Surface defect detection method and system for silicon carbide coating graphite plate

Abstract

The invention relates to the technical field of coating materials and detection, in particular to a surface defect detection method and system for a silicon carbide coating graphite disc. Performing multi-dimensional image data acquisition on the surface of the graphite disc; noise removal is carried out through an adaptive multi-scale convolutional neural network, and detail and edge information is retained in combination with image gradient constraints; high-frequency details are recovered by using an image enhancement technology, geometric correction is performed through a deep learning spatial transformation network, and data consistency is ensured; an authentication authority sequence is introduced, so that the authenticity and the generality of image data are ensured; an improved YOLOv9 model is adopted, an attention mechanism is introduced to enhance the extraction capability of target features, a Soft-NMS algorithm is combined to optimize a positioning frame, the overlapping detection problem is solved, and a traditional convolutional layer is replaced with an SCM module; a detection result is fed back in real time, a detailed detection report is generated, and decision support is provided. The surface defect detection precision and efficiency of the silicon carbide coating graphite disc are improved.

Description

Technical Field

[0001] The present invention relates to the technical fields of coating materials and detection technologies, and particularly relates to a method and a system for detecting surface defects of a silicon carbide coated graphite disk. Background Art

[0002] With the continuous development of advanced manufacturing, the requirements for material properties are becoming increasingly stringent. Especially in high-precision fields such as aerospace and semiconductors, silicon carbide coated graphite disks are widely used as important high-performance materials. However, due to their complex surface structures and high-performance requirements, various surface defects are likely to occur during the manufacturing and use of silicon carbide coated graphite disks. If these defects are not detected in time, they will seriously affect their performance and even lead to the failure of the entire system.

[0003] A Chinese invention patent with the publication number CN117890304A discloses a surface defect detection system for steel plates, including a conveyor main body, a computer, a detection adjustment component, a position correction component, a defect sample collection module, a sample screening module, a sample marking module, a sample set production module, a sample training module, a model establishment module, an environment configuration module, a network model module, and a model testing module. This invention can automatically record and control the surface quality of steel plates efficiently and with high quality, reduce the labor intensity of personnel, and improve the accuracy and efficiency of surface quality detection. By using the detection adjustment component, when the detection device is used to detect steel plates with different thicknesses, the detection device can be adjusted in angle so that the imaging edge of the detection device is aligned with the edge of the steel plate, enabling the entire steel plate to be within the imaging range, thereby recording the details of the steel plate surface and achieving real-time detection, improving the defect detection effect of the steel plate.

[0004] Most of the existing defect detection methods rely on manual inspection or traditional machine vision technologies, and there are problems such as low detection efficiency, insufficient accuracy, and inability to adapt to complex environments, which cannot meet the requirements in actual production. Therefore, how to efficiently and accurately detect the surface defects of silicon carbide coated graphite disks has become a technical problem to be solved urgently. Summary of the Invention

[0005] The purpose of the present invention is to propose a method and a system for detecting surface defects of a silicon carbide coated graphite disk in view of the problems in the background art.

[0006] The technical solution of the present invention: A method for detecting surface defects of a silicon carbide coated graphite disk includes the following specific implementation steps:

[0007] S1. Collect image data of the surface of the silicon carbide coated graphite disk;

[0008] S2. Dynamically adjust the filtering scale through an adaptive multi-scale convolutional neural network for targeted denoising, combine gradient constraints to retain edge information, use multiple enhanced convolutions to restore high-frequency details, optimize the denoising strategy based on adaptive learning, and use a spatial transformation network to achieve geometric correction. Construct an authentication permission sequence through a permission matching code and a sequence generation factor;

[0009] S3. Ensure the authenticity and comprehensiveness of image data by verifying the authentication permission sequence, construct a defect recognition model, based on the YOLOv9 network, embed the SimAM attention mechanism to suppress background interference for enhanced feature extraction, use Soft-NMS combined with a Gaussian function to optimize the confidence of overlapping prediction boxes, construct an SCM module to integrate Sobel edge detection and pooling branch features, fuse spatial and edge information through 3D group convolutions, improve the detection accuracy of surface defects of silicon carbide-coated graphite disks in complex scenarios, and finally output defect type and location information;

[0010] S4. Real-time feedback the detection results and repair suggestions to the operation interface, trigger an alarm for serious defects and recommend production suspension, and automatically generate a detection report including defect type, location, severity, and repair suggestions.

[0011] Preferably, the process of dynamically adjusting the filtering scale through an adaptive multi-scale convolutional neural network for targeted denoising is as follows:

[0012] S21. Based on the noise removal method of the adaptive multi-scale convolutional neural network, dynamically adjust the size of the filter according to the local features of the image to remove different types of noise:

[0013]

[0014] In the formula, I denoised represents the denoised image; I i represents the local area in the original image, the i-th layer feature map; w i represents the weighting coefficient of the i-th layer; f CNN (I i , W i ) represents the denoising operation of the i-th layer in the convolutional neural network; W i represents the convolutional kernel of this layer; N represents the number of layers used in the multi-scale convolutional neural network;

[0015] S22. During the noise removal process, adopt an image gradient constraint strategy, combine noise removal with gradient information to ensure the retention of edge information:

[0016]

[0017] In the formula, represents the gradient value of the denoised image at the (x, y) position, that is, the edge strength at this position; and respectively represent the gradients of the denoised image in the x and y directions;

[0018] S23. On the basis of denoising, combined with image enhancement technology, through high-frequency information restoration, construct a multiple enhanced convolution operation model:

[0019] I enhanced (x,y) = I denoised (x,y) + α·[f edge (x,y) - f denoised (x,y)];

[0020] In the formula, I enhanced (x,y) represents the pixel value of the image after enhancement processing; f edge (x,y) represents the image after edge enhancement operation; f denoised (x,y) represents the denoised image; α represents the enhancement coefficient, which is used to control the intensity of enhancement;

[0021] S24. Adopt a noise removal method based on adaptive learning. By training a deep neural network, it can automatically select an appropriate denoising strategy according to the noise type of the input image and can adaptively adjust the denoising parameters in different noise environments:

[0022]

[0023] In the formula, represents the final denoised image; f adaptive represents the adaptive denoising model; θ represents the learning parameter of the model; f j (I,θ j ) represents the denoising operation adjusted according to the noise type; w' j represents the weight of each denoising method; K represents the number of denoising methods.

[0024] Preferably, the generation process of the authentication permission sequence is as follows:

[0025] S31. Select a random number s ∈ Z p , and calculate the permission matching code C = g s ;

[0026] In the formula, Z p represents the integer ring modulo p; g represents a generator of the predefined group G; p is a prime number and is the order of the group G;

[0027] S32. Convert the image data Image into a binary string sequence data;

[0028] S33. Calculate the sequence generation factor f = H(data) × (fg) + s mod p;

[0029] In the formula, fg represents a predefined auxiliary sequence generation factor, and f g ∈Z p ; H is a predefined hash function;

[0030] S34. Generate the authentication permission sequence GS = (permission matching code C, sequence generation factor f).

[0031] Preferably, the verification process for ensuring the authenticity and comprehensiveness of image data by verifying the authentication permission sequence is as follows:

[0032] S41. Convert the received image data Image into a binary sequence Idata;

[0033] S42. Calculate the sequence verification code AS = g f × (fa) -H(Idata) ;

[0034] Among them, fa represents a predefined permission parsing factor, fa ∈ G, fa = g fg ; H represents a predefined hash function;

[0035] S43. If AS = C, the verification passes, indicating that the received image data is authentic and comprehensive.

[0036] Preferably, the SimAM attention mechanism calculates the attention weight using its energy function, reduces the interference of the remaining information in the background on the defect features, and highlights the key features of the defects. The specific calculation process is as follows:

[0037]

[0038]

[0039] In the formula, μ represents the average value of each channel in the input feature map, that is, the mean value of all neuron values, used to measure the overall brightness or intensity; σ 2 represents the variance of each channel in the input smoke feature map; Q represents the variance of each channel in the input smoke feature map; x i represents the value of the i-th neuron in the input feature map; E represents the energy function of each neuron, and the lower the energy, the more important the neuron; λ represents a hyperparameter used to adjust the sensitivity and stability of the energy function; t represents the value of the target neuron, that is, the value of the currently concerned neuron; represents the enhanced feature value; sigmoid() represents the sigmoid function; X represents the original input feature value; Represents the element-wise multiplication operation, that is, applying the attention weights output by the sigmoid function to the original feature map.

[0040] Preferably, the SCM module is used to replace the first two convolutional layers of the YOLOv9 backbone network.

[0041] Preferably, the working process of the SCM module is as follows:

[0042] S71. Input image processing: The input image first passes through the Sobel Conv branch to extract edge features. The Sobel filter generates an edge feature map by detecting sharp changes in intensity in the image. At the same time, the input image also passes through the pooling branch to extract spatial information;

[0043] S72. Feature fusion: After being processed by the Sobel Conv branch and the pooling branch, the generated feature maps are integrated. Through 3D group convolution technology, the features of the two branches are effectively fused to form a comprehensive feature map containing edge and spatial information;

[0044] S73. Subsequent processing: The generated comprehensive feature map will be passed as input to the subsequent network structure of YOLOv9 for detection.

[0045] The technical solution of the present invention: A surface defect detection system for a silicon carbide coated graphite disk, which is used to execute the above-mentioned surface defect detection method for a silicon carbide coated graphite disk, includes:

[0046] A surface image acquisition module for acquiring the surface image of the silicon carbide coated graphite disk;

[0047] A data preprocessing module for performing noise removal, image enhancement, and geometric correction processing on the collected original image data;

[0048] A defect recognition module for constructing a defect recognition model to perform spatio-temporal feature modeling and classification prediction of defects;

[0049] A result output and decision support module for providing real-time defect detection results to the operator and generating a report.

[0050] Compared with the prior art, the above technical solution of the present invention has the following beneficial technical effects:

[0051] The present invention designs a surface defect detection method and system for a silicon carbide coated graphite disk, which solves the problems of insufficient accuracy, large noise interference, and poor adaptability to complex scenarios existing in the existing defect detection methods, significantly improves the effect of surface defect detection of the silicon carbide coated graphite disk, and has strong industrial application value. Specifically:

[0052] (1) Improve the accuracy of defect detection: By introducing the improved YOLOv9 model, the attention mechanism module SimAM, and the Soft-NMS algorithm, the present invention can significantly enhance the accuracy and precision of defect detection. The YOLOv9 model effectively enhances the network's ability to extract key features by integrating the attention mechanism, reduces background interference, and thus can accurately identify the surface defects of the graphite disk with a silicon carbide coating in complex scenarios. At the same time, the Soft-NMS algorithm effectively solves the problem of missed detections in the traditional NMS method when dealing with high intersection over union (IoU) situations, further improving the accuracy of the detection box;

[0053] (2) Optimize the quality of image data: Using the adaptive multi-scale convolutional neural network (AM-CNN) combined with the image gradient constraint strategy for noise removal can retain the edge information and detail features of the image while removing noise, avoiding the loss of important details in traditional noise removal methods. In addition, the multiple enhancement convolution operation (MEC) further improves the sharpness and detail recovery effect of the image by enhancing high-frequency information, providing clearer image data for subsequent defect recognition;

[0054] (3) Ensure data consistency and reliability: Through the spatial transformation network (STN) of deep learning for geometric correction and normalization of images, effectively solve the problems of perspective distortion and scale inconsistency that may occur during the multi-sensor data acquisition process, so as to ensure that the subsequent processing algorithms can be analyzed under unified scales and consistent data, improving the stability and reliability of the overall system;

[0055] (4) Enhance the model's detection ability for complex scenarios and small targets: The SCM module introduced in the present invention improves the model's ability to extract image edge features and spatial information through the feature fusion of the Sobel Conv branch and the pooling branch. Especially for the detection of complex scenarios and small targets, the SCM module can effectively capture the detail information in the image, enhance the recognition ability of tiny defects, and thus improve the comprehensiveness of detection;

[0056] (5) High-efficiency authentication of the authenticity and comprehensiveness of image data: The present invention ensures the authenticity and integrity of image data by introducing an authentication permission sequence, preventing data tampering and abuse. Using the hash algorithm to verify data enhances the security of data exchange, supports data auditing and traceability, and ensures data compliance during the production process. This technology effectively guarantees the security and reliability of image data, improves the overall credibility of the system, and meets the high-standard data protection requirements. Brief Description of the Drawings

[0057] Figure 1 It is the system architecture diagram of a surface defect detection system for a graphite disk with a silicon carbide coating proposed by the present invention;

[0058] Figure 2 The network structure diagram of the SCM module constructed for the present invention;

[0059] Figure 3 The network structure diagram of the Sobel Conv branch. Detailed implementation manners

[0060] Example 1. As Figure 1 shown, a surface defect detection system for a silicon carbide coated graphite disk proposed by the present invention includes: a surface image acquisition module, a data preprocessing module, a defect recognition module, and a result output and decision support module.

[0061] The surface image acquisition module is configured with high-precision sensors for data acquisition (including but not limited to high-definition cameras, laser sensors) to acquire the surface image of the silicon carbide coated graphite disk;

[0062] The data preprocessing module performs noise removal, image enhancement, and geometric correction processing on the acquired original image or data;

[0063] The defect recognition module constructs a defect recognition model to perform spatio-temporal feature modeling and classification prediction of defects;

[0064] The result output and decision support module is used to provide real-time defect detection results to the operator and generate a report to support subsequent decision-making and processing.

[0065] Example 2. A surface defect detection method for a silicon carbide coated graphite disk proposed by the present invention is applied to a surface defect detection system for a silicon carbide coated graphite disk proposed in Example 1. The specific implementation steps are as follows:

[0066] S1. The surface image acquisition module uses a plurality of high-precision sensors (including but not limited to high-definition cameras, laser scanners, infrared imaging devices) to perform omnidirectional real-time scanning on the surface of the silicon carbide coated graphite disk to obtain multi-dimensional image data, including but not limited to texture and structure, and thereby obtain high-resolution image data.

[0067] S2. The data preprocessing module performs noise removal, image enhancement, and geometric correction processing on the acquired original image or data. The specific implementation process is as follows:

[0068] S21. Based on the image noise removal method, by combining the adaptive multi-scale convolutional neural network (AM-CNN) with the image enhancement model, the efficiency of noise removal and the ability to retain image details are improved. Adaptive multi-scale CNN, gradient constraint, image enhancement, and adaptive learning techniques are introduced to effectively solve the contradiction between noise removal and image detail retention, and improve the effect and efficiency of noise removal. The specific implementation process is as follows:

[0069] S2101. A noise removal method based on an Adaptive Multi-Scale Convolutional Neural Network (AM-CNN) dynamically adjusts the size of the filter according to the local features of the image, so as to better remove different types of noise:

[0070]

[0071] In the formula, I denoised represents the denoised image; I i represents the local region in the original image, the feature map of the i-th layer; w i represents the weighting coefficient of the i-th layer, which is used to dynamically adjust the influence of different scale features; f CNN (I i ,W i ) represents the denoising operation of the i-th layer in the convolutional neural network; W i represents the convolutional kernel of this layer; N represents the number of layers used in the multi-scale convolutional neural network;

[0072] It should be noted that since different noise types (including but not limited to Gaussian noise, salt and pepper noise) will show different frequency characteristics in the image, the adaptive multi-scale convolutional neural network adjusts the filtering scale of each layer according to the local information of the image, so as to achieve targeted removal of different noise types. In addition, the weighting coefficient w i is adaptively adjusted according to the change of the local features of the image, avoiding the limitations of the fixed filter in the traditional method;

[0073] S2102. During the noise removal process, in addition to considering the pixel values of the image, the edge information and gradient information of the image are also crucial. If the image is directly smoothed, important details and edge information may be lost. Therefore, an image gradient constraint strategy is adopted to combine noise removal with gradient information to ensure the retention of edge information:

[0074]

[0075] In the formula, represents the gradient value of the denoised image at the position (x, y), that is, the edge strength at this position; and respectively represent the gradients of the denoised image in the x and y directions;

[0076] It should be noted that the gradient calculation of the image is used to evaluate the edge information in the image. For noise removal, it is crucial to retain the edge features. By constraining the gradient change during the noise removal process, it can ensure that the details and edges in the image are effectively retained while removing the noise;

[0077] S2103. On the basis of denoising, combined with image enhancement technology, through high-frequency information restoration, further improve the denoising effect, construct a multi-enhancement convolution operation model (Multi-enhancement Convolution, MEC), and enhance the sharpness and detail restoration effect of the image by enhancing the high-frequency information in the image:

[0078] I enhanced (x,y) = I denoised (x,y) + α·[f edge (x,y) - f denoised (x,y)];

[0079] In the formula, I enhanced (x,y) represents the pixel value of the image after enhancement processing; f edge (x,y) represents the image after edge enhancement operation; f denoised (x,y) represents the image after denoising; α represents the enhancement coefficient, which is used to control the intensity of enhancement;

[0080] It should be noted that the multi-enhancement convolution operation is an image sharpening strategy based on deep learning. By combining the edge information in the image with the denoised image, the sharpness of the image is adjusted using the enhancement coefficient α to ensure the removal of noise while restoring the high-frequency details of the image, thereby effectively removing most of the noise and restoring the clarity and details of the image;

[0081] S2104. To further optimize the denoising effect, an adaptive learning-based noise removal method is adopted. By training a deep neural network, it can automatically select an appropriate denoising strategy according to the noise type of the input image and can adaptively adjust the denoising parameters in different noise environments:

[0082]

[0083] In the formula, represents the final denoised image; f adaptive represents the adaptive denoising model; θ represents the learning parameter of the model; f j (I,θ j ) represents the denoising operation adjusted according to the noise type; w j represents the weight of each denoising method; K represents the number of denoising methods;

[0084] It should be noted that this adaptive learning method trains on samples of different noise types, enabling the denoising network to select the most appropriate denoising method according to the characteristics of the input image. This method can be flexibly adjusted for different noise types, improving the accuracy of the denoising effect while ensuring the retention of image details. The optimization of the adaptive model enables the system to respond in real time in a dynamic environment and adapt to various complex noise patterns;

[0085] S22. The multi-sensor data collected may be distorted. Especially when images are acquired from different perspectives and different sensors, it is easy to cause spatial inconsistencies. Therefore, geometric correction must be performed to ensure the consistency of the image data. In addition, to ensure that subsequent algorithms can process data of a unified scale, the collected data is standardized:

[0086] Construct a spatial transformation network (Spatial Transformer Network, STN) based on deep learning to perform spatial transformation learning, automatically learning the geometric transformation relationship between images and sensor data. STN can perform flexible geometric transformations on images (including but not limited to rotation, translation, and scaling), thereby achieving efficient and accurate geometric correction;

[0087] S23. To ensure the authenticity and comprehensiveness of the image data, an authentication permission sequence is generated for the processed image data, and the generation process is as follows:

[0088] S2301. Select a random number s ∈ Z p , and calculate the permission matching code C = g s ;

[0089] In the formula, Z p represents the integer ring modulo p; g represents a generator of a predefined group G; p is a prime number and is the order of the group G;

[0090] S2302. Convert the image data Image into a binary string sequence data;

[0091] S2303. Calculate the sequence generation factor f = H(data) × (fg) + s mod p;

[0092] In the formula, fg represents a predefined auxiliary sequence generation factor, f g ∈ Z p ; H is a predefined hash function;

[0093] S2304. Generate the authentication permission sequence GS = (permission matching code C, sequence generation factor f);

[0094] S24. Transmit the {authentication permission sequence GS = (permission matching code C, sequence generation factor f), image data Image} to the defect recognition module.

[0095] S3. The defect recognition module constructs a defect recognition model, based on the defect detection algorithm YOLO-TS of improved YOLOv9 (You Only Look Once Version 9). YOLO-TS strengthens the defect feature extraction ability of the backbone network by introducing the attention mechanism module SimAM, and uses the Soft-NMS algorithm to avoid the possibility of missing detections when filtering overlapping bounding boxes. And it constructs the SCM module to replace the first two convolutional layers. Its specific structure is as follows:

[0096] S31. Receive the {authentication permission sequence GS = (permission matching code C, sequence generation factor f), image data Image}, and verify the authenticity and integrity of the image data. The verification process is as follows:

[0097] S3101. Convert the received image data Image into a binary sequence Idata;

[0098] S3102. Calculate the sequence check code AS = g f ×(fa) -H(Idata) ;

[0099] where fa represents the predefined permission parsing factor, fa ∈ G, fa = g fg ;

[0100] S3103. If AS = C, the verification passes, indicating that the received image data has authenticity and integrity;

[0101] S32. To strengthen the defect feature extraction ability of the network structure, embed the attention mechanism module SimAM into the network structure of YOLOv9. SimAM calculates the attention weight using its energy function, so that the active neurons used to identify defects will inhibit the neurons with scarce surrounding information, reducing the interference of the remaining information in the background on the defect features, thereby highlighting the key features of the defects and enhancing the defect feature extraction ability. The specific calculation process is as follows:

[0102]

[0103] In the formula, μ represents the average value of each channel in the input feature map, that is, the mean value of all neuron values, used to measure the overall brightness or intensity; σ 2 represents the variance of each channel in the input smoke feature map; Q represents the variance of each channel in the input smoke feature map; x irepresents the value of the \(i\)-th neuron in the input feature map; \(E\) represents the energy function of each neuron, and the lower the energy, the more important the neuron; \(\lambda\) represents a hyperparameter used to adjust the sensitivity and stability of the energy function; \(t\) represents the value of the target neuron, that is, the value of the currently concerned neuron; represents the enhanced feature value; sigmoid() represents the sigmoid function, which is used to limit the input value between (0,1) to control the output range; \(X\) represents the original input feature value; represents the element-wise multiplication operation (Hadamard product), that is, applying the attention weights output by the sigmoid function to the original feature map;

[0104] It should be noted that SimAM is a lightweight general attention module. Different from existing channel / space attention modules, it does not increase the number of network parameters, has the characteristics of plug-and-play, and can make the refined features better focus on the main target while avoiding additional parameter calculations;

[0105] S33. Introduce Soft-NMS. For prediction boxes with a high intersection over union (IoU), Soft-NMS does not directly delete them, but sets a penalty function to reduce their confidence. The higher the overlap degree of the prediction box, the more severely its score is reduced. To solve the problem of function continuity, the Gaussian function is selected as the weight function to implement the score reset function of the Soft-NMS algorithm:

[0106]

[0107] In the formula, \(M\) represents the prediction box with the highest confidence; \(b\) i represents the \(i\)-th prediction box; \(s\) i represents the confidence of the \(i\)-th prediction box; \(\sigma'\) represents the variance of the Gaussian function; \(N\) t represents the set threshold; \(e\) represents the natural constant;

[0108] S34. Construct the SCM module, as Figure 2 shown, which is used to replace the first two convolutional layers of the backbone network. By effectively extracting the edge features and spatial information in the image, it improves the model's detection ability for complex scenes and small targets. Combining the Sobel convolution (Sobel Conv) branch (as Figure 3 shown) and the pooling branch with the feature fusion module aims to comprehensively capture the detailed information of the image;

[0109] Among them, Conv represents the convolution operation; MaxPool represents the max pooling operation; Concat represents the concatenation operation, which fuses the features of the Sobel Conv branch and the pooling branch to form a comprehensive feature map containing rich information. In this embodiment, 3D group convolution technology is adopted, which can efficiently integrate the features from the two branches. Furthermore, the model can utilize both edge features and spatial information simultaneously, thereby enhancing the detection ability for surface defects of silicon carbide coated graphite discs; Input represents the input data; The Sobel Conv branch uses the Sobel filter for edge detection and realizes fast calculation through convolution to extract the edge features in the image. The Sobel Conv branch generates an edge feature map by performing a convolution operation on the input image, highlighting the regions with drastic intensity changes in the image, thereby enhancing the sensitivity of the model to detailed features; Sobel-y is used to calculate the gradient of the image in the vertical direction; Sobel-x is used to calculate the gradient of the image in the horizontal direction;

[0110] The working process of the SCM module is as follows:

[0111] (1) Input image processing: The input image first passes through the Sobel convolution (Sobel Conv) branch to extract edge features. The Sobel filter generates an edge feature map by detecting the drastic changes in intensity in the image. At the same time, the input image also passes through the pooling branch to extract spatial information. The pooling operation retains the most prominent features in the image through max pooling;

[0112] (2) Feature fusion: After being processed by the Sobel Conv branch and the pooling branch, the generated feature maps are integrated. This module effectively fuses the features of the two branches through 3D group convolution technology to form a comprehensive feature map containing edge and spatial information;

[0113] (3) Subsequent processing: The generated comprehensive feature map will be passed as input to the subsequent network structure of YOLOv9 for detection. This feature map not only contains rich edge information but also combines spatial context, providing comprehensive feature support for subsequent detection;

[0114] Accordingly, the constructed defect recognition model outputs defect information {defect type, defect location}.

[0115] S4. The result output and decision support module gives real-time feedback on the detected defect information {defect type, defect location} and outputs a defect detection report. The specific steps are as follows:

[0116] Real-time feedback: When the defect detection is completed, the detection results are immediately fed back to the operator through the interface and repair suggestions are provided when needed;

[0117] For example, if serious defects are found, an automatic alarm is triggered and it is recommended to suspend the production line;

[0118] Report generation: Automatically generate a detailed inspection report, including but not limited to defect type, location, severity, and repair suggestions. The report can be sent to relevant personnel via email, printing, or system sharing;

[0119] Decision support: Integrate with the material management system and the production line control system to achieve intelligent decision support;

[0120] For example, when a certain number or level of defects are detected, the system can automatically initiate a maintenance or repair process.

[0121] The embodiments of the present invention have been described in detail above in conjunction with the accompanying drawings. However, the present invention is not limited thereto. Various changes can be made without departing from the spirit of the present invention within the scope of knowledge possessed by those skilled in the art.

Claims

1. A method for detecting surface defects of a silicon carbide coated graphite disk, characterized in that: The specific implementation steps include the following: S1, collecting image data of the surface of the silicon carbide coated graphite disk; S2, dynamically adjust the filter scale through adaptive multi-scale convolutional neural network for targeted denoising, combine gradient constraints to retain edge information, use multiple enhanced convolutions to restore high-frequency details, optimize the denoising strategy based on adaptive learning, and use spatial transformation network to achieve geometric correction, and construct the authentication authority sequence through authority matching code and sequence generation factor; S3. Ensure the authenticity and comprehensiveness of image data by verifying the authentication permission sequence, build a defect recognition model, embed the SimAM attention mechanism based on the YOLOv9 network to suppress background interference to enhance feature extraction, use Soft-NMS combined with Gaussian function to optimize the confidence of overlapping prediction boxes, build an SCM module to integrate Sobel edge detection and pooling branch features, and fuse spatial and edge information through 3D group convolution to improve the surface defect detection accuracy of silicon carbide coated graphite disks in complex scenarios, and finally output defect type and location information; S4. By feeding back the test results and repair suggestions to the operation interface in real time, serious defects trigger alarms and suggest production suspension, and automatically generate a test report containing the defect type, location, severity and repair suggestions.

2. The surface defect detection method of a silicon carbide coated graphite disk according to claim 1, characterized in that: The process of adaptive multi-scale convolutional neural network dynamically adjusting the filter scale for targeted denoising is as follows: S21. A noise removal method based on an adaptive multi-scale convolutional neural network dynamically adjusts the size of the filter according to the local features of the image to remove different types of noise: In the formula, I denoised I represents the denoised image; i Represents the local area in the original image, the i-th layer feature map; w i represents the weight coefficient of the i-th layer; f CNN (I i ,W i ) represents the denoising operation of the i-th layer in the convolutional neural network; W i Represents the convolution kernel of this layer; N represents the number of layers used in the multi-scale convolutional neural network; S22. In the process of noise removal, the image gradient constraint strategy is adopted to combine noise removal with gradient information to ensure the retention of edge information: In the formula, Represents the gradient value of the denoised image at the (x, y) position, that is, the edge strength at that position; and Represent the gradients of the denoised image in the x and y directions respectively; S23. Based on denoising, combined with image enhancement technology, a multiple enhanced convolution operation model is constructed by restoring high-frequency information: I enhanced (x,y)=I denoised (x,y)+α·[f edge (x,y)-f denoised (x,y)]; In the formula, I enhanced (x, y) represents the pixel value of the image after enhancement; f edge (x, y) represents the image after edge enhancement operation; f denoised (x, y) represents the denoised image; α represents the enhancement coefficient, which is used to control the intensity of enhancement; S24, adopting a noise removal method based on adaptive learning, by training a deep neural network, so that it can automatically select the appropriate denoising strategy according to the noise type of the input image, and can adaptively adjust the denoising parameters under different noise environments: In the formula, represents the final denoised image; f adaptive represents the adaptive denoising model; θ represents the learning parameters of the model; f j (I,θ j ) represents the denoising operation adjusted according to the noise type; w' j represents the weight of each denoising method; K represents the number of denoising methods.

3. The surface defect detection method of a silicon carbide coated graphite disk according to claim 1, characterized in that: The generation process of the authentication permission sequence is as follows: S31, select a random number s∈Z p , calculate the permission matching code C = g s ; In the formula, Z p represents the ring of integers modulo p; g represents a generator of a predefined group G; p is a prime number and is the order of the group G; S32, converting the image data Image into a binary string sequence data; S33, calculate the sequence generation factor f = H (data) × (fg) + s mod p; Where fg represents the predefined auxiliary sequence generation factor, f g ∈Z p ; H is a predefined hash function; S34, generate an authentication authority sequence GS = (authority matching code C, sequence generation factor f).

4. The surface defect detection method of a silicon carbide coated graphite disk according to claim 3, characterized in that: The verification process to ensure the authenticity and comprehensiveness of image data by verifying the authentication permission sequence is as follows: S41, converting the received image data Image into a binary sequence Idata; S42, calculate sequence check code AS = g f ×(fa) -H(Idata) ; Where fa represents the predefined permission resolution factor, fa∈G, fa=g fg ; H represents a predefined hash function; S43. If AS=C, the verification is passed, indicating that the received image data is authentic and comprehensive.

5. The surface defect detection method of a silicon carbide coated graphite disk according to claim 1, characterized in that: The SimAM attention mechanism uses its energy function to calculate the attention weight, reduce the interference of other information in the background on the defect characteristics, and highlight the key features of the defect. The specific calculation process is as follows: Where μ represents the average value of each channel in the input feature map, that is, the mean of all neuron values, which is used to measure the overall brightness or intensity; σ 2 represents the variance of each channel in the input smoke feature map; Q represents the variance of each channel in the input smoke feature map; x i represents the value of the i-th neuron in the input feature map; E represents the energy function of each neuron. The lower the energy, the more important the neuron; λ represents a hyperparameter used to adjust the sensitivity and stability of the energy function; t represents the value of the target neuron, that is, the value of the neuron currently being focused on; represents the enhanced eigenvalue; sigmoid() represents the sigmoid function; X represents the original input eigenvalue; Represents an element-by-bit multiplication operation, which applies the attention weights output by the sigmoid function to the original feature map.

6. The surface defect detection method of a silicon carbide coated graphite disk according to claim 1, characterized in that: The SCM module is used to replace the first two convolutional layers of the YOLOv9 backbone network.

7. The surface defect detection method of a silicon carbide coated graphite disk according to claim 6, characterized in that: The workflow of the SCM module is as follows: S71, input image processing: The input image first passes through the Sobel Conv branch to extract edge features. The Sobel filter generates an edge feature map by detecting the sharp change of intensity in the image. At the same time, the input image also passes through the pooling branch to extract spatial information. S72, feature fusion: After being processed by the Sobel Conv branch and the pooling branch, the generated feature maps are integrated, and the features of the two branches are effectively fused through the 3D group convolution technology to form a comprehensive feature map containing edge and spatial information; S73, subsequent processing: The generated comprehensive feature map will be passed as input to the subsequent network structure of YOLOv9 for detection.

8. A surface defect detection system for a silicon carbide coated graphite disk, used to execute a surface defect detection method for a silicon carbide coated graphite disk according to any one of claims 1 to 7, characterized in that: include: A surface image acquisition module, used for acquiring the surface image of the silicon carbide coated graphite disk; A data preprocessing module is used to perform noise removal, image enhancement and geometric correction on the collected raw image data; Defect recognition module, used to build defect recognition model, model the spatiotemporal characteristics of defects and perform classification prediction; Result output and decision support module, used to provide real-time defect detection results to operators and generate reports.

Images