Metal material fracture morphology feature intelligent identification method and system thereof

Abstract

This invention discloses an intelligent identification method and system for fracture morphology features of metallic materials, belonging to the field of failure analysis and intelligent image recognition technology of metallic materials. The method performs multi-scale feature extraction on fracture morphology images acquired by scanning electron microscope, identifies eight types of micro-features such as dimples, cleavage surfaces, and fatigue striations, determines five types of fractures such as ductile fracture, brittle fracture, and fatigue fracture through a multi-label classification network, calculates quantitative parameters such as dimple size distribution and fatigue striation spacing, and establishes a correlation model between fracture features and fracture mechanical parameters.

Description

Technical Field

[0001] This invention relates to the field of failure analysis and intelligent image recognition technology for metallic materials, specifically to an intelligent recognition method and system for fracture morphology features of metallic materials. Background Technology

[0002] Fracture morphology analysis of metallic materials is a core technique in the field of materials failure analysis. By observing and analyzing the microscopic morphological characteristics of the fracture surface, the fracture mechanism, failure cause, and service conditions of the material can be revealed. Scanning electron microscopy (SEM), as the primary tool for observing fracture morphology, can acquire high-resolution microscopic images of the fracture surface, providing rich morphological information for fracture analysis. However, traditional fracture morphology analysis mainly relies on manual interpretation by experts, which suffers from low efficiency, high subjectivity, and poor consistency, making it difficult to meet the needs of large-scale, high-efficiency failure analysis in industrial applications.

[0003] Fracture morphology contains a variety of typical microscopic features, which are closely related to the fracture mechanism of materials. Dimples are a typical feature of ductile fracture, formed by the nucleation, growth, and aggregation of micropores; the size and morphology of dimples reflect the material's plastic deformation capacity. Cleavage planes are a typical feature of brittle fracture, manifesting as smooth fracture surfaces along specific crystal planes, often accompanied by river patterns and tongue-like patterns. Fatigue striations are a typical feature of fatigue fracture, formed by cyclic plastic deformation at the crack tip; the striation spacing is directly related to the crack propagation rate. Intergranular fracture manifests as fracture paths along grain boundaries and is commonly seen in failure modes such as hydrogen embrittlement, stress corrosion cracking, and high-temperature creep. The accurate identification and quantitative analysis of these microscopic features are of great significance for determining failure causes and improving material design.

[0004] In recent years, deep learning technology has made significant progress in the field of image recognition, providing new technical approaches for the automated analysis of fracture morphology. Researchers have begun to explore the application of techniques such as convolutional neural networks and semantic segmentation to the classification and feature recognition of fracture images. Compared with traditional image processing methods, these methods can better handle complex and varied fracture morphologies and extract more discriminative feature representations.

[0005] Chinese invention CN113935989A discloses a deep learning-based method for identifying and segmenting fatigue stripes on fracture surfaces of metallic materials. This technical solution first establishes a mathematical model of fatigue stripes on the fracture surface, then simulates the extension trajectory of the fatigue stripes using an improved RRT algorithm to obtain texture curve images of the fatigue stripes. For dataset construction, scanning electron microscopy is used to acquire and annotate fracture surface images, which are then combined with simulated texture curve images for dataset augmentation. In terms of model architecture, a U-net network structure is selected to achieve end-to-end semantic segmentation, and fatigue stripe identification and segmentation are completed through an encoder-decoder structure and skip connections.

[0006] However, the aforementioned existing technologies have the following shortcomings. First, regarding the identification object, existing technologies only identify a single type of fracture feature—fatigue striations—and cannot cover other important fracture micro-features such as dimples, cleavage planes, quasi-cleavage, intergranular fracture, transgranular fracture, river patterns, and herringbone patterns, thus limiting the identification range. Second, regarding fracture type determination, existing technologies have not established a correlation between fracture features and fracture types, and cannot automatically determine multiple fracture types such as ductile fracture, brittle fracture, stress corrosion fracture, and hydrogen embrittlement fracture, especially failing to identify mixed fracture modes commonly seen in practical engineering. Third, regarding quantitative analysis, existing technologies lack the ability to calculate quantitative parameters for fracture features, and cannot output key quantitative data such as dimple size distribution, fatigue striation spacing, and cleavage plane orientation distribution, making it difficult to support quantitative fracture mechanics analysis. Finally, regarding feature extraction, existing technologies use a single-scale U-net network for feature extraction, failing to fully utilize the multi-scale characteristics of fracture morphology, resulting in insufficient accuracy in identifying fracture features with large scale variations.

[0007] Therefore, there is an urgent need for an intelligent identification method and system that can cover the identification of microscopic features of multiple types of fracture surfaces, support the automatic determination of multiple fracture types, have the ability to perform feature quantification analysis, and establish the correlation between fracture surface features and fracture mechanical parameters. Summary of the Invention

[0008] The purpose of this invention is to provide an intelligent identification method and system for fracture morphology features of metallic materials, in order to solve the technical problems of existing technologies such as single fracture feature identification type, lack of fracture type determination, insufficient feature quantification analysis, and limited multi-scale feature extraction capability.

[0009] To achieve the above objectives, this invention provides an intelligent recognition method for fracture morphology features of metallic materials, comprising the following steps: preprocessing fracture morphology images acquired by scanning electron microscopy, the preprocessing including grayscale normalization, contrast enhancement, and noise suppression, and outputting a preprocessed fracture morphology image; performing multi-scale feature extraction on the preprocessed fracture morphology image, constructing a feature pyramid containing multiple scale levels, extracting fracture texture features and morphological features at each scale level, and outputting a multi-scale feature map; performing feature fusion and microscopic feature recognition on the multi-scale feature map, weighting and fusing features at different scales through a hierarchical attention mechanism, and identifying fracture microscopic features based on the fused features, the fracture microscopic features including dimples, cleavage planes, quasi-cleavage, fatigue striations, intergranular fracture, transgranular fracture, river patterns, and fishbone patterns, and outputting the spatial location of each type of microscopic feature. The system identifies the location, category label, and confidence level of the micro-features. Based on these labels and confidence levels, it determines the fracture type using a multi-label classification network. The fracture types include ductile fracture, brittle fracture, fatigue fracture, stress corrosion fracture, and hydrogen embrittlement fracture. The system outputs the fracture type determination results and the probability distribution of each type. Based on the spatial location of the micro-features and the fracture type determination results, it calculates feature quantification parameters, including dimple size distribution parameters, fatigue fringe spacing parameters, and cleavage plane orientation distribution parameters. The system outputs feature quantification data. Based on the feature quantification data and the fracture type determination results, it predicts fracture toughness, fatigue life, and stress intensity factor using a correlation model between fracture features and fracture mechanics parameters. The system outputs a comprehensive analysis report including annotated fracture feature images, fracture type determination results, feature quantification data, and failure cause analysis suggestions.

[0010] Preferably, the feature pyramid contains four scale levels, with scale factors of 1, 0.5, 0.25 and 0.125 for each scale level, and spatial correspondence between adjacent scale levels is established through downsampling operations.

[0011] Preferably, the hierarchical attention mechanism calculates attention weights based on the spatial location of features at each scale level. The attention weights are used to adjust the contribution of features at different scales in the fusion process. When the confidence level is lower than a preset confidence threshold, the attention weights of features at adjacent scale levels are increased.

[0012] Preferably, the dimple size distribution parameters include the average dimple diameter, the dimple depth-to-width ratio, and the dimple area percentage; the fatigue fringe spacing parameters include the average fringe spacing and the spacing variation coefficient; and the cleavage surface orientation distribution parameters include the principal orientation angle of the cleavage surface and the orientation dispersion.

[0013] Preferably, the multi-scale feature extraction employs a residual convolutional network, with a batch normalization layer and a nonlinear activation layer set after the convolutional layers at each scale level, and low-level detailed features are passed to high-level semantic features through skip connections.

[0014] Preferably, the multi-label classification network includes a global average pooling layer, a fully connected layer, and a Sigmoid activation layer, and supports the simultaneous output of probability values ​​for multiple fracture types. When the probability values ​​for multiple fracture types all exceed a preset type determination threshold, it is determined to be a mixed fracture mode.

[0015] Preferably, the association model adopts a multi-task regression network structure, using the feature quantification data as input and fracture toughness, fatigue life and stress intensity factor as multi-task outputs, and extracts the association information between features through a shared feature representation layer.

[0016] Preferably, the preprocessing further includes image block processing, which divides the fracture morphology image into image blocks of a preset size, performs grayscale normalization, contrast enhancement and noise suppression on each image block, and then stitches the processed image blocks together to form a complete image.

[0017] Preferably, the failure cause analysis suggestion is generated based on the fracture type determination result and the feature quantification data. When it is determined to be fatigue fracture, the fatigue load amplitude range is estimated based on the fatigue stripe spacing parameter. When it is determined to be hydrogen embrittlement fracture, the degree of hydrogen embrittlement is evaluated based on the proportion of intergranular fracture region.

[0018] This invention also provides an intelligent recognition system for fracture morphology features of metallic materials, comprising: an image preprocessing module for performing grayscale normalization, contrast enhancement, and noise suppression on fracture morphology images acquired by scanning electron microscopy, and outputting a preprocessed fracture morphology image; a multi-scale feature extraction module for constructing a feature pyramid containing multiple scale levels on the preprocessed fracture morphology image, extracting fracture texture features and morphological features at each scale level, and outputting a multi-scale feature map; and a feature fusion and recognition module for weighted fusion of features at different scales through a hierarchical attention mechanism, and recognizing dimples, cleavage surfaces, quasi-cleavage, and fatigue bars based on the fused features. The system includes several microscopic features, including fracture patterns, intergranular fracture, transgranular fracture, river patterns, and fishbone patterns. It outputs the spatial location, category label, and confidence level of each feature. A fracture type determination module uses a multi-label classification network to identify ductile fracture, brittle fracture, fatigue fracture, stress corrosion cracking, or hydrogen embrittlement fracture, outputting the fracture type determination results and probability distributions for each type. A feature quantization module calculates dimple size distribution parameters, fatigue fringe spacing parameters, and cleavage plane orientation distribution parameters, outputting feature quantification data. A correlation analysis module predicts fracture toughness, fatigue life, and stress intensity factor using a correlation model between fracture features and fracture mechanics parameters, outputting a comprehensive analysis report.

[0019] The intelligent identification method and system for fracture morphology features of metallic materials provided by this invention have the following beneficial effects:

[0020] First, in terms of the coverage of fracture feature recognition, the present invention can simultaneously identify eight typical fracture micro-features, including dimples, cleavage surfaces, quasi-cleavage, fatigue striations, intergranular fracture, transgranular fracture, river patterns, and fishbone patterns. Compared with the existing technology, which only identifies the single feature of fatigue striations, the recognition type coverage is improved to a complete fracture morphology feature system.

[0021] Secondly, in terms of fracture type determination capability, this invention achieves automatic determination of five fracture types, namely ductile fracture, brittle fracture, fatigue fracture, stress corrosion fracture, and hydrogen embrittlement fracture, through a multi-label classification network, and supports the identification of mixed fracture modes, thus solving the problem that existing technologies cannot automatically determine fracture types.

[0022] Third, in terms of feature quantification analysis, this invention can automatically calculate quantitative parameters such as dimple size distribution, fatigue stripe spacing, and cleavage surface orientation distribution, and establish a correlation model between fracture features and fracture mechanics parameters, thus achieving a technological leap from qualitative identification to quantitative analysis.

[0023] Fourth, in terms of multi-scale feature extraction, this invention constructs a feature pyramid containing four scale levels and achieves adaptive fusion of multi-scale features through a hierarchical attention mechanism, which effectively improves the recognition accuracy of fracture features at different scales. Attached Figure Description

[0024] Figure 1 This is a flowchart of the intelligent identification method for fracture morphology features of metallic materials provided in the embodiments of the present invention.

[0025] Figure 2 This is an architecture diagram of the intelligent recognition system for fracture morphology features of metallic materials provided in an embodiment of the present invention. Detailed Implementation

[0026] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0027] This embodiment provides an intelligent recognition method for fracture morphology features of metallic materials, such as... Figure 1 As shown, it includes the following steps.

[0028] Step S1: Preprocessing of fracture surface morphology image.

[0029] This step preprocesses the fracture morphology image acquired by scanning electron microscope. The preprocessing includes grayscale normalization, contrast enhancement, and noise suppression, and outputs the preprocessed fracture morphology image.

[0030] In this embodiment, the fracture morphology images acquired by scanning electron microscopy are typically 8-bit or 16-bit grayscale images, with image resolutions ranging from 512×512 pixels to 4096×4096 pixels and magnifications ranging from 100x to 10000x. Since the grayscale distribution, contrast, and noise levels of images vary under different acquisition conditions, preprocessing is required to ensure the consistency and stability of subsequent feature extraction.

[0031] Gray-level normalization processing employs an adaptive gray-level mapping method, mapping the gray-level values ​​of the original image to a standard gray-level range. Let the original image be... ,in Representing spatial coordinates (i.e., image pixel positions), the image after grayscale normalization. Calculated using the following formula: ,in: This represents the minimum grayscale value of the original image. 255 represents the maximum grayscale value of the original image; 255 represents the maximum grayscale level of the 8-bit grayscale image. In this embodiment, to avoid the influence of extreme grayscale values ​​on the normalization result, the 1% quantile and 99% quantile of the grayscale histogram are used as... and The value of .

[0032] Contrast enhancement employs a contrast-limited adaptive histogram equalization method, dividing the image into multiple sub-regions and performing histogram equalization within each sub-region. Bilinear interpolation is then used to eliminate discontinuities at the sub-region boundaries. The image after contrast enhancement is shown. Calculated using the following formula: ,in: For the first Gray-scale transformation function for adjacent sub-regions; For the first Each sub-region for pixels The interpolation weights are determined by the distances from the pixel to the centers of its four neighboring sub-regions; the summation range... The values ​​from 1 to 4 represent the contributions of four adjacent sub-regions. In this embodiment, the size of the sub-region is set to 64×64 pixels, and the contrast limiting factor is set to 3.0. This parameter value can enhance image details while avoiding excessive amplification of noise.

[0033] Noise suppression employs a combination of Gaussian filtering and bilateral filtering. First, Gaussian filtering is used to suppress high-frequency random noise, then bilateral filtering is used to smooth the noise while preserving edge information. The image after noise suppression is shown. As the output of the preprocessing, the Gaussian filter has a convolution kernel size of 5×5 and a standard deviation of 1.0. The bilateral filter has a spatial standard deviation of 10 and a grayscale standard deviation of 30. These parameter values ​​were selected to maintain the clarity of the fracture morphology edges while sufficiently suppressing typical noise in scanning electron microscope images.

[0034] In this embodiment, the preprocessing also includes an image block processing step. For large-sized fracture topography images with a resolution exceeding 2048×2048 pixels, they are divided into 512×512 pixel image blocks, with a 64-pixel overlap between each block to ensure the continuity of feature extraction at the boundaries. Each image block is then subjected to grayscale normalization, contrast enhancement, and noise suppression. After processing, the overlapping areas are merged using a weighted average method, and then stitched together to form a complete preprocessed image.

[0035] Step S2: Multi-scale feature extraction.

[0036] This step performs multi-scale feature extraction on the preprocessed fracture morphology image, constructs a feature pyramid containing multiple scale levels, extracts fracture texture features and morphological features at each scale level, and outputs a multi-scale feature map.

[0037] In this embodiment, an adaptive multi-scale pyramid feature extraction algorithm is used to construct a feature pyramid. The microscopic features in fracture morphology exhibit significant multi-scale characteristics: the typical size range of dimples is 1 micrometer to 50 micrometers, the spacing of fatigue striations ranges from 0.1 micrometers to 10 micrometers, and the size of cleavage surfaces ranges from 10 micrometers to 500 micrometers. To effectively capture fracture features at different scales, the feature pyramid contains four scale levels, with each scale level having a scale factor... Set them to 1, 0.5, 0.25, and 0.125 respectively, where the subscripts... Indicates the scale level number. The value range is from 1 to 4.

[0038] Let the size of the preprocessed fracture morphology image be... ,in Image height, The image width. The feature map size of the layer-scale level is Spatial correspondence is established between adjacent scale levels through max pooling with a step size of 2, and the spatial size of the feature map is halved with each downsampling.

[0039] Multi-scale feature extraction employs a residual convolutional network structure. At each scale level, the feature extraction process includes a cascaded combination of convolutional layers, batch normalization layers, and nonlinear activation layers. Feature maps at the layer scale Calculated using the following formula: ,in: This indicates a convolution operation with a kernel size of 3×3 and the number of kernels at the four scale levels being 64, 128, 256, and 512, respectively. This indicates a batch normalization operation, used to accelerate network training and improve the stability of feature extraction; Represents the ReLU nonlinear activation function; This indicates a jump connection operation, which passes low-level detailed features to higher levels to preserve fine structural information about the fracture morphology. Represents the feature map of the previous scale level, when hour, This is the image of the fracture surface morphology after preprocessing.

[0040] At each scale level, fracture texture and morphological features are extracted simultaneously. Fracture texture features describe the microscopic undulations of the fracture surface, and multi-directional, multi-frequency texture responses are extracted using a Gabor filter bank. The Gabor filter has eight directions: 0°, 22.5°, 45°, 67.5°, 90°, 112.5°, 135°, and 157.5°, with four geometrically ordered frequency parameters. Fracture morphological features describe the geometry of the microscopic features, and edge contour information is extracted using a morphological gradient operator.

[0041] Multi-scale feature maps As the output of this step, it is passed to the feature fusion and recognition step. The number of channels in the feature maps at each scale level are 64, 128, 256 and 512, respectively, and the spatial size decreases with the scale factor.

[0042] In this embodiment, the network parameters for multi-scale feature extraction are configured as follows. The first scale layer uses two cascaded 3×3 convolutional layers, each with 64 output channels, a stride of 1, and padding of 1. The second scale layer adds a max-pooling layer with a stride of 2 for downsampling based on the first scale layer, followed by two 3×3 convolutional layers, resulting in 128 output channels. The structures of the third and fourth scale layers are similar to the second scale layer, with 256 and 512 output channels, respectively. The weights of each convolutional layer are initialized using the Kaiming initialization method, and the biases are initialized to zero.

[0043] The parameter configuration of the Gabor filter bank has a significant impact on the texture feature extraction effect. In this embodiment, the wavelength parameter of the Gabor filter is set to four values: 4 pixels, 8 pixels, 16 pixels, and 32 pixels, covering the feature scale range from fine to coarse texture in the fracture morphology. The orientation parameter is set to eight values: 0 degrees, 22.5 degrees, 45 degrees, 67.5 degrees, 90 degrees, 112.5 degrees, 135 degrees, and 157.5 degrees, which can detect texture features in various directions. The bandwidth parameter is adaptively adjusted according to the wavelength and is set to 0.56 times the wavelength. This parameter setting achieves a balance between frequency domain selectivity and spatial domain positioning accuracy.

[0044] Morphological gradient operators are used to extract the edge contours of microscopic features of the fracture surface. In this embodiment, the morphological gradient is defined as the difference between the dilated image and the eroded image, and the structuring element adopts a 3×3 cross-shaped structure, which responds well to linear edge features in the fracture morphology. By fusing the edge information extracted by the morphological gradient operator with the semantic features extracted by the convolutional network, the boundaries of microscopic features can be located more accurately.

[0045] Step S3: Feature fusion and micro-feature recognition.

[0046] This step performs feature fusion and micro-feature recognition on multi-scale feature maps. It uses a hierarchical attention mechanism to weightedly fuse features at different scales, and identifies the micro-features of the fracture surface based on the fused features, outputting the spatial location, category label, and confidence level of various micro-features.

[0047] In this embodiment, a hierarchical attention fusion algorithm based on the microscopic features of the fracture surface is used to achieve adaptive fusion of multi-scale features. This algorithm dynamically calculates the weights of features at each scale level during the fusion process based on their semantic importance and spatial relevance.

[0048] The calculation process for hierarchical attention weights is as follows. First, global average pooling is performed on the feature maps of each scale level to obtain the global feature vectors of each level. : ,in: and The first The height and width of the layer-scale hierarchical feature map; For feature map at location The feature vector at that location; summation is performed to traverse all spatial locations of the feature map at that level.

[0049] The global feature vectors from each level are concatenated and then input into the attention weight calculation network to obtain the attention weights for each scale level. : ,in: This is the attention weight matrix, with dimension 1. , The number of feature channels; The bias vector is used; the denominator is the normalized coefficient of the attention scores at the four scale levels, ensuring that the sum of the attention weights is 1. In this embodiment, when the confidence of a feature at a certain scale level is lower than the preset confidence threshold of 0.5, the attention weights of features at adjacent scale levels are automatically increased to supplement the feature representation of uncertain regions using multi-scale information.

[0050] Feature fusion employs a bidirectional fusion strategy combining bottom-up and top-down approaches. The bottom-up path progressively passes high-resolution detail features from lower levels to higher levels, while the top-down path progressively decodes semantic features from higher levels back to lower levels. The fused multi-scale features... Calculated using the following formula: ,in: This indicates an upsampling operation, which upsamples feature maps from each scale level to the target scale. The target scale is set to the original image scale, i.e. Summation is used to weight and fuse features from the four scale levels.

[0051] Based on the fused multi-scale features, fracture surface micro-features are identified. The fracture surface micro-features identified in this embodiment include eight categories: dimples, cleavage, quasi-cleavage, fatigue striations, intergranular fracture, transgranular fracture, river pattern, and herringbone pattern.

[0052] Microscopic feature recognition employs a semantic segmentation network structure. The fused feature map is input into a classification head network, which outputs the class probability distribution for each pixel location. The classification head network consists of a 1×1 convolutional layer and a softmax activation layer, with 9 output channels corresponding to 8 microscopic feature classes plus 1 background class. For each pixel location... The output includes category labels. and confidence level The category label is the category with the highest probability, and the confidence score is the probability value of that category.

[0053] Based on pixel-level classification results, the spatial locations and bounding boxes of various micro-features are extracted through connected component analysis. Let the... The detected microscopic feature regions are Its output includes: category labels Confidence level (Average confidence score of pixels within the region), bounding box coordinates and area .

[0054] In this embodiment, the connected component analysis uses the 8-neighborhood connectivity rule, i.e., pixel... Pixels with the same category label as the surrounding eight pixels (top, bottom, left, right, and four diagonal directions) are considered connected. The minimum area threshold for connected component analysis is set to 100 square pixels; regions smaller than this threshold are considered noise and filtered out. The bounding box is obtained by calculating the minimum bounding rectangle of the connected components and is used to label the location of each micro-feature in the visualization results.

[0055] The post-processing for micro-feature recognition also includes non-maximum suppression (NMS) to eliminate overlapping detection results. When the intersection-union ratio (IUU) of two similar micro-feature regions exceeds a preset threshold of 0.5, the region with higher confidence is retained, while the region with lower confidence is deleted. This operation effectively reduces redundant detections and improves the accuracy and readability of the recognition results.

[0056] To address the morphological characteristics of different types of microscopic features, this embodiment also sets category-specific post-processing rules. For dimple features, a roundness index greater than 0.6 is required, where roundness is defined as the ratio of the region area to the area of ​​the circumcircle; regions that do not meet the roundness requirement are reclassified to other categories. For fatigue stripe features, an aspect ratio greater than 3 is required to distinguish true stripe features from other linear structures; regions that do not meet the aspect ratio requirement are manually verified and marked. For cleavage surface features, an area greater than 500 square pixels is required to eliminate interference from small-scale noise.

[0057] Step S4: Determine the fracture type.

[0058] This step determines the fracture type based on the category label and confidence level of micro-features. A multi-label classification network is used to determine the fracture type, and the fracture type determination results and probability distribution of each type are output.

[0059] In this embodiment, a multi-label probability determination algorithm for fracture type is used to automatically determine the fracture type. This algorithm takes into account common mixed fracture modes in actual engineering and supports the simultaneous output of probability values ​​for multiple fracture types. The fracture types determined in this embodiment include five categories: ductile fracture, brittle fracture, fatigue fracture, stress corrosion cracking, and hydrogen embrittlement.

[0060] The input to a multi-label classification network is a feature statistics vector of the micro-feature recognition result. This vector contains the following components: the area proportion of various micro-features. The number of various micro-features Average confidence level of various micro-features And the statistics of the global feature distribution. The dimension of the feature statistics vector is 48.

[0061] The multi-label classification network structure includes a global average pooling layer, two fully connected layers, and a sigmoid activation layer. The first fully connected layer maps the 48-dimensional input features to 128-dimensional latent features, and the second fully connected layer maps the 128-dimensional latent features to a 5-dimensional output, corresponding to five fracture types. The sigmoid activation layer maps the output values ​​to probability values ​​between 0 and 1.

[0062] Probability values ​​of each fracture type Calculated using the following formula: ,in: This is the weight matrix of the first fully connected layer, with dimension 1. ; This is the first-level bias vector, with a dimension of 128; This is the weight matrix of the second fully connected layer, with dimension 1. ; This is the second-layer bias vector, with a dimension of 5; It is a non-linear activation function; Use the Sigmoid activation function; This is the fracture type number, with a value ranging from 1 to 5.

[0063] The fracture type determination rules are as follows. Let the preset type determination threshold be... In this embodiment When the first Probability value of fracture type Exceed When the probability values ​​of multiple fracture types all exceed a certain threshold, the existence of that fracture type is determined. When a mixed fracture mode is identified, the output includes all fracture types with probability values ​​exceeding the threshold and their probability values. The correspondence between fracture types and microscopic features is as follows: ductile fracture is mainly characterized by dimples; brittle fracture is mainly characterized by cleavage planes, quasi-cleavage, and transgranular fracture; fatigue fracture is mainly characterized by fatigue striations; stress corrosion fracture is mainly characterized by intergranular fracture and river pattern; hydrogen embrittlement fracture is mainly characterized by intergranular fracture and herringbone pattern.

[0064] In this embodiment, the confidence assessment of fracture type determination employs the Monte Carlo dropout method. During the inference phase, the Dropout layer in the network remains active, and multiple forward propagations are performed on the same input (20 times in this embodiment). The mean and variance of the output results are then calculated. The mean is used as the final probability prediction value, and the variance is used as a measure of prediction uncertainty. When the prediction uncertainty for a certain fracture type exceeds a preset threshold of 0.15, the output result is marked as having high uncertainty, and manual review is recommended.

[0065] The fracture type determination module also integrates a knowledge rule constraint mechanism to improve the physical rationality of the determination results. The constraint rules include: when the dimple area accounts for more than 70% and the cleavage surface area accounts for less than 10%, it is forcibly determined as ductile fracture; when the cleavage surface area accounts for more than 60% and the dimple area accounts for less than 20%, it is forcibly determined as brittle fracture; when fatigue striations are present and arranged in a regular, equidistant pattern, it is forcibly determined as fatigue fracture. These knowledge rules are based on expert experience in fracture surface analysis and can effectively correct any abnormal determination results that the neural network may produce.

[0066] For the identification of mixed fracture modes, this embodiment adopts a hierarchical determination strategy. First, the primary fracture type, i.e., the fracture type with the highest probability value, is determined; then, the secondary fracture types, i.e., those with a probability value exceeding a threshold but lower than the primary fracture type, are determined. The output results are arranged according to the probability value from high to low, and the relative proportions between each type are marked. For certain typical mixed fracture modes, such as fatigue-ductile mixed fracture (fatigue crack propagation zone plus dimple fracture and final failure zone) and stress corrosion-brittle mixed fracture, the system can automatically identify and generate targeted analysis reports.

[0067] Step S5: Calculate the feature quantization parameters.

[0068] This step calculates feature quantification parameters and outputs feature quantification data based on the spatial location of micro-features and the fracture type determination results.

[0069] In this embodiment, the feature quantification parameters include three categories: dimple size distribution parameters, fatigue stripe spacing parameters, and cleavage surface orientation distribution parameters.

[0070] The dimple size distribution parameters were calculated using a statistical algorithm for dimple size distribution. First, the boundary contours of all dimple regions were extracted based on the microscopic feature recognition results. For the first... The dimple region Calculate its equivalent diameter Aspect Ratio and area Equivalent diameter Defined as the diameter of a circle with the same area as the dimple region: ,in: For the first The area of ​​each dimple region, in pixels; Pi is the mathematical constant for a circle. When converting pixel area to actual area, the pixel size needs to be determined based on the scale information of the scanning electron microscope. (Unit: micrometers per pixel), actual equivalent diameter .

[0071] The ratio of dimple depth to width Defined as the ratio of the major axis length to the minor axis length of the dimple region. The major axis length is obtained by fitting an ellipse to the dimple region. and minor axis length The formula for calculating the aspect ratio is: The proportion of dimple area Defined as the ratio of the total area of ​​all dimple regions to the total area of ​​the fracture surface image: ,in: This represents the total number of dimples detected. and This represents the height and width of the fracture image.

[0072] The output of dimple size distribution parameters includes: average dimple diameter Diameter standard deviation Average aspect ratio Percentage of the area of ​​the pit .

[0073] The fatigue fringe spacing parameter is calculated using a fatigue fringe spacing measurement algorithm. This algorithm is based on frequency domain analysis and extracts the periodic spacing of the fringes from the fatigue fringe region. First, the fatigue fringe region is extracted from the microscopic feature identification results. A two-dimensional Fourier transform is performed on the fatigue fringe region to obtain the spectrum. : ,in: Frequency domain coordinates; Spatial coordinates; and These represent the width and height of the fatigue stripe region, respectively. The imaginary unit; is the base of the natural logarithm.

[0074] In the spectrum analysis, the peak position is detected, and the frequency corresponding to the peak is inversely proportional to the stripe spacing. Let the detected main peak frequency be... Average stripe spacing Calculated using the following formula: ,in: The pixel size (in micrometers per pixel) is determined by the scale information from the scanning electron microscope.

[0075] Spacing variation coefficient The uniformity of the fringe spacing is reflected and defined as the ratio of the standard deviation of the spacing to the average spacing. The spacing is calculated within multiple local windows of the fatigue fringe region, and the standard deviation of the spacing is statistically analyzed. The formula for calculating the coefficient of variation is: The output of fatigue stripe spacing parameters includes: average stripe spacing. (Unit: micrometers) and spacing variation coefficient .

[0076] The calculation of cleavage surface orientation distribution parameters is based on image gradient analysis. First, the cleavage surface regions are extracted from the microscopic feature recognition results. The image gradient is calculated over the cleavage surface region to obtain the gradient direction at each pixel location. : ,in: and The image gradients in the horizontal and vertical directions are calculated using the Sobel operator. The gradient direction is perpendicular to the normal direction of the cleavage surface, therefore the orientation angle of the cleavage surface... .

[0077] Principal orientation angle of cleavage plane The orientation angle is defined as the angle value with the highest frequency in the orientation angle histogram. The orientation angle range from 0° to 180° is divided into 36 intervals, and the number of pixels in each interval is counted. The center value of the interval with the highest frequency is the main orientation angle.

[0078] Orientation Dispersion The degree of dispersion of cleavage plane orientation is defined as the circular standard deviation of the orientation angle: ,in: ; ; This represents the number of pixels in the cleavage surface region. For the first The orientation angle of the cleavage surface of each pixel.

[0079] The output of the cleavage surface orientation distribution parameters includes: principal orientation angles of the cleavage surface. (Unit: degrees) and orientation dispersion (Unit: degrees).

[0080] In this embodiment, the calculation of feature quantization parameters also considers the impact of image magnification on measurement accuracy. Scanning electron microscope images are usually labeled with scale information. The system automatically reads the scale values ​​and units using optical character recognition technology to calculate the pixel size. For images with missing or unrecognizable scale information, the system prompts the user to manually input the magnification or pixel size parameters. The accuracy of the pixel size directly affects the absolute values ​​of quantization parameters such as dimple diameter and fatigue fringe spacing; therefore, the output results also indicate the pixel size parameters used and their source.

[0081] Statistical analysis of dimple morphology also includes fitting the dimple size distribution. In this embodiment, a log-normal distribution is used to fit the dimple diameter distribution, with fitting parameters including the geometric mean and geometric standard deviation. The log-normal distribution is a commonly used model to describe dimple size distribution; its physical basis is that the dimple formation process involves the product effect of multiple independent random factors. The goodness of fit is evaluated using the Kolmogorov-Smirnov test. When the test statistic exceeds the critical value, it indicates that the dimple size distribution may deviate from the log-normal distribution, requiring consideration of other distribution models or the existence of a multimodal distribution.

[0082] The fatigue fringe spacing is measured in multiple local regions to assess its spatial variation. According to fatigue crack propagation theory, the fringe spacing is proportional to the crack propagation rate, which increases with the stress intensity factor. Therefore, the fatigue fringe spacing typically increases from the crack initiation point towards the final fracture zone. In this embodiment, multiple measurement windows are set along the fatigue crack propagation direction, and the fringe spacing within each window is calculated, outputting a curve showing the spacing as a function of location. The slope of this curve reflects the degree of crack propagation acceleration, providing important reference information for fatigue life assessment.

[0083] Step S6: Correlation analysis between fracture surface characteristics and fracture mechanics parameters.

[0084] This step, based on feature quantification data and fracture type determination results, predicts fracture toughness, fatigue life, and stress intensity factor through a correlation model between fracture surface characteristics and fracture mechanics parameters, and outputs a comprehensive analysis report.

[0085] In this embodiment, a fracture feature-mechanical parameter correlation regression algorithm is used to establish the mapping relationship between fracture features and fracture mechanical parameters. This algorithm employs a multi-task regression network structure, using feature quantification data as input and fracture toughness, fatigue life, and stress intensity factor as multi-task outputs.

[0086] Input feature vector of the association model Includes: dimple size distribution parameters (average diameter) Diameter standard deviation Average aspect ratio Area percentage Fatigue stripe spacing parameters (average spacing) Coefficient of variation ), Cleavage plane orientation distribution parameters (principal orientation angle) Orientation Dispersion ) and the probability distribution of fracture types (probability values ​​of five fracture types) to The input feature vector has a dimension of 13.

[0087] The multi-task regression network employs a structure consisting of a shared feature representation layer and a task-specific output layer. The shared feature representation layer comprises two fully connected layers, mapping the 13-dimensional input features to 64-dimensional and 32-dimensional latent features respectively. The task-specific output layer consists of three parallel single-layer fully connected networks, each outputting a predicted fracture toughness value. Fatigue life prediction value and predicted values ​​of stress intensity factor .

[0088] Fracture toughness prediction value The fracture toughness is closely related to the dimple size distribution parameter. According to microscopic fracture mechanics theory, the average diameter of the dimples is positively correlated with the fracture toughness. The correlation model learns this physical law and outputs the predicted value of the fracture toughness and its confidence interval. The unit of fracture toughness is MPa·m. .

[0089] Fatigue life prediction value The fatigue life is closely related to the fatigue fringe spacing parameter. According to the Paris formula, the fatigue fringe spacing is proportional to the crack propagation rate, and fatigue life can be estimated by integration. The correlation model takes the average fatigue fringe spacing and coefficient of variation as input and outputs the predicted fatigue life and its confidence interval. The unit of fatigue life is the number of cycles.

[0090] Stress intensity factor prediction value This reflects the stress field intensity at the crack tip. Based on the empirical relationship between fatigue stripe spacing and the range of stress intensity factors: ,in: and These are material constants; This refers to the range of stress intensity factors. The correlation model establishes a regression mapping from quantified feature data to stress intensity factors by learning the correspondence between a large amount of fracture surface data and mechanical parameters. The unit of stress intensity factor is MPa·m. .

[0091] The comprehensive analysis report output includes the following: fracture feature-annotated images, which overlay the segmentation results of various micro-features onto the original fracture image, with different colors used for different categories; fracture type determination results, including the main fracture types and probability values ​​for each type, and listing all detected fracture types for mixed fracture modes; feature quantification data, which outputs the specific values ​​of dimple size distribution parameters, fatigue stripe spacing parameters, and cleavage plane orientation distribution parameters in tabular form; and predicted values ​​of fracture mechanical parameters, including predicted values ​​and confidence intervals for fracture toughness, fatigue life, and stress intensity factor.

[0092] Failure cause analysis suggestions are automatically generated based on the fracture type determination results and characteristic quantification data. When fatigue fracture is determined, the fatigue load amplitude range is estimated based on the fatigue striation spacing parameter, and the crack propagation stage (steady-state propagation or rapid propagation) is assessed. When hydrogen embrittlement fracture is determined, the degree of hydrogen embrittlement (mild, moderate, or severe) is assessed based on the proportion of intergranular fracture regions. When stress corrosion fracture is determined, the type of corrosive medium and stress state are analyzed based on the morphological characteristics of the river pattern. When a mixed fracture mode is determined, the primary and secondary relationships and interactions of each fracture mechanism are analyzed.

[0093] In this embodiment, the training data for the correlation model comes from publicly available databases of metallic material fracture surfaces and laboratory-accumulated test data. The training dataset contains 1800 samples, each including a fracture morphology image, microscopic feature annotations, fracture type labels, and corresponding fracture mechanics parameter test values. Fracture toughness is measured using the compact tensile test, fatigue life is measured using the rotating bending fatigue test, and stress intensity factor is calculated using strain gauge measurement. The training process uses the Adam optimizer with a learning rate of 0.001, a batch size of 32, and 100 training epochs.

[0094] The loss function for the correlation model uses multi-task weighted mean squared error, with the weights of each task set according to their difficulty and importance. The weight for fracture toughness prediction is set to 1.0, the weight for fatigue life prediction is set to 0.8 (due to the greater difficulty in predicting fatigue life, the weight is appropriately reduced to avoid overfitting), and the weight for stress intensity factor prediction is set to 1.0. Furthermore, an L2 regularization term is added to the loss function with a regularization coefficient of 0.0001 to prevent model overfitting.

[0095] The comprehensive analysis report is designed with engineering applications in mind. The first page contains a summary, including sample number, analysis date, main fracture types, and key quantitative parameters. The main body sequentially displays fracture surface feature-annotated images, statistical analysis of microscopic feature distribution, details of fracture type determination, a list of quantitative parameters, and predicted mechanical parameters. The appendix provides original images, intermediate processing results, and algorithm parameter configuration information. The report supports both PDF and HTML output formats; the HTML format supports interactive image zooming and feature annotation highlighting.

[0096] This embodiment provides an intelligent recognition system for fracture morphology features of metallic materials, such as... Figure 2 As shown, the system corresponds one-to-one with the steps of the above method embodiments.

[0097] Image preprocessing module 1 is used to perform grayscale normalization, contrast enhancement, and noise suppression on fracture morphology images acquired by scanning electron microscope, and outputs the preprocessed fracture morphology image. The specific implementation of this module is consistent with the description in step one of the method embodiments, employing adaptive grayscale mapping, contrast-limited adaptive histogram equalization, and a combination of Gaussian filtering and bilateral filtering. The image preprocessing module receives raw fracture image data from the scanning electron microscope, in 8-bit or 16-bit grayscale image format, with a resolution supporting 512×512 to 4096×4096 pixels. For large images, this module automatically performs image block processing, dividing the image into 512×512 pixel sub-blocks for parallel processing, and then stitching them together to form a complete image.

[0098] The multi-scale feature extraction module 2 is used to construct a feature pyramid containing multiple scale levels from the preprocessed fracture morphology image. It extracts fracture texture and morphological features at each scale level and outputs a multi-scale feature map. The specific implementation of this module is consistent with step two in the method embodiment, employing a residual convolutional network structure to construct a feature pyramid containing four scale levels with scale factors of 1, 0.5, 0.25, and 0.125. The data interface between the multi-scale feature extraction module and the image preprocessing module is the preprocessed fracture morphology image, and the output is a feature map with four scale levels and channel numbers of 64, 128, 256, and 512, respectively.

[0099] Feature fusion and recognition module 3 is used to perform weighted fusion of features at different scales through a hierarchical attention mechanism. Based on the fused features, it identifies dimples, cleavage surfaces, quasi-cleavage, fatigue striations, intergranular fractures, transgranular fractures, river patterns, and herringbone patterns, outputting the spatial location, category label, and confidence score of various microscopic features. The specific implementation of this module is consistent with the description in step three of the method embodiment, employing a hierarchical attention fusion algorithm for fracture microscopic features and a semantic segmentation network structure. The feature fusion and recognition module receives four-scale hierarchical feature maps from the multi-scale feature extraction module and outputs the category label, confidence score, and bounding box coordinates of each microscopic feature region.

[0100] The fracture type determination module 4 is used to determine the fracture type (ductile fracture, brittle fracture, fatigue fracture, stress corrosion fracture, or hydrogen embrittlement fracture) through a multi-label classification network, and outputs the fracture type determination result and the probability distribution of each type. The specific implementation of this module is consistent with the description in step four of the method embodiment, employing a multi-label probability determination algorithm for fracture types, and supporting the identification of mixed fracture modes. The fracture type determination module receives the microscopic feature recognition results from the feature fusion and recognition module, extracts the feature statistical vector as input to the multi-label classification network, and outputs the probability values ​​and determination results for the five fracture types.

[0101] Feature quantization module 5 is used to calculate dimple size distribution parameters, fatigue stripe spacing parameters, and cleavage surface orientation distribution parameters, and outputs feature quantization data. The specific implementation of this module is consistent with the description in step five of the method embodiment, employing a dimple size distribution statistical algorithm, a fatigue stripe spacing measurement algorithm, and an image gradient analysis method, respectively. The feature quantization module receives microscopic feature spatial location information from the feature fusion and recognition module and fracture type determination results from the fracture type determination module. Based on different fracture types, it selectively calculates the corresponding quantization parameters and outputs a complete feature quantization data vector.

[0102] The correlation analysis module 6 is used to predict fracture toughness, fatigue life, and stress intensity factor through a correlation model between fracture features and fracture mechanical parameters, and outputs a comprehensive analysis report. The specific implementation of this module is consistent with the description in step six of the method embodiment, employing a fracture feature-mechanical parameter correlation regression algorithm to establish a mapping relationship between features and mechanical parameters based on a multi-task regression network structure. The correlation analysis module receives feature quantification data from the feature quantification module and fracture type probability distribution from the fracture type determination module, outputs predicted fracture mechanical parameters, fracture feature annotation images, and failure cause analysis suggestions, and summarizes these to generate a comprehensive analysis report.

[0103] In this embodiment, the modules of the system communicate with each other through a standardized data interface. The image preprocessing module 1 and the multi-scale feature extraction module 2 exchange preprocessed image data in 32-bit floating-point tensors with dimensions 1×H×W, where H and W are the image height and width. The multi-scale feature extraction module 2 and the feature fusion and recognition module 3 exchange multi-scale feature maps in four 32-bit floating-point tensors, corresponding to the four scale levels. The feature fusion and recognition module 3 and the fracture type determination module 4 exchange microscopic feature recognition results in a JSON object containing the category, confidence level, bounding box, and area information of each feature region.

[0104] The system supports both batch processing and single-image processing modes. In batch processing mode, users can import multiple fracture surface images at once. The system processes each image sequentially and generates an independent analysis report, while also generating a batch processing summary report that statistically analyzes the distribution of different fracture types and the key quantitative parameters. In single-image processing mode, users import a single fracture surface image, and the system performs interactive processing, allowing users to view the processing results, adjust parameter configurations, and perform manual corrections at intermediate stages.

[0105] The system's user interface is designed with simplicity and efficiency in mind. The main interface is divided into three parts: an image display area, a parameter configuration area, and a results display area. The image display area supports image zooming, panning, and region selection, facilitating observation of detailed fracture morphology features. The parameter configuration area provides adjustment options for key algorithm parameters, including confidence threshold, minimum feature area, and magnification. The results display area organizes various output results in a tabbed format, including feature annotation maps, fracture type determination, quantification parameters, and analysis reports.

[0106] The system's data management functions support the storage, retrieval, and export of analysis results. Analysis results are stored in a structured database and can be retrieved by sample number, fracture type, analysis date, and other criteria. Users can export analysis results as Excel spreadsheets, PDF reports, or raw data files for easier subsequent data analysis and report writing. The system also provides a historical data comparison function, allowing users to compare current analysis results with historical samples, helping them identify differences between material batches or the evolution trends of failure modes.

[0107] This embodiment verifies the technical effectiveness on a test dataset containing 2500 scanning electron microscope fracture morphology images. The test dataset covers various metallic materials such as steel, aluminum alloys, titanium alloys, and nickel-based superalloys, and includes fracture samples of ductile fracture, brittle fracture, fatigue fracture, stress corrosion fracture, hydrogen embrittlement fracture, and their mixed modes. The test equipment is a workstation equipped with an NVIDIA RTX 4090 graphics card and running Ubuntu 22.04.

[0108] In terms of the accuracy of microscopic feature recognition, the average recognition accuracy of eight types of fracture surface microscopic features reached 92.6%, with dimple recognition accuracy at 94.2%, cleavage surface recognition accuracy at 93.8%, fatigue stripe recognition accuracy at 91.5%, and intergranular fracture recognition accuracy at 90.3%. Compared with existing technologies that can only identify a single feature of fatigue stripes, the recognition type coverage of this invention is improved to a complete fracture morphology feature system.

[0109] Regarding the accuracy of fracture type identification, the average accuracy rate for the five fracture types reached 89.4%, with ductile fracture identification accuracy at 93.1%, brittle fracture at 91.2%, fatigue fracture at 88.7%, stress corrosion fracture at 86.5%, and hydrogen embrittlement fracture at 87.5%. For mixed fracture modes, the identification accuracy rate was 84.2%, demonstrating the ability to correctly identify and distinguish multiple coexisting fracture mechanisms.

[0110] Regarding the accuracy of feature quantification, the relative error for measuring the average diameter of dimples is 6.8%, the relative error for measuring the average spacing of fatigue striations is 8.2%, and the measurement error for the principal orientation angle of cleavage planes is 4.5 degrees. These quantification accuracies meet the requirements of engineering failure analysis and can provide reliable input data for estimating fracture mechanics parameters.

[0111] Regarding the accuracy of fracture mechanics parameter predictions, the average relative error for fracture toughness prediction was 12.3%, the average relative error for fatigue life prediction was 18.5% (logarithmic scale), and the average relative error for stress intensity factor prediction was 15.2%. These prediction accuracies are within an acceptable range for engineering applications and can provide a reference for failure analysis and material performance evaluation.

[0112] In terms of processing efficiency, for a 1024×1024 pixel fracture image, the complete processing time from image input to output of a comprehensive analysis report is 2.3 seconds. Compared to the 30 minutes to 2 hours typically required for human expert interpretation, the processing efficiency is improved by more than two orders of magnitude.

[0113] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent recognition of fracture morphology features of metallic materials, characterized in that, Includes the following steps: The fracture morphology image acquired by scanning electron microscope is preprocessed, including grayscale normalization, contrast enhancement and noise suppression, and the preprocessed fracture morphology image is output. Multi-scale feature extraction is performed on the preprocessed fracture morphology image to construct a feature pyramid containing multiple scale levels. Fracture texture features and morphological features are extracted at each scale level to output a multi-scale feature map. The multi-scale feature map is subjected to feature fusion and micro-feature recognition. Features of different scales are weighted and fused through a hierarchical attention mechanism. Based on the fused features, the micro-features of the fracture surface are identified. The micro-features of the fracture surface include dimples, cleavage planes, quasi-cleavage, fatigue striations, intergranular fracture, transgranular fracture, river patterns and fishbone patterns. The spatial location, category label and confidence level of each type of micro-feature are output. The fracture type is determined based on the category label and confidence level of the micro-features. The fracture type is determined by a multi-label classification network. The fracture types include ductile fracture, brittle fracture, fatigue fracture, stress corrosion fracture, and hydrogen embrittlement fracture. The fracture type determination result and the probability distribution of each type are output. Based on the spatial location of the micro-features and the fracture type determination result, feature quantification parameters are calculated. The feature quantification parameters include dimple size distribution parameters, fatigue stripe spacing parameters, and cleavage surface orientation distribution parameters. Feature quantification data is then output. Based on the quantified feature data and the fracture type determination result, fracture toughness, fatigue life, and stress intensity factor are predicted by a correlation model between fracture features and fracture mechanical parameters. The correlation model adopts a multi-task regression network structure, takes the quantified feature data as input, and takes fracture toughness, fatigue life, and stress intensity factor as multi-task outputs. By extracting the correlation information between features through a shared feature representation layer, a comprehensive analysis report is output, including fracture feature labeled images, fracture type determination results, quantified feature data, and failure cause analysis suggestions.

2. The intelligent identification method for fracture morphology features of metallic materials according to claim 1, characterized in that, The feature pyramid contains four scale levels, with scale factors of 1, 0.5, 0.25 and 0.125 for each scale level. Spatial correspondence is established between adjacent scale levels through downsampling operations.

3. The intelligent identification method for fracture morphology features of metallic materials according to claim 1, characterized in that, The hierarchical attention mechanism calculates attention weights based on the spatial location of features at each scale. These attention weights are used to adjust the contribution of features at different scales during the fusion process. When the confidence level is lower than a preset confidence threshold, the attention weights of features at adjacent scales are increased.

4. The intelligent identification method for fracture morphology features of metallic materials according to claim 1, characterized in that, The dimple size distribution parameters include the average dimple diameter, the dimple depth-to-width ratio, and the dimple area percentage; the fatigue fringe spacing parameters include the average fringe spacing and the spacing variation coefficient; and the cleavage surface orientation distribution parameters include the principal orientation angle of the cleavage surface and the orientation dispersion.

5. The intelligent identification method for fracture morphology features of metallic materials according to claim 1, characterized in that, The multi-scale feature extraction employs a residual convolutional network, with batch normalization layers and nonlinear activation layers set after the convolutional layers at each scale level. Low-level detailed features are passed to high-level semantic features through skip connections.

6. The intelligent identification method for fracture morphology features of metallic materials according to claim 1, characterized in that, The multi-label classification network includes a global average pooling layer, a fully connected layer, and a Sigmoid activation layer. It supports the simultaneous output of probability values ​​for multiple fracture types. When the probability values ​​for multiple fracture types all exceed a preset type determination threshold, it is determined to be a mixed fracture mode.

7. The intelligent identification method for fracture morphology features of metallic materials according to claim 1, characterized in that, The preprocessing also includes image block processing, which divides the fracture morphology image into image blocks of a preset size, performs grayscale normalization, contrast enhancement and noise suppression on each image block, and then stitches the processed image blocks together to form a complete image.

8. The intelligent identification method for fracture morphology features of metallic materials according to claim 1, characterized in that, The failure cause analysis suggestions are generated based on the fracture type determination results and the characteristic quantification data. When the fracture is determined to be fatigue fracture, the fatigue load amplitude range is estimated based on the fatigue stripe spacing parameter. When the fracture is determined to be hydrogen embrittlement fracture, the degree of hydrogen embrittlement is evaluated based on the proportion of intergranular fracture region.

9. An intelligent recognition system for fracture morphology features of metallic materials, used to implement the intelligent recognition method for fracture morphology features of metallic materials as described in any one of claims 1-8, characterized in that, include: The image preprocessing module is used to perform grayscale normalization, contrast enhancement, and noise suppression on the fracture morphology images acquired by scanning electron microscope, and output the preprocessed fracture morphology images. The multi-scale feature extraction module is used to construct a feature pyramid containing multiple scale levels for the preprocessed fracture morphology image, extract fracture texture features and morphological features at each scale level, and output a multi-scale feature map. The feature fusion and recognition module is used to weightedly fuse features of different scales through a hierarchical attention mechanism. Based on the fused features, it identifies dimples, cleavage surfaces, quasi-cleavage, fatigue striations, intergranular fracture, transgranular fracture, river patterns, and fishbone patterns, and outputs the spatial location, category label, and confidence level of various microscopic features. The fracture type determination module is used to determine the fracture type, such as ductile fracture, brittle fracture, fatigue fracture, stress corrosion fracture, or hydrogen embrittlement fracture, through a multi-label classification network, and outputs the fracture type determination results and the probability distribution of each type. The feature quantization module is used to calculate dimple size distribution parameters, fatigue stripe spacing parameters, and cleavage surface orientation distribution parameters, and output feature quantization data. The correlation analysis module is used to predict fracture toughness, fatigue life, and stress intensity factor through a correlation model between fracture features and fracture mechanics parameters. The correlation model adopts a multi-task regression network structure, takes the feature quantification data as input, and takes fracture toughness, fatigue life, and stress intensity factor as multi-task outputs. It extracts the correlation information between features through a shared feature representation layer and outputs a comprehensive analysis report.

Images