as various types of displays, speakers, etc.; storage unit 18, such as disk, optical disk, etc.; and communication unit 19, such as network card, modem, wireless transceiver, etc. Communication unit 19 allows electronic device 10 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0184] Processor 11 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 11 include, but are not limited to, central processing unit (CPU), graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, digital signal processors (DSPs), and any suitable processor, controller, microcontroller, etc. Processor 11 performs the conductive deposition evaluation method 200 based on copper-sulfur analysis described above.

[0185] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0186] The above embodiments are merely illustrative examples and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A conductive deposition evaluation system based on copper-sulfur analysis, characterized in that, include: The basic testing module is used to perform conductive deposition tests on PCB board samples and to acquire resistance timing data and cumulative distribution function (CDF) timing data related to the deposition process in real time. The elemental analysis module is used to analyze the PCB board samples and lubricant samples after the basic detection module has been tested, to obtain the concentration-related data of copper and sulfur elements, and to collect SEM images of dendrite morphology and sulfur and oxygen element distribution maps on the PCB board surface. The multimodal data preprocessing module is used to normalize the concentration-related data of copper and sulfur elements and enhance the dendritic morphology SEM image. The multimodal data preprocessing module uses a diffusion model with a condition-guided adaptation mechanism and an improved Koschmieder light scattering model to enhance the dendritic morphology SEM image, thereby improving the clarity of dendritic morphology and elemental distribution. The conditional guidance adaptation mechanism includes: constructing four guidance schemes based on the combination of "guiding variable-guiding domain", wherein the guiding variable is selected from the intermediate noise image in the diffusion process. Or a clear image predicted by it. The guidance domain is selected from either the natural image domain or the SEM image domain; by comparing the enhancement effects and stability of four schemes, the optimal guidance scheme is selected; in the optimal guidance scheme, a pseudo-label image is generated by processing the original SEM image using an image enhancement algorithm. ; Calculate the sharp image predicted by the diffusion model. With the pseudo-label image The fusion loss; applying the fusion loss to the predicted sharp image. The negative gradient is used as a guiding signal to correct the mean of the reverse process of the diffusion model, thereby guiding the generation of the enhanced SEM image. The multimodal data preprocessing module also employs the Koschmieder light scattering model to directly perform physical restoration and image enhancement on the original SEM image through its reverse restoration process. The parameter optimization method for the improved Koschmieder light scattering model includes: estimating the global background light A from the original SEM image using Gaussian blurring, including: performing... After Gaussian kernel blurring, select the brightness level. The mean value of the pixels is used as the initial value of the global background light A; the medium transmission map T is initialized based on the generalized dark channel prior algorithm; the original SEM image y and the improved Koschmieder degradation model are used to output the image. The mean squared error is used as the loss, and the global background light A and the medium transmission map T are dynamically and iteratively optimized using the gradient descent method; the optimized global background light A and the medium transmission map T are then substituted into the restoration formula of the improved Koschmieder light scattering model. Where y is the original SEM image and J is the restored clear image, the enhanced clear image J is calculated; The deep learning fusion assessment module is used to extract key temporal pattern features from preprocessed resistance time series data and cumulative distribution function (CDF) time series data; extract dynamic evolution image features from enhanced current and historical dendrite morphology SEM images; and fuse normalized copper and sulfur concentration correlation data, temporal pattern features, and image features into a trained machine learning model to obtain the risk assessment results of conductive deposition.

2. The system of claim 1, wherein, The deep learning fusion evaluation module includes: The timing feature extraction unit uses the Shapelet Bottleneck model to extract features from the resistance timing data and the cumulative distribution function (CDF) timing data, and outputs a timing feature vector. The image feature extraction unit employs a self-attention interaction mechanism based on search frames and memory frames to extract features from the enhanced current SEM image and historical SEM images from historical data, and outputs an image feature vector. The conductive deposition risk assessment unit is used to concatenate the concentration-related data of copper and sulfur elements, time-series feature vectors, and image feature vectors, and input them into a multilayer perceptron model. The model adaptively learns the contribution weight of each feature to the deposition state and outputs a quantified conductive deposition risk score. Combining the risk score with a preset judgment threshold, a comprehensive assessment result is output.

3. The system of claim 2, wherein, The time-series feature extraction unit uses the Shapelet-Bottleneck model to extract features from the resistance time-series data and the cumulative distribution function (CDF) time-series data, and outputs a time-series feature vector, including: Define a multi-length shapelet set H, and use Gaussian radial basis functions to convert time series data into logical predicates to capture key patterns; The Shapelet diversity loss function is introduced into model training to avoid learning redundant features. Its formula is as follows: ； in, : Shapelet diversity loss function; K is the number of shapeslets; M is the number of time-series variables; |H| is the set of shapelet lengths. The number of length types included. and They represent the m-th variable of length l. and the One Shapelet; The output includes a 3D time-series feature vector containing the average rate of change of resistance, CDF peak value, and fault occurrence time.

4. The system of claim 2, wherein, The self-attention interaction mechanism based on search frames and memory frames specifically includes: The current SEM image to be analyzed is used as the search frame, and the associated SEM image selected from historical data is used as the memory frame to form a dual-frame input structure. The search frame and the memory frame are respectively encoded by a convolutional neural network sharing weights to obtain a search frame feature map and a memory frame feature map ; The search frame feature map is subjected to self-attention enhancement, correlation weights between pixels in the search frame are calculated, and a search frame feature map subjected to self-attention enhancement is generated , and the feature expression of the dendrite core region and the sulfur-oxygen aggregation region is strengthened. The association weights between the search frame and memory frame features are calculated, and the memory frame feature maps are weighted and fused based on the association weights to generate a memory feature map related to the search frame that contains spatiotemporal evolution information. ; Feature map of the search frame after interaction and memory feature map Channel attention processing and spatial attention processing are performed sequentially, and then the images are stitched together to form the final image feature vector. Channel attention is used to select key feature channels related to sedimentary evolution, while spatial attention is used to enhance the target region and suppress background interference.

5. The system of claim 4, wherein, The calculation of the association weight between the features of the search frame and the memory frame, and the weighted fusion of the memory frame feature maps based on the association weight, specifically includes: Calculate the similarity of each pixel in the search frame feature map and each pixel in the memory frame feature map to form a correlation weight matrix ; ； in, Search frame number The pixel and the memory frame The similarity weight of each pixel is used to quantify the correlation strength between the features of two pixels. Search frame number The feature vector of each pixel contains features related to dendrite morphology and sulfur and oxygen element distribution at that location. : Memory Frame The feature vector of each pixel records the deposition-related features at that location at a historical moment; The transpose operation is used to convert a feature vector into a dimension suitable for dot product operations. Vector dot product operation; Feature channel number, which is the dimension of the feature vector of each pixel, represents the richness of the features; : Memory frame pixel index, value range and Consistent, used to iterate through all memory frame pixels to calculate the normalized denominator; : Normalized summation term, ensuring that all memory frame pixels correspond to the search frame's first pixel. The sum of the weights of each pixel is 1; The memory frame feature map is subjected to a weighted average operation by using the correlation weight matrix to obtain a memory feature map related to the search frame and containing space-time evolution information : .

6. The system according to claim 2, characterized in that, The comprehensive assessment results output by the conductive deposition risk assessment unit include: The deposition risk level is determined by combining the conductive deposition risk score with traditional criteria that include at least one of the resistance threshold and CDF distribution characteristics. By analyzing the weight ratio of each input feature in the multilayer perceptron model, the main causes of conductive deposition are identified.

7. A method for evaluating electroconductive deposition based on copper sulfur analysis, characterized by, The method, applied to the system as described in any one of claims 1-6, comprises: Conductive deposition testing is performed using a basic detection module to obtain resistance timing data and CDF timing data. The sample was analyzed by the elemental analysis module to obtain the concentration data of copper and sulfur elements, and SEM images of dendrite morphology and sulfur and oxygen element distribution spectrum were collected on the PCB board surface. The concentration correlation data of copper and sulfur elements were normalized, and the SEM image of dendrite morphology was enhanced. Key temporal pattern features are extracted from preprocessed resistance time-series data and cumulative distribution function (CDF) time-series data; dynamic evolution image features are extracted from enhanced current and historical dendrite morphology SEM images; and the concentration correlation data of copper and sulfur elements after normalization, temporal pattern features, and image features are fused and input into a trained machine learning model to obtain the risk assessment results of conductive deposition.

Images