Method for detecting internal defects of ceramic glazed tile based on image recognition
By constructing a three-dimensional digital twin and a defect-cause knowledge graph, the problem of low efficiency in manual inspection of ceramic glazed tile defects has been solved, achieving high-precision defect detection and automated production optimization.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HUANGGANG QICHUN COUNTY XINTIANDI CERAMIC IND CO LTD
- Filing Date
- 2025-07-25
- Publication Date
- 2026-07-31
AI Technical Summary
Existing methods for detecting defects in ceramic glazed tiles rely on manual inspection, which is inefficient and easily affected by subjective factors. They also lack intelligent closed-loop management of diagnosis, feedback, and control, making it difficult to meet the demands of high-quality production.
By acquiring multimodal data to construct a three-dimensional digital twin, combining it with a three-dimensional segmentation network to identify internal defects, and using a defect-cause knowledge graph to perform physical causal reasoning, production process control instructions are generated to achieve closed-loop defect detection.
It achieves high-precision defect detection, accurately locates the cause of defects and generates automated control commands, thus improving the production quality and inspection efficiency of ceramic glazed tiles.
Smart Images
Figure CN121010551B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of industrial vision technology, specifically to a method for detecting internal defects in ceramic glazed tiles based on image recognition. Background Technology
[0002] In the production of ceramic glazed tiles, product quality control is of paramount importance, and various defect detection methods are typically deployed on the production line. However, existing detection methods still have many limitations and are insufficient to meet the demands of high-quality production.
[0003] Traditional quality control relies heavily on manual visual inspection, which is not only labor-intensive and inefficient, but also susceptible to subjective influences on test results, and fails to establish an intelligent closed-loop management system of inspection-feedback-control. Furthermore, current research on ceramic glazed tile defect detection technology is limited in terms of data depth, intelligent diagnosis of defect causes, and closed-loop linkage with production processes. Therefore, a systematic solution is needed that can integrate multimodal information, intelligently trace the root causes of defects, and automatically optimize production process parameters.
[0004] To address this, an image recognition-based method for detecting internal defects in ceramic glazed tiles is proposed. Summary of the Invention
[0005] The purpose of this invention is to provide an image recognition-based method for detecting internal defects in ceramic glazed tiles. This method acquires multimodal data including X-ray computed tomography sequences, surface thermal images, visible light images, and production process parameters. Based on the X-ray data, a three-dimensional digital twin is constructed, fusing surface image information, and a three-dimensional segmentation network is used to accurately identify the three-dimensional geometric features of internal and underglaze defects. Through a constructed defect-cause knowledge graph reasoning model, physical causal reasoning is performed on the geometric features of the defects, intelligently matching and identifying the physical causes of the defects. Combined with targeted correlation analysis and source tracing of the physical causes and production process parameters, the method accurately locates the production link leading to the defect and generates control instructions for optimizing production. This invention achieves a closed-loop defect detection system of "diagnosis-source tracing-control," enabling precise location of defect problems and providing automated decision support for optimizing production and improving product quality.
[0006] To achieve the above objectives, the present invention provides the following technical solution:
[0007] An image recognition-based method for detecting internal defects in ceramic glazed tiles includes:
[0008] S1. Acquire multimodal data of the ceramic glazed tile to be tested, the multimodal data including X-ray tomography sequence, surface thermal image sequence, visible light image and production process parameters;
[0009] S2. Construct a three-dimensional digital twin based on the scanning sequence, and register the thermal image sequence and visible light image onto the surface of the twin to obtain a fused twin; segment the fused twin according to a three-dimensional segmentation network, identify the three-dimensional geometric features of internal and underglaze defects, and obtain the first defect feature data;
[0010] S3. Construct a defect-cause knowledge graph reasoning model to perform physical causal reasoning on the first defect feature data, associate and match the three-dimensional geometric features with preset physical causes, and generate second defect feature data with additional physical cause identifiers.
[0011] S4. Based on the second defect feature data and the production process parameters, perform targeted production defect tracing, and perform targeted correlation analysis on the production process parameters according to the physical cause identifier to locate the cause of the defect and generate production process control parameters.
[0012] Preferably, the construction process of the three-dimensional digital twin is as follows: First, the X-ray tomography sequence is preprocessed to filter out noise and correct artifacts; second, the two-dimensional sequence image is reconstructed into a three-dimensional voxel grayscale data field using synchronous algebraic reconstruction technology; then, the glazed tile body is segmented from the data field using a three-dimensional region growing algorithm; finally, the segmented glazed tile body voxel data is converted into a three-dimensional surface mesh model using a traveling cube algorithm.
[0013] Preferably, the process for obtaining the fused twin is as follows:
[0014] The camera that acquires visible light images and the thermal imaging device that acquires surface thermal image sequences are jointly calibrated to determine the spatial transformation relationship between their coordinate systems.
[0015] The camera pose is solved by matching the visible light image with the local invariant feature descriptors on the 3D digital twin rendering image using the perspective n-point algorithm.
[0016] Using spatial transformation relationships and camera pose, the visible light image and surface thermal image sequence are used as textures and projected onto the surface of the twin.
[0017] Preferably, the three-dimensional geometric features are quantized into a set of descriptors, which include: the three-dimensional coordinates of the defect centroid, the volume of the defect, the size of the minimum circumscribed cuboid, the sphericity value used to distinguish between bubbles and cracks, the elongation and flattening rates describing the defect morphology, and the principal axis unit vector indicating the spatial orientation of the defect; the first defect feature data is a structured set of records; each record uniquely corresponds to an identified defect, and the record contains multiple fields that store the unique defect identifier, the three-dimensional coordinates of the centroid, the volume, the size of the minimum circumscribed cuboid, the sphericity, the elongation, the flattening rate, and the quantized three-dimensional geometric features of the principal axis unit vector.
[0018] Preferably, the defect-cause knowledge graph reasoning model includes a knowledge graph construction layer, a graph attention network-based reasoning layer, and an output layer;
[0019] The knowledge graph construction layer is used to construct a first node set, a second node set, a third node set, a fourth node set, and a directed edge set. Each node in the first node set represents a quantized geometric descriptor contained in the first defect feature data. Each node in the second node set is a preset defect physical type identifier. Each node in the third node set is a preset defect physical cause identifier. Each node in the fourth node set is a production process parameter identifier. The directed edge set is used to establish directed paths for causal relationships between the four node sets.
[0020] The inference layer based on the graph attention network maps a set of quantized geometric descriptors from the first defect feature data to the initial feature vectors of the corresponding nodes in the first type of node set in the knowledge graph; and drives the graph attention network to propagate information in multiple layers along the directional edges, starting from the node that has completed initialization; in each layer of propagation, the graph attention network performs weighted aggregation of the feature vectors of adjacent nodes according to the weights calculated by the attention mechanism, so as to update the feature vector of the current node;
[0021] The output layer is used to calculate the target confidence score for each node in the third type of node set after the information propagation is completed, and output the physical cause identifier corresponding to the node with the highest confidence score as the reasoning result.
[0022] Preferably, the process of obtaining the second defect feature data is as follows: First, each first defect feature data record is input into the defect-cause knowledge graph reasoning model to perform reasoning; then, the model outputs a physical cause with the highest confidence for each record; finally, this physical cause identifier and its corresponding confidence score are used as new fields and appended to the original first defect feature data record to form the second defect feature data.
[0023] Preferably, the specific process of targeted production defect tracing includes:
[0024] Association Validation and Quantification: For high-frequency physical causes, a gradient boosting decision tree model is trained. The decision tree model uses production process parameters associated with high-frequency physical causes in the knowledge graph as input features and whether defects occur as labels. The nonlinear relationship between production process parameters and defects is quantified, validated, and learned.
[0025] Key parameter identification: The Shapley and interpretive analysis methods are applied to analyze the trained model to identify and quantify the production process parameters that contribute most to the high-frequency physical causes.
[0026] Instruction generation: Generate production process parameter control instructions based on the production process parameters that contribute the most.
[0027] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0028] 1. By constructing a fused digital twin, this invention, for the first time, unifies and registers X-ray CT data describing the internal structure of an object, thermal image data, and visible light data recording the appearance texture in a high-fidelity three-dimensional space; it can perform cross-modal correlation diagnosis. This method of correlation analysis between internal geometry and external physical properties provides comprehensive diagnostic depth and evidence chain, which can effectively improve the accuracy of glazed tile defect detection.
[0029] 2. This invention constructs a defect-cause knowledge graph reasoning model, enabling defects to establish a strong association with physical causes. When a defect that matches this characteristic is detected, it can perform logical reasoning along the path of the knowledge graph and output the physical cause with the highest confidence. This makes the defect diagnosis process quantifiable and traceable, and can effectively improve the accuracy of glazed tile defect detection.
[0030] 3. This invention transforms quantitative attribution and optimization solutions into a command-based control strategy. This method uses analysis tools such as SHAP to correlate physical causes with specific process parameters and calculates optimized adjustment values, thereby constructing a closed-loop path from data acquisition and analysis to process control. This solves the problem of loose connection between existing quality control and production control systems and difficulty in achieving automated closed-loop adjustment, and can effectively improve the accuracy of glazed tile defect detection.
[0031] 4. This invention first constructs a "fusion twin," unifying X-ray data describing the internal structure with thermal images and visible light images recording surface physical properties in three-dimensional space. This achieves holographic perception from "surface" to "interior," providing a cross-modal evidence chain for deep diagnosis that traditional technologies cannot match. Based on this, a "defect-cause knowledge graph reasoning model" is used to intelligently match and logically reason with the quantified defects in the fusion twin and the physical causes combined with expert experience and data mining, thus completing the reasoning from "what is the defect" to "why does the defect occur." Finally, through a "targeted production defect tracing" mechanism, abstract diagnostic conclusions are quantitatively correlated with specific production process parameters, and optimization algorithms are used to automatically generate precise control commands to directly intervene in the production system. In summary, this invention organically combines "holographic perception," "deep cognition," and "closed-loop feedback adjustment" to form a highly efficient and holistic defect detection method, effectively improving the accuracy of glazed tile defect detection. Attached Figure Description
[0032] Figure 1 This is a structural schematic diagram illustrating the process of an image recognition-based method for detecting internal defects in ceramic glazed tiles, as provided in an embodiment of the present invention.
[0033] Figure 2 This is a schematic diagram of the structure of a defect-cause knowledge graph reasoning model provided in an embodiment of the present invention. Detailed Implementation
[0034] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0035] Example 1
[0036] To improve the accuracy of defect detection in ceramic glazed tiles, an image recognition-based method for detecting internal defects in ceramic glazed tiles was applied, such as... Figure 1 This is a flowchart illustrating an image recognition-based method for detecting internal defects in ceramic glazed tiles, as provided in this embodiment. The method includes:
[0037] S1. Acquire multimodal data of the ceramic glazed tile to be tested, the multimodal data including X-ray tomography sequence, surface thermal image sequence, visible light image and production process parameters;
[0038] Furthermore, after leaving the forming process, ceramic glazed tile A is conveyed via a constant-speed conveyor belt into an online X-ray CT scanning system. This system performs a 360-degree rotational scan around the ceramic glazed tile, acquiring one two-dimensional X-ray projection image for each degree of rotation, resulting in a sequence of 360 images with a resolution of 1024x1024 pixels. While the ceramic glazed tile is still in the high-temperature cooling stage after leaving the sintering kiln, the conveyor belt passes directly above it by an industrial-grade infrared thermal imager (such as the FLIRA700). The thermal imager continuously captures images at a frequency of 5 frames per second, recording the temperature distribution changes on the surface of the ceramic glazed tile during the cooling process, forming a dynamic sequence of surface thermal images. After the ceramic glazed tile has cooled, an industrial-grade high-resolution camera captures a color image to record the final appearance, color, and texture of the ceramic glazed tile surface.
[0039] All process parameters of the A ceramic glazed tile production process are obtained through the SCADA system, including raw material information, molding parameters, firing parameters, and cooling parameters;
[0040] The raw material information includes raw material batch number, supplier information, moisture content, etc.; the molding parameters include molding pressure, holding time, vacuum chamber pressure, green body moisture content, mold temperature, pressing cycle time, etc.; the firing parameters include maximum firing zone temperature, holding time, etc.; the cooling parameters include cooling parameters, kiln exit temperature, etc.
[0041] S2. Construct a three-dimensional digital twin based on the scanning sequence, and register the thermal image sequence and visible light image onto the surface of the twin to obtain a fused twin; segment the fused twin according to a three-dimensional segmentation network, identify the three-dimensional geometric features of internal and underglaze defects, and obtain the first defect feature data;
[0042] Furthermore, the construction process of the three-dimensional digital twin is as follows: First, the X-ray tomography sequence is preprocessed to filter out noise and correct artifacts; second, the two-dimensional sequence image is reconstructed into a three-dimensional voxel grayscale data field using the Simultaneous Algebraic Reconstruction (SIRT) technique; then, the glazed tile body is segmented from the data field using a three-dimensional region growing algorithm; finally, the segmented glazed tile body voxel data is converted into a three-dimensional surface mesh model using the traveling cube algorithm.
[0043] Furthermore, the process for obtaining the fused twin is as follows:
[0044] The camera that acquires visible light images and the thermal imaging device that acquires surface thermal image sequences are jointly calibrated to determine the spatial transformation relationship between their coordinate systems.
[0045] Furthermore, by using a calibration plate that simultaneously possesses visible light and thermal characteristics, the visible light camera and the infrared thermal imager are jointly calibrated to obtain the spatial transformation matrix between their coordinate systems.
[0046] The camera pose is solved by matching the visible light image with the local invariant feature descriptors on the 3D digital twin rendering image using the perspective n-point algorithm.
[0047] Furthermore, by matching SIFT feature points on the visible light image with corresponding feature points in the rendered image of the 3D digital twin in the virtual camera, the perspective n-point (PnP) algorithm is used to solve for the precise 3D spatial pose of the real visible light camera relative to the twin model.
[0048] Using spatial transformation relationships and camera pose, the visible light image and surface thermal image sequence are used as textures and projected onto the surface of the twin.
[0049] Using the camera pose obtained in the previous step and the transformation relationship obtained from joint calibration, the visible light image and the surface thermal image obtained after averaging the surface thermal image sequence in the time dimension are used as textures and "attached" to the surface of the three-dimensional digital twin through projection mapping to obtain a fused twin that has both internal three-dimensional structure and real surface appearance and temperature distribution.
[0050] This embodiment, by constructing a fused digital twin, for the first time unifies and registers X-ray CT data describing the internal structure of an object, thermal image data, and visible light data recording the appearance texture in a high-fidelity three-dimensional space; it can perform cross-modal correlation diagnosis. This method of correlation analysis between internal geometry and external physical properties provides comprehensive diagnostic depth and evidence chain, which can effectively improve the accuracy of glazed tile defect detection.
[0051] Furthermore, a pre-trained 3D U-Net segmentation network is input with a three-dimensional voxel grayscale data field. The segmentation network classifies each voxel to obtain a segmentation mask, and the voxels are labeled as "background," "healthy glazed tile," or "defect." Connectivity analysis is performed on all voxel clusters labeled as "defects" in the segmentation mask. For each defect, three-dimensional geometric features are obtained, and these three-dimensional geometric features are used as the first defect feature data.
[0052] The construction, training, and application process of the 3D U-Net segmentation network are as follows:
[0053] 1. Construction and labeling of the training dataset:
[0054] First, 500 samples were collected from production line A of ceramic glazed tiles. These samples contained at least one internal defect (such as bubbles, cracks, etc.). X-ray tomography was performed on these 500 samples to obtain their three-dimensional voxel grayscale data field. The defect areas in the three-dimensional data field were segmented and labeled layer by layer to generate a high-precision three-dimensional binary mask corresponding to the original data. This three-dimensional binary mask was then used as training labels.
[0055] 2. Specific architecture and training of the 3D U-Net network:
[0056] The 3D U-Net segmentation network used in this embodiment includes an encoder path and a decoder path, each with 5 resolution levels. Each level consists of two consecutive 3x3x3 convolutional layers, followed by a batch normalization layer and a Rectified Linear Unit (ReLU) activation function. Downsampling in the encoder path is achieved through 3x3x3 max pooling with a stride of 2, while upsampling in the decoder path is accomplished through 2x2x2 deconvolution.
[0057] Using this architecture, the dataset obtained in step 1 is divided into training and validation sets in an 8:2 ratio. The network is trained using the Dice similarity coefficient loss function and the Adam optimizer; the batch size is 4, and the training run is 200 epochs in total. During training, if the loss on the validation set does not decrease for 10 consecutive epochs, training is terminated early to prevent overfitting.
[0058] 3. Defect identification and feature extraction:
[0059] The 3D voxel grayscale data field of the fused twin of the glazed tile to be detected is input into the pre-trained 3D U-Net network. The network outputs a 3D segmentation mask, in which each voxel is classified as 'background', 'healthy tile', or 'defect'. Subsequently, connected component analysis is performed on the voxel clusters labeled 'defect' to assign a unique identifier to each individual defect, and its 3D geometric features are calculated to form the first defect feature data.
[0060] Furthermore, the three-dimensional geometric features are quantized into a set of descriptors, which include: the three-dimensional coordinates of the defect centroid, the volume of the defect, the size of the minimum circumscribed cuboid, the sphericity value used to distinguish between bubbles and cracks, the elongation and flattening rates describing the defect morphology, and the principal axis unit vector indicating the spatial orientation of the defect; the first defect feature data is a structured set of records; each record uniquely corresponds to an identified defect, and the record contains multiple fields that store the unique defect identifier, the three-dimensional coordinates of the centroid, the volume, the size of the minimum circumscribed cuboid, the sphericity, the elongation, the flattening rate, and the quantized three-dimensional geometric features of the principal axis unit vector.
[0061] S3. Construct a defect-cause knowledge graph reasoning model to perform physical causal reasoning on the first defect feature data, associate and match the three-dimensional geometric features with preset physical causes, and generate second defect feature data with additional physical cause identifiers.
[0062] Furthermore, the defect-cause knowledge graph reasoning model includes a knowledge graph construction layer, a graph attention network-based reasoning layer, and an output layer; Figure 2 This is a schematic diagram of the structure of a defect-cause knowledge graph reasoning model provided in an embodiment of the present invention;
[0063] The knowledge graph construction layer is used to construct a first node set, a second node set, a third node set, a fourth node set, and a directed edge set. Each node in the first node set represents a quantized geometric descriptor contained in the first defect feature data. Each node in the second node set is a preset defect physical type identifier. Each node in the third node set is a preset defect physical cause identifier. Each node in the fourth node set is a production process parameter identifier. The directed edge set is used to establish directed paths for causal relationships between the four node sets.
[0064] Furthermore, the knowledge graph construction layer specifically includes:
[0065] 1. Construction of initial causal paths based on expert knowledge:
[0066] Based on the practical experience of a panel of senior ceramic process engineers and materials scientists, initial causal relationship paths were defined. Using the above method, these initial causal relationship paths were obtained, and a knowledge graph was constructed based on them.
[0067] 2. Association verification based on data mining:
[0068] Historical production data collected by the SCADA system (containing product defect detection results and corresponding production process parameters for over 20,000 production batches) was used. For this data, the Apriori association rule mining algorithm was employed, with a minimum support of 0.05 and a minimum confidence of 0.8. Rules with a confidence higher than 0.8 were transformed into directed edges in a knowledge graph.
[0069] Based on the four types of nodes and the directional edges connecting each node, the final knowledge graph is obtained.
[0070] The inference layer based on the graph attention network maps a set of quantized geometric descriptors from the first defect feature data to the initial feature vectors of the corresponding nodes in the first type of node set in the knowledge graph; and drives the graph attention network to propagate information in multiple layers along the directional edges, starting from the node that has completed initialization; in each layer of propagation, the graph attention network performs weighted aggregation of the feature vectors of adjacent nodes according to the weights calculated by the attention mechanism, so as to update the feature vector of the current node;
[0071] The output layer is used to calculate the target confidence score for each node in the third type of node set after the information propagation is completed, and output the physical cause identifier corresponding to the node with the highest confidence score as the reasoning result.
[0072] Furthermore, the process of obtaining the second defect feature data is as follows: First, each first defect feature data record is input into the defect-cause knowledge graph reasoning model to perform reasoning; then, the model outputs a physical cause with the highest confidence for each record; finally, this physical cause identifier and its corresponding confidence score are used as new fields and appended to the original first defect feature data record to form the second defect feature data.
[0073] S4. Based on the second defect feature data and the production process parameters, perform targeted production defect tracing, and perform targeted correlation analysis on the production process parameters according to the physical cause identifier to locate the cause of the defect and generate production process control parameters.
[0074] Furthermore, the specific process of targeted production defect tracing includes:
[0075] Association Validation and Quantification: For high-frequency physical causes, a gradient boosting decision tree model is trained. The decision tree model uses production process parameters associated with high-frequency physical causes in the knowledge graph as input features and whether defects occur as labels. The nonlinear relationship between production process parameters and defects is quantified, validated, and learned.
[0076] Key parameter identification: The Shapley and interpretive analysis methods are applied to analyze the trained model to identify and quantify the production process parameters that contribute most to the high-frequency physical causes.
[0077] Instruction generation: Generate production process parameter control instructions based on the production process parameters that contribute the most.
[0078] Furthermore, correlation verification and quantification are used to conduct specific analyses of the frequently diagnosed physical causes.
[0079] Process: A large number of production records related to this cause were selected from the historical database. Using the associated production parameters "molding pressure", "vacuum chamber pressure" and "moisture content of the billet" as features, and "whether this type of defect occurs" as a label, a gradient boosting decision tree (GBDT) model was trained.
[0080] This embodiment constructs a defect-cause knowledge graph reasoning model, enabling defects to establish a strong association with their physical causes. When a defect matching this characteristic is detected, it can perform logical reasoning along the path of the knowledge graph and output the physical cause with the highest confidence. This makes the defect diagnosis process quantifiable and traceable, effectively improving the accuracy of glazed tile defect detection.
[0081] 1. Key Parameter Identification: The SHAP analysis method is applied to interpret the trained GBDT model; the production process parameter with the highest SHAP value contribution is selected as the key parameter; for example, in the scenario of "insufficient billet venting", the SHAP value contribution of "vacuum chamber pressure" far exceeds that of other parameters, so "vacuum chamber pressure" is defined as the key parameter.
[0082] 2. Control command generation: If the SHAP values of "vacuum chamber pressure" are all negative, it indicates that the lower the parameter value, the higher the probability of defects occurring. Therefore, the vacuum chamber pressure is adjusted to "increase".
[0083] 3. Optimization Solution: With the goal of "minimizing the defect probability", the GBDT model is used for optimization solution.
[0084] 4. Command Generation: Automatically generates a standardized JSON format control command;
[0085] Furthermore, the optimization solution and instruction generation process include:
[0086] The production process parameters are used as optimization variables; the optimization objective is to minimize the probability of defect occurrence predicted by the gradient boosting decision tree (GBDT) model.
[0087] Define constraints: Based on the process safety regulations for the A ceramic glazed tile production line, constrain the production process parameters;
[0088] The Brent method is used for optimization; based on the optimization objective of minimizing the defect probability and the constraints, the algorithm is iteratively calculated until the production process parameter value that minimizes the defect probability predicted by the GBDT model is found within the constraints.
[0089] 3. Generate and issue control commands:
[0090] The system will automatically generate standardized JSON format control instructions based on the optimized production process parameter values; these instructions will be sent to the factory's SCADA system to complete the closed-loop control and adjustment of the production process.
[0091] This embodiment relates to the field of industrial vision technology, specifically a method for detecting internal defects in ceramic glazed tiles based on image recognition. It acquires multimodal data including X-ray tomography sequences, surface thermal images, visible light images, and production process parameters. A three-dimensional digital twin is constructed based on the X-ray data, fusing surface image information, and a three-dimensional segmentation network is used to accurately identify the three-dimensional geometric features of internal and underglaze defects. Through a constructed defect-cause knowledge graph reasoning model, physical causal reasoning is performed on the geometric features of the defects, intelligently matching and identifying the physical causes of the defects. Targeted correlation analysis and source tracing are performed by combining the physical causes with production process parameters to accurately locate the production link leading to the defect and generate control instructions for optimizing production. This invention achieves a closed-loop defect detection system of "diagnosis-source tracing-control," enabling precise location of defect problems and providing automated decision support for optimizing production and improving product quality.
[0092] This embodiment first constructs a "fusion twin," unifying X-ray data describing the internal structure with thermal images and visible light images recording surface physical properties in three-dimensional space. This achieves holographic perception from "surface" to "interior," providing a cross-modal evidence chain for deep diagnosis that traditional technologies cannot match. Based on this, a "defect-cause knowledge graph reasoning model" is used to intelligently match and logically reason with the quantified defect "geometric fingerprint" from the fusion twin and the physical causes combined with expert experience and data mining. This completes the reasoning from "what is the defect" to "why does the defect occur." Finally, through a "targeted production defect tracing" mechanism, abstract diagnostic conclusions are quantitatively correlated with specific production process parameters, and optimization algorithms automatically generate precise control commands to directly intervene in the production system. In summary, this invention organically combines "holographic perception," "deep cognition," and "closed-loop feedback adjustment" to form a highly efficient and holistic defect detection method, effectively improving the accuracy of glazed tile defect detection.
[0093] Example 2
[0094] To improve the accuracy of defect detection in B-type ceramic glazed tiles, an image recognition-based method for detecting internal defects in ceramic glazed tiles was applied, such as... Figure 1 This is a flowchart illustrating an image recognition-based method for detecting internal defects in ceramic glazed tiles, as provided in this embodiment. The method includes:
[0095] An image recognition-based method for detecting internal defects in ceramic glazed tiles includes:
[0096] S1. Acquire multimodal data of the ceramic glazed tile to be tested, the multimodal data including X-ray tomography sequence, surface thermal image sequence, visible light image and production process parameters;
[0097] S2. Construct a three-dimensional digital twin based on the scanning sequence, and register the thermal image sequence and visible light image onto the surface of the twin to obtain a fused twin; segment the fused twin according to a three-dimensional segmentation network, identify the three-dimensional geometric features of internal and underglaze defects, and obtain the first defect feature data;
[0098] S3. Construct a defect-cause knowledge graph reasoning model to perform physical causal reasoning on the first defect feature data, associate and match the three-dimensional geometric features with preset physical causes, and generate second defect feature data with additional physical cause identifiers.
[0099] S4. Based on the second defect feature data and the production process parameters, perform targeted production defect tracing, and perform targeted correlation analysis on the production process parameters according to the physical cause identifier to locate the cause of the defect and generate production process control parameters.
[0100] Furthermore, the construction process of the three-dimensional digital twin is as follows: First, the X-ray tomography sequence is preprocessed to filter out noise and correct artifacts; second, the two-dimensional sequence image is reconstructed into a three-dimensional voxel grayscale data field using synchronous algebraic reconstruction technology; then, the glazed tile body is segmented from the data field using a three-dimensional region growing algorithm; finally, the segmented glazed tile body voxel data is converted into a three-dimensional surface mesh model using the traveling cube algorithm.
[0101] Furthermore, the process for obtaining the fused twin is as follows:
[0102] The camera that acquires visible light images and the thermal imaging device that acquires surface thermal image sequences are jointly calibrated to determine the spatial transformation relationship between their coordinate systems.
[0103] The camera pose is solved by matching the visible light image with the local invariant feature descriptors on the 3D digital twin rendering image using the perspective n-point algorithm.
[0104] Using spatial transformation relationships and camera pose, the visible light image and surface thermal image sequence are used as textures and projected onto the surface of the twin.
[0105] Furthermore, the three-dimensional geometric features are quantized into a set of descriptors, which include: the three-dimensional coordinates of the defect centroid, the volume of the defect, the size of the minimum circumscribed cuboid, the sphericity value used to distinguish between bubbles and cracks, the elongation and flattening rates describing the defect morphology, and the principal axis unit vector indicating the spatial orientation of the defect; the first defect feature data is a structured set of records; each record uniquely corresponds to an identified defect, and the record contains multiple fields that store the unique defect identifier, the three-dimensional coordinates of the centroid, the volume, the size of the minimum circumscribed cuboid, the sphericity, the elongation, the flattening rate, and the quantized three-dimensional geometric features of the principal axis unit vector.
[0106] Furthermore, the defect-cause knowledge graph reasoning model includes a knowledge graph construction layer, a graph attention network-based reasoning layer, and an output layer; Figure 2 This is a schematic diagram of the structure of a defect-cause knowledge graph reasoning model provided in an embodiment of the present invention;
[0107] The knowledge graph construction layer is used to construct a first node set, a second node set, a third node set, a fourth node set, and a directed edge set. Each node in the first node set represents a quantized geometric descriptor contained in the first defect feature data. Each node in the second node set is a preset defect physical type identifier. Each node in the third node set is a preset defect physical cause identifier. Each node in the fourth node set is a production process parameter identifier. The directed edge set is used to establish directed paths for causal relationships between the four node sets.
[0108] The inference layer based on the graph attention network maps a set of quantized geometric descriptors from the first defect feature data to the initial feature vectors of the corresponding nodes in the first type of node set in the knowledge graph; and drives the graph attention network to propagate information in multiple layers along the directional edges, starting from the node that has completed initialization; in each layer of propagation, the graph attention network performs weighted aggregation of the feature vectors of adjacent nodes according to the weights calculated by the attention mechanism, so as to update the feature vector of the current node;
[0109] The output layer is used to calculate the target confidence score for each node in the third type of node set after the information propagation is completed, and output the physical cause identifier corresponding to the node with the highest confidence score as the reasoning result.
[0110] Furthermore, the process of obtaining the second defect feature data is as follows: First, each first defect feature data record is input into the defect-cause knowledge graph reasoning model to perform reasoning; then, the model outputs a physical cause with the highest confidence for each record; finally, this physical cause identifier and its corresponding confidence score are used as new fields and appended to the original first defect feature data record to form the second defect feature data.
[0111] Furthermore, the specific process of targeted production defect tracing includes:
[0112] Association Validation and Quantification: For high-frequency physical causes, a gradient boosting decision tree model is trained. The decision tree model uses production process parameters associated with high-frequency physical causes in the knowledge graph as input features and whether defects occur as labels. The nonlinear relationship between production process parameters and defects is quantified, validated, and learned.
[0113] Key parameter identification: The Shapley and interpretive analysis methods are applied to analyze the trained model to identify and quantify the production process parameters that contribute most to the high-frequency physical causes.
[0114] Instruction generation: Generate production process parameter control instructions based on the production process parameters that contribute the most.
[0115] This embodiment first constructs a "fusion twin," unifying X-ray data describing the internal structure with thermal images and visible light images recording surface physical properties in three-dimensional space. This achieves holographic perception from "surface" to "interior," providing a cross-modal evidence chain for deep diagnosis that traditional technologies cannot match. Based on this, a "defect-cause knowledge graph reasoning model" is used to intelligently match and logically reason with the quantified defect "geometric fingerprint" from the fusion twin and the physical causes combined with expert experience and data mining. This completes the reasoning from "what is the defect" to "why does the defect occur." Finally, through a "targeted production defect tracing" mechanism, abstract diagnostic conclusions are quantitatively correlated with specific production process parameters, and optimization algorithms automatically generate precise control commands to directly intervene in the production system. In summary, this invention organically combines "holographic perception," "deep cognition," and "closed-loop feedback adjustment" to form a highly efficient and holistic defect detection method, effectively improving the accuracy of glazed tile defect detection.
[0116] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A method for detecting internal defects of ceramic glazed tiles based on image recognition, characterized in that, include: S1. Acquire multimodal data of the ceramic glazed tile to be tested, the multimodal data including X-ray tomography sequence, surface thermal image sequence, visible light image and production process parameters; S2. Construct a three-dimensional digital twin based on the scanning sequence, and register the thermal image sequence and visible light image onto the surface of the twin to obtain a fused twin; segment the fused twin according to a three-dimensional segmentation network, identify the three-dimensional geometric features of internal and underglaze defects, and obtain the first defect feature data; S3. Construct a defect-cause knowledge graph reasoning model to perform physical causal reasoning on the first defect feature data, associate and match the three-dimensional geometric features with preset physical causes, and generate second defect feature data with additional physical cause identifiers. The defect-cause knowledge graph reasoning model includes a knowledge graph construction layer, a graph attention network-based reasoning layer, and an output layer. The knowledge graph construction layer is used to construct a first node set, a second node set, a third node set, a fourth node set, and a directed edge set. Each node in the first node set represents a quantized geometric descriptor contained in the first defect feature data. Each node in the second node set is a preset defect physical type identifier. Each node in the third node set is a preset defect physical cause identifier. Each node in the fourth node set is a production process parameter identifier. The directed edge set is used to establish directed paths for causal relationships between the four node sets. The inference layer based on the graph attention network maps a set of quantized geometric descriptors from the first defect feature data to the initial feature vectors of the corresponding nodes in the first type of node set in the knowledge graph; and drives the graph attention network to propagate information in multiple layers along the directional edges, starting from the node that has completed initialization; in each layer of propagation, the graph attention network performs weighted aggregation of the feature vectors of adjacent nodes according to the weights calculated by the attention mechanism, and updates the feature vector of the current node; The output layer is used to calculate the target confidence score for each node in the third type of node set after the information propagation is completed, and output the physical cause identifier corresponding to the node with the highest confidence score as the reasoning result. S4. Based on the second defect feature data and the production process parameters, perform targeted production defect tracing, perform targeted correlation analysis on the production process parameters according to the physical cause identifier, locate the cause of the defect, and generate production process parameter control instructions based on the cause of the defect. The specific process for tracing the source of targeted production defects includes: Association Validation and Quantification: For high-frequency physical causes, a gradient boosting decision tree model is trained. The decision tree model uses production process parameters associated with high-frequency physical causes in the knowledge graph as input features and whether defects occur as labels. The nonlinear relationship between production process parameters and defects is quantified, validated, and learned. Key parameter identification: The Shapley and interpretive analysis methods are applied to analyze the trained model to identify and quantify the production process parameters that contribute most to the high-frequency physical causes. Instruction generation: Generate production process parameter control instructions based on the production process parameters that contribute the most.
2. The image recognition-based method for detecting internal defects of ceramic glazed tiles according to claim 1, characterized in that: The construction process of the three-dimensional digital twin is as follows: the X-ray tomography sequence is preprocessed to filter out noise and correct artifacts; the two-dimensional sequence image is reconstructed into a three-dimensional voxel grayscale data field using synchronous algebraic reconstruction technology; then, the glazed tile body is segmented from the data field using a three-dimensional region growing algorithm; and the segmented glazed tile body voxel data is converted into a three-dimensional surface mesh model using a traveling cube algorithm.
3. The image recognition-based method for detecting internal defects of ceramic glazed tiles according to claim 1, characterized in that: The process for obtaining the fused twin is as follows: The camera that acquires visible light images and the thermal imaging device that acquires surface thermal image sequences are jointly calibrated to determine the spatial transformation relationship between their coordinate systems. The camera pose is solved by matching the visible light image with the local invariant feature descriptors on the 3D digital twin rendering image using the perspective n-point algorithm. Using spatial transformation relationships and camera pose, the visible light image and surface thermal image sequence are used as textures and projected onto the surface of the twin.
4. The image recognition-based method for detecting internal defects of ceramic glazed tiles according to claim 1, characterized in that: The three-dimensional geometric features are quantized into a set of descriptors, which include: the three-dimensional coordinates of the defect centroid, the volume of the defect, the size of the minimum circumscribed cuboid, the sphericity value used to distinguish between bubbles and cracks, the elongation and flattening rates describing the defect morphology, and the principal axis unit vector indicating the spatial orientation of the defect. The first defect feature data is a structured set of records. Each record uniquely corresponds to an identified defect. The record contains multiple fields that store the unique defect identifier, the three-dimensional coordinates of the centroid, the volume, the size of the minimum circumscribed cuboid, the sphericity, the elongation, the flattening rate, and the quantized three-dimensional geometric features of the principal axis unit vector.
5. The method for detecting internal defects in ceramic glazed tiles based on image recognition according to claim 1, characterized in that: The process of obtaining the second defect feature data is as follows: each first defect feature data record is input into the defect-cause knowledge graph reasoning model for reasoning; the model outputs the physical cause with the highest confidence for each record; the physical cause identifier and its corresponding confidence score are used as new fields and appended to the original first defect feature data record to form the second defect feature data.