Ceramic fracture morphology image feature extraction and fracture mode recognition method and system

Abstract

This invention relates to the field of fracture failure analysis technology for ceramic materials, and discloses a method and system for extracting features from ceramic fracture surface images and recognizing fracture modes. The method includes: multi-scale fracture surface image acquisition and panoramic stitching; fracture feature region detection and localization based on a deep convolutional neural network; quantitative analysis of fracture surface roughness based on gray-level co-occurrence matrix and shape recovery technology; fracture mode recognition based on a multi-label classification network; and fracture failure source analysis and report generation. This invention achieves automatic extraction of ceramic fracture surface features and accurate recognition of multiple fracture modes, significantly improving the efficiency and accuracy of fracture analysis.

Description

Technical Field

[0001] This invention relates to the field of fracture failure analysis technology for ceramic materials, specifically to a method and system for extracting features from ceramic fracture surface images and recognizing fracture modes. Background Technology

[0002] Ceramic materials, due to their excellent high-temperature stability, wear and corrosion resistance, and high hardness, have been widely used in high-end applications such as hot-end components of aerospace engines, electronic packaging substrates, biomedical implants, and advanced manufacturing tools. However, the inherent brittleness of ceramic materials makes them highly susceptible to fracture failure during actual service due to factors such as internal defect propagation, thermal shock stress concentration, or excessive external mechanical loads, leading to serious safety accidents and economic losses. Fracture morphology analysis, as one of the core methods of material failure analysis, can effectively trace the fracture initiation location, determine the fracture propagation mechanism, and infer the cause of failure by observing and analyzing the microscopic morphological characteristics of the fracture surface. This has significant engineering value for improving the quality and reliability of ceramic products.

[0003] In traditional ceramic fracture analysis, analysts typically use scanning electron microscopes (SEMs) at different magnifications to observe the microscopic morphology of the fracture surface region by region. They rely on expert experience to identify typical fracture morphology features such as river patterns, mirror-like areas, hazy areas, serrated steps, and transgranular cleavage planes, and then comprehensively judge the fracture mode and infer the cause of failure. This analytical method, which heavily relies on human experience, has several limitations. First, ceramic fracture morphology is complex and diverse, with different fracture modes often overlapping and mixed, making precise quantitative differentiation difficult to achieve solely through visual observation. Second, obtaining complete fracture information is challenging. Due to the limited field of view of a single SEM, analysts need to repeatedly switch between low and high magnification to balance macroscopic positioning and microscopic details, resulting in low operational efficiency and a risk of missing crucial information. Third, differences in the experience levels and subjective judgments of different analysts lead to insufficient consistency and repeatability of analytical results, making it difficult to meet the standardized requirements of fracture analysis for large-scale ceramic products in industrial production.

[0004] In recent years, deep learning technology has made some progress in the fields of material surface defect detection and fracture morphology analysis. For example, Chinese patent CN120411056A discloses a multimodal fusion learning system for alloy fracture surfaces. This system uses an improved YOLOv5 model to perform semantic segmentation and target detection of dimple fracture and brittle fracture regions in aluminum alloy fracture images, and uses the XGBoost model to establish a regression prediction relationship between material microstructure parameters and fracture ratio. However, this technical solution has the following shortcomings: First, its research object is metallic alloy materials, and the fracture morphology features (dimples, brittleness) are relatively simple, which cannot be directly transferred to the complex fracture morphology feature system unique to ceramic materials (including river patterns, mirror areas, hazy areas, serrated steps, transgranular cleavage planes, etc.). Second, the system lacks multi-scale fracture image acquisition and panoramic stitching capabilities, and only performs imaging at a single fixed magnification, which cannot simultaneously acquire the macroscopic overall picture and microscopic details of the fracture surface. Third, the system lacks quantitative analysis of fracture surface roughness and does not incorporate quantification methods that fuse texture features with three-dimensional morphology features. Fourth, the system employs a binary regression prediction model and lacks the ability to classify and identify multiple fracture modes using multi-label methods. Fifth, the system does not include a fracture failure tracing analysis function and cannot automatically infer possible failure causes based on fracture characteristics.

[0005] In summary, existing technologies still have significant gaps in areas such as automatic extraction of ceramic fracture morphology features, complete acquisition of multi-scale image information, quantitative characterization of fracture surface roughness, accurate identification of multiple fracture modes, and intelligent tracing of fracture failure causes. A systematic fracture analysis technology solution for ceramic materials is urgently needed to address these issues. Furthermore, existing fracture analysis technologies generally lack a multi-level feature fusion mechanism from macroscopic feature distribution to microscopic texture parameters, and also lack the automated tracing capability to link fracture morphology analysis results with a failure cause knowledge base, failing to form a complete closed-loop analysis process from image acquisition, feature extraction, pattern recognition to failure tracing. Therefore, there is an urgent need to develop an end-to-end intelligent analysis method and system for ceramic material fractures to achieve automation, quantification, and standardization of the entire fracture analysis process, reduce reliance on human experience, and improve analysis efficiency and result consistency. Summary of the Invention

[0006] To address the technical problems in existing technologies, such as reliance on manual experience for extracting ceramic fracture morphology features, incomplete fracture information acquisition, lack of automated means for fracture mode determination, and difficulty in tracing the cause of failure,

[0007] This invention provides a method for extracting features from ceramic fracture surface images and recognizing fracture patterns, comprising the following steps:

[0008] Step S1, Multi-scale fracture morphology image acquisition and panoramic stitching: The ceramic fracture sample to be analyzed is imaged by scanning electron microscopy at multiple magnifications to acquire macroscopic and microscopic fracture morphology images, locate the fracture source region and output the fracture source location result; Feature point detection and matching are performed on multiple images with overlapping areas acquired at the same magnification, and high-resolution panoramic images are generated by image registration and fusion stitching.

[0009] Step S2, Fracture Feature Region Detection and Localization: The panoramic image is input into a fracture feature region detection network constructed based on a deep convolutional neural network to automatically identify the feature regions on the fracture surface and output the bounding box coordinates and detection confidence of each feature region; wherein, the feature regions include river pattern regions, mirror regions, fog-like regions, serrated step regions, and transgranular cleavage surface regions.

[0010] Step S3, quantitative analysis of fracture surface roughness: For each feature region, extract texture energy, texture entropy value and texture contrast parameters based on gray-level co-occurrence matrix, estimate the local three-dimensional surface undulation height by combining light and dark recovery shape technology, and normalize and fuse the texture parameters and undulation height to obtain a multi-dimensional roughness feature vector.

[0011] Step S4, fracture mode multi-label classification and recognition: The bounding box coordinates of the feature region are concatenated with the detection confidence and multidimensional roughness feature vector, and input into the multi-label classification network to determine the fracture mode, and the probability of each fracture mode is output; the fracture modes include brittle fracture, quasi-cleavage fracture, intergranular fracture and mixed fracture.

[0012] Step S5, Fracture Failure Source Tracing Analysis and Report Generation: Based on the fracture source location results output in Step S1 and the spatial distribution relationship determined by the bounding box coordinates of each feature region output in Step S2, the crack propagation path direction is determined. Combining the fracture mode attribution probability and the roughness change trend calculated based on the multidimensional roughness feature vector output in Step S3, a matching inference is performed based on the preset failure cause knowledge base to generate a structured fracture analysis report.

[0013] This invention also provides a ceramic fracture morphology image feature extraction and fracture mode recognition system, comprising: a multi-scale image acquisition and stitching module for imaging and stitching panoramic images in low-magnification and high-magnification modes respectively; a fracture feature region detection module for automatically identifying typical morphological feature regions of the fracture surface; a roughness quantification analysis module for extracting texture parameters and estimating the height of three-dimensional surface undulations to obtain a multi-dimensional roughness feature vector; a fracture mode classification module for performing multi-label classification to determine the fracture mode type; and a failure tracing analysis module for inferring the failure cause based on fracture features and a failure cause knowledge base and generating an analysis report.

[0014] The beneficial effects of this invention are as follows: Through multi-scale image acquisition and panoramic stitching strategies, it fully preserves both macroscopic structural information and microscopic detail information of the fracture surface, overcoming the problem of limited field of view in single-magnification imaging; through a deep learning fracture feature region detection network, it achieves automatic identification and precise positioning of ceramic-specific fracture morphology features such as river patterns, mirror areas, hazy areas, serrated steps, and transgranular cleavage surfaces, eliminating the subjective differences of manual identification; through the innovative integration of gray-level co-occurrence matrix texture analysis and light and dark shape recovery technology, it achieves multi-dimensional quantitative characterization of fracture surface roughness, thus mitigating the limitations of traditional methods. This approach overcomes the limitations of traditional methods that can only qualitatively describe surface morphology. By employing a multi-label classification network and an asymmetric weighted loss function, it enables the simultaneous determination of multiple fracture modes, including brittle fracture, quasi-cleavage fracture, intergranular fracture, and mixed fracture, thus adapting to the practical characteristics of multiple fracture modes coexisting in ceramics. Through matching and inference between the failure tracing analysis module and the knowledge base, it achieves automated intelligent inference of fracture failure causes and standardized report output, significantly improving the efficiency, accuracy, and repeatability of ceramic fracture analysis and providing reliable technical support for ceramic product quality control and process optimization. Attached Figure Description

[0015] Figure 1 This is a flowchart of the method for extracting features from ceramic fracture surface images and recognizing fracture patterns provided in an embodiment of the present invention.

[0016] Figure 2 This is an architecture diagram of the ceramic fracture morphology image feature extraction and fracture pattern recognition system provided in the embodiments of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of protection of this invention.

[0018] See Figure 1The ceramic fracture morphology image feature extraction and fracture mode recognition method provided in this embodiment of the invention includes steps S1 to S5, forming a data-driven, deeply coupled closed-loop collaborative architecture. Specifically, the panoramic image output of step S1 serves as the core input of step S2, the feature region detected in step S2 provides a precise spatial positioning basis for roughness analysis in step S3, and the outputs of steps S2 and S3 together constitute the multi-dimensional feature input for fracture mode recognition in step S4. The determination result of step S4 and the roughness trend information of step S3 provide key evidence for failure tracing in step S5. Preferably, the fracture mode determination result of step S4 is also fed back to step S1 to adaptively adjust the subsequent acquisition strategy, and the failure tracing conclusion of step S5 can be fed back to step S2 to dynamically optimize the detection weight, thereby forming a complete positive analysis and reverse optimization closed loop.

[0019] Step S1: Multi-scale fracture morphology image acquisition and panoramic stitching. In this embodiment of the invention, considering that the fracture morphology of ceramic materials simultaneously contains millimeter-level macroscopic structural information and micrometer-level microscopic detail information, the system is designed with a dual-mode multi-scale acquisition strategy to ensure that both global positioning accuracy and local detail resolution are taken into account in a complete fracture analysis process.

[0020] First, the ceramic fracture sample is imaged using a full-field scanning method in low-magnification mode. Preferably, the magnification of the low-magnification mode is set between 20x and 100x; in one embodiment of the invention, 50x is used as the default magnification. The accelerating voltage of the scanning electron microscope is set to 10kV to 20kV, and the working distance is set to 10mm to 15mm to obtain clear images within a large depth of field. In this magnification mode, the system performs a raster scan of the fracture surface using a preset path planning strategy, maintaining an overlap of at least 25% between adjacent images; preferably, the overlap is set to 30%. In one embodiment of the invention, N1 images are acquired for a ceramic fracture sample with a diameter of approximately 10mm in 50x magnification mode, where N1 typically ranges from 16 to 36 images, and the specific value is adaptively determined based on the size and regularity of the fracture area and shape. Each image has a resolution of 2048×2048 pixels. After acquisition, the system preprocesses the low-magnification image sequence, including grayscale equalization and adaptive contrast enhancement, to eliminate the uneven brightness in different areas caused by the difference in the incident angle of the electron beam during scanning electron microscopy imaging.

[0021] After the low-magnification panoramic image is constructed, the system automatically locates the fracture source region on the low-magnification panoramic image. The fracture source localization adopts a gradient convergence direction analysis method, that is, the fracture initiation position is determined by calculating the convergence direction of river patterns or radial textures in the low-magnification panoramic image. Specifically, the system performs Gabor filtering on the low-magnification panoramic image to extract directional texture information, calculates the local directional gradient on the filtered response map, and then performs reverse tracing along the gradient direction. The intersection area of ​​multiple tracing paths is identified as the fracture source candidate region. Preferably, the range of the fracture source candidate region is extended to 1.5 to 2.0 times the outer edge of the original candidate region to ensure that subsequent high-magnification acquisition can cover the complete morphological information of the fracture source and its neighboring areas.

[0022] Subsequently, the system switches to high-magnification mode to perform detailed acquisition of the fracture source area and its surroundings. The magnification of the high-magnification mode is set in the range of 500x to 5000x, and in one embodiment of the present invention, 1000x to 2000x is used as a typical magnification range. In high-magnification mode, the image resolution is not less than 1024×1024 pixels, preferably set to 2048×2048 pixels. The system performs intensive acquisition in the fracture source area and key areas along the crack propagation direction. The number of high-magnification images N2 is generally greater than the number of low-magnification images N1. In one embodiment of the present invention, N2 is set to 24 to 64 images. The acquisition range of high-magnification images is jointly determined by the fracture source location determined in the low-magnification panoramic analysis and the initially identified crack propagation path, reflecting the guiding and constraining role of the low-magnification analysis results in step S1 on the high-magnification acquisition strategy.

[0023] After completing multi-scale image acquisition, the core task of this step is to stitch together multiple images with overlapping areas under the same magnification mode into a complete panoramic image. The panoramic stitching algorithm used in this embodiment of the invention includes three stages: feature point detection and matching, geometric transformation estimation, and image fusion.

[0024] In the feature point detection and matching stage, the system uses the Scale Invariant Feature Transform (SIFT) algorithm to extract feature points and their 128-dimensional feature descriptors for each image. The SIFT algorithm detects key points by constructing a Gaussian difference-of-scale space and generates rotation-invariant feature descriptors based on the gradient orientation histogram of the key point's neighborhood. This property provides good robustness to geometric distortions caused by minute changes in the scanning angle in fracture images. For adjacent image pairs, the system uses the nearest neighbor distance ratio method for feature point matching, where the distance ratio threshold... The value is set between 0.6 and 0.8; in one embodiment of the present invention, it is taken as... The physical meaning of this threshold is: for a given feature point, the ratio of its nearest neighbor matching distance to its second nearest neighbor matching distance must be less than [a certain value]. Only those pairs are accepted as valid matches, thus filtering out false matches.

[0025] In the geometric transformation estimation stage, the system uses the Random Sample Consensus Algorithm (RANSAC) to estimate the homography transformation matrix between adjacent images from feature point matching pairs. The RANSAC algorithm calculates candidate transformation matrices by randomly selecting at least four pairs of matching points, and then uses a projection error threshold. The process involves filtering inliers and iteratively selecting the transformation matrix with the highest proportion of inliers as the final estimation result. In one embodiment of this invention, the maximum number of RANSAC iterations is set to 2000, and the projection error threshold is... Set to 3.0 pixels.

[0026] During the image fusion stage, the system employs a weighted average fusion strategy to smooth the transition of overlapping regions. For each pixel location within the overlapping region... Its final fused grayscale value Calculate using the following formula: ,in: For the first image at position The grayscale value at the location ranges from 0 to 255 and is dimensionless. The second image after homography transformation is located at... The grayscale value at the location ranges from 0 to 255 and is dimensionless. For the first image at position The fusion weight at the point has a value range of 0 to 1 and is dimensionless. For the second image at position The fusion weights at the point satisfy Weight The normalized distance of the pixel position from the center of the first image decreases linearly. With corresponding increases, this linear gradient weight allocation strategy can effectively eliminate stitching seams, allowing the brightness and contrast of panoramic images to transition naturally in overlapping areas.

[0027] Through the above process, step S1 ultimately outputs a low-magnification panoramic image. and high-magnification panoramic images These serve as the basis for macroscopic and microscopic analysis in subsequent steps, respectively. In one embodiment of the present invention, for an alumina ceramic fracture sample, a low-magnification panoramic image covers a fracture area of ​​approximately 8mm × 6mm, while a high-magnification panoramic image covers a core area of ​​approximately 1.5mm × 1.2mm surrounding the fracture source, with a panoramic stitching accuracy better than 2 pixels.

[0028] Step S2: Fracture Feature Region Detection and Localization. This step receives the panoramic images (including low-magnification and high-magnification panoramic images) output from Step S1. Using a fracture feature region detection network based on a deep convolutional neural network, it automatically identifies and locates typical morphological feature regions on the fracture surface. Typical morphological features of ceramic material fractures include five categories: river patterns, mirror-like areas, hazy areas, serrated steps, and transgranular cleavage planes. Each type of feature exhibits different texture morphology and distribution patterns on the fracture surface, providing crucial indication for determining the fracture mechanism.

[0029] In this embodiment of the invention, the overall architecture of the fracture feature region detection network includes three core components: a feature extraction backbone network, a multi-scale feature pyramid neck network, and a detection head network.

[0030] The feature extraction backbone network employs a deep convolutional neural network structure based on residual connections. In one embodiment of the invention, the backbone network uses ResNet-50 as its basic architecture, comprising four residual stages, each outputting multi-scale feature maps with strides of 4, 8, 16, and 32, respectively. Each residual block contains a bottleneck structure of 1×1 convolutions, 3×3 convolutions, and 1×1 convolutions, effectively alleviating the gradient vanishing problem in deep networks through skip connections using identity mapping. Preferably, a 3×3 convolutional preprocessing layer is added at the input end of the backbone network to expand the single-channel grayscale image into a 3-channel pseudo-color representation, thereby fully utilizing the feature extraction capabilities of the pre-trained model. The input image size of the backbone network is uniformly 800×800 pixels. For panoramic images exceeding this size, the system employs a sliding window strategy for block detection, with the window stride set to 75% of the image size. The detection results are merged through global coordinate mapping and non-maximum suppression.

[0031] The design purpose of the multi-scale feature pyramid neck network is to fuse semantic features output from different levels of the backbone network to enhance the detection sensitivity of fracture feature regions at different scales. This embodiment of the invention employs a bidirectional feature pyramid network structure, comprising two feature propagation channels: a top-down path and a bottom-up path. The top-down path starts from the deepest feature map, progressively transmitting high-level semantic information to the shallow layers through 2x upsampling and lateral connections, enabling the shallow features to acquire rich semantic context. The bottom-up path, on the other hand, progressively transmits high-resolution spatial details from the shallow layers to the deeper layers, allowing the deep features to retain accurate spatial localization information. At each feature fusion node, a 1×1 convolution is used to uniformly adjust the number of channels to 256 dimensions, and feature merging is completed through element-wise addition.

[0032] In one embodiment of the present invention, the feature pyramid network outputs five feature layers of different scales (denoted as P2 to P6), with corresponding downsampling step sizes of 4, 8, 16, 32, and 64 pixels, respectively. Among them, layer P2 has high sensitivity for detecting small-scale fracture features (such as tiny transgranular cleavage plane regions), while layers P5 and P6 have good detection performance for large-scale fracture features (such as complete mirror-like regions). Through multi-level joint detection, comprehensive coverage of fracture feature regions of different scales is achieved.

[0033] The detection head network is responsible for target region proposal and classification regression based on the feature maps output from each layer of the feature pyramid. The detection head used in this embodiment of the invention comprises two parallel branches: a classification branch and a regression branch. The classification branch predicts the probability that each anchor box position belongs to one of five types of breakpoint features using four 3×3 convolutional layers and one 3×3 convolutional layer with a sigmoid activation function. The regression branch predicts the center offset and width / height scaling factor of the anchor box relative to the true target region using convolutional layers of the same depth.

[0034] During the model training phase, constructing a fracture feature annotation dataset is fundamental to ensuring detection accuracy. In one embodiment of this invention, the dataset contains 1200 scanning electron microscope images of ceramic fracture surfaces, covering four typical ceramic materials: alumina, silicon nitride, silicon carbide, and zirconium oxide. Each image was annotated by a materials engineer with fracture analysis experience using rectangular bounding boxes to mark identifiable fracture feature regions, with at least 200 annotated samples for each feature region. The dataset was divided into training, validation, and test sets in a 7:1:2 ratio. Model training employed FocalLoss and SmoothL1 loss functions to optimize the classification and regression branches, respectively. The initial learning rate was set to 0.01, and cosine annealing was used for learning rate decay. The training epochs were set to 120, and the batch size was set to 8.

[0035] During the model inference phase, the detection confidence threshold is set between 0.5 and 0.8. In one embodiment of the present invention, 0.6 is used as the default threshold. Detection results below this threshold are filtered to ensure that the output feature region detection results have sufficient reliability. The cross-union threshold for non-maximum suppression (NMS) is set between 0.3 and 0.5. In one embodiment of the present invention, 0.4 is used to effectively eliminate duplicate detection boxes for the same feature region.

[0036] The output of step S2 is a set of detection results for each feature region. Each detection result includes a feature category label (river pattern / mirror area / fog area / serrated step / transgranular cleavage surface) and bounding box coordinates. and detection confidence (Value range 0 to 1). These detection results will be simultaneously transmitted to step S3 as the spatial positioning basis for roughness analysis, and to step S4 as the macroscopic feature input for fracture mode recognition.

[0037] Step S3: Quantitative Analysis of Fracture Surface Roughness. This step receives the detection results of each feature region output from Step S2 and performs quantitative analysis of the fracture surface roughness for each detected feature region. Fracture surface roughness is an important physical quantity reflecting the degree of microscopic deformation of the material and the energy dissipation state of crack propagation during fracture. Significant differences exist in the fracture surface roughness corresponding to different fracture modes. This embodiment of the invention constructs a multidimensional roughness feature vector through the innovative fusion of gray-level co-occurrence matrix texture analysis and shading shape recovery technology, providing rich quantitative feature inputs for subsequent fracture mode recognition.

[0038] In terms of gray-level co-occurrence matrix (GLCM) texture analysis, the system first crops local images of each feature region from the high-magnification panoramic image based on the bounding box coordinates output in step S2. For each local image, its GLCM is calculated. (GLCM is then described in the original text.) Describes the situation at a given pixel spacing and direction angle Under certain conditions, grayscale values ​​in the image The frequency distribution of co-occurrence. In one embodiment of the present invention, the calculation window size of the gray-level co-occurrence matrix is ​​set to... pixels, of which The value range is from 5 to 15, with the preferred value being... The gray level is set to... Level, with a value range of 8 to 64, preferably selected. Pixel pitch Set to 1 to 3 pixels, preferably Calculation direction The calculation results are taken in four directions: 0°, 45°, 90° and 135°, and the average value of the calculation results in the four directions is used to eliminate the direction dependence.

[0039] Based on the gray-level co-occurrence matrix, the system extracts the following three core texture parameters: texture energy. This characterizes the uniformity and regularity of the surface texture of the fracture surface, and its calculation formula is as follows: ,in: The normalized element values ​​of the gray-level co-occurrence matrix, i.e., gray-level value pairs. The probability of occurrence is dimensionless, ranging from 0 to 1. The gray level is 32 in this embodiment, which is dimensionless; subscript and These represent the row and column indices of the gray-level co-occurrence matrix, respectively, both ranging from 0 to... . The range of values ​​is The value ranges from 1 to 1, with higher values ​​indicating more uniform and regular textures. The texture energy in the mirror-like area is typically higher than 0.15, while the texture energy in the jagged step area is typically lower than 0.05. This difference provides an important quantitative basis for fracture pattern recognition.

[0040] Texture entropy The complexity and information content of the fracture surface texture are characterized by the following formula: ,in: The meaning is the same as above; To prevent overflow of extremely small positive numbers in logarithmic operations, the value is set to... Dimensionless; Represents the natural logarithm operation. The value range is 0 to A higher value indicates a more complex texture. The texture entropy value of the foggy area is usually higher than that of the river pattern area because the surface of the foggy area exhibits a fine and disordered micro-morphology.

[0041] Texture contrast The grayscale difference between neighboring pixels on the fracture surface is characterized by the following formula: ,in: It is the squared difference between grayscale value pairs, reflecting the local grayscale contrast intensity, and is dimensionless; The value range is 0 to A higher value indicates more pronounced surface undulations. Transgranular cleavage planes typically exhibit higher texture contrast due to the presence of cleavage steps, while mirror-like areas show lower texture contrast due to their smooth and flat surfaces.

[0042] In terms of shape recovery technology, the system estimates the local three-dimensional undulation height of the fracture surface based on the grayscale distribution information of a single scanning electron microscope image. In the secondary electron imaging mode of a scanning electron microscope, there is an approximate cosine relationship between the image grayscale value and the local tilt angle of the surface. This invention utilizes this physical principle to establish a mapping relationship from grayscale to surface normal vector, and then obtains the height distribution through the integration of the surface normal vector.

[0043] For the location within the feature region The local surface tilt component at a given location is estimated using the following formula: , ,in: For position The local surface height at a given location, expressed in μm; For position The grayscale value at the location ranges from 0 to 255 and is dimensionless. This represents the actual physical distance between adjacent pixels, measured in micrometers (µm) per pixel. It is determined by both the magnification and image resolution, for example, in a 1000x magnification mode with a resolution of 2048×2048 pixels. um / pixel; The grayscale-tilt conversion coefficient for shape recovery is dimensionless and ranges from 0.5 to 2.0. In one embodiment of the present invention, it is obtained by calibration using a known standard roughness sample. . and They are respectively direction and The surface tilt component in the direction is dimensionless.

[0044] Based on the surface tilt component, the Poisson equation solver is used to... and Integral reconstruction is performed to obtain the local three-dimensional surface height distribution of the feature region. In one embodiment of the present invention, the frequency domain Poisson integration method is employed, and the specific steps are as follows: For... Perform a two-dimensional discrete Fourier transform, divide the result by the frequency response corresponding to the Laplace operator in the frequency domain, and then perform an inverse transform to obtain the height field.

[0045] After obtaining the texture parameters and height distribution, the system calculates the multi-scale roughness quantification index for each feature region. The calculation formula is as follows: ,in: , , These are the texture energy, texture entropy, and texture contrast values ​​after being normalized to the [0,1] interval by Min-Max, respectively, and are dimensionless. This is the normalized value of the standard deviation of the height distribution of the characteristic region, i.e. ,in This represents the standard deviation (in μm) of the height field in this characteristic region. This represents the maximum standard deviation of height across all feature regions (in μm). The value ranges from 0 to 1 and is dimensionless. , , , Let be the weight coefficients of each component, dimensionless, satisfying . In one embodiment of the present invention, , , , The weighting scheme was obtained through grid search optimization on the validation set. The weight of the height distribution standard deviation was slightly higher than that of the texture parameter, reflecting the important contribution of 3D undulation information to roughness characterization.

[0046] The final output of step S3 is the multidimensional roughness feature vector of each feature region. This vector simultaneously contains the original texture parameters, three-dimensional height statistics, and a fused roughness index, with a dimension of 5. These multi-dimensional roughness feature vectors will be passed to step S4 as important input features for fracture pattern classification.

[0047] Step S4: Multi-label classification and identification of fracture modes. This step receives the location distribution features and detection confidence scores of each feature region output from step S2, as well as the multi-dimensional roughness feature vector output from step S3. A multi-label classification network is then used to comprehensively determine the fracture mode type of the ceramic fracture surface. Ceramic materials exhibit diverse and mixed fracture modes; multiple fracture modes often coexist on the same fracture surface. Therefore, this invention employs a multi-label classification architecture instead of the traditional single-label classification architecture to more accurately reflect the complexity of the actual fracture process.

[0048] In this embodiment of the invention, the input features of the multi-label classification network consist of the following three parts: the first part is the feature vector of the feature region location distribution. The first part consists of the quantity statistics, area proportion, and spatial distribution entropy of various feature regions output from step S2, with a dimension of 15 (3 statistics for each feature category × 5 feature categories); the second part is the feature region detection confidence vector. The first part consists of the average and maximum detection confidence scores of various feature regions, with a dimension of 10 (2 confidence indices per feature category × 5 feature categories); the second part is a multidimensional roughness statistical feature vector. It consists of the average multidimensional roughness feature vector and standard deviation vector of various feature regions, with a dimension of 50 (10 roughness statistics for each feature category × 5 feature categories). The three feature vectors are concatenated to form a comprehensive feature vector with a total dimension of 75. The multi-label classification network adopts a fully connected neural network structure, containing three hidden layers. In one embodiment of the present invention, the number of neurons in each hidden layer is 256, 128, and 64, respectively. The ReLU activation function is used, and a Dropout layer is set after each hidden layer to prevent overfitting. The Dropout probability is set to 0.3. The output layer contains four neurons, corresponding to four fracture modes: brittle fracture, quasi-cleavage fracture, intergranular fracture, and mixed fracture. Each output neuron independently outputs the probability of belonging to the fracture mode using the Sigmoid activation function. ( The value range is 0 to 1. When Greater than the preset judgment threshold At that time, it was determined that the fracture surface had a first fault. Fracture modes. In one embodiment of the present invention, the default determination threshold for all four fracture modes is set to 0.5.

[0049] During the model training phase, considering that the label distribution of various fracture modes in actual fracture surface datasets is usually unbalanced (e.g., there are significantly more labeled samples of brittle fracture than mixed fracture), this invention employs an asymmetric weighted loss function for optimization. This loss function sets different focusing parameters for positive samples (label 1) and negative samples (label 0). and The calculation formula is as follows: ,in: This represents the total number of fracture mode categories, with a value of 4, and is dimensionless. For the first The true label value of the fracture-like pattern, which takes the value of 0 or 1 and is dimensionless; For the network prediction of the first The probability of classifying a fracture mode, ranging from 0 to 1, is dimensionless. This is the value after applying a hard threshold to the predicted probability of negative samples. This is the hard threshold margin parameter, ranging from 0 to 0.3. In this embodiment, it is set to... This parameter is used to ignore the contribution of easily separable negative samples to the gradient; This is the positive sample focusing parameter, with a value ranging from 0 to 4. In this embodiment, it is set to... Dimensionless; This is the negative sample focusing parameter, with a value ranging from 0 to 4. In this embodiment, it is set to... Dimensionless, large This makes the network pay more attention to the difficult-to-classify negative samples, reducing the interference of easily classified negative samples on the training process, thus effectively alleviating the label imbalance problem. To prevent overflow of extremely small positive numbers in logarithmic operations, the value is set to... , dimensionless.

[0050] Preferably, the multi-label classification network also incorporates a label smoothing regularization strategy during training to smooth out hard labels. Convert to soft tags ,in The label smoothing coefficient ranges from 0 to 0.2. In one embodiment of the present invention, it is taken as... This strategy can prevent the model from becoming overconfident in the training data and improve generalization performance.

[0051] The model was trained using the Adam optimizer, with an initial learning rate set to... The weight decay coefficient is The training iterations consist of 200 epochs, with a batch size of 32. In one embodiment of the present invention, on a test set containing 800 labeled fracture samples, the multi-label classification network achieves average accuracies of 94.2%, 89.7%, 91.3%, and 87.5% for brittle fracture, quasi-cleavage fracture, intergranular fracture, and mixed fracture, respectively, with a macro-average F1 score of 90.1%.

[0052] Step S5: Fracture Failure Source Tracing Analysis and Report Generation. This step receives the output results from steps S1 to S4, comprehensively infers the source of fracture failure, and generates a structured fracture surface analysis report. Fracture failure source tracing is the final goal of the entire analysis process. Its core logic is to match the spatial distribution pattern of fracture morphology features, the combination characteristics of fracture modes, and the trend of roughness variation with typical fracture surface feature patterns of known failure causes, thereby inferring the possible types of failure causes.

[0053] First, based on the fracture source localization results on the low-magnification panoramic image in step S1 and the spatial distribution relationship of each feature region in step S2, the system determines the direction of the crack propagation path. The method for determining the crack propagation path is as follows: starting from the fracture source location, a main crack propagation path is established from the fracture source to the fracture edge along the opposite direction of the river pattern convergence or along the spatial transition direction of the mirror area-fog area-serrated step. In one embodiment of the present invention, the crack propagation path consists of a series of ordered path nodes. It means that, among them This represents the number of path nodes, with the spacing between adjacent nodes ranging from approximately 50 to 100 pixels.

[0054] Subsequently, the system calculates the roughness variation trend along the crack propagation path. Specifically, for each node position on the path... Take the local region centered on this node (with a radius of ). Pixel, in this embodiment ), calculate the multi-scale roughness index of this local region. This forms a roughness variation curve along the crack propagation path. The roughness variation trends along the crack propagation path of the fracture process caused by different failure causes show significant differences: thermal shock failure usually shows a monotonically increasing trend of roughness from a low level at the fracture source (mirror area characteristics) to a sharp increase at the edge (serrated step characteristics); mechanical overload failure usually shows a nonlinear increasing characteristic of roughness change that is slow at first and then rapid; while fatigue accumulation failure often shows a periodic fluctuation characteristic of roughness with the propagation distance.

[0055] Based on this, the system performs matching inferences based on a preset failure cause knowledge base. In this embodiment of the invention, the failure cause knowledge base includes the following four failure cause types and their corresponding feature matching rules:

[0056] The first type is material defect-type failure. Its characteristic matching rules are: the fracture source is located inside the fracture surface (rather than on the surface), there are pores or inclusions near the fracture source, the fracture mode is mainly brittle fracture, and the area of ​​the mirror-like region accounts for more than 15% of the total fracture surface area.

[0057] The second type is thermal shock failure. Its characteristic matching rules are: the fracture source is located on the fracture surface or edge, there are multiple fracture sources, the fracture mode is mainly intergranular fracture, the roughness along the crack propagation path shows a monotonically increasing trend, and the river pattern distribution is radially divergent.

[0058] The third type is mechanical overload failure. Its characteristic matching rules are: the fracture source is located in the stress concentration area (usually a surface notch or geometric change point), the fracture mode is mainly brittle fracture or quasi-cleavage fracture, the mirror area-fog area-serrated step presents a complete spatial transition sequence, and the roughness shows a nonlinear increase that is first slow and then rapid.

[0059] The fourth type is fatigue cumulative failure. Its characteristic matching rules are: the fracture source is located on or near the surface, the fracture mode is determined to be a mixed fracture, the roughness along the crack propagation path shows periodic fluctuation characteristics, and fatigue arc-shaped stripes or shell-like morphology can be identified.

[0060] For each type of failure cause, the system calculates a matching confidence score. ( (Corresponding to the four failure types mentioned above), the calculation formula is as follows: ,in: For the first The number of matching rule entries for each failure type is dimensionless. In this embodiment, the number of rule entries for each failure type is 4 to 5. For the first The weight coefficients of the matching rules are dimensionless, ranging from 0 to 1, and satisfy the following conditions: Rule entries with higher weights correspond to features with stronger diagnostic discrimination (for example, the weight of fracture source location features is usually set to 1.2 to 1.5). For the first The matching degree of a rule, with a value ranging from 0 to 1, is dimensionless. It is 1 when the rule is a complete match, a fuzzy value between 0 and 1 when the rule is a partial match, and 0 when the rule is a complete mismatch. The value ranges from 0 to 1, with a value closer to 1 indicating a higher degree of match between the fracture characteristics and the failure type. The system selects... The most significant failure type is used as the primary inference result, and all conditions that satisfy this condition are also output. The failure type is used as a candidate inference result.

[0061] Finally, step S5 generates a structured fracture analysis report, which includes: a panoramic image of the fracture surface and a visualization of overlaid feature regions; area statistics and spatial distribution information of various feature regions; multidimensional roughness parameter data for each feature region; fracture mode determination results and the probability of each mode being assigned; a schematic diagram of crack propagation paths and roughness variation curves; failure cause inference results and matching confidence scores; and improvement suggestions for this failure type. The report is output simultaneously in structured data format (JSON or XML) and visual document format (PDF) for easy subsequent data management and engineering applications.

[0062] Preferably, the fracture mode determination result in step S4 can be fed back to step S1 to adaptively adjust the acquisition area and magnification of the high-magnification mode. For example, when step S4 determines that a mixed fracture exists on the fracture surface, the system automatically expands the coverage area of ​​the high-magnification acquisition and increases the acquisition density in the transition area between different fracture modes. The failure tracing result in step S5 can be fed back to step S2 to dynamically adjust the detection weights of various feature regions. For example, when step S5 initially infers a thermal shock failure, the system increases the detection sensitivity of intergranular fracture features and multi-fracture source features to obtain more accurate feature distribution information in subsequent re-analysis. This closed-loop mechanism of forward analysis and reverse optimization enables the system to gradually improve analysis accuracy and tracing reliability through iterative optimization.

[0063] See Figure 2 This invention also provides a ceramic fracture morphology image feature extraction and fracture mode recognition system. This system corresponds one-to-one with steps S1 to S5 in the above method embodiments, forming a complete modular software architecture, including a multi-scale image acquisition and stitching module, a fracture feature region detection module, a roughness quantification analysis module, a fracture mode classification module, and a failure source analysis module. The data flow and collaborative relationship between each module are completely consistent with the relationship between each step in the method embodiments. That is, the output of the multi-scale image acquisition and stitching module is transmitted to the fracture feature region detection module, the output of the fracture feature region detection module is transmitted to both the roughness quantification analysis module and the fracture mode classification module, the output of the roughness quantification analysis module is transmitted to the fracture mode classification module, and the output of the fracture mode classification module is transmitted to the failure source analysis module. Preferably, the failure source analysis module can also feed back to the fracture feature region detection module to dynamically adjust the detection weight, and the fracture mode classification module can also feed back to the multi-scale image acquisition and stitching module to adjust the acquisition strategy, thereby forming a closed-loop mechanism of forward analysis and reverse optimization.

[0064] The multi-scale image acquisition and stitching module, corresponding to step S1, is used to perform scanning electron microscopy imaging of the ceramic fracture sample under low magnification (20x to 100x) and high magnification (500x to 5000x) modes on the sample. SIFT feature point detection and matching are performed on the acquired images with overlapping areas. The RANSAC algorithm is used to estimate the homography transformation matrix between images, and a weighted average fusion strategy is employed to complete the panoramic stitching, ultimately generating high-resolution low-magnification and high-magnification panoramic images covering the complete fracture area. This module integrates a preliminary fracture source localization function, automatically locating candidate fracture source regions on the low-magnification panoramic image through Gabor filtering and gradient convergence direction analysis, and planning the spatial range and path for high-magnification acquisition accordingly. Preferably, this module also includes an image quality assessment subunit, which evaluates the image sharpness and contrast in real time during acquisition, automatically triggering re-acquisition for images that do not meet quality requirements, ensuring the reliability of the images used for subsequent analysis.

[0065] The fracture feature region detection module corresponds to step S2. It inputs the panoramic image into a fracture feature region detection network built on a ResNet-50 backbone network and a bidirectional feature pyramid structure. Through multi-scale feature extraction and region proposal and classification regression mechanisms, it automatically identifies river-pattern regions, mirror regions, hazy regions, serrated step regions, and transgranular cleavage surface regions on the fracture surface, outputting the category label, bounding box coordinates, and detection confidence score for each feature region. This module pre-loads the detection model weight parameters obtained through the training process described in the above method embodiments, supporting inference on both GPU and CPU hardware environments. Preferably, this module employs a sliding window block detection and global coordinate merging strategy for large-size panoramic images, eliminating duplicate detections through NMS post-processing to ensure feature coverage detection of the complete fracture region.

[0066] The roughness quantification analysis module corresponds to step S3, which is used to extract texture energy based on the gray-level co-occurrence matrix for each feature region detected by the fracture feature region detection module. Texture entropy and texture contrast Three core texture parameters, combined with shading and shape restoration techniques, are used to estimate the height of local 3D surface undulations based on grayscale gradient distribution. And calculate the standard deviation of the height distribution. This module normalizes and fuses texture parameters with undulation height, according to the multi-scale roughness quantification index described in the above method embodiments. The calculation formula yields the multidimensional roughness feature vectors of each feature region. This module also includes a calibration database that stores grayscale-tilt conversion coefficients for different scanning electron microscope models and imaging parameters. Calibration values ​​are used to ensure the comparability and consistency of roughness estimation results under different imaging conditions.

[0067] The fracture pattern classification module corresponds to step S4, which is used to classify the feature vector of the feature region location distribution output by the fracture feature region detection module. Detection confidence vector The multidimensional roughness statistical feature vector output by the roughness quantification analysis module The features are concatenated to form a comprehensive feature vector with a total dimension of 75, which is then input into a fully connected multi-label classification network containing three hidden layers. This classification network is trained using the asymmetric weighted loss function described in the above-mentioned method embodiments. The four sigmoid neurons in the output layer output the probability of attribution for brittle fracture, quasi-cleavage fracture, intergranular fracture, and mixed fracture, respectively. Preferably, this module also integrates model interpretability analysis functions, calculating the gradient contribution of each input feature to the classification result and outputting the ranking information of key features affecting fracture mode determination, assisting analysts in understanding and verifying the rationality of the automatic determination results.

[0068] The failure tracing analysis module, corresponding to step S5, is used to comprehensively analyze the fracture source location determined by the multi-scale image acquisition and stitching module, the spatial distribution relationship of each feature region output by the fracture feature region detection module, the fracture mode classification probability output by the fracture mode classification module, and the roughness change trend output by the roughness quantification analysis module along the crack propagation path. Based on the built-in failure cause knowledge base, feature matching and confidence score calculation are performed to ultimately infer the possible failure cause types and generate a structured fracture analysis report. Internally, this module uses the matching confidence score formula described in the above method embodiment to calculate the matching degree of each failure type and output it in a sorted order. The failure cause knowledge base of this module supports expansion and maintenance; analysts can add new failure types and their feature matching rules based on actual engineering experience, allowing the system's tracing analysis capabilities to continuously improve with accumulated engineering practice. Preferably, this module also includes a human-computer interaction interface. When the highest matching confidence score of automatic tracing is lower than a preset confidence threshold (e.g., 0.6), the system automatically prompts the analyst to conduct manual review and supplementary judgment to ensure the reliability of the tracing conclusions. The report generation subunit of this module supports outputting structured data in JSON format and visual documents in PDF format. The visual documents include feature area annotations superimposed on the panoramic image of the fracture surface, schematic lines of crack propagation path, roughness change curves, and textual explanations of failure cause inferences and improvement suggestions, which are convenient for engineering technicians to refer to and archive directly.

[0069] Furthermore, the collaborative working mechanism among the five modules in this embodiment exhibits significant deep coupling characteristics. The multi-scale image acquisition and stitching module, as the data source module, directly impacts the analysis accuracy of all subsequent modules by outputting panoramic images of quality. The fracture feature region detection module, as the core hub module, provides spatial positioning for both the roughness quantification analysis module and the fracture pattern classification module, achieving an efficient data utilization mode of one-time detection and multiple reuse. The roughness quantification analysis module supplements the fracture pattern classification module with texture depth and three-dimensional shape information that are difficult to capture by traditional target detection frameworks; the feature fusion of the two produces a synergistic enhancement effect that surpasses a single feature source. The fracture pattern classification module and the failure source analysis module achieve a cognitive leap from phenomenon description to cause inference. In addition, the feedback channels from the fracture pattern classification module to the multi-scale image acquisition and stitching module, and from the failure source analysis module to the fracture feature region detection module, enable the system to have adaptive optimization capabilities, continuously improving the processing accuracy and collaborative efficiency of each module during iterative analysis.

[0070] In the testing and verification phase, the system provided in this invention underwent comprehensive performance evaluation using fracture samples of 200 different ceramic materials (including alumina, silicon nitride, silicon carbide, and zirconium oxide). Test results showed that the fracture feature region detection module achieved an average detection accuracy (mAP@0.5) of 88.6% for five typical fracture morphology features. The highest detection accuracy was observed in the specular region (93.2%), while the detection accuracy for transgranular cleavage surfaces was relatively lower (84.1%), but still significantly better than the consistency level of human visual identification. The fracture mode classification module achieved a macro-average F1 score of 90.1% for four fracture modes, an improvement of approximately 5.8 percentage points compared to a control scheme using only an image classification network without introducing roughness features, validating the feature enhancement effect brought about by the fusion of gray-level co-occurrence matrix texture analysis and light and dark shape recovery technology. The failure source analysis module achieved a primary inference hit rate of 82.5% and a cumulative hit rate of 94.0% for the top two candidate inferences. The entire system takes about 3 to 8 minutes to complete the analysis of a single fracture sample (depending on the fracture area and the number of images acquired). Compared with the traditional manual analysis method (which usually requires experienced analysts to spend 30 to 120 minutes), the efficiency is improved by about 10 to 15 times, which has significant engineering application value.

[0071] The embodiments of the present invention are not limited to the specific embodiments described above. Those skilled in the art can make various equivalent changes or substitutions based on the technical solutions of the present invention, and all such changes or substitutions should be included within the protection scope of the present invention.

Claims

1. A method for ceramic fracture morphology image feature extraction and fracture mode recognition, characterized in that, Includes the following steps: Step S1, Multi-scale fracture morphology image acquisition and panoramic stitching: The ceramic fracture sample to be analyzed is imaged by scanning electron microscopy at multiple magnifications to acquire macroscopic and microscopic fracture morphology images, locate the fracture source region and output the fracture source location result; Feature point detection and matching are performed on multiple images with overlapping areas acquired at the same magnification, and high-resolution panoramic images are generated by image registration and fusion stitching. Step S2, Fracture Feature Region Detection and Localization: The panoramic image is input into a fracture feature region detection network constructed based on a deep convolutional neural network to automatically identify the feature regions on the fracture surface and output the bounding box coordinates and detection confidence of each feature region; wherein, the feature regions include river pattern regions, mirror regions, fog-like regions, serrated step regions, and transgranular cleavage surface regions; Step S3, quantitative analysis of fracture surface roughness: For each feature region detected in step S2, texture energy, texture entropy value and texture contrast parameters are extracted based on gray-level co-occurrence matrix. Combined with the light and dark recovery shape technique, the local three-dimensional surface undulation height of each feature region is estimated according to the gray-level gradient distribution of the image. The texture parameters and the undulation height are normalized and fused to obtain a multi-dimensional roughness feature vector characterizing the surface roughness of each feature region. Step S4, fracture mode multi-label classification and recognition: The bounding box coordinates of each feature region obtained in step S2 are concatenated with the detection confidence and the multidimensional roughness feature vector obtained in step S3, and then input into the multi-label classification network to determine the fracture mode, and output the probability of each fracture mode; the fracture modes include brittle fracture, quasi-cleavage fracture, intergranular fracture and mixed fracture.

2. The method of claim 1, wherein, In step S1, the magnification range of the low-magnification mode is 20x to 100x, and the overlap rate between adjacent images is not less than 25%; the magnification range of the high-magnification mode is 500x to 5000x, and the image resolution is not less than 1024×1024 pixels; the number of images acquired in the low-magnification mode is N1, and the number of images acquired in the high-magnification mode is N2, where N2 is greater than N1.

3. The method of claim 1, wherein, In step S2, the detection confidence threshold of the detection network is set between 0.5 and 0.8, and the crossover ratio threshold of non-maximum suppression is set between 0.3 and 0.5; in the fracture feature annotation dataset used by the detection network during the training phase, the number of annotated samples for each type of feature region is not less than 200.

4. The method of claim 1, wherein, In step S3, the calculation window size of the gray-level co-occurrence matrix is ​​5×5 pixels to 15×15 pixels, the gray-level quantization level is 8 to 64 levels, and the calculation directions include four directions: 0°, 45°, 90° and 135°. The calculation results of the four directions are averaged to eliminate the direction dependency.

5. The method of claim 1, wherein, In step S1, the feature point detection and matching, as well as the image registration and fusion stitching, specifically include: extracting feature points and generating feature descriptors from adjacent images using a scale-invariant feature transformation algorithm; matching feature points using the nearest neighbor distance ratio method; estimating the homography transformation matrix between images using a random sampling consensus algorithm; and mixing pixel values ​​in overlapping areas using a weighted average fusion strategy.

6. The method of claim 1, wherein, In step S4, the multi-label classification network is trained using an asymmetric weighted loss function. The asymmetric weighted loss function sets different focusing parameters for positive and negative samples to solve the problem of uneven label distribution of fracture patterns. The output layer of the multi-label classification network uses a sigmoid activation function to independently output the attribution probability for each fracture pattern.

7. The method of claim 1, wherein, The process also includes step S5, fracture failure tracing analysis and report generation: Based on the fracture source location results output in step S1 and the spatial distribution relationship determined by the bounding box coordinates of each feature region output in step S2, the crack propagation path direction is determined. Combining the fracture mode attribution probability output in step S4 and the roughness change trend along the crack propagation path calculated based on the multidimensional roughness feature vector output in step S3, matching inference is performed based on a preset failure cause knowledge base to determine the possible failure cause types and corresponding confidence scores. A structured fracture analysis report is generated, which includes panoramic image annotation of the fracture surface, feature region distribution, fracture mode determination results, and failure cause inference. The preset failure cause knowledge base includes four failure cause types: material defect failure, thermal shock failure, mechanical overload failure, and fatigue accumulation failure. Each failure cause type corresponds to a predefined fracture mode combination feature and crack propagation path feature matching rule.

8. The method of claim 7, wherein, In step S2, the fracture feature region detection network includes a feature extraction backbone network, a multi-scale feature pyramid neck network, and a detection head network. The feature extraction backbone network of the fracture feature region detection network adopts a residual network structure, and the multi-scale feature pyramid neck network of the fracture feature region detection network includes a bidirectional feature fusion channel with a top-down path and a bottom-up path.

9. The method of claim 8, wherein, The fracture mode determination result in step S4 is fed back to step S1 to adaptively adjust the acquisition area range and magnification of the high-magnification mode; the failure tracing result in step S5 is fed back to step S2 to dynamically adjust the detection weight of various feature regions.

10. A system for ceramic fracture topography image feature extraction and fracture mode recognition for implementing the method of claim 7, characterized in that, include: The multi-scale image acquisition and stitching module is used to perform scanning electron microscopy imaging of the ceramic fracture sample to be analyzed in low-magnification and high-magnification modes respectively. It performs feature point detection and matching, image registration and fusion stitching on multiple images with overlapping areas to generate a high-resolution panoramic image covering the complete fracture area. The fracture feature region detection module is used to input the panoramic image into the fracture feature region detection network based on a deep convolutional neural network. Through a multi-scale feature pyramid structure and a region proposal and classification regression mechanism, it automatically identifies the river pattern region, mirror region, fog region, sawtooth step region and transgranular cleavage surface region on the fracture surface, and outputs the bounding box coordinates and detection confidence of each feature region. The roughness quantification analysis module is used to extract texture parameters based on the gray-level co-occurrence matrix for each feature region detected by the fracture feature region detection module, estimate the local three-dimensional surface undulation height by combining the light and dark recovery shape technology, and normalize and fuse the texture parameters and undulation height to obtain a multi-dimensional roughness feature vector. The fracture mode classification module is used to concatenate the location distribution features and detection confidence output by the fracture feature region detection module and the multidimensional roughness feature vector output by the roughness measurement and analysis module, and then input them into a multi-label classification network to determine the fracture mode and output the probability of each fracture mode. The failure source analysis module is used to perform matching inference based on the fracture source area location results, the spatial distribution relationship of characteristic areas, the fracture mode attribution probability and the roughness change trend, and generate a structured fracture analysis report.

Images