Intelligent analysis method and device for microstructure of corundum spinel material

Abstract

The invention discloses a corundum spinel material microstructure intelligent analysis method and device, and the method comprises the steps: obtaining an original microstructure image of a material, carrying out the standardization preprocessing, and inputting the image into a pre-trained material analysis model, so as to obtain a sintering temperature type and a prediction probability; the model comprises an input receiving module, a double-branch feature extraction module, a multi-modal feature fusion module and a classification decision module, wherein the double-branch feature extraction module comprises a local feature extraction branch and a global feature extraction branch which are arranged in parallel; and the model performs classification decision by fusing the local detail features and the global structure features which are extracted in parallel. The method also generates a decision thermodynamic diagram based on a model classification result, and reveals a microstructure basis of a model sintering temperature classification decision. According to the method, the problems of feature extraction limitation and black box decision making of an existing single model are solved, high-precision and interpretable intelligent analysis of the sintering temperature of the corundum spinel material is achieved, and reliable guidance is provided for material process optimization.

Description

Technical Field

[0001] This invention relates to the fields of materials science and computer vision technology, specifically to an intelligent analysis method and device for the microstructure of corundum spinel material. Background Technology

[0002] Corundum spinel, as a key refractory material, is widely used in high-temperature industrial fields such as metallurgy and building materials. Its service performance exhibits a strong nonlinear correlation with its microstructure (such as pore morphology, grain distribution, and crack network). Traditional methods for analyzing the microstructure of materials mainly rely on researchers' subjective experience or analyze parameters such as pore number and area using simple statistical models to establish a correlation between microstructure and preparation processes (such as sintering temperature). These methods generally suffer from inherent defects such as low analytical efficiency, large subjective bias, and difficulty in capturing morphological details and complex nonlinear laws, making them insufficient to meet the demands of modern high-throughput research and precision manufacturing of materials. With the penetration of artificial intelligence technology into materials science, deep learning-based microstructure analysis has become an important approach to optimizing material performance.

[0003] In current material microstructure analysis, Convolutional Neural Networks (CNNs) and Transformers are two mainstream models. Deep convolutional networks, such as ResNet and DenseNet, leverage their local connectivity and weight sharing to efficiently extract local details from images, demonstrating potential in identifying local features like grain boundaries and micropores in material microscopic images. Existing research has applied them to the classification and recognition of material microstructures. Self-attention models, such as ViT and ViM, segment images into fixed-size patches and model global dependencies, enabling them to understand the overall structure and contextual information of the image. They exhibit significant advantages in capturing global features such as phase distribution uniformity and crack network topology.

[0004] However, in the specific scenario of microstructure analysis of corundum spinel materials, both of the above-mentioned single-model-based technical solutions have significant limitations and cannot meet the high precision and high stability requirements of industrial-grade analysis.

[0005] (1) Limitations of feature extraction: A single CNN model is difficult to capture long-distance dependencies in microstructures (such as cross-regional crack networks and phase distribution correlations), and is not sensitive enough to details of sub-pixel level cracks and irregular pores; a single Transformer model is inefficient in capturing local details and has extremely high requirements for data volume and computing resources, which can easily lead to training efficiency bottlenecks.

[0006] (2) Data quality has a great impact: In actual industrial testing, micro-computed tomography (μ-CT) imaging often has problems such as missing regional signals, redundancy of adjacent slices, and incomplete slices due to sample preparation. Existing technologies lack systematic preprocessing schemes, which can easily lead to overfitting of training models and a decrease in classification accuracy.

[0007] (3) Poor interpretability: Most existing deep learning models are "black box" predictions, which cannot establish the physical relationship between the model decision mechanism and the thermodynamic and kinetic evolution of materials. For example, the model cannot explain the specific microstructure basis for its judgment of a certain sintering temperature, and it is difficult to explain the core nonlinear relationship such as "sintering temperature-pore morphology". This seriously weakens the credibility and application value of the analysis results in guiding actual process development and optimization.

[0008] (4) Weak generalization ability: The heterogeneity of microstructure data caused by different sintering temperatures and preparation processes makes it easy for existing models to fluctuate in performance on new samples, making it difficult to adapt to the needs of multi-scenario analysis. Summary of the Invention

[0009] To address the shortcomings of the existing technologies, the present invention aims to provide an intelligent analysis method and apparatus for the microstructure of corundum spinel materials, thereby achieving high-precision, high-stability, and highly interpretable analysis of the microstructure of corundum spinel materials, and providing efficient and reliable intelligent analysis support for the optimization of corundum spinel material processes and performance prediction.

[0010] To achieve the above objectives, the present invention adopts the following technical solution:

[0011] This invention proposes an intelligent analysis method for the microstructure of corundum spinel materials, comprising the following steps:

[0012] Obtain the original microstructure image of the corundum spinel material to be analyzed;

[0013] The original microstructure image is subjected to standardization preprocessing to obtain the preprocessed image;

[0014] The preprocessed image is input into a pre-trained material analysis model to obtain the sintering temperature category and corresponding predicted probability of the corundum spinel material. The material analysis model includes an input receiving module, a dual-branch feature extraction module, a multimodal feature fusion module, and a classification decision module. The dual-branch feature extraction module includes parallel local feature extraction branches and global feature extraction branches. The material analysis model performs classification decisions by fusing local detail features and global structural features extracted in parallel from the preprocessed image.

[0015] Based on the sintering temperature classification results of the material analysis model, a decision heatmap corresponding to the preprocessed image is generated. The decision heatmap reveals the microstructural basis for the model's classification decision.

[0016] Preferably, before inputting the preprocessed image into a pre-trained material analysis model to obtain the sintering temperature category and corresponding predicted probability of the corundum spinel material, the method further includes:

[0017] To obtain original microstructure image data of corundum spinel materials at different sintering temperatures;

[0018] The original microstructure image is standardized and preprocessed, and the preprocessed image data is divided into training set, validation set and test set.

[0019] The material analysis model initially constructed is trained using the training set. The model parameters are optimized by minimizing the cross-entropy loss and using the backpropagation algorithm to obtain the trained material analysis model.

[0020] Preferably, the original microstructure image is obtained by micro / nano computed tomography scanning of the target material sample, and the normalization preprocessing includes:

[0021] Invalid data removal processing includes removing images whose signal intensity at the beginning and end is lower than a set threshold from the original microstructure image sequence obtained from sample scanning;

[0022] Image integrity filtering processing includes identifying and removing images with incomplete upper and lower edge data due to uneven physical preparation of the sample;

[0023] Edge signal processing includes cropping edge regions of each retained image to remove low signal-to-noise ratio or no-signal areas at the edges of the image.

[0024] Size standardization processing includes adjusting the processed image to a preset size;

[0025] When the standardization preprocessing is used for image data processing during the model training phase, data augmentation processing is included before the size standardization processing. The data augmentation processing includes random cropping of a single image.

[0026] Preferably, the material analysis model makes classification decisions by fusing local detail features and global structural features extracted in parallel from the preprocessed image, including:

[0027] The preprocessed image is received using the input receiving module.

[0028] The dual-branch feature extraction module is used to extract local detail features and global structural features from the preprocessed image in parallel, including extracting local detail features from the preprocessed image using the local feature extraction branch and extracting global structural features from the preprocessed image using the global feature extraction branch.

[0029] The local detail features and the global structural features are fused using the multimodal feature fusion module to obtain fused features;

[0030] The fusion features are processed using the classification decision module to output the sintering temperature category and corresponding prediction probability of the corundum spinel material.

[0031] Preferably, the local feature extraction branch is constructed based on a structure-adjusted ResNet50 network; the extraction of local detail features from the preprocessed image using the local feature extraction branch includes:

[0032] The preprocessed image is then converted into a three-channel image adapted for the ResNet50 network.

[0033] The three-channel image is input into the structure-adjusted ResNet50 network for feature extraction to obtain a deep feature map, wherein the structure adjustment is to remove the fully connected layer at the end of the ResNet50 network;

[0034] The deep feature map is subjected to global average pooling and flattening to obtain the local detail features.

[0035] Preferably, the global feature extraction branch is constructed based on the ViT architecture; the step of extracting global structural features from the preprocessed image using the global feature extraction branch includes:

[0036] The preprocessed image is segmented into image blocks and converted into feature sequences;

[0037] The feature sequence is linearly projected and positionally encoded to obtain the encoded feature sequence;

[0038] The encoded feature sequence is input into a multi-layer Transformer encoder, and long-distance dependencies are captured through a self-attention mechanism to obtain a deep feature sequence;

[0039] The deep feature sequence is subjected to mean pooling and linear projection to obtain the global structural features.

[0040] Preferably, the multimodal feature fusion module includes a feature concatenation unit and a fusion mapping unit; the step of fusing the local detail features and the global structural features using the multimodal feature fusion module to obtain fused features includes:

[0041] The feature splicing unit is used to splice the local detail features and the global structural features along the channel dimension to obtain the spliced ​​features.

[0042] The spliced ​​features are sequentially input into at least two fully connected layers through the fusion mapping unit for nonlinear transformation and dimensionality reduction to obtain the fusion features used for subsequent classification.

[0043] Preferably, the classification decision module is constructed based on a multilayer perceptron; the step of processing the fused features using the classification decision module to output the sintering temperature category and corresponding predicted probability of the corundum spinel material includes:

[0044] Receive the fused features output by the multimodal feature fusion module;

[0045] The fused features are sequentially passed through at least two fully connected layers of the classification decision module for feature transformation and dimensionality compression to obtain category scores corresponding to different sintering temperature categories.

[0046] The category scores are converted into a probability distribution using the Softmax function, and the sintering temperature category and its corresponding predicted probability are output.

[0047] Preferably, the decision heatmap is generated using Grad-CAM technology, which calculates and generates the decision heatmap based on the feature map output by the last convolutional layer in the local feature extraction branch and the corresponding gradient information.

[0048] In another aspect, this invention also proposes an intelligent analysis device for the microstructure of corundum spinel material, comprising:

[0049] The data acquisition unit is used to acquire the original microstructure image of the corundum spinel material to be analyzed.

[0050] The data preprocessing unit is used to perform standardized preprocessing on the original microstructure image to obtain a preprocessed image.

[0051] The model analysis unit is used to input the preprocessed image into a pre-trained material analysis model to obtain the sintering temperature category and corresponding prediction probability of the corundum spinel material. The material analysis model includes an input receiving module, a dual-branch feature extraction module, a multimodal feature fusion module, and a classification decision module. The dual-branch feature extraction module includes parallel local feature extraction branches and global feature extraction branches. The material analysis model performs classification decisions by fusing local detail features and global structural features extracted in parallel from the preprocessed image.

[0052] The visualization unit is used to generate a decision heatmap corresponding to the preprocessed image based on the sintering temperature classification results of the material analysis model, and to reveal the microstructural basis for the model to make classification decisions through the decision heatmap.

[0053] The beneficial effects of this invention are as follows:

[0054] (1) This invention proposes a novel ResVi hybrid deep learning model, which for the first time deeply integrates the local feature extraction capability of ResNet50 with the global dependency modeling capability of ViT. Through the dual-branch architecture, it overcomes the performance limitations of a single CNN or Transformer, and provides a new method for the accurate analysis of the microstructure of corundum spinel.

[0055] (2) This invention constructs a high-quality corundum spinel microstructure dataset covering a temperature gradient of 110℃-1600℃, and designs a standardized preprocessing process including invalid data removal, integrity filtering, edge processing, and data augmentation, which effectively improves data quality, eliminates the interference of imaging noise and redundant information on model performance to the greatest extent, and improves the stability of the model.

[0056] (3) Based on the advantages of the above model architecture and training dataset, the ResVi model constructed in this invention shows significant advantages in multiple performance indicators. In terms of classification accuracy, the ResVi model achieves a stable accuracy of over 0.98 in the microstructure classification task of corundum spinel, with the best reaching 0.9974, which is significantly higher than traditional CNN models (such as ConvNeXt with an accuracy of 0.74-0.89) and single Transformer models (such as ViT with an accuracy of 0.91-0.93). In terms of training efficiency and stability, the ResVi model converges 3-4 times faster than ViT (it only needs 12 epochs to achieve an accuracy of 0.9 when the batch size is 8, while ViT needs 40 epochs), and the training loss value can be as low as 0.008, with small fluctuations in the training curve, effectively avoiding overfitting and loss rebound. In terms of generalization ability, the ResVi model can effectively resist the influence of data heterogeneity and maintain stable performance in samples with different batches and different imaging parameters, thus solving the problem of insufficient generalization ability of existing models.

[0057] (4) This invention uses thermal visualization technology to intuitively display the key micro-regions that the model focuses on in decision-making. Combined with the microstructure analysis of the image, it reveals the microstructure basis of the model in the sintering temperature classification decision-making process, overcomes the defect of the existing model's "black box" decision-making, and provides clear guidance for material process optimization, thereby improving the credibility of industrial applications.

[0058] (5) The training time of the ResVi model of this invention is effectively controlled within 69-87 minutes (Batch Size=4-32), which significantly shortens the model development and iteration cycle while ensuring high performance. It takes into account both high performance and high efficiency, so that it can be efficiently deployed and applied to the process optimization and performance prediction of industrial-grade corundum spinel materials, which helps the industry to greatly improve R&D efficiency and save R&D costs. Attached Figure Description

[0059] Figure 1 This is a flowchart illustrating the intelligent analysis method for the microstructure of corundum spinel material in Embodiment 1 of the present invention.

[0060] Figure 2 This is a diagram of the structure of the material analysis model in Embodiment 1 of the present invention;

[0061] Figure 3 This is a schematic diagram illustrating the process by which the material analysis model analyzes the preprocessed input image in Embodiment 1 of the present invention.

[0062] Figure 4 This is a comparison of the thermograms of the material analysis model in Embodiment 2 of the present invention at five different sintering temperatures. Detailed Implementation

[0063] To make the objectives, techniques, and advantages of this invention clearer, the invention will be described clearly and completely below with reference to specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0064] Terminology Explanation:

[0065] Corundum spinel material: a refractory material widely used in high-temperature industries. Its microstructure (such as pore size, grain boundary morphology, phase distribution, etc.) directly determines the core service performance such as high-temperature volume stability and thermal shock resistance.

[0066] ResNet50: A deep convolutional neural network (CNN) whose core innovation lies in the introduction of a "residual connection" structure. This structure directly passes the input to the deep layers through shortcut connections, effectively alleviating the gradient vanishing and degradation problems in extremely deep networks, enabling the network depth to reach 50 layers. ResNet50 can efficiently extract local features from low to high levels in images through multi-layer convolution and pooling operations.

[0067] Vision Transformer (ViT): An image classification model based on a pure Transformer architecture. Its core is to segment the input image into a sequence of fixed-size image patches, convert them into vectors through linear projection, and introduce positional encoding. By utilizing the self-attention mechanism in the Transformer encoder, ViT can model the global dependency between any two image patches in the image, thereby capturing the overall structure and long-range contextual information of the image.

[0068] ResVi Model: The dual-branch hybrid deep learning model proposed in this invention integrates the local feature extraction capability of ResNet50 with the global dependency modeling capability of Vision Transformer (ViT) for microstructure image classification and temperature-microstructure relationship analysis of corundum spinel materials.

[0069] μ-CT imaging: a high-resolution omnidirectional micro-nano computed tomography technique that can acquire slice images of the three-dimensional microstructure of materials. The slice spacing and imaging parameters can affect the data quality.

[0070] Multimodal feature fusion module: Used to integrate the local features output by the ResNet50 branch and the global features output by the ViT branch, and generate fused feature representation through operations such as channel concatenation and linear transformation.

[0071] Heatmap visualization: Based on gradient-weighted class activation mapping (Grad-CAM) technology, it is used to intuitively display the key microstructural regions that the model is interested in, thereby improving the interpretability of model decisions.

[0072] Pore ​​Count: The total number of unique pore structures identified.

[0073] Total Area: The total pixel area occupied by all porous regions.

[0074] Perim: The total length of the profile of all pore edges.

[0075] Circularity (Circ): Characterizes the degree to which the shape of the pore is close to an ideal circle (calculated as 4π × area / perimeter²).

[0076] Solidity: Characterizes the density of the internal structure of the porous region (calculated as the ratio of the pore area to the area of ​​its convex hull).

[0077] Integral density (IntDen): The product of the average gray intensity and area within the pore region, reflecting the local density of the material.

[0078] Example 1

[0079] See Figure 1 and Figure 2 This embodiment provides an intelligent analysis method for the microstructure of corundum spinel material, which specifically includes the following steps:

[0080] S1, Obtain the original microstructure image of the corundum spinel material to be analyzed.

[0081] In this step, the corundum spinel material to be analyzed is first prepared into a standard-sized sample (e.g., 20×20×20 mm³). Then, the sample is scanned using high-resolution omnidirectional micro-nano computed tomography (μ-CT) technology to obtain a series of slice images of the sample at a set high resolution. These slice images are the original microstructure images of the corundum spinel material to be analyzed.

[0082] S2 performs standardization preprocessing on the original microstructure image to obtain the preprocessed image.

[0083] Because raw μ-CT images inherently contain imaging defects and noise, directly using them for model analysis and prediction will affect the accuracy of the model. Therefore, this embodiment designs a standardized preprocessing workflow to eliminate invalid information and provide the model with high-quality, standardized input. It should be noted that this workflow is applicable to both dataset construction during the training phase and single-sample analysis during the application phase. The specific processing includes:

[0084] (1) Invalid data removal. Invalid slices generated in the scanned sequence due to unstable initial signal or attenuation of ending signal are identified and removed. Specifically, the overall gray value or signal intensity of the slice can be detected to remove slices with signal intensity lower than a set threshold at both ends of the sequence; preferably, about 5% of the slices at the beginning and end are removed.

[0085] (2) Image integrity filtering. To eliminate incomplete slice data caused by uneven physical preparation of samples, the pixel distribution characteristics of the top and bottom edge regions of the slice image can be analyzed to identify and remove slices with incomplete top and bottom edges, ensuring that all retained slices have complete and consistent data regions, thereby guaranteeing the uniformity of data dimensions.

[0086] (3) Edge signal processing. To eliminate imaging edge noise, the edge region of each retained slice is cropped. Preferably, a low signal-to-noise ratio or no-signal region of about 70 pixels wide around the periphery of each slice can be removed.

[0087] (4) Size standardization. To adapt to the input requirements of the model, the sliced ​​images processed above are adjusted to a preset size. In a preferred embodiment of the present invention, the images are uniformly adjusted to a size of 448 pixels × 448 pixels.

[0088] Finally, one or more sliced ​​images are randomly selected from the standardized preprocessed images above as the final input images for the samples.

[0089] S3, the preprocessed image is input into a pre-trained material analysis model to obtain the sintering temperature category and corresponding prediction probability of the corundum spinel material; the material analysis model includes an input receiving module, a dual-branch feature extraction module, a multimodal feature fusion module, and a classification decision module, wherein the dual-branch feature extraction module includes a parallel local feature extraction branch and a global feature extraction branch; the material analysis model performs classification decision by fusing the local detail features and global structural features extracted in parallel from the preprocessed image.

[0090] It should be noted that the material analysis model in this embodiment of the invention is named the ResVi model.

[0091] Specifically, such as Figure 3 The process described in this embodiment for obtaining the sintering temperature category and corresponding predicted probability of the target corundum spinel material through ResVi model analysis includes:

[0092] S301 uses the input receiving module to receive the preprocessed image.

[0093] The input receiving module is configured to receive the image after normalization preprocessing in step S2. In a preferred embodiment, the input receiving module receives a single-channel grayscale microstructure image with a size of 448×448 pixels, the input form of which can be represented as follows: Where B is the batch size, C=1 is the number of channels, and H and W are the height and width of the image, respectively, both of which are 448 in this embodiment.

[0094] S302 utilizes a dual-branch feature extraction module to extract local detail features and global structural features from the preprocessed image in parallel, including extracting local detail features from the preprocessed image using the local feature extraction branch and extracting global structural features from the preprocessed image using the global feature extraction branch.

[0095] S3021 utilizes the local feature extraction branch to extract local detail features from the preprocessed image.

[0096] The local feature extraction branch is used to extract local detail features of the image. In a preferred embodiment, this branch can be implemented using a feature extraction structure based on a convolutional neural network (CNN), the core function of which is to capture detailed information of local regions in the image through convolution operations.

[0097] As a specific example, this embodiment preferentially uses the ResNet50 network as the foundation for constructing the local feature extraction branch. It should be understood that the above selection is merely illustrative and not a limitation on this branch structure. In other implementations, other networks in the ResNet series, such as ResNet34 and ResNet101, can also be used, or replaced with CNN architectures such as EfficientNet that have similar local feature extraction capabilities. As long as this branch retains the core function of local feature extraction based on CNN, it does not deviate from the basic architecture of this dual-branch feature extraction module.

[0098] Specifically, this branch uses a single-channel image with a size of 448×448 pixels. As input, its workflow mainly includes three core stages: channel adaptation, deep feature extraction, and feature vectorization.

[0099] (1) Channel Adaptation. Since the ResNet50 network used in this embodiment is preset to input a three-channel (RGB) image, the channel dimension of the input single-channel features is first expanded. This expansion is achieved through a channel copy operation, which does not introduce any trainable parameters. The channel copy operation can be expressed as:

[0100]

[0101] in, For raw single-channel input, This is the adapted three-channel input. Input characteristics. After replication, a three-channel feature map is formed. This satisfies the input requirements of subsequent networks.

[0102] (2) Deep feature extraction. It should be noted that the structure of the ResNet50 network has been adapted in this embodiment. The fully connected classification layer at the end of the original network has been removed, and only its convolutional feature extraction part has been retained, so as to focus on extracting local detail features in the image, such as pore edges, grain boundaries or micron-level cracks.

[0103] Specifically, using the aforementioned three-channel feature map that has undergone channel replication... As input, the feature map first passes through an initial convolutional layer with a large receptive field (7×7 kernel, stride 2, padding 3) for preliminary feature mapping and spatial downsampling, reducing the input size from 448×448 to 224×224. The output is a feature map with a size of 224×224 and 64 channels. .

[0104] Then, the convolution output After batch normalization (BN) and ReLU activation, a max pooling layer (3×3 pooling kernel, stride 2, padding 1) is used for non-linear enhancement and further spatial compression, halving the feature map size and outputting a feature map of size 112×112. .

[0105] Subsequently, feature maps Deep feature extraction is performed using a backbone network consisting of four layers (Layer 1 to Layer 4) stacked together, with each layer composed of several residual blocks with bottleneck structures. This backbone network, through progressively increasing network depth and feature channel count, achieves a gradual, coarse-to-fine abstraction and refinement of local image details (such as edges and textures). Specifically,

[0106] The first layer (Layer 1) consists of three residual blocks connected sequentially. This layer does not perform spatial downsampling on the feature map. Its core function is to increase the number of channels in the input feature map from 64 to 256 while maintaining the spatial resolution (112×112 pixels). The output feature map is denoted as... .

[0107] The second layer (Layer 2) consists of four residual blocks. The first residual block in this layer introduces downsampling with a stride of 2, thereby halving the spatial size of the feature map to 56×56 pixels, while increasing the number of feature channels to 512. The output feature map of this layer is denoted as... .

[0108] The third layer (Layer 3) consists of 6 residual blocks. Similar to Layer 2, its first residual block also performs a downsampling operation with a stride of 2, further compressing the feature map size to 28×28 pixels and expanding the number of channels to 1024. The output feature map of this layer is denoted as... .

[0109] Layer 4 consists of three residual blocks. In its first residual block, the feature map undergoes a final downsampling with a stride of 2, resulting in a final spatial size of 14×14 pixels and an increase in the number of channels to 2048. The deepest feature map output is denoted as... .

[0110] Each residual block (Bottleneck) contains three convolutional layers. The first 1×1 convolutional layer (Conv1) with batch normalization (BN1) reduces the channel dimension; the second 3×3 convolutional layer (Conv2) with batch normalization (BN2) extracts core spatial features; and the third 1×1 convolutional layer (Conv3) with batch normalization (BN3) restores or increases the channel dimension to the target number. Finally, a ReLU activation function is applied, and the input is directly added to the output through residual connections. It should be noted that in the first residual block from Layer 2 to Layer 4, to match potential differences in the number of channels and spatial dimensions between the input and output, the residual connections adjust the input through downsampling operations (usually a 1×1 convolution with a stride of 2) to ensure dimensionality consistency.

[0111] (3) Feature vectorization. This involves vectorizing the deep feature map. To convert to a fixed-length feature vector, first... Adaptive Average Pooling (GAP) is performed to aggregate the 14×14 spatial features of each channel into a single scalar value, resulting in a 2048-dimensional spatially compressed feature. Finally, a flattening operation is used to remove the spatial dimension, forming the final local detail feature vector. This process can be represented as:

[0112]

[0113]

[0114] in, (·) indicates global average pooling, and Flatten(·) indicates flattening operation.

[0115] The resulting 2048-dimensional vector This is the final output of the local feature extraction branch. It should be noted that in this embodiment, the fully connected classification layer at the end of the original ResNet50 network has been replaced with an identity mapping, meaning no parameter transformation or dimensionality mapping is performed. Therefore, this 2048-dimensional vector is directly used as the branch output, containing deep local detail features extracted from the input image, which can be used by the subsequent feature fusion module.

[0116] S3022 utilizes the global feature extraction branch to extract global structural features from the preprocessed image.

[0117] The global feature extraction branch is used to extract global structural features of the image. In a preferred embodiment, this branch is constructed using a Vision Transformer (ViT)-based architecture, capturing long-range dependencies between arbitrary locations in the image through a self-attention mechanism, thereby extracting deep features that characterize the overall distribution and global patterns of the material's microstructure. This branch uses a single-channel image with a size of 448×448 pixels. As input, its processing flow includes the following steps in sequence:

[0118] (1) Image Blocking and Serialization. To convert a two-dimensional image into a one-dimensional sequence that can be processed by the Transformer, this branch divides the input 448×448 single-channel image into 32×32 pixel image blocks in a non-overlapping manner. Since 448 / 32=14, the image is uniformly divided into 14×14=196 image blocks. For each image block, its pixel values ​​are flattened into a vector of length 1024 (32×32×1). This operation is implemented through tensor rearrangement, converting the input image into a sequence of image blocks. This process can be represented as:

[0119]

[0120] in, For the input image, For the output sequence, This refers to the batch size; Tensor rearrangement is an operation that reorganizes an image from a spatial dimension to a sequence dimension.

[0121] This operation does not introduce trainable parameters, but only changes the data organization to obtain a sequence containing features of 196 image patches. Each image block has a feature dimension of 1024.

[0122] It should be noted that the 32×32 pixel image block division method described above is a specific example of this embodiment. In other embodiments of the present invention, the image block size can be adjusted to other specifications (e.g., 16×16 or 64×64 pixels) according to the resolution of the input image and specific analysis requirements. When the image block size changes, the input dimension of the linear projection layer, the positional encoding length, and some parameters of the Transformer encoder (such as the attention head dimension, the feedforward network intermediate layer dimension, etc.) need to be adjusted accordingly. These adjustments are conventional adaptations by those skilled in the art based on the known ViT architecture principles. As long as this branch maintains the core process of extracting global features through image block division, serialization, and self-attention mechanism, it falls within the protection scope of the present invention.

[0123] (2) Linear Projection and Position Encoding. This step first maps each 1024-dimensional image patch vector to the hidden dimension space of the Transformer through a trainable linear projection layer. Then, a learnable class token (CLS token) is added at the beginning of the sequence, and a learnable position encoding vector is added to each position in the sequence to embed spatial position information, compensate for the Transformer's insensitivity to spatial position information, and form the initial input sequence of the encoder. .

[0124] (3) Global dependency modeling is performed using stacked Transformer encoders. In this embodiment, the feature sequence... Depth feature transformation is performed using a 6-layer Transformer encoder. Each encoder layer employs layer normalization (LSN). The standard structure is "Multi-Head Self-Attention - Residual Connections - Layer Normalization - Feed-Forward Network (FFN) - Residual Connections". For the ... Layer encoder ( Its input is the output of the previous layer. The processing procedure can be uniformly represented as:

[0125] Layer normalization and multi-head self-attention. First, layer normalization is performed on the input sequence. Then, the normalized features are passed through three trainable linear projection layers to generate corresponding query (Q) matrices, key (K) matrices, and value (V) matrices. To extract features from different subspaces in parallel, in a preferred embodiment, each matrix is ​​divided into three segments in the feature dimension (1024 dimensions), each segment being 512 dimensions. Then each segment is further divided into 8 attention heads, each head having a dimension of For each attention head, its scaled dot product attention is calculated, then normalized using the Softmax function, and the value matrix is ​​weighted and aggregated to obtain the output of that head. Finally, the outputs of all attention heads are concatenated along the feature dimension and fused into a multi-head attention output through a trainable linear layer.

[0126] Residual connection and normalization: Connect the above multi-head attention output to the encoder input. The residuals are summed to obtain intermediate features. These intermediate features are then subjected to layer normalization again.

[0127] • Feedforward Network and Residual Output: The normalized features are fed into a feedforward network, which consists of two linear layers and an intermediate GELU activation function to introduce a nonlinear transformation. Finally, the output of the feedforward network is residually concatenated with the aforementioned intermediate features to obtain the final output of this encoder layer. .

[0128] Through the progressive stacking and interaction of six encoder layers, features at any position in the sequence can be fully integrated with global contextual information, ultimately outputting a deep feature sequence containing rich long-range dependencies. .

[0129] (4) Feature aggregation and output projection. This refers to the feature sequence output from the encoder. To extract global information representing the entire image, this step first performs mean pooling on the features at all locations in the sequence. Specifically, the 1024-dimensional vector corresponding to each of the 197 locations in the feature sequence is summed and averaged, thus condensing the sequence information into a single 1024-dimensional feature vector. This operation, instead of traditional class token pooling, integrates the feature contributions from all locations, forming a more robust global feature representation.

[0130] Subsequently, the feature vectors obtained by pooling are subjected to layer normalization to stabilize their numerical distribution; then, a trainable linear projection layer is used to further refine and transform the features.

[0131] Finally, the global feature extraction branch outputs a 1024-dimensional global feature vector. This vector integrates the overall structural information and long-range semantic association of the input image and is used to feed it into the subsequent multimodal feature fusion module.

[0132] S303 utilizes a multimodal feature fusion module to fuse local detail features with global structural features to obtain fused features.

[0133] The multimodal feature fusion module is responsible for effectively fusing the features output from the local feature extraction branch and the global feature extraction branch to form a unified feature representation that combines local details with global macroscopic information. In a preferred implementation, the multimodal feature fusion module includes a feature concatenation unit and a fusion mapping unit, which perform the following operations in sequence:

[0134] (1) The local detail features and the global structural features are spliced ​​together in the channel dimension by the feature splicing unit to obtain the spliced ​​features.

[0135] First, the 2048-dimensional feature vector output by the local feature extraction branch is processed. The 1024-dimensional feature vector output by the global feature extraction branch By concatenating along the feature dimensions, a 3072-dimensional fused feature vector is obtained. , can be represented as:

[0136]

[0137] Here, || indicates that splicing is performed on the last dimension.

[0138] This operation integrates complementary information from local details (such as pore edges and grain boundary morphology) and the overall structure (such as phase distribution and grain arrangement).

[0139] (2) The spliced ​​features are sequentially input into at least two fully connected layers through the fusion mapping unit for nonlinear transformation and dimensionality reduction to obtain fusion features for subsequent classification.

[0140] To map high-dimensional spliced ​​features to a lower-dimensional space more suitable for classification tasks while preserving effective information, this module employs a fusion network containing two fully connected layers. Nonlinear transformation and dimensionality reduction are performed, where,

[0141] The first fully connected layer maps the 3072-dimensional input to 1024 dimensions, introduces nonlinearity using the ReLU activation function, and applies a dropout rate (e.g., 0.2) to enhance generalization ability. It can be represented as:

[0142]

[0143] in, This is the weight matrix of the first fully connected layer. Let be the corresponding bias vector, and ReLU(·) be the activation function.

[0144] The second fully connected layer further maps the 1024-dimensional features to 512 dimensions, also using the ReLU activation function and Dropout (e.g., 0.2), which can be represented as:

[0145]

[0146] in, This is the weight matrix of the second fully connected layer. Let be the corresponding bias vector, and ReLU(·) be the activation function.

[0147] After the above two transformations, a 512-dimensional fused feature vector is finally output. This vector integrates the microscopic details captured by the local feature extraction branch and the macroscopic structural features extracted by the global feature extraction branch, forming a complete, multi-scale feature representation of the input image, providing a rich and robust feature foundation for subsequent temperature classification decisions.

[0148] It should be noted that the above-described fusion method, which involves feature concatenation followed by mapping to a fully connected layer, is only a preferred embodiment. In other embodiments of the present invention, the multimodal feature fusion module may also employ other fusion strategies capable of effectively integrating local and global features, such as weighted summation of the two feature vectors or the introduction of a cross-attention mechanism for interactive fusion. As long as the module can effectively combine the complementary feature information extracted from the two branches to form a unified enhanced feature representation, similar technical effects can be achieved, and the invention falls within its scope of protection.

[0149] S304 uses the classification decision module to process the fused features and outputs the sintering temperature category and corresponding prediction probability for the corundum spinel material.

[0150] As a preferred implementation, the classification decision module in this embodiment is built based on a multilayer perceptron (MLP) to map the 512-dimensional fused features output by the multimodal feature fusion module to specific sintering temperature categories and output the predicted probabilities for each category. This module adopts a hierarchical structure of fully connected layer-activation function-random discarding-fully connected layer-probability transformation to progressively complete feature compression and classification decision-making. Specifically, it includes the following steps:

[0151] (1) Feature transformation and compression. The 512-dimensional fused features are transformed and compressed. The input is fed into the first fully connected layer, which maps it to a 256-dimensional feature space and introduces a nonlinear transformation using the ReLU activation function. Subsequently, a Dropout operation is applied to the output of this layer, randomly dropping elements with a set probability (e.g., 0.1) to suppress overfitting.

[0152] (2) Calculation of category scores. The 256-dimensional features after the above processing are input into the second fully connected layer and mapped to the dimension corresponding to the number of categories (in this embodiment, there are 5 sintering temperatures, namely 110℃, 800℃, 1100℃, 1400℃, and 1600℃), to obtain the unnormalized category scores.

[0153] (3) Probability Transformation and Output. The scores of each category are transformed into a probability distribution using the Softmax function, so that the predicted probability of each category is between 0 and 1, and the sum of the probabilities of all categories is 1. Finally, a probability vector is output, where each dimension represents the predicted probability of the input sample belonging to the corresponding sintering temperature category, thus completing the decision mapping from the microstructure image to the sintering temperature classification.

[0154] S4 generates a decision heatmap corresponding to the preprocessed image based on the sintering temperature classification results of the material analysis model. The decision heatmap reveals the microstructural basis for the model's classification decision.

[0155] This step uses visualization technology to reveal the image regions that the model focuses on for classification decisions and analyzes their microstructure, transforming the original "black box" prediction into a physically verifiable "white box" analysis, thereby providing a physical explanation of the microstructure for the classification results of sintering temperature.

[0156] As a preferred implementation, this embodiment uses Gradient Weighted Class Activation Mapping (Grad-CAM) to generate a decision heatmap. It should be noted that Grad-CAM primarily generates a heatmap reflecting the contribution of different image regions to the final decision by calculating the gradient of the target class score relative to the feature map of a specific convolutional layer (generally the last convolutional layer). Considering that this model uses a dual-branch structure, where the global feature extraction branch (ViT branch) does not contain convolutional layers, the Grad-CAM calculation in this embodiment is based only on the feature map and corresponding gradient information generated by the last convolutional layer of ResNet50 in the local feature extraction branch (i.e., the last convolutional layer of the last residual block of Layer 4). Specifically, it includes the following steps:

[0157] (1) It should be understood that for a pre-trained material analysis model, the system saves the pre-trained weight parameter file of the model. When using Grad-CAM to generate a decision heatmap, a Grad-CAM instance is created using tools such as Python and the pre-trained model is loaded. This instance automatically registers a hook function in the target convolutional layer (the last convolutional layer of ResNet50 mentioned above). During the forward propagation of the model on the preprocessed input image, the hook function captures the feature map output by the convolutional layer in real time; subsequently, the model backpropagates the predicted sintering temperature category score, and the hook function synchronously captures the gradient relative to the feature map.

[0158] (2) Perform global average pooling on the acquired gradient data in the spatial dimensions (width and height) to obtain a weight vector. Each weight value in this vector represents the importance of the corresponding feature map channel to the sintering temperature category predicted by the model.

[0159] (3) The weight vector calculated in step (2) is weighted and summed with the corresponding feature map to generate a preliminary class activation map. Then, the ReLU function is applied to this activation map to filter out regions that have a negative contribution to the prediction result, and only regions that have a positive impact on the target class are retained. The filtered activation map is upsampled to the same size as the input preprocessed image using methods such as bilinear interpolation, so as to obtain the original class activation heatmap that is spatially aligned with the input preprocessed image.

[0160] (4) Perform pseudo-color mapping on the original heatmap generated in step (3) to enhance the visual effect. Specifically, the applyColorMap function of OpenCV can be used with JET color maps, where red represents high activation / high contribution regions and blue represents low activation regions. Convert the generated BGR format color heatmap to RGB format, and then overlay it onto the corresponding preprocessed image using image overlay techniques such as alpha fusion to finally generate an intuitive decision heatmap.

[0161] The decision heatmap clearly indicates the image regions that the model focuses on when classifying sintering temperatures. By analyzing the microstructure in these regions, a nonlinear relationship between sintering temperature and microstructure can be established, thus providing a direct physical explanation for the model's classification decisions.

[0162] Furthermore, to transform the aforementioned visual focus into a quantifiable materials science explanation, in some embodiments, the image regions of particular interest in the heatmap can be correlated with microstructural parameters extracted from the preprocessed image of the input model. Preferably, binarization processing can be used to extract the following six core performance parameters of the microstructure: pore count, total surface area, perimeter, circularity, solidity, and integral density (IntDen), providing a quantitative basis for model interpretability analysis.

[0163] Furthermore, using the aforementioned generated decision heatmap, an attention threshold can be set to identify the key regions upon which the model relies for classification decisions. Specifically, the attention values ​​of all pixels in the heatmap can be sorted in descending order. For example, the pixel regions with values ​​in the top 20% can be defined as high-attention regions, and the remaining regions as low-attention regions. Then, the values ​​of the six microstructural parameters mentioned above are calculated within the high-attention regions and across the entire heatmap. For each parameter, the relative difference rate between the high-attention regions and the global region is calculated, and a statistical significance test is performed.

[0164] Through the above comparison, a mapping relationship between specific sintering temperature categories and quantitative characteristics of microstructure is established. For example, if the model predicts the sintering category at 1600℃, and the average roundness of its high-interest region is significantly lower than the global average, and the boundary perimeter is significantly higher, it can be explained that the model's decision is mainly based on those irregularly shaped and complex-boundary pore features, which are typical morphologies formed by abnormal grain growth and pore merging during ultra-high temperature sintering.

[0165] This step, by combining heatmap visualization with qualitative or quantitative analysis of microstructure, not only enhances the credibility and interpretability of model decisions, but also provides a clear direction for microstructure control in material process optimization.

[0166] Furthermore, in a preferred embodiment, the material analysis model (ResVi model) of this embodiment is obtained through a training process including the following steps:

[0167] (1) Obtain original microstructure image data of corundum spinel materials at different sintering temperatures.

[0168] Preferably, the original image data was acquired through independent experiments in this embodiment. Specifically, five groups of corundum spinel material samples prepared at typical sintering temperatures (110℃, 800℃, 1100℃, 1400℃, and 1600℃) were selected. The standard samples of 20×20×20 mm³ were continuously scanned using the Diondo D2 high-resolution omnidirectional micro-nano computed tomography (μ-CT) imaging system to acquire high-resolution slice image data with a total capacity of 22.98 GB, forming an original microstructure image dataset covering the temperature gradient.

[0169] It should be noted that, in some implementations, microstructure image data of corundum spinel materials under different sintering processes can also be obtained from publicly available scientific databases, shared research data platforms, or published academic literature to expand the scale and diversity of the dataset.

[0170] (2) Standardize the original microstructure images and divide the preprocessed image data into training set, validation set and test set.

[0171] Because raw μ-CT images inherently contain imaging defects and noise, directly using them for model training would severely impact their accuracy and generalization ability. Therefore, the standardized preprocessing procedure defined in step S2 above is performed on all raw slice images of the samples. This procedure sequentially includes invalid data removal, image integrity filtering, edge signal processing, and size normalization. Notably, in the data processing for the model training phase, before size normalization, an additional random cropping operation of a single image is introduced to enhance data diversity and the model's generalization ability, while suppressing overfitting.

[0172] Ultimately, a total of 2852 uniformly sized, high-quality single-channel grayscale images were obtained. Each image is clearly labeled with a sintering temperature and clearly represents the typical microstructure at the corresponding temperature (such as pore morphology and grain morphology). The above image data were randomly divided into training, validation, and test sets according to a preset ratio for model training, optimization, and evaluation.

[0173] (3) ResVi model training and optimization.

[0174] The initialized ResVi model was trained using the partitioned training and validation sets. The training process aimed to minimize the cross-entropy loss, employing the Adam optimizer and backpropagation algorithm to iteratively update the model parameters. The batch size was adjustable between 4 and 32. After training, the model performance was evaluated on an independent test set to ensure high accuracy in sintering temperature classification and good stability. Ultimately, an accurate correlation model was established between the microstructural characteristics of corundum spinel and sintering temperature categories.

[0175] It should be noted that the ResVi model constructed in this embodiment has its training time effectively controlled within 69-87 minutes, and its convergence speed is 3-4 times faster than ViT. When the batch size is 8, it only requires 12 iterations (epochs) to achieve an accuracy of 0.9, while ViT requires 40 epochs. Moreover, the loss value is as low as 0.008, and the training curve fluctuates little.

[0176] Furthermore, the trained ResVi model consistently achieved an accuracy of over 0.98 (with an optimal accuracy of 0.9974) for different batches and imaging parameters in the microstructure classification task of corundum spinel, which is significantly higher than traditional CNN models (such as ConvNeXt with an accuracy of 0.74-0.89) and single Transformer models (such as ViT with an accuracy of 0.91-0.93).

[0177] Example 2

[0178] To further verify the performance and interpretability of the ResVi model, this embodiment designs a comprehensive evaluation experiment. The experiment selects test samples covering five sintering temperatures (110°C, 800°C, 1100°C, 1400°C, and 1600°C) from the dataset, and uses ResVi models independently trained under four different batch sizes (4, 8, 16, and 32) to perform inference on these samples and generate decision heatmaps.

[0179] It should be noted that a series of continuous slice images were obtained for each sample at each sintering temperature using μ-CT. This sequence reflects the complete structure of the sample in the spatial dimension, and the slice images at different locations in the sequence show differences in microscopic morphology. To comprehensively examine the model's ability to recognize features at different spatial locations at the same temperature, this experiment selected four representative slices from the corresponding image sequence for each sintering temperature, including the slice at the first 1 / 4 of the sequence (Q1), the slice at the middle (1 / 2) of the sequence (Q2), the slice at the last 1 / 4 of the sequence (Q3), and the slice at the end of the sequence (max).

[0180] Using the methods described above, a diverse test sample set was constructed, encompassing different sintering temperatures and different spatial locations at the same temperature. Subsequently, multiple models trained with different batch sizes were used for inference and heatmap visualization, allowing for multi-dimensional and multi-level analysis of the models' classification consistency, feature capture ability, and the rationality of their decision-making criteria.

[0181] Experimental results show that all ResVi models trained under different batch sizes can accurately classify the sintering temperature of all test samples, proving that the models have stable and reliable sintering temperature recognition capabilities under different training configurations and their performance does not degrade due to minor fluctuations in training settings.

[0182] The heatmap visualization results on which each model's decision is based are as follows: Figure 4 As shown in the figure, the ResVi model exhibits unique edge-aware characteristics at temperatures of 110°C and 1400°C. Despite significant differences in temperature and microstructure, and variations in batch size, the model consistently focuses on the edge regions of the image. This phenomenon reveals the model's boundary sensitivity under temperature conditions, indicating its ability to effectively identify residual pristine grain boundaries at low temperatures and initial grain boundary features formed at 1400°C. These edge details provide crucial information for distinguishing intermediate temperature states.

[0183] Furthermore, within the intermediate-temperature sintering range of 800°C to 1100°C, the ResVi model exhibits a significant feature learning gradient. For the low quartile samples (Q1, Q2), the model primarily focuses on edge structures; while for the high quartile samples (Q3, max), especially at batch sizes of 8, 16, and 32, the model's attention span rapidly expands to the entire image region. This indicates that ResVi can comprehensively capture the continuous evolution from isolated pores to the formation of connected networks under these temperature conditions.

[0184] Furthermore, the ResVi model exhibits stable classification performance and adaptive feature learning capabilities across different batch sizes, with a significantly enhanced feature coverage as the batch size increases. The model focuses more on key local details during mini-batch training. When the batch size increases from 4 to 32, the model's thermal response gradually expands from local hotspots to covering the entire image region, indicating that large-batch training effectively improves the model's overall understanding of the complex porous network formed during high-temperature sintering, thereby comprehensively capturing the unique densification structure and phase distribution characteristics under high-temperature conditions.

[0185] The above findings demonstrate that the ResVi model not only maintains stable classification performance under different training conditions, but its decision-making process can also be traced back to stable and interpretable microstructural features through heatmaps, reflecting the intrinsic physical laws of material sintering. This proves that the ResVi model obtained in this embodiment not only has good performance, but also a transparent and reliable decision-making mechanism with excellent interpretability. It can establish a reliable correlation mapping between "sintering temperature - microstructure - model decision," providing a high-performance and highly reliable solution for intelligent materials analysis.

[0186] Example 3

[0187] This embodiment further provides an intelligent analysis device for the microstructure of corundum spinel material, specifically including:

[0188] The data acquisition unit is used to acquire the original microstructure image of the corundum spinel material to be analyzed.

[0189] The data preprocessing unit is used to perform standardized preprocessing on the original microstructure image to obtain a preprocessed image.

[0190] The model analysis unit is used to input the preprocessed image into a pre-trained material analysis model to obtain the sintering temperature category and corresponding prediction probability of the corundum spinel material. The material analysis model includes an input receiving module, a dual-branch feature extraction module, a multimodal feature fusion module, and a classification decision module. The dual-branch feature extraction module includes parallel local feature extraction branches and global feature extraction branches. The material analysis model performs classification decisions by fusing local detail features and global structural features extracted in parallel from the preprocessed image.

[0191] The visualization unit is used to generate a decision heatmap corresponding to the preprocessed image based on the sintering temperature classification results of the material analysis model, and to reveal the microstructural basis for the model to make classification decisions through the decision heatmap.

[0192] It should be noted that the structure and function of each unit in the device have been described in detail in the above embodiments, and will not be repeated here.

Claims

1. A method for intelligent analysis of the microstructure of corundum spinel material, characterized in that, Includes the following steps: Obtain the original microstructure image of the corundum spinel material to be analyzed; The original microstructure image is subjected to standardization preprocessing to obtain the preprocessed image; The preprocessed image is input into a pre-trained material analysis model to obtain the sintering temperature category and corresponding predicted probability of the corundum spinel material. The material analysis model includes an input receiving module, a dual-branch feature extraction module, a multimodal feature fusion module, and a classification decision module. The dual-branch feature extraction module includes parallel local feature extraction branches and global feature extraction branches. The material analysis model performs classification decisions by fusing local detail features and global structural features extracted in parallel from the preprocessed image. Based on the sintering temperature classification results of the material analysis model, a decision heatmap corresponding to the preprocessed image is generated. The decision heatmap reveals the microstructural basis for the model's classification decision.

2. The intelligent analysis method for the microstructure of corundum spinel material according to claim 1, characterized in that, Before inputting the preprocessed image into a pre-trained material analysis model to obtain the sintering temperature category and corresponding predicted probability of the corundum spinel material, the process further includes: To obtain original microstructure image data of corundum spinel materials at different sintering temperatures; The original microstructure image is standardized and preprocessed, and the preprocessed image data is divided into training set, validation set and test set. The material analysis model initially constructed is trained using the training set. The model parameters are optimized by minimizing the cross-entropy loss and using the backpropagation algorithm to obtain the trained material analysis model.

3. The intelligent analysis method for the microstructure of corundum spinel material according to claim 1 or 2, characterized in that, The original microstructure image was obtained by micro-nano computed tomography scanning of the target material sample, and the normalization preprocessing includes: Invalid data removal processing includes removing images whose signal intensity at the beginning and end is lower than a set threshold from the original microstructure image sequence obtained from sample scanning; Image integrity filtering processing includes identifying and removing images with incomplete upper and lower edge data due to uneven physical preparation of the sample; Edge signal processing includes cropping edge regions of each retained image to remove low signal-to-noise ratio or no-signal areas at the edges of the image. Size standardization processing includes adjusting the processed image to a preset size; When the standardization preprocessing is used for image data processing during the model training phase, data augmentation processing is included before the size standardization processing. The data augmentation processing includes random cropping of a single image.

4. The intelligent analysis method for the microstructure of corundum spinel material according to claim 1, characterized in that, The material analysis model makes classification decisions by fusing local detail features and global structural features extracted in parallel from the preprocessed image, including: The preprocessed image is received using the input receiving module. The dual-branch feature extraction module is used to extract local detail features and global structural features from the preprocessed image in parallel, including extracting local detail features from the preprocessed image using the local feature extraction branch and extracting global structural features from the preprocessed image using the global feature extraction branch. The local detail features and the global structural features are fused using the multimodal feature fusion module to obtain fused features; The fusion features are processed using the classification decision module to output the sintering temperature category and corresponding prediction probability of the corundum spinel material.

5. The intelligent analysis method for the microstructure of corundum spinel material according to claim 4, characterized in that, The local feature extraction branch is constructed based on a structure-adjusted ResNet50 network; the extraction of local detail features from the preprocessed image using the local feature extraction branch includes: The preprocessed image is then converted into a three-channel image adapted for the ResNet50 network. The three-channel image is input into the structure-adjusted ResNet50 network for feature extraction to obtain a deep feature map, wherein the structure adjustment is to remove the fully connected layer at the end of the ResNet50 network; The deep feature map is subjected to global average pooling and flattening to obtain the local detail features.

6. The intelligent analysis method for the microstructure of corundum spinel material according to claim 4, characterized in that, The global feature extraction branch is built based on the ViT architecture; the extraction of global structural features from the preprocessed image using the global feature extraction branch includes: The preprocessed image is segmented into image blocks and converted into feature sequences; The feature sequence is linearly projected and positionally encoded to obtain the encoded feature sequence; The encoded feature sequence is input into a multi-layer Transformer encoder, and long-distance dependencies are captured through a self-attention mechanism to obtain a deep feature sequence; The deep feature sequence is subjected to mean pooling and linear projection to obtain the global structural features.

7. The intelligent analysis method for the microstructure of corundum spinel material according to claim 4, characterized in that, The multimodal feature fusion module includes a feature concatenation unit and a fusion mapping unit; the process of fusing the local detail features and the global structural features using the multimodal feature fusion module to obtain fused features includes: The feature splicing unit is used to splice the local detail features and the global structural features along the channel dimension to obtain the spliced ​​features. The spliced ​​features are sequentially input into at least two fully connected layers through the fusion mapping unit for nonlinear transformation and dimensionality reduction to obtain the fusion features used for subsequent classification.

8. The intelligent analysis method for the microstructure of corundum spinel material according to claim 4, characterized in that, The classification decision module is built based on a multilayer perceptron; the process of using the classification decision module to process the fused features and output the sintering temperature category and corresponding predicted probability of the corundum spinel material includes: Receive the fused features output by the multimodal feature fusion module; The fused features are sequentially passed through at least two fully connected layers of the classification decision module for feature transformation and dimensionality compression to obtain category scores corresponding to different sintering temperature categories. The category scores are converted into a probability distribution using the Softmax function, and the sintering temperature category and its corresponding predicted probability are output.

9. The intelligent analysis method for the microstructure of corundum spinel material according to claim 1, characterized in that, The decision heatmap is generated using Grad-CAM technology, which calculates and generates the decision heatmap based on the feature map output by the last convolutional layer in the local feature extraction branch and the corresponding gradient information.

10. A smart analysis device for the microstructure of corundum spinel material, characterized in that, include: The data acquisition unit is used to acquire the original microstructure image of the corundum spinel material to be analyzed. The data preprocessing unit is used to perform standardized preprocessing on the original microstructure image to obtain a preprocessed image. The model analysis unit is used to input the preprocessed image into a pre-trained material analysis model to obtain the sintering temperature category and corresponding prediction probability of the corundum spinel material. The material analysis model includes an input receiving module, a dual-branch feature extraction module, a multimodal feature fusion module, and a classification decision module. The dual-branch feature extraction module includes parallel local feature extraction branches and global feature extraction branches. The material analysis model performs classification decisions by fusing local detail features and global structural features extracted in parallel from the preprocessed image. The visualization unit is used to generate a decision heatmap corresponding to the preprocessed image based on the sintering temperature classification results of the material analysis model, and to reveal the microstructural basis for the model to make classification decisions through the decision heatmap.

Images