A digital main line-based electronic equipment assembly representation enhancement method, electronic device and readable storage medium

Abstract

The application belongs to the technical field of electronic equipment assembly, and specifically discloses an electronic equipment assembly representation enhancement method based on a digital main line, an electronic device and a readable storage medium, which comprises the following steps: S1, collecting assembly data of key assembly nodes, building an assembly process digital main line, and obtaining various types of samples, wherein each arbitrary sample in the various types of samples is an original sample; S2, constructing a category guide; S3, stripping generalized assembly features from the various types of samples, and introducing different types of assembly information to simulate random changes in production; S4, using a style transfer algorithm to disturb the features of the original sample, and obtaining cross-category disturbance samples; S5, constructing a style diffusion model based on a conditional guide; S6, constructing a U-Net noise prediction network; S7, processing the cross-category disturbance samples in step S4 to obtain enhanced style samples; and S8, mixing the enhanced style samples and the original samples according to a preset ratio to form an amplified training data set.

Description

Technical Field

[0001] This invention belongs to the field of electronic equipment assembly technology, specifically relating to an electronic equipment assembly characterization enhancement method based on digital thread, an electronic device, and a readable storage medium. Background Technology

[0002] With the rapid development of digital manufacturing in the industrial field, data-driven deep learning models have received widespread attention in the field of quality control of complex product assembly. Methods such as generative adversarial networks, variational autoencoders, and diffusion models not only reduce the reliance on human experience in interpretation, but also automatically extract quality-related features from massive process data. They have been gradually applied to assembly quality analysis and fault diagnosis tasks in industries such as aerospace, shipbuilding, and precision electronics.

[0003] While existing data augmentation methods have shown satisfactory performance in sample augmentation in some industrial scenarios, their effectiveness largely depends on the following conditions: 1) a sufficient number of training samples with a balanced distribution across categories, enabling the model to learn a complete distribution of category features; 2) training and testing data following the principle of independent and identically distributed (IOD), meaning there is no significant difference in distribution between them. However, electronic equipment assembly, especially phased array antenna assembly, is a typical "multi-variety, small-batch" production scenario with extremely scarce total data, a very low proportion of anomalous samples, and a long-tailed distribution, making it difficult for existing models to learn sufficient generalized assembly representations. At the same time, random factors such as parameter changes and component replacements in actual production lines cause significant fluctuations in data distribution. Traditional methods struggle to generate augmented samples with both diversity and mechanistic consistency with limited samples, and samples generated using traditional augmentation methods are prone to augmentation boundary distortion, lacking realistic physical feasibility.

[0004] In response, Li et al. integrated a hybrid loss function into a multi-scale residual network to generate high-quality samples; Lin et al. combined feature selection and data augmentation to solve the class imbalance problem in vibration prediction of aero-engine assembly; Tian et al. used a combination of VAE and GAN to achieve directional generation of fault samples; and Ren et al. proposed a knowledge-image synthesis process for industrial defect detection. However, all of the above methods treat each category of samples as an independent generation target, lacking effective utilization of the feature correlation between categories, and it is difficult to guarantee the generalized representation ability of the generated samples when samples are extremely scarce.

[0005] In recent years, as diffusion models have demonstrated superior generation quality and stability compared to GANs in fields such as image generation, Hu et al. proposed the AnomalyDiffusion framework, and Karras et al. optimized its training dynamics, leading scholars to introduce it into industrial data augmentation. However, standard diffusion models still have the following significant drawbacks in assembly scenarios: the reverse denoising process lacks effective conditional constraints, and the augmentation boundary is prone to distortion under scarce samples; moreover, it fails to utilize the feature association between categories, making it difficult for generated samples to cover the complete feature space of real working conditions.

[0006] Furthermore, digital thread technology, as a key means to achieve data connectivity throughout the entire product lifecycle, has been initially explored in the field of aerospace assembly in recent years. Zhang et al. proposed a digital twin framework modeling method for aircraft assembly systems based on digital thread, realizing the structured organization of assembly process data. Jia et al. established a general mathematical model framework for digital twins oriented towards assembly accuracy, systematically expressing assembly process data according to process logic. However, existing research on digital thread mainly focuses on data connectivity and visualization, and has not yet combined it with deep generative models to solve the problem of sample amplification under the condition of scarce assembly data.

[0007] In summary, existing data augmentation technologies have the following three main shortcomings when facing electronic equipment assembly, especially phased array antenna assembly scenarios: (1) lack of effective organization and structured expression of multi-source heterogeneous data in the assembly process, resulting in the data augmentation process being deviated from the physical semantic constraints of the assembly process; (2) serious homogenization of generated samples, lack of full utilization of feature correlation information between categories, and difficulty in solving the problem of effective amplification under extreme long-tail distribution; (3) standard diffusion models lack effective category condition guidance in scenarios with few samples, making it difficult to guarantee the generation quality and easily causing enhancement boundary distortion. Summary of the Invention

[0008] To address the aforementioned technical problems, this invention provides a method for enhancing the characterization of electronic equipment assembly based on a digital thread, an electronic device, and a readable storage medium.

[0009] To achieve the above objectives, the present invention is implemented through the following technical solution:

[0010] The first aspect of this invention provides a method for enhancing the characterization of electronic equipment assembly based on a digital thread, comprising the following steps:

[0011] S1. Collect assembly data from key assembly nodes in the electronic equipment assembly line, especially in the phased array antenna assembly line, build a digital master line for the assembly process, obtain samples of each category, and each arbitrary sample in each category is the original sample.

[0012] The assembly data includes equipment control parameters, assembly process response data, and quality inspection results. The steps for building the digital master line of the assembly process include: defining a digital master line data set for the electronic equipment assembly process; taking key assembly nodes as basic organizational units, dividing them according to data function performance around the process flow, and reintegrating the discrete data in the original assembly process into a data link that runs through the assembly process, thus completing the construction of the digital master line of the assembly process and obtaining samples of various categories. Each arbitrary sample in each category is an original sample. The original sample refers to the real assembly data sample directly obtained from the digital master line without any enhancement processing, that is, the real assembly process data (force-displacement curves, torque values, displacement values, etc.) collected from key nodes of the assembly production line.

[0013] S2, Building Category Guidelines It is used to augment data for samples of various categories.

[0014] S3. Extract generalized assembly features from samples of each category to reduce the impact of random factors on sample fluctuations. Introduce assembly information of different categories to simulate random changes in production and avoid overall learning bias in the network due to the scarcity of samples of a certain category.

[0015] S4. A style transfer algorithm is used to perturb the original sample features by giving mixed sample feature information to obtain cross-class perturbation samples, which are used to simulate the cross-influence relationship of various random factors on data trend changes in actual assembly scenarios.

[0016] S5. Construct a condition-guided style diffusion model to obtain data samples with diverse assembly representations.

[0017] S6. Construct a U-Net noise prediction network to recover complete noise as data representation with original class assembly features.

[0018] S7. Process the cross-class perturbation samples in step S4 to obtain enhanced style samples.

[0019] S8. The enhanced style samples and the original samples are mixed according to a preset ratio to form an amplified training dataset, which is used to realize intelligent discrimination and prediction of electronic equipment assembly quality; the amplified training dataset is used for training downstream assembly quality classification or assembly accuracy regression prediction models; the preset ratio is adaptively adjusted according to the category distribution of downstream verification tasks.

[0020] According to the above-mentioned electronic equipment assembly characterization enhancement method based on digital thread, preferably, in step S2, the construction category guidance... The steps for data augmentation of samples in each category include:

[0021] Specifically, the Fuzzy C-means (FCM) algorithm is used to cluster the sample clusters. This involves iteratively optimizing the objective function to minimize the objective function, thereby obtaining the optimal category feature centers and membership matrix. Let the sample set of the assembly data be... The feature center clusters of each category are Each data point exists among multiple feature centers with different membership degrees (between 0 and 1), reflecting the feature ambiguity of the data and the overlap of features between categories; the center of each sample cluster is determined by both eigenvalues ​​and membership degrees, and is based on a weighted average of the feature space, rather than the geometric center of the data, expressed as:

[0022] (1)

[0023] in, This represents the objective function of the fuzzy C-means clustering algorithm. This represents the sum of squared weighted Euclidean distances from all samples to each feature center. Let j represent the j-th sample. It is the feature center of the i-th category. It is the membership degree of the j-th sample to the i-th feature center. and These represent the number of feature centers and the number of samples, respectively. It is a fuzzification function, and its value range is... ;

[0024] Therefore, the sample With feature center The Euclidean distance between them is expressed as:

[0025] (2)

[0026] in, It is a sample or feature center The dimension;

[0027] The objective function is solved using the Lagrange multiplier method to obtain the membership degrees. With feature center The iterative process; when the sample Distance from feature center The closer the sample is to the feature center, the higher its membership degree to that feature center, expressed as:

[0028] (3)

[0029] (4)

[0030] Among them, feature center It is a weighted average of the data points, with the weights being the m-th power of the membership degree. Indicates the first Feature centers of each category;

[0031] This allows us to obtain the feature center clusters for each category of assembly data. , feature center cluster In addition to serving as a guiding indicator for subsequent data augmentation, it is also used for the identification of unlabeled samples; the unlabeled samples refer to newly collected assembly process data samples that have not yet been manually labeled or quality judged during the actual assembly process; in this invention, in addition to serving as a condition guide for the diffusion model, the feature center cluster is also used for the automatic classification of such unlabeled samples, that is, by calculating the cosine similarity between the unlabeled sample and each type of feature center cluster, it is classified into the category with the highest similarity.

[0032] By calculating the unlabeled samples in the assembly data and the feature center clusters of each class... Cosine similarity between Unlabeled samples are then categorized and assigned to the feature center clusters with the highest similarity, as follows:

[0033] (5)

[0034] in, It is a sample or feature center Dimensions For the feature vector of the unlabeled sample, For a certain category of feature center vectors, Unlabeled samples In the Values ​​in each dimension For feature center In the The values ​​can be taken in each dimension.

[0035] According to the above-mentioned electronic equipment assembly characterization enhancement method based on digital thread, preferably, the specific steps of step S3 include:

[0036] First, the samples of each category are summarized, and then the mean and variance of each feature center sample cluster are calculated in turn according to the assembly state category.

[0037] (6)

[0038] in, It is the average of the mean values ​​calculated from the k-th feature center. This represents the number of feature centers, where N represents the total number of samples contained in the current feature center.

[0039] This leads to the characteristic central variance of the assembly response data for each category;

[0040] (7)

[0041] in, It is the standard deviation calculated from the k-th feature center. Indicates the first The sample at the th The values ​​can be taken in each feature dimension.

[0042] According to the above-mentioned electronic equipment assembly characterization enhancement method based on digital thread, preferably, the specific steps of step S4 include:

[0043] Introducing disturbance immunity coefficient The mean of the k-th group of samples with mixed characteristics With variance Positioned as:

[0044] (8)

[0045] (9)

[0046] in, This is the characteristic robustness coefficient; the smaller the value, the better the performance of the original sample. The stronger the disturbance;

[0047] Finally, cross-class perturbation samples were obtained. , represented as:

[0048] (10)

[0049] in, It is a scaling factor, which is essentially the ratio of the target standard deviation to the original standard deviation. Using cross-class perturbation samples Replace the original sample Input the subsequent style diffusion model to generate style samples; This represents the original sample.

[0050] According to the above-mentioned method for enhancing the characterization of electronic equipment assembly based on digital mainline, preferably, in step S5, the training process of the style diffusion model includes a forward noise addition process and a condition-guided reverse noise reduction process.

[0051] The forward noise addition process is used to gradually add random noise to the original samples input to the network until they become complete noise, thereby losing their original assembly representation; the noisy complete noise... With the noise added at step t Represented as:

[0052] (11)

[0053] (12)

[0054] in, This represents the noise variance at step t, which typically increases with t. Standard Gaussian noise was added; Represents the original sample after diffusion at step t. The "retention rate";

[0055] Based on the known initial data With complete noise Therefore, it can be concluded that noise exists. Able to directly realize the original sample To complete noise The transformation of , its noise distribution and functional expression are as follows:

[0056] (13)

[0057] (14)

[0058] in, Original sample The initial noise representation, Denotes the marginal variance, and , This represents the identity matrix, indicating that the noise in each dimension is independent and has the same variance.

[0059] According to the above-mentioned electronic equipment assembly characterization enhancement method based on digital thread, preferably, the inverse denoising process is a complete noise reduction process. The process of progressive denoising, which involves t steps of denoising to generate augmented samples, is represented by the noise in this stage as follows:

[0060] (15)

[0061] After reverse denoising, the original sample at step t-1. Represented as:

[0062] (16)

[0063] in, For the predicted noise, Gaussian noise was added.

[0064] According to the above-mentioned electronic equipment assembly characterization enhancement method based on digital thread, preferably, in step S6, the step of constructing the U-Net noise prediction network includes: introducing feature centers as category guides during the reverse denoising process. The U-Net noise prediction network, i.e., U-shaped convolutional network (U-Net), is used as a conditional variable input to the diffusion process.

[0065] The category guide This is used to constrain the U-Net noise prediction network's prediction process for specific noise, enabling it to generate noise estimates associated with the target class, and to gradually remove that specific noise to make the complete noise... The data representation is gradually restored to have the original category assembly characteristics; the conditional constraints of the data distribution in this reverse process are expressed as follows:

[0066] (17)

[0067] (18)

[0068] in, This represents the predicted distribution of the style diffusion model;

[0069] During the reverse engineering process, the category guidance introduced by the feature center cluster is utilized. The prediction noise is evaluated, and the estimation loss of the prediction noise is minimized. During the training of the diffusion model, the loss is estimated based on the mathematical expectation of the difference between the predicted noise and the actual noise. The expected expression is:

[0070] (19)

[0071] in, It is the time step of the diffusion process, ranging from 0 to T; It is standard Gaussian noise; It is the predicted noise; This indicates category guidance information.

[0072] According to the above-mentioned electronic equipment assembly characterization enhancement method based on digital mainline, preferably, in step S7, the step of obtaining the enhanced style sample includes:

[0073] Cross-class perturbation samples While performing reverse denoising, the noise removal process is also being improved. and the corresponding estimated loss Synchronous application A complete noise denoising process is performed to generate augmented data based on style samples from other categories. ;

[0074] (20)

[0075] in, Indicates the center sample of features of other categories, For input samples The fully noisy representation, At step t, the state is in the intermediate noise stage; when t=T, it is completely noisy. A fully noisy representation of the feature center samples of other categories;

[0076] Introducing enhanced boundary coefficients Enhanced samples generated after inverse denoising of target category samples using a diffusion model. Other category feature center samples Weighted representation is performed to obtain enhanced style samples. ;

[0077] ;(twenty one)

[0078] in, Indicates the enhancement boundary coefficient. Indicates the number of feature centers.

[0079] A second aspect of the present invention provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement any step of the electronic equipment assembly characterization enhancement method based on digital thread as described in the first aspect.

[0080] A third aspect of the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a computer processor, implements any step in the electronic equipment assembly characterization enhancement method based on a digital thread as described in the first aspect.

[0081] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0082] (1) This invention uses digital thread technology to structure and organize multi-source heterogeneous data in the assembly process. On this basis, it constructs the feature centers of each category of samples by fuzzy C-means clustering as category guidance for the generation process. It introduces a cross-category feature perturbation mechanism to simulate the influence of random factors in real scenarios. It also generates enhanced samples with both feature fidelity and diversity through a designed condition-guided style diffusion model. This solves the problem that the quality prediction model is difficult to train effectively in the assembly process of electronic equipment, especially in the assembly process of phased array antennas, due to the scarcity of data and the serious imbalance of category distribution.

[0083] (2) In view of the problem of ambiguous sample category boundaries in the scenario with few labels, the present invention uses fuzzy C-means soft clustering to construct feature center clusters of samples of each category, and embeds them as condition variables into the reverse denoising process of style diffusion model, so that the generation has clear category orientation, while undertaking the automatic classification of unlabeled samples, reducing the dependence on labeled data.

[0084] (3) This invention addresses the problem that the extreme scarcity of abnormal category samples leads to traditional methods being able to perform only minor local deformations. It utilizes statistical information of different categories of samples to perturb the mean and variance of the scarce category, so that the generated samples retain the core assembly trend while incorporating cross-category generalized feature fluctuations, effectively simulating the impact of random factors such as parameter changes and parts replacements in actual production lines.

[0085] (4) In view of the problem that existing generative models are unstable in training and difficult to guarantee generation quality in scenarios with few samples, this invention constructs a style diffusion model based on U-Net, and inputs feature center conditional variables and cross-class perturbation samples together. Through end-to-end training of forward noise addition and conditional reverse noise removal, style samples with diverse assembly representations are generated.

[0086] (5) This invention constructs a category feature center representation and guidance mechanism for assembly data augmentation. In view of the problem that traditional data augmentation methods lack effective category constraints and generate sample category boundaries in the case of scarce samples, this invention uses the fuzzy C-means soft clustering method to construct feature center clusters of samples of each category and embeds them as conditional variables into the reverse denoising process of the style diffusion model. This mechanism enables the generation process to have clear category guidance, solves the problem of class distinction between samples with blurred boundaries such as normal assembly and assembly tilt, and avoids the phenomenon of sample category confusion in unconditional generation from the root. At the same time, the feature center also undertakes the function of automatic classification of unlabeled samples, reducing the dependence on labeled data.

[0087] (6) The sample diversity amplification strategy of cross-category perturbation in this invention: In view of the problem that the extreme scarcity of abnormal category samples such as tilt and fault in assembly data leads to the traditional enhancement method being able to only perform minor local deformation and unable to cover the real working condition distribution, this invention uses the statistical information of different category samples to perturb the mean and variance of scarce category samples, so that the generated samples can retain the core assembly trend while incorporating cross-category generalized feature fluctuations; the strategy effectively simulates the data changes caused by random factors such as parameter changes and parts replacement in the actual production line, and solves the technical problem of homogenization and boundary distortion of enhanced samples under extreme long-tail distribution.

[0088] (7) The present invention is based on a condition-guided style diffusion model. In view of the problems that existing generative adversarial networks and standard diffusion models are unstable in training and difficult to guarantee generation quality in scenarios with few samples, a style diffusion model is constructed based on U-Net. The feature center conditional variable and cross-class perturbation samples are input into the network in a coordinated manner. Through end-to-end training of forward noise addition and conditional reverse noise removal, style samples with diverse assembly representations are generated.

[0089] (8) This invention first addresses the problem of class imbalance and insufficient representation of assembly deviation in key assembly nodes, leading to the homogenization of generated representations, and proposes a style diffusion-based data augmentation strategy. This strategy introduces multi-domain sample information to perturb the node data across categories during the training phase, reducing the guidance drift of the network's overall learning caused by random factors in scarce class samples. Secondly, it constructs feature centers for samples of different categories to measure the sufficiency of feature expression in generated samples. Finally, it uses style transfer to extract generalized inherent features from samples of different state categories and enhances the cross-domain representation of generated samples. This invention can significantly reduce the randomness of generated samples and improve the richness of representation in scenarios with few labels and class imbalance. Finally, it is verified on two key assembly nodes, the T / R component and the conduction network. By comparing with traditional augmentation methods such as generative adversarial networks and variational autoencoders, the results show that this invention significantly improves the recognition accuracy of scarce categories in downstream assembly quality classification tasks and can effectively suppress the phenomenon of augmentation boundary distortion. Attached Figure Description

[0090] Figure 1 This is an overall flowchart of the present invention in Example 1;

[0091] Figure 2 This is a schematic diagram of the digital master line for the assembly process in Example 1;

[0092] Figure 3 This is a schematic diagram of the t-SNE visualization results of the feature centers of the T / R component in Example 1;

[0093] Figure 4 This is a schematic diagram of the t-SNE visualization results of the conduction network feature centers in Example 1;

[0094] Figure 5 This is a schematic diagram showing the performance comparison of each comparison method in Example 1 in the verification task;

[0095] Figure 6 This is a schematic diagram of the verification results of style samples under each task in Example 1;

[0096] Figure 7 This is a schematic diagram of the visualization results of t-SNE features in Tasks 1-3 of Example 1;

[0097] Figure 8 This is a schematic diagram of the visualization results of t-SNE features in Tasks 4-6 of Example 1. Detailed Implementation

[0098] The present invention will be further illustrated by specific embodiments below, but this does not limit the scope of the invention.

[0099] Example 1

[0100] A method for enhancing the characterization of electronic equipment assembly based on digital thread, the flowchart of which is as follows: Figure 1 As shown, it includes the following steps:

[0101] S1. Collect assembly data from key assembly nodes in the electronic equipment assembly line, especially in the phased array antenna assembly line, build a digital master line for the assembly process, obtain samples of each category, and each arbitrary sample in each category is the original sample.

[0102] The assembly data includes equipment control parameters, assembly process response data, and quality inspection results, etc.; the steps for building the digital master line of the assembly process include: defining a digital master line data set for the electronic equipment assembly process; taking key assembly nodes as basic organizational units, dividing them according to data function performance around the process flow, and re-integrating the discrete data in the original assembly process into a data link that runs through the assembly process, completing the construction of the digital master line of the assembly process, obtaining samples of each category, and each arbitrary sample in each category is an original sample, such as... Figure 2 As shown.

[0103] S2, Building Category Guidelines It is used to augment data for samples of various categories.

[0104] The Fuzzy C-means (FCM) algorithm is used to cluster the sample clusters. This involves iteratively optimizing the objective function to minimize the objective function, thereby obtaining the optimal category feature centers and membership matrix. Let the sample set of the assembly data be... The feature center clusters of each category are Each data point exists among multiple feature centers with different membership degrees (between 0 and 1), reflecting the feature ambiguity of the data and the overlap of features between categories; the center of each sample cluster is determined by both eigenvalues ​​and membership degrees, and is based on a weighted average of the feature space, rather than the geometric center of the data, expressed as:

[0105] (1)

[0106] in, This represents the objective function of the fuzzy C-means clustering algorithm. This represents the sum of squared weighted Euclidean distances from all samples to each feature center. Let j represent the j-th sample. It is the feature center of the i-th category. It is the membership degree of the j-th sample to the i-th feature center. and These represent the number of feature centers and the number of samples, respectively. It is a fuzzification function, and its value range is... ;

[0107] Therefore, the sample With feature center The Euclidean distance between them is expressed as:

[0108] (2)

[0109] in, It is a sample or feature center The dimension;

[0110] The objective function is solved using the Lagrange multiplier method to obtain the membership degrees. With feature center The iterative process; when the sample Distance from feature center The closer the sample is to the feature center, the higher its membership degree to that feature center, expressed as:

[0111] (3)

[0112] (4)

[0113] Among them, feature center It is a weighted average of the data points, with the weights being the m-th power of the membership degree. Indicates the first Feature centers of each category;

[0114] This allows us to obtain the feature center clusters for each category of assembly data. , feature center cluster In addition to serving as a guiding beacon for subsequent data augmentation, it has also been used for the identification of unlabeled samples;

[0115] By calculating the unlabeled samples in the assembly data and the feature center clusters of each class... Cosine similarity between Unlabeled samples are then grouped into clusters of feature centers with the highest similarity, as follows:

[0116] (5)

[0117] in, It is a sample or feature center Feature dimensions, For the feature vector of the unlabeled sample, For a certain category of feature center vectors, Unlabeled samples In the Values ​​in each dimension For feature center In the The values ​​can be taken in each dimension.

[0118] S3. Extract generalized assembly features from samples of each category to reduce the impact of random factors on sample fluctuations. Introduce assembly information of different categories to simulate random changes in production and avoid overall learning bias in the network due to the scarcity of samples of a certain category.

[0119] First, the samples of each category are summarized, and then the mean and variance of each feature center sample cluster are calculated in turn according to the assembly state category.

[0120] (6)

[0121] in, It is the average of the mean values ​​calculated from the k-th feature center. This represents the number of feature centers, where N represents the total number of samples contained in the current feature center.

[0122] This leads to the characteristic central variance of the assembly response data for each category;

[0123] (7)

[0124] in, It is the standard deviation calculated from the k-th feature center. Indicates the first The sample at the th The values ​​can be taken in each feature dimension.

[0125] S4. A style transfer algorithm is used to perturb the original sample features by giving mixed sample feature information to obtain cross-class perturbation samples, which are used to simulate the cross-influence relationship of various random factors on data trend changes in actual assembly scenarios.

[0126] Introducing disturbance immunity coefficient The mean of the k-th group of samples with mixed characteristics With variance Positioned as:

[0127] (8)

[0128] (9)

[0129] in, This is the characteristic robustness coefficient; the smaller the value, the better the performance of the original sample. The stronger the disturbance;

[0130] Finally, cross-class perturbation samples were obtained. , represented as:

[0131] (10)

[0132] in, It is a scaling factor, which is essentially the ratio of the target standard deviation to the original standard deviation. Using cross-class perturbation samples Replace the original sample Input the subsequent style diffusion model to generate style samples; This represents the original sample.

[0133] S5. Construct a condition-guided style diffusion model to obtain data samples with diverse assembly representations.

[0134] The training process of the style diffusion model includes a forward noise addition process and a conditionally guided reverse noise reduction process.

[0135] The forward noise addition process is used to gradually add random noise to the original samples input to the network until they become complete noise, thereby losing their original assembly representation; the noisy complete noise... With the noise added at step t Represented as:

[0136] (11)

[0137] (12)

[0138] in, This represents the noise variance at step t, which typically increases with t. Standard Gaussian noise was added; Represents the original sample after diffusion at step t. The "retention rate";

[0139] Based on the known initial data With complete noise Therefore, it can be concluded that noise exists. Able to directly realize the original sample To complete noise The transformation of , its noise distribution and functional expression are as follows:

[0140] (13)

[0141] (14)

[0142] in, Original sample The initial noise representation, Denotes the marginal variance, and , It is an identity matrix, representing that the noise in each dimension is independent and has the same variance.

[0143] The reverse denoising process is the complete noise reduction process. The process of progressive denoising, which involves t steps of denoising to generate augmented samples, is represented by the noise in this stage as follows:

[0144] (15)

[0145] After reverse denoising, the original sample at step t-1. Represented as:

[0146] (16)

[0147] in, For the predicted noise, Gaussian noise was added.

[0148] S6. Construct a U-Net noise prediction network to recover complete noise as data representation with original class assembly features.

[0149] The steps to construct the U-Net noise prediction network include: introducing feature centers as category guides during the inverse denoising process. The U-Net noise prediction network, also known as the U-shaped convolutional network (U-Net), is used as a conditional variable input to the diffusion process.

[0150] The category guide This is used to constrain the U-Net noise prediction network's prediction process for specific noise, enabling it to generate noise estimates associated with the target class, and to achieve complete noise reduction by progressively removing that specific noise. The data representation is gradually restored to have the original category assembly characteristics; the conditional constraints of the data distribution in this reverse process are expressed as follows:

[0151] (17)

[0152] (18)

[0153] in, This represents the predicted distribution of the style diffusion model;

[0154] During the reverse engineering process, the category guidance introduced by the feature center cluster is utilized. The prediction noise is evaluated, and the estimation loss of the prediction noise is minimized. Estimate the loss The expected expression is:

[0155] (19)

[0156] in, It is the time step of the diffusion process, ranging from 0 to T; It is standard Gaussian noise; It is the predicted noise; This indicates category guidance information.

[0157] S7. Process the cross-class perturbation samples in step S4 to obtain enhanced style samples.

[0158] The steps to obtain enhanced style samples include: perturbing samples across categories. While performing reverse denoising, the noise removal process is also being improved. and the corresponding estimated loss Synchronous application A complete noise denoising process is performed to generate augmented data based on style samples from other categories. ;

[0159] (20)

[0160] in, Indicates the center sample of features of other categories, For input samples The fully noisy representation, A fully noisy representation of the feature center samples of other categories;

[0161] Introducing enhanced boundary coefficients Enhanced samples generated after inverse denoising of target category samples using a diffusion model. Other category feature center samples Weighted representation is performed to obtain enhanced style samples. ;

[0162] ;(twenty one)

[0163] in, Indicates the enhancement boundary coefficient. Indicates the number of feature centers.

[0164] S8. The enhanced style samples and the original samples are mixed according to a preset ratio to form an amplified training dataset, which is used to realize intelligent discrimination and prediction of electronic equipment assembly quality; the amplified training dataset is used for training downstream assembly quality classification or assembly accuracy regression prediction models; the preset ratio is adaptively adjusted according to the category distribution of downstream verification tasks.

[0165] Verification Example: The invention was verified using datasets from two key assembly nodes: the T / R component and the conductive network. Specifically, the T / R component dataset was collected by pressure sensors on the assembly press. The experiment simulated the press assembly process under three process parameter settings: assembly stroke of 64.2mm, 64.4mm, and 64.7mm. The samples covered three assembly states: normal assembly (NC), assembly tilt (AT), and assembly failure (AF). The conductive network was set to a process stroke of 56.4mm, and the data types also covered the three assembly states: NC, AT, and AF.

[0166] Table 1: Task settings for enhancing style samples of key assembly nodes;

[0167]

[0168] Table 2: Downstream validation task settings for data augmentation experiments;

[0169]

[0170] For the three different data categories of samples under the two assembly components of T / R component and conduction network, a total of three sets of data augmentation tasks (task A, B, and C) were constructed. The number of samples of each category under each task is shown in Table 1.

[0171] To systematically evaluate the quality of the style samples generated by the proposed method and verify whether they fully possess the generalized fault characteristics in the original samples, a downstream verification task for data augmentation experiments is added, as shown in Table 2. Six downstream verification tasks are set up, of which tasks 1, 2 and 5 are trained by mixing the original samples and style samples, with the ratio of style samples to original samples being 1:5, 1:10 and 1:10, respectively; tasks 3, 4 and 6 consist entirely of style samples as the training set and are directly applied to verify the assembly results of real samples.

[0172] Table 3: Details of hyperparameter settings for data augmentation and validation experiments;

[0173]

[0174] The training configurations and hyperparameter settings involved in the style sample enhancement experiment and downstream validation experiment are shown in Table 3. To evaluate the superiority of this embodiment, the following methods are used for comparison:

[0175] (1) Baseline method: Based on the style diffusion model set in this embodiment, no other strategies are introduced;

[0176] (2) GAN: Based on the traditional adversarial game mechanism, without optimizing the network structure and training method;

[0177] (3) Mixup-Transformer: Based on the Transformer model, virtual training samples are constructed using a dynamic mixing (Mixup) method to enhance the training process;

[0178] (4) EnhDiffusion: Based on the Diff-Model, attention-enhanced residual blocks replace the original residual structure, and the original fixed hyperparameters are dynamically adapted and optimized to adapt to multi-domain features.

[0179] (5) WM-CVAE: Based on Varia-AE, the data samples are enhanced and weight calculation is redefined by kernel mean matching to reduce the negative impact between samples;

[0180] (6) GeCNN: It enhances data samples based on CNN, reduces the training sample requirements by focusing on various subsequences, and integrates multiple pooling operations to mitigate the loss of sample features.

[0181] (7) Ablation comparison method: Ablation experiments are usually used to explore the importance of key modules in the proposed strategy; therefore, this verification example designs relevant ablation comparison methods, as shown in Table 4; among them, sample perturbation means that different types of samples are used to perturb the mean and variance of the original samples, condition guidance means that the feature center is selected as the target center of the reverse denoising process to guide the sample generation, and style diffusion means that the training process contains information features of different types of samples.

[0182] Table 4: Ablation Experiment Setup Information;

[0183]

[0184] The experimental verification phase is mainly divided into three core parts: The first part is to address the problem of insufficient labeled samples in the training data by using the FCM clustering algorithm to perform cluster analysis on the samples and formulate feature center curves for each category; The second part is to address the problem of "long-tail distribution" of samples in actual engineering scenarios by using the CGS-DiffNet network to generate style samples; The third part is to verify that the augmented data has sufficient generalized fault characterization capabilities by performing related downstream assembly quality tasks.

[0185] The first step: Construction and cluster analysis of feature centers in assembly data; soft clustering of samples from the conduction network and T / R component datasets, and solving for feature centers of samples within the same category to obtain feature center curves between different categories; low-dimensional visualization analysis of feature centers using t-SNE technology, such as... Figure 3-4 As shown, the results indicate that there are significant feature differences among the sample categories of the T / R component. After dimensionality reduction and projection, the boundaries of each category are clear, and the assembly failure samples show obvious sample clustering, indicating that the current node data has sufficient fault characterization capability.

[0186] Meanwhile, the feature representation of the assembly tilt sample AT is between the normal assembly NC sample cluster and the assembly failure AF sample cluster, proving that it is not a typical "complete assembly failure", but a transitional sample with assembly deviation but no direct impact on subsequent assembly. In addition, through the dimensionality reduction visualization of the feature center of the conductive network, it can be clearly seen that there is a cross-over between the features of the NC sample and the AT sample, indicating that the AT sample does not form a completely independent generalized feature distribution and the sample cluster boundary is blurred, which significantly increases the difficulty of quality analysis.

[0187] The second step: style sample generation and feature distance verification; Table 5 shows the average distance of the style samples generated from the three sets of data augmentation experiments from the feature centers of each category. It can be seen that the generated NC, AT and AF samples show clear differences in category features in the feature space. The average distance of the assembly tilt AT sample from the feature center in augmentation experiment A is 4.3894, while the distances from the feature centers of NC and AF categories are 6.3077 and 14.3213, respectively. This indicates that there are significant differences among the augmented samples of each category, proving that the style samples augmented by the method of this embodiment (CGS-DiffNet) retain the generalized representation of each category in the real samples.

[0188] Furthermore, the consistency performance of augmentation experiments A and B verifies the stability of the augmented samples generated by the proposed method. In augmentation experiment C, the mean distance of style samples to the center of each category decreases significantly, indicating that the style samples generated in this task are clustering towards the feature center. It can be concluded that the style samples after data augmentation not only have the same generalized category representation ability as the real samples, but also exhibit a diversified feature distribution trend. This effectively alleviates the problem of homogenization of augmented samples caused by the long-tailed distribution of the training dataset in the data augmentation model.

[0189] Table 5: Comparison of mean distances from style samples to each feature center;

[0190]

[0191] The third step: evaluation of downstream validation tasks; Table 6 shows the mean and standard deviation of classification accuracy obtained by different comparison methods after 10 experiments in 6 groups of downstream validation tasks. By comparison, it can be seen that the enhanced samples generated by the method of this embodiment (CGS-DiffNet) achieved the best performance in downstream Tasks 1, 2, 3 and 4, and also showed satisfactory results in Tasks 4 and 6, with downstream validation accuracies of 92.72% and 81.08%, respectively; In Task 1, the proposed method achieved a sample classification accuracy of up to 98.50%, which is significantly better than other popular methods. The next best methods in this task, GAN and M3 ablation, only achieved 93.33% and 95.89%, respectively.

[0192] Table 6: Comparison of results (%) of different methods in downstream validation tasks.

[0193]

[0194] Compared to traditional data augmentation model GAN ​​samples, style samples improved the overall performance of Tasks 1, 2, and 5 by 5.17%, 4.92%, and 5.83%, respectively, and showed more stable performance than GANs, with a diagnostic standard deviation of only 1.52% under the best task. Task 3, as a task without real sample training, showed an average accuracy improvement of 3-5% compared to popular methods, indicating that the style differences generated by the proposed method (CGS-DiffNet) have a certain robustness and can more fully simulate sample variation in real assembly scenarios. For Task 4, which involves multiple working conditions, the proposed method's validation results were slightly lower than WM-CVAE by 0.23%, which may be due to performance fluctuations caused by differences in data distribution or hyperparameters. Notably, the sample classification accuracy of the proposed method in Task 6 was only 81.08%, significantly lower than WM-CVAE's 90.42%. This indicates that style diffusion had a negative effect on this task, possibly because the real sample space was relatively concentrated, causing the style diffusion process to destroy the key generalized feature representations.

[0195] To further verify the role of each functional module in the method (CGS-DiffNet) of this embodiment, an ablation experiment comparison group consisting of M1-M3 was used for analysis and discussion. In Tasks 1 and 2, the proposed method achieved a diagnostic accuracy of over 98.5%, which is far higher than that of other ablation experiment methods. As can be seen from Task 2, M3 outperformed M1 and M2 by 17.17% and 0.84%, respectively. After introducing style diffusion, the performance improved by 5.97%, reaching 99.14%, indicating that the three functional modules have a significant synergistic gain effect in this task. From the results of Task 3, it can be concluded that when the network model M2, which is not guided by feature centers, is trained entirely by augmented samples, the generated samples cannot fully represent the assembly quality information in the real data, resulting in a 19.16% decrease in representation performance. Although the further performance improvement of the method of this invention is limited to only 0.89%, it helps to increase the diversity of augmented samples.

[0196] Comprehensive analysis reveals that the conditional guidance module constructed using category feature centers is the most critical performance driver, resulting in a performance improvement of over 9.78% across five downstream validation tasks. Sample perturbation demonstrates a significant sample advantage in some generalization tasks (such as Task 2, Task 3, and Task 5). The proposed style diffusion module generally exhibits feature enhancement in downstream validation tasks, but shows a significant negative gain in Task 6. This indicates that when facing assembly components that do not change much with process parameters or operating conditions during assembly, the generated samples under style amplification often contain invalid and chaotic assembly representations. This causes diverse style samples to mislead the feature learning direction of the downstream validation network, requiring further verification of the performance of diverse style samples in assembly quality analysis tasks.

[0197] The performance of 10 experiments conducted on 6 tasks using 2 basic methods, 4 popular methods, and the method of this embodiment was analyzed and compared. Figure 5 As shown, the method in this embodiment (CGS-DiffNet) exhibits the best and most stable overall performance across all six tasks. Its bar mean is mostly ranked first, and the distribution of discrete points is more convergent, indicating that the generated samples outperform other popular and basic methods in terms of feature representation and distribution consistency, demonstrating superior cross-task generalization performance. Compared with GeCNN and WM-CVAE, the method in this embodiment (CGS-DiffNet) often exhibits a smaller standard deviation on most tasks, proving that the generated style samples have better feature representation capabilities, thereby significantly reducing network performance fluctuations. Furthermore, GAN and Baseline perform poorly on different validation tasks and exhibit significant validation fluctuations, showing high sensitivity to changes in data distribution. This indicates that the two basic methods are not suitable for the actual assembly tasks of this product.

[0198] like Figure 6 As shown, the distribution of diagnostic accuracy of the proposed method in downstream tasks is illustrated. It is evident that the results for each task exhibit high accuracy and low dispersion. For Tasks 1 and 2, where the training and test domains are in the same or near-domain, the diagnostic results are mainly distributed in the 95%-100% range, with compact scatter points and narrow envelopes, demonstrating strong stability and consistency. With increasing validation difficulty, the accuracy of the proposed method (CGS-DiffNet) drops to 88%-95% when facing a task where the training domain consists entirely of augmented samples classifying real samples. Although the diagnostic results show a significant downward shift, the overall performance remains convergent. As the most challenging task scenarios, the diagnostic results for Tasks 5 and 6 exhibit higher uncertainty compared to other tasks. Task 5 shows three sets of abnormal diagnostic results, but thanks to the overall performance, its average accuracy still reaches 95.83%. In contrast, the proposed method (CGS-DiffNet) shows lower accuracy in Tasks 5 and 6. The diagnostic results in Task 6 were more stable, with a standard deviation of only 2.22%, but the average accuracy was slightly lower than that in Task 5, at 81.08%.

[0199] To characterize the feature separability of the proposed method under different working conditions, high-dimensional representations were extracted from the training samples of six tasks, and t-SNE technology was used to reduce the dimensionality and visualize the representations. The results are as follows: Figure 7 and 8 As shown; Figure 7-8 The results show that in Task 1 and Task 2, the NC, AT, and AF sample families have clear outlines, clusters of samples within the same class are clustered together, and the boundaries between each class are large, presenting an approximately linearly separable structure. Furthermore, Task 3 and Task 5 also show high separability. In particular, in Task 5, the three class sample clusters are arranged in a band-like or strip-like pattern, with balanced intra-class spacing and clear class boundaries, indicating that the sample representation under this task provides the most sufficient support for validating the model's discrimination boundary.

[0200] In comparison, although the intra-class sample spacing in Task 4 and Task 6 is relatively compact, misclassification of samples with blurred class boundaries occurs between classes. Specifically, in Task 4, there is a small amount of overlap between normal samples NC and assembled tilted samples AT, and in Task 6, the boundary between NC and AT categories is relatively blurred, with some samples exhibiting incorrect clustering. However, the main clustering axis of the sample clusters still maintains a separation trend, and the overall separability is not compromised.

[0201] In summary, the six sets of visualization results collectively demonstrate that the style samples generated by the method in this embodiment (CGS-DiffNet) contain sufficient generalized fault representations. This significantly improves the feature recognition and sample classification capabilities of the downstream validation model. When facing tasks with few samples or long-tailed distributions, it helps generate a sufficient number of diverse style samples to expand the training dataset, avoiding the problems of enhanced sample boundary distortion and sample homogenization, which is consistent with the performance in the previous accuracy distribution, confusion matrix, and convergence curve. At the same time, the results corroborate that the generated assembly node style samples have sufficient real sample fault representations and contain deeper assembly quality information, which is of significance for further research and analysis.

[0202] In the optimization objective of the method in this embodiment (CGS-DiffNet), there are two key trade-off parameters: the disturbance resistance coefficient and the perturbation resistance coefficient. and enhanced boundary coefficient These factors directly impact data trends and style transfer processes, respectively. Taking Task 1 and Task 5 as examples, the sensitivity of different combinations of these two sets of hyperparameters to model performance is analyzed. Specifically, the perturbation resistance coefficient is set... The value range is [0.60, 1.00], which is the enhanced boundary coefficient. The value range is [0.60, 1.00], and the sensitivity precision of both sets of hyperparameters is set to 0.02; when More than 0.64 and When the value exceeds 0.70, the accuracy of Task 1 is generally in the 92%-100% range, therefore the optimal parameter range is [missing value]. The value ranges from [0.75, 0.96]. The values ​​[0.75, 0.80] and [0.85, 0.95] are relatively insensitive to the two sets of hyperparameters under this task, and there is a relatively tolerant usable range.

[0203] In contrast, Task 5 is more sensitive to changes in both sets of hyperparameters. Below 0.66, network training falls into a low-performance range, and when When the value exceeds 0.72, the accuracy rate increases rapidly and... The value is between [0.72, 0.92]. When the value is between [0.82, 0.95], continuous performance peaks appear, and the local contour lines fluctuate sharply.

[0204] In summary, The value range is set to [0.85, 0.95]. The value range is set to [0.75, 0.92], and fine-tuned within the above range according to the difficulty of specific tasks and domain differences, in order to enhance the discriminativeness and stability of the model and avoid performance degradation caused by hyperparameter sensitivity.

[0205] In summary, the electronic equipment assembly characterization enhancement method based on digital thread proposed in this invention demonstrates superior classification accuracy and stability compared to existing methods in multiple downstream verification tasks at two key assembly nodes: the T / R assembly and the conduction network. The method in this embodiment accurately replicates the assembly characterization in the original domain samples while mitigating style biases in multi-class samples, thus improving the feature richness of the enhanced samples and effectively avoiding enhancement redundancy and boundary distortion. This verifies the effectiveness and superiority of the method in scenarios where phased array antenna assembly data is scarce and class imbalanced, providing sufficient and high-quality data support for the construction of subsequent assembly quality prediction models.

[0206] Example 2

[0207] An electronic device includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement any step in the method for enhancing the characterization of electronic equipment assembly based on a digital thread as described in Embodiment 1.

[0208] Furthermore, the process of the electronic equipment assembly characterization enhancement method based on a digital thread described in Embodiment 1 can be implemented as a computer software program. For example, this embodiment includes a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the method. In such an embodiment, the computer program can be downloaded and installed from a network, and / or installed from a removable medium. When the computer program is executed by a processor, it performs the functions defined in the method of this application.

[0209] Example 3

[0210] A computer-readable storage medium storing a computer program that, when executed by a processor, implements any step of the method for enhancing the characterization of electronic equipment assembly based on a digital thread, as described in Embodiment 1.

[0211] The computer-readable medium described in this application may be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this application, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media can also be any computer-readable medium other than computer-readable storage media, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0212] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Python and C++, as well as conventional procedural programming languages ​​or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0213] In this embodiment, the computer-readable storage medium can be accelerated using hardware such as a GPU. The parallel computing advantage of the GPU is used to accelerate the processing of any step in the implementation of the electronic equipment assembly characterization enhancement method based on digital mainline as described in Embodiment 1.

[0214] In summary, this invention effectively overcomes the shortcomings of the prior art and has high industrial applicability. The above embodiments are intended to illustrate the substantive content of this invention, but are not intended to limit the scope of protection of this invention. Those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of this invention without departing from the essence and scope of protection of this invention.

[0215] The above embodiments are specific implementations of the present invention, but the implementation of the present invention is not limited to the above embodiments. Any other combination, change, modification, substitution, or simplification that does not exceed the design concept of the present invention shall fall within the protection scope of the present invention.

Claims

1. A method for enhancing the characterization of electronic equipment assembly based on a digital thread, characterized in that, Includes the following steps: S1. Collect assembly data of key assembly nodes in the assembly production line, build a digital main line of the assembly process, obtain samples of each category, and each arbitrary sample in each category is the original sample. S2. Construct category guidelines to augment data for samples in each category; S3. Extract generalized assembly features from samples of each category and introduce assembly information of different categories to simulate random changes in production; S4. Using a style transfer algorithm, the original sample features are perturbed by the given mixed sample feature information to obtain cross-class perturbed samples; S5. Construct a style diffusion model based on conditional guidance; S6. Construct a U-Net noise prediction network to recover complete noise into data representations with original class assembly features; S7. Process the cross-class perturbation samples in step S4 to obtain enhanced style samples; S8. Mix the enhanced style samples with the original samples according to a preset ratio to form the amplified training dataset.

2. The electronic equipment assembly characterization enhancement method based on digital thread according to claim 1, characterized in that, In step S2, the construction category guide The steps for data augmentation of samples in each category include: Specifically, the fuzzy C-means algorithm is used to cluster the sample clusters, that is, the objective function is minimized through iterative optimization, thereby obtaining the optimal category feature centers and membership matrix; let the sample set of the assembly data be... The feature center clusters of each category are Each data point exists among multiple feature centers with different membership degrees; the center of each sample cluster is determined by both the eigenvalue and the membership degree, as follows: ；（1） in, This represents the objective function of the fuzzy C-means clustering algorithm. This represents the sum of squared weighted Euclidean distances from all samples to each feature center. Let j represent the j-th sample. It is the feature center of the i-th category. It is the membership degree of the j-th sample to the i-th feature center. and These represent the number of feature centers and the number of samples, respectively. It is a fuzzification function; Therefore, the sample With feature center The Euclidean distance between them is expressed as: ；（2） in, It is a sample or feature center The dimension; The objective function is solved using the Lagrange multiplier method to obtain the membership degrees. With feature center The iterative process; when the sample Distance from feature center The closer the sample is to the feature center, the higher its membership degree to that feature center, expressed as: ；（3） ；（4） Among them, feature center It is a weighted average of the data points, with the weights being the m-th power of the membership degree. Indicates the first Feature centers of each category; This allows us to obtain the feature center clusters for each category of assembly data. ; By calculating the unlabeled samples in the assembly data and the feature center clusters of each class... Cosine similarity between Unlabeled samples are then categorized and assigned to the feature center clusters with the highest similarity, as follows: ；（5） in, It is a sample or feature center Feature dimensions, The feature vector of the unlabeled sample. For a certain category of feature center vectors, Unlabeled samples In the Values ​​in each dimension For feature center In the The values ​​can be taken in each dimension.

3. The electronic equipment assembly characterization enhancement method based on digital thread according to claim 1, characterized in that, The specific steps of step S3 include: First, the samples of each category are summarized, and then the mean and variance of each feature center sample cluster are calculated in turn according to the assembly state category. ；（6） in, It is the average of the mean values ​​calculated from the k-th feature center. This represents the number of feature centers, where N represents the total number of samples contained in the current feature center. This leads to the characteristic central variance of the assembly response data for each category; ；（7） in, It is the standard deviation calculated from the k-th feature center. Indicates the first The sample at the th The values ​​can be taken in each feature dimension.

4. The electronic equipment assembly characterization enhancement method based on digital thread according to claim 1, characterized in that, The specific steps of step S4 include: Introducing disturbance immunity coefficient The mean of the k-th group of samples with mixed characteristics With variance Positioned as: ；（8） ；（9） in, It is the characteristic disturbance rejection coefficient; Finally, cross-class perturbation samples were obtained. , represented as: ；（10） in, It is the scaling factor; This represents the original sample.

5. The electronic equipment assembly characterization enhancement method based on digital thread according to claim 1, characterized in that, In step S5, the training process of the style diffusion model includes a forward noise addition process and a conditionally guided reverse noise reduction process. The forward noise addition process is used to gradually add random noise to the original samples input to the network until they are transformed into complete noise, thereby losing their original assembly representation; the complete noise With the noise added at step t Represented as: ；（11） ； （12） in, This represents the noise variance at step t; Standard Gaussian noise was added; Represents the original sample after diffusion at step t. The "retention rate"; Based on the known initial data With complete noise Therefore, it can be concluded that noise exists. Able to directly realize the original sample To complete noise The transformation of , its noise distribution and functional expression are as follows: ；（13） ；（14） in, Original sample The initial noise representation, Represents marginal variance. It is an identity matrix, representing that the noise in each dimension is independent and has the same variance.

6. The electronic equipment assembly characterization enhancement method based on digital thread according to claim 5, characterized in that, The reverse denoising process is the complete noise reduction process. The process of progressive denoising, which involves t steps of denoising to generate augmented samples, is represented by the noise in this stage as follows: ； （15） After reverse denoising, the original sample at step t-1. Represented as: ；（16） in, For the predicted noise, Gaussian noise was added.

7. The electronic equipment assembly characterization enhancement method based on digital thread according to claim 1, characterized in that, In step S6, the steps of constructing the U-Net noise prediction network include: introducing feature centers as category guides during the inverse denoising process. And it is used as a conditional variable input into the U-Net noise prediction network for the diffusion process; The category guide This is used to constrain the U-Net noise prediction network's prediction process for specific noise, enabling it to generate noise estimates associated with the target class, and to achieve complete noise reduction by progressively removing that specific noise. The data representation is gradually restored to have the original category assembly characteristics; the conditional constraints of the data distribution in this reverse process are expressed as follows: ；（17） ；（18） in, This represents the predicted distribution of the style diffusion model; During the reverse engineering process, the category guidance introduced by the feature center cluster is utilized. The prediction noise is evaluated, and the estimation loss of the prediction noise is minimized. Estimate the loss The expected expression is: ；（19） in, It is the time step of the diffusion process, ranging from 0 to... ; It is standard Gaussian noise; It is the predicted noise; This indicates category guidance information.

8. The method for enhancing the characterization of electronic equipment assembly based on digital thread according to claim 1, characterized in that, In step S7, the steps for obtaining the enhanced style sample include: Cross-class perturbation samples While performing reverse denoising, the noise removal process is also being improved. and the corresponding estimated loss Synchronous application The complete noise denoising process generates augmented data based on style samples from other categories. ; ；（20） in, Indicates the center sample of features of other categories, For input samples The fully noisy representation, A fully noisy representation of the feature center samples of other categories; Introducing enhanced boundary coefficients Enhanced samples generated after inverse denoising of target category samples using a diffusion model. Other category feature center samples Weighted representation is performed to obtain enhanced style samples. ; ；（21） in, Indicates the enhancement boundary coefficient. Indicates the number of feature centers.

9. An electronic device, comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements any step in the electronic equipment assembly characterization enhancement method based on digital mainline as described in any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a computer processor, implements any step in the electronic equipment assembly characterization enhancement method based on a digital mainline as described in any one of claims 1-8.

Images