An intelligent manufacturing-based industrial big data analysis method

Abstract

The application discloses an industrial big data analysis method based on intelligent manufacturing, and relates to the technical field of industrial big data analysis.The stress field driving defect generation mechanism is used, and virtual defects strictly follow the material mechanics response law, for example, cracks expand along the direction of the maximum principal stress, and bubble distribution is constrained by surface tension, so that the sample distortion risk caused by the traditional generative adversarial network (GAN) due to the separation from the physical mechanism is avoided; combined with dynamic updating of the virtual defect library, the synthesized samples cover the typical failure modes under the real working conditions; the meta-learning framework extracts the cross-device common law through the shared feature layer, the task adapter layer injects the device-specific knowledge, one-time pre-training and light fine-tuning are realized, and fast migration is realized; the physical constraint function is used as a priori regularization term to limit the overfitting tendency of the model to the noise of the limited samples, and the detection stability when the production line switches to a new product is improved.

Description

Technical Field

[0001] This invention relates to the field of industrial big data analysis technology, and in particular to an industrial big data analysis method based on intelligent manufacturing. Background Technology

[0002] In the field of intelligent manufacturing, industrial big data analysis technology is gradually being applied to defect detection scenarios in high-end equipment, such as precision components like aero-engine blades and semiconductor wafers. In such scenarios, due to limitations in process maturity and quality control levels, defect samples are naturally scarce, with very few defect samples that can be collected in actual production, averaging less than 100 per month. Moreover, defect morphology is strongly constrained by physical laws, such as crack propagation direction and bubble distribution patterns. Current mainstream solutions adopt deep learning-based visual detection models, relying on massive amounts of data for training to achieve high-precision recognition.

[0003] Common approaches focus on few-shot learning and transfer learning, such as Meta-Learning: reusing pre-trained models across devices and fine-tuning the target domain through task adaptation; and Generative Adversarial Networks (GANs): synthesizing defect samples to expand the dataset. However, these approaches lack modeling of material stress distribution and thermal deformation mechanisms, causing synthesized defects to deviate from actual physical constraints, such as randomly generated cracks that violate the direction of metal fatigue propagation. In addition, during cross-device transfer, the difference in working conditions between the source domain (e.g., ordinary machine tools) and the target domain (e.g., five-axis machining centers) causes feature distribution shifts, leading to overfitting noise in the model after fine-tuning. Black-box models struggle to correlate defect features with process parameters, such as cutting force fluctuations, and cannot guide process improvement.

[0004] To address the aforementioned issues, most current solutions should introduce physical priors, using material mechanics equations as loss function constraints, or synthesize defects based on the process causal chain. However, such methods require precise physical equations or full-chain process models, which are extremely costly to model in complex scenarios with multivariate coupling, such as composite material molding. At the same time, the models tend to memorize the noise characteristics of limited samples rather than learn generalized defect laws, resulting in a sharp drop in detection performance when switching production lines to new products. Summary of the Invention

[0005] In view of the aforementioned existing problems, the present invention is proposed.

[0006] This invention provides an industrial big data analysis method based on intelligent manufacturing to solve the problems of physical law mismatch, poor cross-domain generalization, and black box modeling in the detection of small samples in high-end manufacturing.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0008] This invention provides an industrial big data analysis method based on intelligent manufacturing, which includes:

[0009] Step S1, construct a virtual defect sample library: based on the mapping relationship between material mechanical properties and process parameters, generate virtual defect images that conform to physical laws;

[0010] Step S2, Pre-trained meta-learning model: Construct a multi-task pre-trained model using cross-device historical defect data and virtual defect samples generated in step S1;

[0011] Step S3, Physical constraint fine-tuning: Input small sample defect data from the target production line into the pre-trained model, and add constraint functions derived from material failure mechanisms to optimize the loss function;

[0012] Step S4, Online Defect Detection: Real-time acquisition of images of production line components, and output of defect classification and location information by fine-tuning the model.

[0013] As a preferred embodiment of the industrial big data analysis method based on intelligent manufacturing described in this invention, the virtual defect image generation in step S1 must satisfy the following:

[0014] The defect morphology extends along the direction of material stress concentration;

[0015] The defect distribution is consistent with the simulation results of the thermo-mechanical coupled field.

[0016] The determination of the stress concentration direction includes:

[0017] Finite element statics simulation was performed based on the 3D model of the component to extract the maximum principal stress distribution field.

[0018] Correct principal stress direction deviations based on material lattice orientation database.

[0019] As a preferred embodiment of the industrial big data analysis method based on intelligent manufacturing described in this invention, in the process of obtaining the stress concentration direction, the coordinate set of the weak area of ​​the component is obtained through finite element analysis, and the virtual defect is generated using the coordinate set as the initial growth point.

[0020] As a preferred embodiment of the industrial big data analysis method based on intelligent manufacturing described in this invention, the acquisition of the weak area coordinate set in step S1 includes:

[0021] Establish a finite element model and solve for the nodal stress field;

[0022] The nodes are measured based on the composite weakening index formed by the von Mises equivalent stress and its spatial gradient mode;

[0023] The composite weakening index is defined as:

[0024] ,

[0025] in, For nodes The weakening index, For the material's yield strength, This is a reference value for the stress gradient. For nodes von Mises equivalent stress, For adaptive weights, dimensionless. It is the stress gradient mode;

[0026] Weight , Adaptive allocation based on the magnitude of stress and gradient;

[0027] use The statistical threshold determines the output of the weak region coordinate set;

[0028] in, To weaken the threshold, dimensionless, and Each for all The mean and standard deviation are dimensionless.

[0029] As a preferred embodiment of the industrial big data analysis method based on intelligent manufacturing described in this invention, the physical constraint function in step S3 includes:

[0030] Crack curvature continuity constraint: By embedding the loss using a variational method, the rate of change of adjacent curvatures of the predicted crack skeleton is limited to not exceeding a material-related threshold.

[0031] Bubble distribution constraints: Markov random fields are used to model the repulsion relationship between adjacent bubbles.

[0032] As a preferred embodiment of the industrial big data analysis method based on intelligent manufacturing described in this invention, step S3, fine-tuning the crack curvature continuity constraint, includes:

[0033] The crack segmentation results output by the model are refined and the skeleton pixel sequence is extracted. Calculate discrete curvature ;

[0034] Define neighborhood curvature difference and set Allowable threshold; where, The curvature difference between adjacent pixels, in units of , The threshold for the allowable curvature difference of the material. As an empirical factor, The fracture toughness of the material is expressed in units of 1. , This is the elastic modulus, expressed in Pa.

[0035] Build based on The curvature continuity loss of the penalty term;

[0036] use A monotonically increasing weight scheduling strategy, wherein, For the first The training round curvature loss weights are dimensionless. This is the maximum weight limit. The growth rate constant is The number of iterations is represented by the unit "times".

[0037] As a preferred embodiment of the industrial big data analysis method based on intelligent manufacturing described in this invention, the multi-task pre-training model in step S2 adopts the following structure:

[0038] Shared feature extraction layer: Cross-device feature encoding is achieved by a residual convolutional network;

[0039] Task adapter layer: Lightweight fully connected layers trained separately for each type of device.

[0040] As a preferred embodiment of the industrial big data analysis method based on intelligent manufacturing described in this invention, the defect location information output in step S4 is associated with the process parameters, specifically including: mapping the defect area to the processing process node; and using KL divergence to measure the deviation of the current process parameters relative to the historical good product parameters, and outputting the deviation value of the key process parameters.

[0041] As a preferred embodiment of the industrial big data analysis method based on intelligent manufacturing described in this invention, it further includes a model iteration mechanism:

[0042] When the online detection confidence level falls below a preset threshold, the virtual defect library is automatically updated.

[0043] The weights of the physical constraint function are dynamically adjusted based on newly added defect samples.

[0044] As a preferred embodiment of the industrial big data analysis method based on intelligent manufacturing described in this invention, the model iteration mechanism includes:

[0045] The average detection confidence level of several defect instances within a batch is used as the online confidence level measure:

[0046] ,

[0047] in, Indicates time The average detection confidence, dimensionless. The number of defect instances used in the confidence score calculation, in units of [number]. For example The confidence level is 0-1;

[0048] when Below the threshold At that time, the defect confidence difference is determined based on the relative difference between the confidence level and the threshold, and virtual defect samples for this round are generated proportionally accordingly. The proportion is adjusted by the magnification within a preset range.

[0049] Simultaneously, the weights of each constraint are proportionally increased according to the defect confidence difference, and limited to their respective maximum values.

[0050] The beneficial effects of this invention are as follows: This invention uses a stress field-driven defect generation mechanism, and the virtual defects strictly follow the material mechanics response laws, such as cracks propagating along the direction of maximum principal stress and bubble distribution being constrained by surface tension, thus avoiding the risk of sample distortion caused by traditional generative adversarial networks (GANs) deviating from physical mechanisms; combined with dynamic updates of the virtual defect library, it ensures that the synthesized samples cover typical failure modes under real working conditions; the meta-learning framework of this invention extracts common laws across devices through a shared feature layer and injects device-specific knowledge into the task adapter layer, achieving rapid transfer with one pre-training and lightweight fine-tuning; physical constraint functions (crack curvature continuity, bubble repulsion) serve as prior regularization terms, limiting the model's tendency to overfit to limited sample noise and improving the detection stability when switching to new products on the production line.

[0051] The defect-process correlation of this invention traces the root cause of defects through knowledge graphs and quantifies process parameter deviations using KL divergence. When a blade crack is detected, it can be located to specific process nodes such as heat treatment temperature control deviations or excessive cutting forces, providing a clear direction for process optimization and overcoming the limitation that black-box models cannot guide production.

[0052] The iterative triggering of this invention adaptively adjusts the virtual sample generation scale and constraint weights based on the online detection confidence level; when the confidence level decreases due to production line switching, it automatically strengthens the physical constraint weights and expands the relevant defect samples, so that the model can continuously adapt to new scenarios and reduce the cost of manual intervention. Attached Figure Description

[0053] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0054] Figure 1 This is a flowchart illustrating the industrial big data analysis method based on intelligent manufacturing in Example 1. Detailed Implementation

[0055] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0056] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0057] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0058] Example 1, referring to Figure 1 This embodiment provides an industrial big data analysis method based on intelligent manufacturing, including the following steps:

[0059] Step S1, construct a virtual defect sample library: based on the mapping relationship between material mechanical properties and process parameters, generate virtual defect images that conform to physical laws;

[0060] The virtual defect image generation in step S1 must meet the following requirements:

[0061] The defect morphology extends along the direction of material stress concentration;

[0062] The defect distribution is consistent with the simulation results of the thermo-mechanical coupled field.

[0063] Determining the direction of stress concentration includes:

[0064] Finite element statics simulation was performed based on the 3D model of the component to extract the maximum principal stress distribution field.

[0065] Correct principal stress direction deviations based on material lattice orientation database;

[0066] During the process of obtaining the stress concentration direction, the coordinate set of the weak area of ​​the component is obtained through finite element analysis, and the virtual defect is generated using this coordinate set as the initial growth point;

[0067] Step S1, the step of obtaining the coordinate set of the weak area of ​​the component through finite element analysis, includes:

[0068] Construct a finite element model and solve for the nodal stresses:

[0069] , ,

[0070] Where is the stiffness matrix, with units of . , This is the nodal displacement vector, in meters. This is the external load vector, in units of N. For nodes Cauchy stress tensor at point, in Pa. The constitutive matrix is ​​Pa. For nodes Displacement-strain transformation matrix, in units of ;

[0071] The formula for calculating the equivalent stress at the nodes is:

[0072] ,

[0073] in, For nodes von Mises equivalent stress, in Pa. This represents the normal stress component, in Pa. The shear stress component is expressed in Pa.

[0074] The formula for calculating the stress gradient modulus is:

[0075] ,

[0076] in, For nodes Equivalent stress gradient modulus, in units of , For spatial gradient operators, units ;

[0077] Construction of composite weakening index:

[0078] ,

[0079] in, For nodes The weakening index is dimensionless. The yield strength of the material is expressed in Pa. This is a reference value for the stress gradient. , Represents a full node 95th percentile, unit , For adaptive weights, dimensionless.

[0080] , ,

[0081] in, For adjustment coefficients, ,in For length features, Yield strength, in Pa;

[0082] Perform threshold determination and output coordinate set:

[0083] , ,

[0084] in, This is the coordinate set of the weak region, in meters. These are the node coordinates, in meters. To weaken the threshold, dimensionless, and Each for all The mean and standard deviation are dimensionless.

[0085] Specifically, this step starts with the static boundary value problem to obtain the complete stress field, laying the foundation for subsequent physical inferences; equivalent stress characterizes absolute strength, and stress gradient reveals local changes. The two are coupled by a weakening exponent to simultaneously capture high-risk areas and transition zones; adaptive weights are calculated in real time by node states to avoid distortion caused by empirical parameter tuning; thresholds are automatically determined using statistical quantiles to reduce manual intervention, and after the coordinate set is output, discrete points are removed by density clustering to make the boundaries of weak areas smooth and coherent; the overall process forms a closed loop of finite element solution gradient analysis exponent evaluation threshold extraction, ensuring that virtual defect growth points have traceable physical evidence and improving the coverage and credibility of subsequent sample synthesis of real failure modes;

[0086] Step S2, Pre-trained meta-learning model: Construct a multi-task pre-trained model using cross-device historical defect data and virtual defect samples generated in step S1;

[0087] The multi-task pre-trained model in step S2 adopts the following structure:

[0088] Shared feature extraction layer: Cross-device feature encoding is achieved by a residual convolutional network;

[0089] Task adapter layer: A lightweight fully connected layer trained separately for each type of device, where lightweight means ≤10 parameters. 6 ;

[0090] Step S3, Physical constraint fine-tuning: Input small sample defect data from the target production line into the pre-trained model, and add constraint functions derived from material failure mechanisms to optimize the loss function;

[0091] The physical constraint functions in step S3 include:

[0092] Crack curvature continuity constraint: Force the rate of change of curvature of adjacent pixels to be lower than a set threshold;

[0093] Bubble distribution constraint: restricting bubble spacing to conform to the fluid surface tension equation;

[0094] The crack curvature continuity constraint is embedded in the loss function through variational method to force the predicted crack curvature derivative to be smooth.

[0095] The bubble distribution constraint is modeled using Markov random fields to model the repulsive effect between adjacent bubbles;

[0096] In step S3, fine-tuning of the crack curvature continuity constraint is performed, specifically including:

[0097] The crack skeleton is extracted, the discrete curvature is calculated, and a thinning operator is used to obtain the pixel sequence from the crack segmentation results output by the model. ;

[0098] ,

[0099] in, Cracks In pixels Discrete curvature, in units of , For three points The resulting angle, measured in rad. and These represent the Euclidean distances between adjacent pixels, in pixels. Cracks No. The center coordinates of each pixel Indicates the pixel number. The number of pixels per crack;

[0100] Define the neighborhood curvature difference and the threshold respectively:

[0101] , ,

[0102] in, The curvature difference between adjacent pixels, in units of , The threshold for the allowable curvature difference of the material. As an empirical factor, The fracture toughness of the material is expressed in units of 1. , This is the elastic modulus, expressed in Pa.

[0103] Define curvature continuity loss:

[0104] ,

[0105] in, For curvature continuity loss, dimensionless, Count the pixel differences for all cracks, dimensionless. The number of cracks is [number];

[0106] Perform weighted adaptive scheduling:

[0107] ,

[0108] in, For the first The training round curvature loss weights are dimensionless. This is the maximum weight limit. The growth rate constant is The number of iterations, in units of iterations;

[0109] The overall fine-tuning goal is:

[0110] ,

[0111] in, The loss is a comprehensive loss, dimensionless. For monitoring of losses, dimensionless.

[0112] Specifically, the curvature continuity constraint explicitly incorporates the local smoothness of the crack into the loss function through discrete geometry definition, avoiding pseudo-textures caused by high-frequency jitter; the threshold is derived from the material toughness and elastic modulus, mapping the physical limitation of crack inflection on energy release rate; the adaptive weight increases with training progress, retaining task loss as the main factor in the early stage and strengthening shape regularity in the later stage, resulting in more stable convergence; the overall curvature calculation-differential judgment-threshold penalty-weight scheduling form a closed loop, making the crack contour curvature change of the fine-tuned model output smooth, conforming to the continuity of fracture mechanics, and improving the consistency between defect location and actual crack propagation path;

[0113] Step S4, Online Defect Detection: Real-time acquisition of images of production line components, and output of defect classification and location information by fine-tuning the model;

[0114] The defect location information output in step S4 is associated with process parameters, including:

[0115] Defect areas are mapped to processing step nodes;

[0116] Output the deviation values ​​of key process parameters that cause defects;

[0117] The output of key process parameter deviation values ​​includes:

[0118] Match causal chains in the process knowledge graph based on defect type;

[0119] Calculate the KL divergence between the current process parameters and the historical good product parameters as a deviation measure;

[0120] It also includes a model iteration mechanism:

[0121] When the online detection confidence level falls below a preset threshold, the virtual defect library is automatically updated.

[0122] The weights of the physical constraint function are dynamically adjusted based on newly added defect samples.

[0123] In the model iteration mechanism, the methods for triggering iterations and updating dynamic weights include:

[0124] Online detection confidence metrics:

[0125] ,

[0126] in, Indicates time The average detection confidence, dimensionless. The number of defect instances used in the confidence score calculation, in units of [number]. For example The confidence level is 0-1;

[0127] Defect confidence difference and trigger determination:

[0128] ,like Then, a virtual defect library update will be initiated.

[0129] in, The confidence difference ratio is dimensionless. The preset threshold is 0-1;

[0130] Number of newly added virtual defect samples:

[0131] ,

[0132] in, This represents the number of virtual defect samples that need to be generated in this round. To enlarge, Dimensionless The baseline number of defect samples, This is the floor operator;

[0133] The physical constraint weights are dynamically adjusted using the following formula:

[0134] ,

[0135] in, To constrain the next round of training The weights are dimensionless. The current weight is dimensionless. This is the weighting growth factor. The dimensionless interval was obtained by Bayesian optimization, which compromised the data increment with the convergence speed. To constrain The maximum allowed weight, dimensionless;

[0136] Specifically, this mechanism introduces the confidence difference ratio to measure the real-time reliability of the model. The measurement method is independent of the input size and has adaptability. When the confidence level is lower than the threshold, the scale of the new sample is determined by the proportional coefficient, and the virtual defect library is expanded in a targeted manner to avoid blindly increasing the number of samples and causing a surge in training costs. The weight update adopts a linear growth with an upper limit truncation. The more obvious the defect, the higher the proportion of physical constraints. This ensures that the shape and distribution regularization are dynamically strengthened as the model's weak links are strengthened, thereby steadily improving the detection consistency and physical interpretability in new scenarios.

[0137] The values ​​of 0.3, 0.05, 0.85, etc. listed in this invention are merely preferred examples. Those skilled in the art can replace or adjust them to other suitable values ​​without creative effort, based on specific equipment operating conditions, data scale, or through conventional parameter tuning methods (such as cross-validation, grid / Bayesian optimization, etc.) without departing from the essence of this invention.

[0138] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An industrial big data analysis method based on intelligent manufacturing, characterized in that, Includes the following steps: Step S1, construct a virtual defect sample library: based on the mapping relationship between material mechanical properties and process parameters, generate virtual defect images that conform to physical laws; Step S2, Pre-trained meta-learning model: Construct a multi-task pre-trained model using cross-device historical defect data and virtual defect samples generated in step S1; Step S3, Physical constraint fine-tuning: Input small sample defect data from the target production line into the pre-trained model, and add constraint functions derived from material failure mechanisms to optimize the loss function; Step S4, Online Defect Detection: Real-time acquisition of images of production line components, and output of defect classification and location information by fine-tuning the model; The virtual defect image generation in step S1 must meet the following requirements: The defect morphology extends along the direction of material stress concentration. During the process of obtaining the stress concentration direction, the coordinate set of the weak area of ​​the component is obtained through finite element analysis, and the virtual defect is generated using this coordinate set as the initial growth point. Obtaining the set of weak region coordinates in step S1 includes: Establish a finite element model and solve for the nodal stress field; The nodes are measured based on the composite weakening index formed by the von Mises equivalent stress and its spatial gradient mode; The composite weakening index is defined as: ， in, For nodes The weakening index, For the material's yield strength, This is a reference value for the stress gradient. For nodes von Mises equivalent stress, For adaptive weights, dimensionless. It is the stress gradient mode; Weight , Adaptive allocation based on the magnitude of stress and gradient; use The statistical threshold determines the coordinate set of the weak area. in, To weaken the threshold, dimensionless, and Each for all The mean and standard deviation are dimensionless. The physical constraint functions in step S3 include: Crack curvature continuity constraint: By embedding the loss using a variational method, the rate of change of adjacent curvatures of the predicted crack skeleton is limited to not exceeding a material-related threshold. Bubble distribution constraints: Markov random fields are used to model the repulsion relationship between adjacent bubbles.

2. The industrial big data analysis method based on intelligent manufacturing as described in claim 1, characterized in that, The virtual defect image generation in step S1 must meet the following requirements: The defect morphology extends along the direction of material stress concentration. The defect distribution is consistent with the simulation results of the thermo-mechanical coupled field. The determination of the stress concentration direction includes: Finite element statics simulation was performed based on the 3D model of the component to extract the maximum principal stress distribution field. Correct principal stress direction deviations based on material lattice orientation database.

3. The industrial big data analysis method based on intelligent manufacturing as described in claim 1, characterized in that, Step S3, fine-tuning the crack curvature continuity constraint, includes: The crack segmentation results output by the model are refined and the skeleton pixel sequence is extracted. Calculate discrete curvature ; Define neighborhood curvature difference and set Allowable threshold; where, The curvature difference between adjacent pixels, in units of , The threshold for the allowable curvature difference of the material. As an empirical factor, The fracture toughness of the material is expressed in units of 1. , This refers to the elastic modulus, expressed in Pa. Build based on The curvature continuity loss of the penalty term; use A monotonically increasing weight scheduling strategy, wherein, For the first The training round curvature loss weights are dimensionless. This is the maximum weight limit. The growth rate constant is The number of iterations is represented by the unit "times".

4. The industrial big data analysis method based on intelligent manufacturing as described in claim 1, characterized in that, The multi-task pre-trained model in step S2 adopts the following structure: Shared feature extraction layer: Cross-device feature encoding is achieved by a residual convolutional network; Task adapter layer: Lightweight fully connected layers trained separately for each type of device.

5. The industrial big data analysis method based on intelligent manufacturing as described in claim 1, characterized in that, The defect location information output in step S4 is associated with the process parameters, specifically including: mapping the defect area to the processing step node; and using KL divergence to measure the deviation of the current process parameters relative to the historical good product parameters, and outputting the deviation value of the key process parameters.

6. The industrial big data analysis method based on intelligent manufacturing as described in claim 1, characterized in that, It also includes a model iteration mechanism: When the online detection confidence level falls below a preset threshold, the virtual defect library is automatically updated. The weights of the physical constraint function are dynamically adjusted based on newly added defect samples.

7. The industrial big data analysis method based on intelligent manufacturing as described in claim 6, characterized in that, The model iteration mechanism includes: The average detection confidence level of several defect instances within a batch is used as the online confidence level measure: ， in, Indicates time The average detection confidence, dimensionless. The number of defect instances used in the confidence score calculation, expressed in units of individual instances. For example The confidence level is 0-1; when Below the threshold At that time, the defect confidence difference is determined based on the relative difference between the confidence level and the threshold, and virtual defect samples for this round are generated proportionally accordingly. The proportion is adjusted by the scaling factor within a preset range. Simultaneously, the weights of each constraint are proportionally increased according to the defect confidence difference, and limited to their respective maximum values.

Images