Graphite bipolar plate surface defect identification method

Abstract

This invention provides a method for identifying surface defects in graphite bipolar plates, belonging to the field of defect detection technology in computer vision. First, it collects and standardizes multi-source heterogeneous image data of aligned flow channels to construct a physically consistent benchmark dataset that excludes environmental noise. Then, based on a self-supervised encoder-decoder architecture, it builds a graphite surface topology repair and reconstruction model. By using masked image modeling, the network learns prior knowledge of flow channel distribution and micro-texture, and a dynamic feature enhancement module repairs missing information in masked areas. Subsequently, it performs residual analysis on the reconstructed reference image and the actual observation image, constructing a defect discrimination feedback loop using pixel-level comparison and structural similarity measurement. This accurately identifies key defect indicators such as microcracks, edge chipping, and scratches without a standard physical template. Finally, it completes defect classification and localization through connected component quantization features, decoupling the residual mapping map into quantifiable geometric parameters, effectively reducing the false alarm rate and improving the level of detection intelligence.

Description

Technical Field

[0001] This invention belongs to the field of defect detection technology in computer vision, and particularly relates to a method for identifying surface defects in graphite bipolar plates. Background Technology

[0002] In low-carbon energy technologies such as hydrogen fuel cells and redox flow batteries, graphite bipolar plates serve as the core framework of the fuel cell stack, undertaking multiple crucial functions including supporting electrodes, collecting current, isolating reactant gases, and distributing fluids. Graphite materials, due to their excellent electrical conductivity, chemical stability, and highly tunable pore structure, have become the mainstream choice for electrode plate manufacturing. However, despite their superior electrical properties, graphite materials are brittle and prone to surface defects such as chipping, scratches, and pores during processing. These surface defects can severely impact their macroscopic electrochemical performance and sealing safety.

[0003] However, in actual large-scale production and quality control, the identification of surface defects in graphite bipolar plates still heavily relies on traditional manual visual inspection or machine vision methods based on simple operators. Inspectors or algorithms typically make qualitative judgments about defects within the plate's flow channels based on preset contrast thresholds or shape features. However, due to the complex flow channel structure of graphite bipolar plates and the influence of machining errors, positioning offsets, and pressing deformation, the microscopic topological characteristics of each plate exhibit nonlinear differences. Traditional forward inspection modes based on standard template differentiation or preset thresholds lack flexible comparison benchmarks and are prone to misjudging normal processing deviations as defects, resulting in an extremely high false alarm rate.

[0004] In recent years, with the development of artificial intelligence technology, some deep learning methods have been proposed for end-to-end industrial surface defect identification. However, due to the deep black color, high light absorption, and complex natural granular texture of graphite materials, functional damage is often highly coupled with the background texture in this low-contrast context. When traditional convolutional neural networks directly perform end-to-end prediction, they are severely lacking in sensitivity to capturing minute defects because they are difficult to effectively separate background noise. In addition, end-to-end deep neural network methods require massive amounts of sample data, but due to the extremely high yield requirements of industrial production, defect samples in real production scenarios are extremely scarce and diverse. Existing supervised learning algorithms rely excessively on labeled negative samples for training, and the generalization ability and robustness of the model will decrease when facing unknown or novel defects. Summary of the Invention

[0005] To address the above problems, this invention proposes a method for identifying surface defects in graphite bipolar plates, comprising the following steps: S1. Take a large-format image of a defect-free electrode plate using an industrial camera, perform standardized image segmentation and receptive field alignment, and obtain standardized sub-image blocks. S2, execute dynamic random mask generation logic injection to construct a polymorphic structured mask image block simulating defect state; then execute multidimensional data augmentation and physical property association again to obtain background statistical feature vectors characterizing the micro-texture and local contrast variance of the graphite plate surface; A training dataset is constructed by using polymorphic structured mask image patches as model input, standardized sub-image patches as reconstruction targets, and background statistical feature vectors as auxiliary attribute labels for samples. S3 constructs and self-supervises trains a graphite surface topology reconstruction network model based on a mask autoencoder, including a flow channel topology feature encoding network that outputs a global topology feature token and a conditional embedding vector for the flow channel; a lightweight pixel decoding and reconstruction reference image generation network that generates a reconstruction reference image that is perfectly aligned with the coordinates of the input image; and a multi-scale discrimination network and a physical consistency verification network that distinguish between genuine and fake reconstruction reference images based on connectivity, texture smoothness, and physical properties at different feature resolutions. S4 acquires the surface image of the graphite bipolar plate to be tested, and after standardization, it is input into the graphite surface topology reconstruction network model in S3. The model outputs the reconstruction reference image, extracts the deviation features between the two, and decouples the defect categories.

[0006] Preferably, the process of injecting the dynamic random mask generation logic is as follows: using an adaptive mask generation operator with defect morphology awareness capability. By using a pre-defined defect prior probability distribution, various mask topologies, including linear, clustered, and diffuse point distributions, are dynamically generated to establish a topological simulation of real physical defects. Linear masks, with extremely high aspect ratios and random orientations, simulate microcracks and scratches distributed across the flow channel, forcing the network to learn the ability to recover texture continuity across regions. Edge cluster masks, using irregular polygonal masks with high fractal dimensions and constraining their spatial position to abrupt gradient changes on the flow channel wall, simulate edge chipping and damage, enhancing the model's accuracy in reconstructing complex geometric boundaries. Diffuse point masks, using Gaussian-distributed fine spot masks, simulate the natural sand holes, pores, and surface contamination of graphite materials. Operators... By logically combining sub-modules, a composite mask layer is generated in the image space, and the pixel values ​​within the region are either zeroed out or filled with noise based on the statistical distribution of graphite background texture. By constructing an input generation method with mask constraints, a polymorphic structured mask image block is output.

[0007] Preferably, the process of performing multidimensional data enhancement and physical property association is as follows: First, for the high light absorption characteristics and non-uniform reflectivity of graphite bipolar plates, adaptive contrast stretching and Gamma nonlinear mapping are performed to simulate the micro-roughness changes on the graphite plate surface caused by differences in oxidation degree, tool wear, and light intensity drift during online detection; Second, gray-level co-occurrence matrix and local binary mode algorithm are introduced to quantitatively extract the micro-texture roughness, energy distribution, and local contrast variance of the graphite plate surface as background statistical feature vectors.

[0008] Preferably, the flow channel topology feature encoding network is an encoder based on the Vision Transformer architecture. First, the input image is segmented into a series of non-overlapping image patches and linearly embedded to transform them into a token sequence. Second, the background statistical feature vector is used as a conditional embedding and concatenated or fused with the image token sequence to inject prior material physical properties into the Transformer. Third, the long-range dependencies between image patch tokens are calculated through a multi-layer self-attention mechanism to capture the periodic distribution and global geometry of the flow channel across the mask region. The encoder only processes visible (i.e., unmasked) image patch tokens and outputs the global topology feature tokens of the flow channel and the conditional embedding vector.

[0009] Preferably, the lightweight pixel decoding and reconstructed reference image generation network is a lightweight pixel decoder, responsible for restoring the low-dimensional semantic features output by the channel topology feature encoding network and the masked token sequence into a high-resolution pixel space image; the decoder visualizes the channel topology structure and background texture prior knowledge learned by the encoder through multi-layer upsampling and convolution operations; when restoring the masked area, the decoder repairs it according to the distribution law of qualified product channels it has learned, reconstructs the defective areas that may exist in the input image into a defect-free structure that conforms to the periodic characteristics of graphite channels, and finally generates a reconstructed reference image that is completely aligned with the coordinates of the input image.

[0010] Preferably, the multi-scale discriminant network and the physical consistency verification network adopt a multi-level feature pyramid structure to distinguish between the reconstructed reference image and the real qualified sample image at the full image scale, local flow channel scale, and micro-texture scale, respectively, so as to optimize the visual quality of the reconstructed image. At the same time, a background statistical feature regression head is embedded at the end of the discriminant network to predict the background texture parameters of the reconstructed reference image, including GLCM roughness, and compare it with the extracted original background statistical feature vector to ensure that the generation network maintains the intrinsic physical properties of the material during the reconstruction process.

[0011] Preferably, in step S4, firstly, a reconstruction residual map calculation based on multi-criteria fusion is performed. By comparing the differences in pixel grayscale and local structure between the real observed image and the reconstruction reference image, a composite residual map containing significant defect features is generated. Secondly, defect saliency enhancement and adaptive binarization segmentation are performed. The Laplacian operator is used to sharpen the edges and adaptive threshold segmentation is combined to filter out background noise, generating a high-quality binarized defect mask. Finally, multidimensional geometric feature quantization and category decoupling of defects are performed. The binarized defect connected components are mathematically represented with high precision, and their topological features are extracted to achieve automatic classification of defect categories. For each identified target, its equivalent area, aspect ratio of the bounding rectangle, roundness factor, contour fractal dimension, and skew angle relative to the flow channel direction are extracted, and multidimensional quantized feature parameters of defects are output. Defect classification is then performed using the quantization indicators.

[0012] Preferably, the calculation of the reconstructed residual map based on multi-criteria fusion specifically involves: on the one hand, calculating the absolute value of the pixel-level grayscale residual. This is used to lock in chipped edges and scratches with obvious contrast; on the other hand, the local structural similarity sliding window is used to calculate the structural deviation map to capture the subtle geometric deformation of the flow channel edge; the two types of residual information are fused at multiple scales to generate a composite residual mapping map containing significant defect features.

[0013] Preferably, the defect classification using quantitative indicators specifically involves: when the extension length of a target exceeds the preset flow channel period width and the shape factor exhibits a slender shape, it is determined to be a scratch; when the fractal dimension of a target exceeds the reference value of a smooth edge and its spatial position is located on the sidewall of the flow channel edge, it is determined to be a chipped edge, such defects usually mean that the mechanical strength of the graphite structure at that location has been damaged; when the roundness of the target is extremely high, the area is small and randomly distributed in the flow channel plane area, it is determined to be a pinhole.

[0014] Preferably, the process also includes the following steps: Based on the obtained multidimensional quantitative characteristic parameters of defects, the defect area, grayscale residual intensity, morphological irregularity, fractal dimension, and spatial topological location are input into a preset quality threshold model, and risk classification rules are established in conjunction with the flow channel structure design parameters; for Level 1 risk, defined as structural damage, if the detected defect crosses the flow channel ridge line, i.e., the scratch exceeds the set threshold, or the chipping depth exceeds the flow channel wall thickness, it is judged as a structural damage defect, the system triggers a scrap-level warning signal, and marks it as a non-circulating product in the MES system; for Level 2 risk, defined as process anomaly, if a non-penetrating defect such as diffuse pinholes is detected, and its total area ratio exceeds the set threshold, it is judged as a process fluctuation type defect, the system marks the sample as requiring manual review level warning, for process quality traceability and equipment status diagnosis.

[0015] The innovative aspects and beneficial effects of this invention include: (1) Design of a self-supervised generative reconstruction network architecture for flow channel background suppression: A generative reconstruction model architecture based on masked image modeling and long-range dependency awareness is proposed. The architecture first constructs a graphite topology feature encoding network, which captures the long-range spatial periodicity of the graphite surface through dilated convolution and self-attention mechanism to overcome complex flow channel geometric interference, forming a strong semantic expression of the healthy flow channel topology. Then, a graphite surface topology repair network is constructed, which adopts a self-supervised reconstruction strategy. By randomly masking the input image and forcing the network to recover damaged pixels, the network dynamically generates a reconstruction reference image that is aligned with the coordinates of the image to be tested by utilizing the learned normal sample distribution law. At the same time, a multi-scale structural integrity discrimination network is constructed to distinguish the authenticity of the connectivity and texture smoothness of the reconstruction image at different feature resolutions, ensuring that the generated reference background is visually close to the defect-free state. (2) Design of Non-contact Residual Analysis and Defect Localization Strategy Based on Virtual Template Generation: A defect judgment strategy is proposed that abandons external physical templates and uses self-healing reconstruction images as dynamic benchmarks. During the inspection process, the generated reconstruction reference image is directly used as the normal standard of the workpiece to eliminate the global alignment problem caused by processing displacement, deformation and clamping errors. By constructing a composite evaluation function including pixel grayscale residual, spatial gradient loss and structural similarity loss, the micro-deformation difference between the original test image and the reconstruction reference image is measured. Under this mechanism, the defect area generates a significant residual gradient because it cannot be perfectly reconstructed by the normal flow channel distribution law. After binarization segmentation and morphological filtering, the gradient is transformed into a defect heat map that can guide industrial decision-making, realizing the accurate decoupling of subtle damage features in complex texture background. (3) Data preprocessing and multidimensional defect quantification evaluation system for graphite material feature perception: In view of the physical characteristics of graphite material such as high light absorption, low contrast and complex texture, a full-link data standardization and quantification evaluation system is established. First, a physical property enhancement benchmark for graphite blackness is constructed. Through adaptive histogram normalization and Laplacian operator resampling, the interference of uneven illumination on the material surface is eliminated. Second, a confidence optimization mechanism based on shape factor and gray level statistics is designed to automatically screen and identify typical defects such as edge chipping, scratches, pores and processing burrs from residual candidates. Most importantly, a special defect quantification characterization network is constructed. With the residual feature vector as input, the nonlinear mapping relationship between defect geometry and battery electrochemical sealing performance is analyzed. The abstract micro-damage is restored into a standardized inspection report including location, depth, area and severity level, realizing a closed-loop identification path from original image acquisition to virtual template generation to accurate defect quantification. Attached Figure Description

[0016] Figure 1 This is a flowchart of the overall method of the present invention.

[0017] Figure 2 This is a diagram of the core architecture of the graphite surface topology reconstruction network model of the present invention.

[0018] Figure 3 This is a flowchart of the defect detection and decision-making logic of the present invention.

[0019] Figure 4 This is a performance evaluation diagram of surface defect detection in the flow channel of a graphite bipolar plate in an embodiment of the present invention.

[0020] Figure 5 This is a performance comparison chart of the model with a small sample size in an embodiment of the present invention.

[0021] Figure 6 This is a flowchart illustrating the entire process of identifying surface defects in a graphite bipolar plate in an embodiment of the present invention.

[0022] Figure 7 This is a comparison chart of inference time at different image sizes in an embodiment of the present invention.

[0023] Figure 8 This is an experimental evaluation diagram of the system module ablation in an embodiment of the present invention. Detailed Implementation

[0024] To achieve accurate identification of surface defects in graphite bipolar plates in a template-free environment, this invention proposes a method for identifying surface defects in graphite bipolar plates based on self-supervised generative reconstruction. The overall process is as follows: Figure 1 As shown: First, the system performs multi-source heterogeneous image data acquisition and standardization alignment of graphite bipolar plate flow channels to construct a physically consistent benchmark dataset that excludes environmental noise interference. Then, based on a self-supervised encoder-decoder architecture, a graphite surface topology repair and reconstruction model for complex flow channel backgrounds is built. The masked image modeling mechanism forces the network to learn prior knowledge of the periodic flow channel distribution and micro-texture of graphite materials, and the dynamic feature enhancement module realizes the logical repair of missing information in the masked areas. Further, a residual analysis strategy based on the reconstructed reference image and the actual observed image is executed. A feedback loop for defect discrimination is constructed using pixel-level comparison and structural similarity measurement to ensure accurate locking of key indicators such as microcracks, chipping, and scratches without a standard physical template. Finally, defect classification and localization based on connected component quantization features are executed, decoupling the abstract residual mapping map into quantifiable geometric parameters, thereby significantly reducing the false alarm rate and improving the intelligence level of the graphite bipolar plate quality inspection process.

[0025] The specific implementation of the present invention will be further described below with reference to specific embodiments.

[0026] S1. Construction of a self-supervised training dataset for graphite flow channels based on topological prior features. Because surface defects in graphite bipolar plates are diverse and difficult to annotate manually, this step aims to utilize the regular geometric features of graphite channel height in graphite bipolar plates to manually inject surface defects into defect-free graphite bipolar plate images. The defect-free graphite bipolar plate image and its corresponding artificially defective graphite bipolar plate image are then used as a sample pair in the dataset. This dataset allows the model to understand the structural regularity of graphite channels in normal graphite bipolar plates. During actual detection, any defects that do not conform to this regularity will be revealed because they cannot be reconstructed by the trained model, thus solving the problems of high manual annotation costs and difficulty in covering rare surface defects. This step aims to utilize the geometric regularity of graphite bipolar plate channel height to construct a structured benchmark library containing only features of qualified samples. By performing image segmentation and mask logic injection, the original complete channel images are transformed into self-supervised training pairs with logic from damage to repair, thereby forcing the model to autonomously learn the normal topological distribution of graphite surfaces without manual annotation. The specific process is as follows: First, standardized image segmentation and receptive field alignment are performed to divide the large-format original electrode image into standardized sub-image blocks that meet the input constraints of the neural network and have continuous edge features. Second, dynamic random mask generation logic injection is performed to construct polymorphic structured mask image blocks simulating information-deficient states using an adaptive mask generation operator with defect morphology perception capabilities. Third, multidimensional data augmentation and physical attribute association are performed to obtain background statistical feature vectors representing the micro-texture and local contrast variance of the graphite plate surface by extracting the gray-level co-occurrence matrix and using local binary mode algorithms. Finally, self-supervised reconstruction dataset encapsulation with multidimensional attribute labels is performed to serialize and store the polymorphic structured mask image blocks, the original complete image blocks, and the background statistical feature vectors to form a serialized self-supervised dataset with multidimensional attribute labels.

[0027] S1-1 Standardized Image Segmentation and Receptive Field Alignment: Large-format electrode images captured by industrial cameras typically exceed the input limitations of neural networks. Before feature extraction, these images need to be standardized and segmented. This step employs an overlapping sliding window algorithm to segment the original image into fixed-size sub-image blocks. Set a preset overlap rate to ensure the continuity of flow channel edge features between adjacent sub-blocks and prevent edge effects from occurring in the model during the inference phase; finally, obtain a standardized sub-image block containing complete flow channel ridge and groove features. S1-2 Dynamic random mask generation logic injection; for each standardized sub-image block after segmentation. This application designs an adaptive mask generation operator with defect shape awareness capability. The operator does not employ a single random masking method, but rather dynamically generates various mask topologies, including linear, clustered, and diffuse point distributions, based on a predefined defect prior probability distribution, to establish a topological simulation of real physical defects. The linear mask, with its extremely high aspect ratio and random orientation, simulates microcracks and scratches distributed across flow channels, forcing the network to learn the ability to recover texture continuity across regions. The edge clustered mask uses irregular polygonal masks with high fractal dimensions and their spatial positions... Constrained at points of abrupt gradient changes on the flow channel wall, this model simulates edge chipping and damage, enhancing its accuracy in reconstructing complex geometric boundaries. A diffuse dot mask, using Gaussian-distributed fine speckle masks, simulates the natural sand holes, pores, and surface contamination of graphite materials. An operator logically combines these sub-modules to generate a composite mask layer in the image space, and either sets the pixel values ​​within this region to zero grayscale or fills them with noise based on the statistical distribution of the graphite background texture. By constructing an input generation method that constrains the mask, a polymorphic structured mask image patch is output. ,use In the image The information gap created by constructing a topological structure that is approximately isomorphic to the actual defect morphology forces the model to structurally reconstruct the suppressed region based on the contextual information of the unoccluded region during the training phase. This allows for deep learning of the periodic skeleton distribution and spatial continuity characteristics of the flow channel region. This mechanism aims to simulate the information occlusion caused by scratches, chipping, or pinholes that may occur in actual production. Through this training strategy, the model can establish an intrinsic representation of the prior distribution of the normal graphite surface structure, providing a stable discrimination benchmark for subsequent anomaly detection based on the reconstructed residuals.

[0028] S1-3 Multidimensional Data Augmentation and Physical Property Correlation: This step aims to construct augmented training pairs with material property awareness capabilities by simulating the complex physical appearance differences and optical imaging fluctuations of graphite materials, thereby improving the model's reconstruction stability under extremely complex backgrounds. The specific process is as follows: First, considering the high light absorption characteristics and non-uniform reflectivity of graphite bipolar plates, adaptive contrast stretching and Gamma nonlinear mapping are performed to correct simulated illumination drift. The mapping formula is: ; in, coordinates The output pixel value after Gamma correction coordinates The original pixel grayscale value at that location; The maximum gray value within the current sub-image block and The minimum grayscale value within the current sub-image block; This is a correction factor, and its value range is defined. This is used to enhance dark details, thereby simulating the high absorptivity background of the graphite bipolar plate surface and forcing the model to achieve robust extraction of flow channel features within an extremely narrow grayscale range. Secondly, a gray-level co-occurrence matrix is ​​introduced for macroscopic texture quantification. The energy and contrast of the graphite plate surface are extracted using GLCM to quantitatively describe the density of graphite materials and processing marks.

[0029] Energy (ASM) reflects texture roughness, and the calculation formula is as follows: ; Contrast (Con) reflects the sharpness of the flow channel edges, and is calculated using the following formula: ; in In spatial direction ,distance Above, grayscale value and The joint probability distribution of simultaneous occurrences. The higher the ASM value, the more uniform the graphite surface texture.

[0030] Finally, in order to capture the minute pitting or scratch features generated during the graphite plate lamination process, a Local Binary Mode (LBP) operator is introduced, and the calculation formula is as follows: ; in The grayscale value of the center pixel. radius of the center pixel The first in the field The value of each neighboring pixel. The total number of neighboring pixels. For symbolic functions, These are the weighting coefficients for binary encoding. This algorithm generates binary codes by comparing local pixel differences, effectively describing the microscopic unevenness of the graphite surface and exhibiting invariance to illumination fluctuations. Finally, the weighting coefficients calculated above are... , as well as The statistical histograms are stitched together to construct a background statistical feature vector that represents the physical properties of the image patch.

[0031] S1-4 encapsulates the self-supervised reconstruction dataset with multi-dimensional attribute labels. This step aims to perform high-dimensional mapping and serialized storage of the polymorphic structured mask image patches extracted in the previous steps and the background statistical feature vectors, constructing a standardized training benchmark, including the enhanced input image, target image, and auxiliary feature labels. First, using the globally unique sample index as the association key, the polymorphic structured mask image patches generated in S1-2 are... Encapsulate the original complete image patch obtained in S1-1 as model input. The background statistical feature vector extracted in S1-3, including the gray-level co-occurrence matrix (GLCM) energy index, texture roughness, and local contrast variance, is encapsulated as an auxiliary attribute label for the sample. This label serves as a conditional constraint during model training, guiding the generative network to adaptively adjust parameters under different graphite textures. The dataset is then randomly divided into training, validation, and test sets according to a preset ratio. Finally, a serialized self-supervised dataset with multi-dimensional attribute labels is output as the input to the graphite surface topology reconstruction network model based on a mask autoencoder in subsequent steps.

[0032] S2. Construct a graphite surface topology reconstruction network model based on a mask autoencoder. The input to this step is the serialized self-supervised dataset with multi-dimensional attribute labels output by S1. The aim is to leverage the long-range dependency modeling capabilities of the Vision Transformer to learn global topological and local texture prior knowledge of graphite channels in a self-supervised mask reconstruction task, thereby constructing a virtual reference map generator with high-fidelity repair capabilities without requiring manual defect annotation. The core network architecture is as follows: Figure 2 As shown: First, a flow channel topology feature encoding network based on Vision Transformer is built. A multi-layer self-attention mechanism is used to capture the periodic distribution and global geometric structure of the flow channel across the mask region, and output the flow channel global topology feature token and conditional embedding vector. Second, a lightweight pixel decoding and reconstruction reference image generation network is constructed. Through multi-layer upsampling and convolution operations, semantic features are restored to high-resolution pixels to generate a reconstruction reference image that is perfectly aligned with the coordinates of the input image. In addition, a multi-scale discrimination network and physical consistency verification system are constructed. Using a multi-level feature pyramid structure and background statistical feature regression head, the connectivity, texture smoothness and physical properties of the reconstruction reference image are judged as true or false at different feature resolutions. Combined with a composite loss function to drive model parameter optimization, the trained graphite surface topology reconstruction model is output.

[0033] S2-1 Construction of a channel topology feature encoding network based on Vision Transformer; This step aims to construct a physically aware encoder that extracts global topological features capable of traversing occlusion regions by using a sparse representation of the flow channel features in a graphite bipolar plate image. The specific process is as follows: Polymorphic structured mask image patches encapsulated in a self-supervised dataset with serialized multidimensional attribute labels. The image is cut into a series of non-overlapping patches and linearly embedded to form a token sequence. To preserve the absolute spatial position information of the periodic arrangement of the graphite channels, absolute position encoding is added to the token sequence.

[0034] Secondly, a physical property prior injection layer is constructed. The background statistical feature vector encapsulated in the serialized self-supervised dataset of multidimensional attribute labels is used as a conditional embedding to inject material physical property priors into the Transformer. Before entering the Transformer, a two-layer fully connected layer is set up. The physical attribute prior injection layer, composed of S1-3, injects background statistical feature vectors extracted from S1-3. Mapping to image The same dimensions are used, and an additive fusion method is employed for injection. The calculation method is as follows: ; Furthermore, long-range dependency modeling is based on sparse self-attention. Multi-layer standard Transformer layers are used to calculate the visible tokens that are not masked. Each layer efficiently captures the periodic distribution of the flow channel across the mask region through a multi-head self-attention mechanism. The initial token sequence after merging; For the first An image token obtained by linearly embedding image blocks; This is for absolute position encoding.

[0035] S2-2 Lightweight Pixel Decoding and Reconstruction Reference Image Generation: A relatively lightweight pixel decoder is constructed to reconstruct a high-resolution pixel space image from the low-dimensional semantic features and masked token sequence output by the S2-1 encoder. The decoder contains only 4 standard Transformer Block layers, with only 1 / 4 the number of parameters of the encoder. Its core task is to use the global topological tokens provided by the encoder to perform preliminary semantic filling on the mask placeholders. At the end of the decoder, multi-layer upsampling and convolution operations are used, and finally, a single layer... The activation function output has the same size as the original image. Image reconstruction. The decoder visualizes the prior knowledge of the flow channel topology and background texture learned by the encoder. Since the encoder forcibly ignores the real content of the masked area during training, the decoder, when reconstructing the masked area, will repair it according to the distribution rules of qualified flow channels it has learned. This forcibly reconstructs defective areas such as chipped edges and scratches in the input image into a perfect structure that conforms to the periodic characteristics of graphite flow channels, and finally generates a reconstructed reference image that is perfectly aligned with the coordinates of the input image. .

[0036] S2-3 multi-scale discriminant network and physical consistency verification; to further improve the reconstruction reference image To enhance realism and physical consistency, a multi-scale discriminant network is introduced. This network employs a multi-level feature pyramid structure, evaluating features at the full-image scale, local flow channel scale, and micro-texture scale. The network performs image authentication against real, qualified sample images to optimize the visual quality of the reconstructed images. Simultaneously, a background statistical feature regression head is embedded at the end of the discrimination network to predict... The background texture parameters, including GLCM roughness, are compared with the original background statistical feature vector extracted by S1-3. This adversarial verification mechanism based on physical properties ensures that the generative network maintains the intrinsic physical properties of the material during the reconstruction process, avoiding the generation of artifacts that do not match the real graphite plate.

[0037] S2-4 Composite Loss Function Design and Self-Supervised Model Training: A composite objective function including reconstruction loss, perceptual loss, adversarial loss, and physical consistency loss is designed to drive iterative optimization of model parameters. The reconstruction loss constrains the difference between the decoder output pixels and the original complete image in the unmasked region, as well as the repair accuracy in the masked region; the perceptual loss utilizes a pre-trained feature extraction network to calculate... The high-order feature distance at the semantic level from the original complete image is used to ensure the visual naturalness of the reconstructed image; the adversarial loss comes from the discriminator's... The authenticity judgment forces the generator to learn to generate high-quality images that are enough to deceive the discriminator; the physical consistency loss constrains the background statistical features predicted by the discriminator's regression head to be consistent with the original features extracted in S1-3; the model uses the serialized dataset encapsulated in S1-4 to perform self-supervised iterative training until the various loss indicators on the validation set converge, and finally obtains a graphite surface topology reconstruction model with high performance in reconstructing reference images.

[0038] S3. Perform quantitative detection of surface defects in graphite bipolar plates based on the reconstruction difference of a mask autoencoder. Because the surface of graphene bipolar plates is covered with channels of varying depths and complex shapes, ordinary visual inspection struggles to distinguish the channel structure and defects on the graphene bipolar plate surface. This step utilizes a mask autoencoder model to perform nonlinear residual analysis on the reconstructed reference image and the actual observed image. The aim is to remove interference from the channel structure, allowing the model to identify pure defect signals and achieve accurate extraction and attribute quantification of minute damages. First, a reconstruction residual map calculation based on multi-criteria fusion is performed. By comparing the differences in pixel grayscale and local structure between the real observed image and the reconstructed reference image, a composite residual map containing significant defect features is generated. Second, defect saliency enhancement and adaptive binarization segmentation are performed. The Laplacian operator is used to sharpen edges, and adaptive threshold segmentation is combined to filter out background noise, generating a high-quality binary defect mask. Finally, multidimensional geometric feature quantization and category decoupling of defects are performed. The equivalent area, shape factor, fractal dimension, and spatial location parameters of the defect connected domains are extracted, and multidimensional quantized defect feature parameters are output to classify scratches, chipping, and pinholes.

[0039] S3-1 Calculation of reconstructed residual map based on multi-criteria fusion; acquisition of resampled graphite flow channel image to be tested. The reconstructed reference image generated by the mask autoencoder model in step S2. Since the mask autoencoder model forcibly repairs abnormal regions based on flow channel priors during the decoding process, this step extracts the deviation features between the two through a dual criterion: on the one hand, it calculates the absolute value of the pixel-level grayscale residual. This is used to identify chips and scratches with significant contrast. The calculation formula is as follows: ; in coordinates Pixel-level grayscale residual at the location; The image to be tested in coordinates The grayscale value at that location; To reconstruct the reference image in coordinates On the one hand, it measures the grayscale value at the location; on the other hand, it uses a local structural similarity sliding window to calculate the structural deviation map, capturing subtle geometric deformations at the flow channel edge. Structural differences are defined as: ; in coordinates Structural difference values ​​at the location; The grayscale mean of a local window of the image under test. To reconstruct the grayscale mean of a local window in the reference image, Let Variance be the grayscale variance of a local window of the image under test. To reconstruct the grayscale variance of a local window in the reference image, The grayscale variance of a local window between the image to be tested and the reconstructed image; and It is a constant; The two types of residual information are fused at multiple scales to generate a composite residual mapping map containing significant defect features. : ; in These are the weighting coefficients. For pixel-level grayscale residuals, This represents the structural difference value, thereby achieving precise decoupling of defect signals in complex background textures.

[0040] S3-2 Defect Saliency Enhancement and Adaptive Binarization Segmentation: Spatial domain feature enhancement is performed on the composite residual map to eliminate the weak background noise caused by the natural graininess of graphite during reconstruction. First, the Laplacian operator is used to sharpen the residual edges to highlight the defect contours. Second, adaptive threshold segmentation based on local mean and variance is performed to convert the residual map into a binary defect mask. Third, morphological closing operation is performed to connect the broken micro-crack pixels and remove random noise points with an area lower than the preset threshold, thereby locking in high-quality defect candidate connected regions.

[0041] S3-3 Defect Multidimensional Geometric Feature Quantization and Category Decoupling: High-precision mathematical representation of the binary defect connected domain is performed, and its topological features are extracted to achieve automatic defect category classification. For each identified target, its equivalent area, aspect ratio of the circumscribed rectangle, roundness factor, contour fractal dimension, and deflection angle relative to the flow channel direction are extracted, outputting multidimensional quantized defect feature parameters. The quantization indicators are then used to execute defect classification logic. When the extension length of a target exceeds the preset flow channel cycle width and the shape factor exhibits a slender shape (i.e., the aspect ratio is greater than 5), it is judged as a scratch. When the fractal dimension of a target exceeds the reference value of a smooth edge and its spatial position is located on the sidewall of the flow channel edge, it is judged as a chipped edge. Such defects usually mean that the mechanical strength of the graphite structure at that location has been damaged. When the roundness of the target is extremely high, the area is small, and it is randomly distributed in the flow channel plane area, it is judged as a pinhole.

[0042] S4. Defect classification, early warning, and online model update based on quality criteria This step aims to transform the multidimensional quantitative feature parameters of defects extracted in step S3 into executable industrial quality judgment decisions, realizing the transformation from image understanding to quality inspection level assessment. It also constructs a data feedback closed-loop mechanism oriented towards the actual production line environment, ensuring that the system can adapt to distribution shifts caused by fluctuations in graphite bipolar plate raw materials, equipment state drift, and changes in production process parameters during long-term operation. This maintains stable and reliable detection performance, outputting a structured inspection report and an updated graphite surface topology reconstruction model. First, a defect risk level assessment based on multidimensional feature space is performed, introducing three types of constraint criteria: structural integrity, functional safety, and process stability. The multidimensional quantitative feature parameters of defects are mapped to risk levels including structural damage and process anomalies. Second, a structured inspection report is generated, encapsulating defect categories, spatial coordinates, geometric parameters, and risk levels into standardized JSON format data for easy integration with industrial systems. Finally, online learning of a self-supervised model based on incremental learning is performed. A closed-loop self-supervised learning mechanism is used to select high-confidence qualified samples to fine-tune the model, enabling it to dynamically adapt to changes in the production line environment. Specifically, as follows... Figure 3 As shown.

[0043] S4-1 Defect Risk Level Assessment Based on Multidimensional Feature Space; Based on the quantitative features obtained in step S3-3, a preset quality threshold model is first established. This model introduces three types of criteria in the multidimensional feature space: structural integrity constraints, functional safety constraints, and process stability constraints, elevating defects from detection results to risk levels. Specifically, the system inputs parameters such as defect area, grayscale residual intensity, morphological irregularity, fractal dimension, and spatial topological location into the preset quality threshold model, and establishes risk classification rules in conjunction with flow channel structure design parameters. For Level 1 risk... Determined as structural damage, if the detected defect crosses the flow channel ridge line, i.e., the scratch exceeds the set threshold, or the chipping depth exceeds the flow channel wall thickness, it indicates that the pressure-bearing structure of the electrode plate or the gas sealing path has been damaged, and is judged as a structural damage defect.

[0044] ; If the scratch extends to a certain length Exceeding the flow channel cycle threshold or edge breakage depth Exceeding wall thickness preset ratio If the defect is detected, it is determined that the gas sealing path or pressure-bearing structure has been compromised. The system immediately triggers a scrap signal and automatically interlocks the MES system to intercept the product. This type of defect directly affects the flow field integrity and fuel cell stack sealing performance. The system triggers a scrap-level warning signal and marks the product as non-recyclable in the MES system. Level 2 risk Defined as a process anomaly, it mainly targets diffuse defects such as pinholes or minor edge chipping: ; in, The total area of ​​all trachoma. Total area of ​​graphite bipolar plates The preset roundness threshold; if the total area of ​​non-penetrating defects exceeds the threshold. and roundness If the sample exhibits characteristics of pinholes, it indicates a process drift in the raw material particle size or molding pressure. The system marks this as requiring manual review to guide equipment maintenance and diagnostics.

[0045] After completing the risk assessment, the system automatically generates a defect distribution map with spatial coordinate markers. The distribution of defects on the electrode surface is presented intuitively through pseudo-color overlay or vector contour, realizing the visualization and traceability of quality status.

[0046] S4-2 Structured Inspection Report Generation: To meet the integration requirements of industrial information systems, this step standardizes and encapsulates the inspection results. The system serializes the defect categories, spatial coordinates, geometric parameters, statistical feature values, and inspection timestamps identified in step S3, generating a structured inspection report in JSON format. This report includes the product's unique identifier (ID), defect type and risk level, defect location coordinates, geometric dimensions and texture feature vectors, judgment threshold and judgment result, equipment number, and inspection time. The structured report can be directly integrated with MES, SPC, or quality traceability systems to achieve production batch correlation analysis and long-term quality statistical modeling.

[0047] S4-3 Online Evolution of Self-Supervised Model Based on Incremental Learning: To address the performance degradation of models caused by ambient light drift, mold wear, batch differences in graphite raw materials, or changes in surface texture during long-term production line operation, this application constructs a closed-loop self-supervised learning mechanism to achieve online adaptive evolution of the graphite surface topology reconstruction network model. During operation, the system continuously monitors the reconstruction residual distribution and decision confidence, automatically filtering low-confidence samples near the decision threshold and archiving them to a calibration sample library. Simultaneously, high-confidence qualified samples from the detection process are used as stable background samples and input into the reconstruction model in step S2 for small-step incremental fine-tuning. This allows the model to dynamically learn the physical background color distribution characteristics of the graphite surface under current production line conditions, thereby suppressing pseudo-anomaly responses caused by environmental changes and improving the ability to distinguish real, subtle defects. This maintains long-term detection threshold stability, forming an adaptive closed loop from detection evaluation to relearning, enabling the system to continuously evolve without frequent manual re-labeling and offline retraining, outputting an updated graphite surface topology reconstruction model.

[0048] Experimental verification and analysis: 1. Comprehensive evaluation of defect detection performance To comprehensively evaluate the performance of the proposed graphite bipolar plate surface defect identification method based on topological prior and self-supervised learning, this study constructed an evaluation dataset containing 200 real-world industrial graphite bipolar plate images. To establish a performance benchmark, all samples were manually labeled by professional quality inspectors for in-depth comparison with traditional supervised models. Unlike conventional approaches that rely on large amounts of labeled data, this method is based on a self-supervised reconstruction mechanism, aiming to achieve accurate identification of minute defects using only the topological distribution patterns of qualified samples.

[0049] like Figure 4 The results show that this invention, based on self-supervised training logic, achieves superior detection performance compared to conventional supervised models using only a dataset constructed with 200 qualified samples. In actual industrial production, defective samples are often extremely difficult to obtain, causing traditional supervised models to frequently fall into overfitting bottlenecks due to a lack of labeled data. This method, by executing the dynamic random mask generation logic in S1-2, transforms the original qualified images into self-supervised training pairs, thereby forcing the model to autonomously extract the periodic skeleton patterns of graphite flow channels without manual annotation.

[0050] This mechanism of predicting completeness from incompleteness significantly improves the model's generalization ability with small sample sizes. Comparative analysis shows that, with the same data scale, this scheme can effectively isolate complex flow channel texture interference, achieving precise decoupling of minute chipping, scratches, and pinholes. Furthermore, by combining the background statistical features extracted in S1-3 as auxiliary constraints, the model can maintain an extremely low false alarm rate even with very small sample sizes, effectively distinguishing between natural micropore noise in graphite and functional physical damage. This fully verifies the technological advancement and practical value of this invention in solving the challenge of small-sample industrial inspection.

[0051] 2. Analysis of the impact of data scale on detection performance To further verify the ability of this invention to address the problem of sample scarcity in industrial applications, this experiment compares the detection accuracy evolution trends of the proposed solution with those of traditional supervised models under different training sample sizes. Figure 5 As shown.

[0052] Experimental data shows that when the available training samples are at an extremely low level, traditional supervised models exhibit significant underfitting due to their heavy reliance on the accumulation of defect feature annotations. Their detection accuracy fluctuates only between 32.0% and 48.5%, failing to meet basic industrial quality inspection needs. In contrast, the proposed solution achieves a detection accuracy of 78.5% with only 20 qualified samples; and the accuracy rapidly increases to 85.2% when the sample size is increased to 50. This fully demonstrates that the self-supervised reconstruction mechanism based on topological priors proposed in this invention can effectively utilize the structural patterns of qualified products to construct judgment criteria, eliminating the dependence on a large number of defect-annotated samples.

[0053] 3. Visualized verification of the entire defect identification process Figure 6 This demonstrates the effectiveness of the proposed solution from input image to output. The reconstructed reference image generated by the mask autoencoder can forcibly repair the input defective region into a perfect structure conforming to topological rules. Utilizing multi-criteria fusion of residual mapping, as shown in the residual map and binarized mask, the system successfully removes complex background textures, achieves high-contrast extraction of defect signals, and finally automatically decouples defect categories based on geometric features.

[0054] 4. System inference efficiency and real-time performance verification This experiment compared the single-frame inference time at different resolutions to verify the system's online detection potential. Figure 7 At low resolutions, the inference time of this scheme is significantly faster than traditional spatial domain matching and conventional supervised models. Even in high-resolution scenes of 1024×1024, the inference time of this scheme is still controlled within 38ms. Experiments show that, thanks to the selective processing mechanism for non-mask tokens in this invention, the computational efficiency is improved by more than 10 times compared to conventional convolutional reconstruction, fully meeting the needs of high-speed online detection in industrial production lines.

[0055] 5. System module ablation experiment evaluation By performing ablation experiments on the core modules of the system, the contribution of each technical step to the final performance was quantitatively analyzed. (See [link to relevant documentation]). Figure 8 Experiments show that the detection rate of the basic reconstruction model is only 72.5%. After introducing the dynamic mask logic in step S1-2, the detection rate increases to 81.3%, demonstrating the crucial role of the incomplete prediction task in learning the flow channel topology features. After introducing the background statistical feature association in step S1-3, the system's ability to distinguish under noise interference is further enhanced, and the detection rate of the complete scheme finally reaches 89.5%. This fully verifies the synergistic effect between the various technical steps of this invention and its robustness in complex physical environments.

[0056] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0057] While the specific embodiments of the present invention have been described above, they are not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for identifying surface defects in a graphite bipolar plate, characterized in that, Includes the following processes: S1. Take a large-format image of a defect-free electrode plate using an industrial camera, perform standardized image segmentation and receptive field alignment, and obtain standardized sub-image blocks. S2, execute dynamic random mask generation logic injection to construct a polymorphic structured mask image block simulating defect states; Multidimensional data augmentation and physical property association were performed again to obtain background statistical feature vectors characterizing the micro-texture and local contrast variance of the graphite plate surface; A training dataset is constructed by using polymorphic structured mask image patches as model input, standardized sub-image patches as reconstruction targets, and background statistical feature vectors as auxiliary attribute labels for samples. S3 constructs and self-supervises trains a graphite surface topology reconstruction network model based on a mask autoencoder, including a flow channel topology feature encoding network that outputs a global topology feature token and a conditional embedding vector for the flow channel; a lightweight pixel decoding and reconstruction reference image generation network that generates a reconstruction reference image that is perfectly aligned with the coordinates of the input image; and a multi-scale discrimination network and a physical consistency verification network that distinguish between genuine and fake reconstruction reference images based on connectivity, texture smoothness, and physical properties at different feature resolutions. S4 acquires the surface image of the graphite bipolar plate to be tested, and after standardization, it is input into the graphite surface topology reconstruction network model in S3. The model outputs the reconstruction reference image, extracts the deviation features between the two, and decouples the defect categories.

2. The method for identifying surface defects in a graphite bipolar plate as described in claim 1, characterized in that: The specific process of injecting the dynamic random mask generation logic is as follows: an adaptive mask generation operator with defect morphology awareness is used. By using a pre-defined defect prior probability distribution, various mask topologies, including linear, clustered, and diffuse point distributions, are dynamically generated to establish a topological simulation of real physical defects. Linear masks, with extremely high aspect ratios and random orientations, simulate microcracks and scratches distributed across the flow channel, forcing the network to learn the ability to recover texture continuity across regions. Edge cluster masks, using irregular polygonal masks with high fractal dimensions and constraining their spatial position to abrupt gradient changes on the flow channel wall, simulate edge chipping and damage, enhancing the model's accuracy in reconstructing complex geometric boundaries. Diffuse point masks, using Gaussian-distributed fine spot masks, simulate the natural sand holes, pores, and surface contamination of graphite materials. Operators... By logically combining sub-modules, a composite mask layer is generated in the image space, and the pixel values ​​within the area are either zeroed out or filled with noise based on the statistical distribution of the graphite background texture. By constructing an input generation method with mask constraints, polymorphic structured mask image blocks are output.

3. The method for identifying surface defects in a graphite bipolar plate as described in claim 1, characterized in that: The process of performing multidimensional data enhancement and physical property association is as follows: First, for the high light absorption characteristics and non-uniform reflectivity of graphite bipolar plates, adaptive contrast stretching and Gamma nonlinear mapping are performed to simulate the changes in micro-roughness of the graphite plate surface caused by differences in oxidation degree and tool wear, as well as the light intensity drift during online detection; Second, gray-level co-occurrence matrix and local binary mode algorithm are introduced to quantitatively extract the micro-texture roughness, energy distribution and local contrast variance of the graphite plate surface as background statistical feature vectors.

4. The method for identifying surface defects in a graphite bipolar plate as described in claim 1, characterized in that: The channel topology feature encoding network is an encoder based on the Vision Transformer architecture. First, the input image is segmented into a series of non-overlapping image patches and linearly embedded to transform them into a token sequence. Second, the background statistical feature vector is used as a conditional embedding and concatenated or fused with the image token sequence to inject prior material physical properties into the Transformer. Third, the long-range dependencies between image patch tokens are calculated through a multi-layer self-attention mechanism to capture the periodic distribution and global geometry of the channel across the mask region. The encoder only processes visible (i.e., unmasked) image patch tokens and outputs the channel global topology feature tokens and conditional embedding vectors.

5. The method for identifying surface defects in a graphite bipolar plate as described in claim 1, characterized in that: The lightweight pixel decoding and reconstruction reference image generation network is a lightweight pixel decoder responsible for restoring the low-dimensional semantic features and masked token sequences output by the channel topology feature encoding network into a high-resolution pixel space image. The decoder visualizes the channel topology structure and background texture prior knowledge learned by the encoder through multi-layer upsampling and convolution operations. When the decoder reconstructs the masked area, it repairs it based on the distribution pattern of qualified product channels that it has learned. It reconstructs the defective areas that may exist in the input image into a defect-free structure that conforms to the periodic characteristics of graphite channels, and finally generates a reconstructed reference image that is completely aligned with the coordinates of the input image.

6. The method for identifying surface defects in a graphite bipolar plate as described in claim 1, characterized in that: The multi-scale discriminant network and physical consistency verification network adopt a multi-level feature pyramid structure to distinguish between the reconstructed reference image and the real qualified sample image at the full image scale, local flow channel scale, and micro-texture scale, respectively, in order to optimize the visual quality of the reconstructed image. At the same time, a background statistical feature regression head is embedded at the end of the discriminant network to predict the background texture parameters of the reconstructed reference image, including GLCM roughness, and compare it with the extracted original background statistical feature vector to ensure that the generator network maintains the intrinsic physical properties of the material during the reconstruction process.

7. The method for identifying surface defects in a graphite bipolar plate as described in claim 1, characterized in that: In step S4, firstly, a reconstruction residual map calculation based on multi-criteria fusion is performed. By comparing the differences in pixel grayscale and local structure between the real observed image and the reconstruction reference image, a composite residual map containing significant defect features is generated. Secondly, defect saliency enhancement and adaptive binarization segmentation are performed. The Laplacian operator is used to sharpen the edges and adaptive threshold segmentation is combined to filter out background noise, generating a high-quality binarized defect mask. Finally, multidimensional geometric feature quantization and category decoupling of defects are performed. The binarized defect connected components are mathematically represented with high precision, and their topological features are extracted to achieve automatic classification of defect categories. For each identified target, its equivalent area, aspect ratio of the bounding rectangle, roundness factor, contour fractal dimension, and skew angle relative to the flow channel direction are extracted, and multidimensional quantized feature parameters of defects are output. Defect classification is then performed using the quantization indicators.

8. The method for identifying surface defects in a graphite bipolar plate as described in claim 7, characterized in that: The calculation of the reconstructed residual map based on multi-criteria fusion specifically involves: on the one hand, calculating the absolute value of the pixel-level grayscale residual. This is used to lock in chipped edges and scratches with obvious contrast; on the other hand, the local structural similarity sliding window is used to calculate the structural deviation map to capture the subtle geometric deformation of the flow channel edge; the two types of residual information are fused at multiple scales to generate a composite residual mapping map containing significant defect features.

9. The method for identifying surface defects in a graphite bipolar plate as described in claim 7, characterized in that: The specific method of using quantitative indicators to classify defects is as follows: when the extension length of a target exceeds the preset flow channel period width and the shape factor shows a slender shape, it is judged as a scratch; when the fractal dimension of a target exceeds the reference value of the smooth edge and its spatial position is located on the side wall of the flow channel edge, it is judged as a chipped edge. Such defects usually mean that the mechanical strength of the graphite structure at that location has been damaged; when the roundness of the target is extremely high, the area is small and it is randomly distributed in the flow channel plane area, it is judged as a pinhole.

10. The method for identifying surface defects in a graphite bipolar plate as described in claim 7, characterized in that, The process also includes the following steps: Based on the obtained multidimensional quantitative characteristic parameters of defects, the defect area, grayscale residual intensity, morphological irregularity, fractal dimension, and spatial topological location are input into a preset quality threshold model, and risk classification rules are established in conjunction with the flow channel structure design parameters. For Level 1 risk, defined as structural damage, if the detected defect crosses the flow channel ridge line, i.e., the scratch exceeds the set threshold, or the chipping depth exceeds the flow channel wall thickness, it is judged as a structural damage defect, the system triggers a scrap-level warning signal, and marks it as a non-circulating product in the MES system. Level 2 risk is defined as process anomaly. If a non-penetrating defect such as diffuse pinholes is detected, and its total area ratio exceeds the set threshold, it is judged as a process fluctuation defect, and the system marks the sample as requiring manual review level warning for process quality traceability and equipment status diagnosis.

Images