Method and system for discriminating microscopic components and hydration process of cement

Abstract

The invention provides a cement microscopic component and hydration process discrimination method and system, and relates to the technical field of cement component identification, and the method comprises the steps: obtaining a low-power SEM image and a high-power SEM image of a cement-based material; respectively inputting the low-power SEM image and the high-power SEM image into a dual-channel microscopic component segmentation network, extracting global features of the low-power SEM image through a low-channel network branch, and extracting local features of the high-power SEM image through a high-channel network branch; fusing the global features and the local features in a multi-head self-attention mechanism to obtain a fused feature map; semantic segmentation is carried out by using the fusion feature map to obtain a multi-scale component segmentation image of the cement-based material; and extracting multi-feature data of an unhydrated particle area proportion, a hydration product area proportion and an average pore diameter from the multi-scale component segmentation image, inputting the multi-feature data into the process classification model, and outputting a hydration process category label to realize classification judgment of the hydration stage of the cement-based material.

Description

Technical Field

[0001] The present disclosure relates to the technical field of cement component identification, and particularly to a method and system for discriminating microscopic components and hydration processes of cement. Background Art

[0002] The statements in this section merely provide background technical information related to the present disclosure and do not necessarily constitute prior art.

[0003] Cement-based materials are widely used in the fields of architecture and civil engineering. The types of their microscopic components and the hydration process directly affect the strength, durability, and other key properties of the materials. Therefore, accurate identification of the cement microstructure and dynamic judgment of the hydration process are of great significance.

[0004] Traditional analysis of cement microscopic components and judgment of the hydration process rely on manual identification of cement microscopic components in scanning electron microscope (SEM) images, such as unhydrated particles like C3S, C2S, C3A, C4AF, etc. and hydration products like C-S-H, CH, and ettringite. This method is not only inefficient, but also requires high professional knowledge and experience of the observer, and it is difficult to ensure consistency and accuracy when identifying complex microstructures.

[0005] In recent years, the development of deep learning technology has provided strong support for image recognition and classification tasks. By constructing a neural network model, it is possible to train on a large amount of data sets to accurately identify the cement microscopic components in SEM images.

[0006] However, the microscopic structural features of cement include both large-scale macroscopic information (such as particle shape, pore distribution, etc.) and microscopic-scale detailed information (such as the microscopic morphology of hydration products). Existing methods usually analyze SEM images at a single magnification, unable to comprehensively capture the multi-scale information of the cement microstructure, resulting in insufficient judgment accuracy of cement components and hydration stages. Summary of the Invention

[0007] To solve the above problems, the present disclosure proposes a method and system for discriminating microscopic components and hydration processes of cement. By designing a dual-channel convolutional neural network, which respectively receives and processes low-magnification images and high-magnification images, introducing a multi-head self-attention mechanism to improve the accuracy of feature maps, calculating key features such as the area ratio of unhydrated particles, the area ratio of hydration products, and the average pore diameter, and classifying and judging the hydration stage to describe the hydration stage of cement materials, an efficient and intelligent judgment of the hydration process is achieved.

[0008] According to some embodiments, the present disclosure adopts the following technical solutions:

[0009] A method for discriminating microscopic components and hydration processes of cement, comprising:

[0010] Obtain low-magnification SEM images and high-magnification SEM images of cement-based materials, and preprocess them;

[0011] Input the preprocessed low-magnification SEM image and high-magnification SEM image into a dual-channel microscopic component segmentation network respectively. Extract the global features of the low-magnification SEM image through the low-channel network branch, and extract the local features of the high-magnification SEM image through the high-channel network branch;

[0012] Fuse the global features and local features in a multi-head self-attention mechanism to obtain a fused feature map;

[0013] Use the fused feature map for semantic segmentation to obtain a multi-scale component segmentation image of the cement-based material;

[0014] Extract multi-feature data such as the area ratio of unhydrated particles, the area ratio of hydration products, and the average pore diameter from the multi-scale component segmentation image. Input the multi-feature data into a process classification model, and output the hydration process category label to realize the classification and judgment of the hydration stage of the cement-based material.

[0015] According to some embodiments, the present disclosure adopts the following technical solutions:

[0016] A cement microscopic component and hydration process discrimination system, comprising:

[0017] An image acquisition module, configured to obtain low-magnification SEM images and high-magnification SEM images of cement-based materials, and preprocess them;

[0018] A feature extraction module, configured to input the preprocessed low-magnification SEM image and high-magnification SEM image into a dual-channel microscopic component segmentation network respectively. Extract the global features of the low-magnification SEM image through the low-channel network branch, and extract the local features of the high-magnification SEM image through the high-channel network branch;

[0019] A feature fusion module, configured to fuse the global features and local features in a multi-head self-attention mechanism to obtain a fused feature map;

[0020] A component segmentation module, configured to use the fused feature map for semantic segmentation to obtain a multi-scale component segmentation image of the cement-based material;

[0021] A process discrimination module, configured to extract multi-feature data such as the area ratio of unhydrated particles, the area ratio of hydration products, and the average pore diameter from the multi-scale component segmentation image. Input the multi-feature data into a process classification model, and output the hydration process category label to realize the classification and judgment of the hydration stage of the cement-based material.

[0022] According to some embodiments, the present disclosure adopts the following technical solutions:

[0023] A computer program product includes a computer program which, when executed by a processor, implements the method for discriminating micro-components and hydration process of cement as described above.

[0024] According to some embodiments, the present disclosure adopts the following technical solutions:

[0025] A non-transitory computer-readable storage medium is used to store computer instructions which, when executed by a processor, implement the method for discriminating micro-components and hydration process of cement as described above.

[0026] According to some embodiments, the present disclosure adopts the following technical solutions:

[0027] An electronic device includes: a processor, a memory, and a computer program; wherein, the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device runs, the processor executes the computer program stored in the memory so that the electronic device executes and implements the method for discriminating micro-components and hydration process of cement as described above.

[0028] Compared with the prior art, the beneficial effects of the present disclosure are:

[0029] The method for discriminating micro-components and hydration process of cement of the present disclosure utilizes multi-scale information of low-magnification (100 - 1000 times) and high-magnification (1000 - 10000 times) SEM images, and respectively extracts features through a two-channel convolutional neural network. The low-magnification image channel is designed to capture macroscopic-scale features, and adopts deeper convolutional layers and large pooling kernels to aggregate large-scale structural information such as particles and pores; the high-magnification image channel focuses on microscopic details, adopts shallower convolutional layers and small pooling kernels, and can finely capture the microscopic features of hydration products. By fusing the two-scale information, the model can comprehensively judge the micro-components of cement, effectively improving the accuracy and stability of recognition.

[0030] The method for discriminating micro-components and hydration process of cement of the present disclosure introduces a multi-head self-attention mechanism to correlate the features extracted by the low-magnification and high-magnification image channels. The self-attention mechanism generates Query, Key, and Value matrices of the low-magnification and high-magnification image features, calculates the weight matrix of the features, and realizes the adaptive enhancement of the features, so that the information of the low-magnification channel and the high-magnification channel has higher correlation during the fusion process. This not only strengthens the model's ability to distinguish the micro-components of cement, but also improves the recognition accuracy of fine structures, ensuring the reliability of the semantic segmentation results.

[0031] A method for discriminating microconstituents and hydration process of cement discloses calculating key features such as the area ratio of unhydrated particles, the area ratio of hydration products, and the average pore diameter to describe the degree of hydration of cementitious materials. By inputting these features into a dynamic weight random forest classification model and designing a time series weight allocation algorithm to dynamically adjust the decision path of the random forest, the model can adjust the feature importance in real time according to the hydration process stage, thereby improving the classification performance and classifying the hydration stage. The hydration process is divided into five stages: pre-induction period, induction period, acceleration period, deceleration period, and stable period, realizing efficient and intelligent judgment of the hydration process. It can effectively monitor the microscopic changes of cement at each stage and provide dynamic feedback support for subsequent material design and construction processes. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] The accompanying drawings, which form a part of this disclosure, are used to provide a further understanding of this disclosure. The schematic embodiments and descriptions thereof of this disclosure are used to explain this disclosure and do not constitute an improper limitation of this disclosure.

[0033] Figure 1 Schematic diagram of the residual block structure for an embodiment of this disclosure;

[0034] Figure 2 Flowchart of the method for an embodiment of this disclosure. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0035] The present disclosure will be further described below in conjunction with the accompanying drawings and embodiments.

[0036] It should be noted that the following detailed descriptions are all illustrative and are intended to provide further explanations of this disclosure. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which this disclosure belongs.

[0037] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to this disclosure. As used herein, unless otherwise clearly specified in the context, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0038] Embodiment 1

[0039] An embodiment of this disclosure provides a method for discriminating microconstituents and hydration process of cement, including the following steps:

[0040] Step 1: Obtain low-magnification SEM images and high-magnification SEM images of cementitious materials and preprocess them;

[0041] Step 2: Input the preprocessed low-magnification SEM image and high-magnification SEM image into the dual-channel microstructure segmentation network respectively. Extract the global features of the low-magnification SEM image through the low-channel network branch, and extract the local features of the high-magnification SEM image through the high-channel network branch;

[0042] Step 3: Fuse the global features and local features in the multi-head self-attention mechanism to obtain a fused feature map;

[0043] Step 4: Use the fused feature map for semantic segmentation to obtain a multi-scale component segmentation image of the cement-based material;

[0044] Step 5: Extract multi-feature data such as the area ratio of unhydrated particles, the area ratio of hydration products, and the average pore diameter from the multi-scale component segmentation image. Input the multi-feature data into the process classification model, and output the hydration process category label to realize the classification and judgment of the hydration stage of the cement-based material.

[0045] As an embodiment, the specific implementation process of a method for discriminating cement microstructures and hydration processes according to the present disclosure is as follows:

[0046] Step 1: Obtain low-magnification SEM images and high-magnification SEM images of the cement-based material and preprocess them;

[0047] Collect cement-based material samples that have experienced different hydration times for SEM scanning, and collect low-magnification SEM images and high-magnification SEM images of each cement sample one by one to construct a data set; among them, the low magnification range is 100-1000 times, and the high magnification range is 1000-10000 times.

[0048] Furthermore, preprocess the low-magnification and high-magnification SEM images of each cement sample, including data augmentation (rotation, flipping), size adjustment, brightness adjustment, etc.

[0049] Furthermore, divide the data set into a training set and a test set. Specifically, the training set accounts for 70% of the total data set, and the test set accounts for 30% of the total data set. Labeling is performed on the training set for neural network model training. The labeling content includes the labeling of unhydrated particles C3S, C2S, C3A, C4AF; hydration products C-S-H, CH, ettringite, and pores.

[0050] Step 2: Construct and train a dual-channel microstructure segmentation network

[0051] Specifically, the dual-channel microscopic component segmentation network is a dual-channel convolutional neural network. The dual-channel microscopic component segmentation network includes a low-channel network branch and a high-channel network branch, which are respectively used to receive low-magnification images and high-magnification images. The low-channel network branch (low-magnification image channel) is suitable for deeper convolutional layers and larger pooling kernels, and is used to extract large-scale global features (macro features). The high-channel network branch (high-magnification image channel) is suitable for shallower convolutional layers and smaller pooling kernels, and is used to extract detailed local features.

[0052] Among them, two residual blocks are designed for each channel network branch, namely residual block 1 and residual block 2. Residual block 1 consists of two convolutional layers with 128 convolutional kernels and a convolutional window of 3×3. Residual block 2 consists of two convolutional layers with 256 convolutional kernels and a convolutional window of 3×3.

[0053] Further, the low-magnification image channel is set as follows: it includes an initial convolutional layer with 64 convolutional kernels and a convolutional window of 7×7; residual block 1 and residual block 2.

[0054] The high-magnification image channel is set as follows: it includes an initial convolutional layer with 64 convolutional kernels and a convolutional window of 3×3; residual block 1 and residual block 2.

[0055] Further, a multi-head self-attention mechanism is introduced after the outputs of the low-magnification image channel and the high-magnification image channel to fuse the feature maps of the low-magnification image channel and the high-magnification image channel, including: after extracting global features and local features, the feature maps of the global features and the feature maps of the local features are adjusted to the same size through interpolation operations, aligned in the spatial dimension, a multi-head attention mechanism is introduced after the two channel network branches, the two aligned feature maps are transformed into sequence form and input into the multi-head attention mechanism, the self-attention weight matrices of the feature maps of the global features and the feature maps of the local features are respectively multiplied by their corresponding Value matrices to obtain the output features after self-attention, and finally the output features after self-attention of each are concatenated along the channel dimension to obtain the fused feature map, and a deconvolution layer or an upsampling layer is used to restore the fused feature map to the size of the input image and perform semantic segmentation to generate a multi-scale component segmentation image.

[0056] Specifically, the feature maps of the global features and the feature maps of the local features after passing through the low-magnification image channel and the high-magnification image channel are respectively obtained. The feature map of the global features is F low , with a size of (B, C, H, W); the feature map of the local features is F high , with a size of (B, C, H, W). Among them, B is the batch size, that is, the number of images processed in one batch, C is the number of channels, which is 3 for RGB image channels, and H and W are the spatial dimensions of the feature map.

[0057] When Flow and F high When their spatial dimensions are different, they are adjusted to the same size through interpolation operations to ensure that F low and F high are aligned in the spatial dimension. To apply the self-attention mechanism, the feature map F needs to be converted into a sequence form by flattening the feature map into a two-dimensional matrix, where each row represents the feature vector of a position. The flattening operation is: F l = reshape(F low , (B, H, W, C)), F h = reshape(F high , (B, H, W, C)). After flattening, the size of F l , F h becomes (B, H×W, C).

[0058] As an embodiment, the fusion process after input into the multi-head self-attention mechanism includes the following steps:

[0059] First, generate Query, Key, and Value matrices for the feature map of the global feature and the feature map of the local feature respectively:

[0060] Q l , Q h = W q F l , W q F h

[0061] K l , K h = W k F l , W k F h

[0062] V l , V h = W v F l , W v F h

[0063] Among them, W q , W k , W v are learnable weight matrices used to generate Query, Key, and Value. After conversion, the dimensions of Query, Key, and Value are all (B, H×W, d), where d is the dimension of the attention head, which is reasonably selected according to the complexity of the task. Usually, 4, 6, or 8 can be selected and adjusted according to the performance of the model.

[0064] Perform a dot product on the Query and Key of each of the low-magnification image channel and the high-magnification image channel to calculate the self-attention weight matrix:

[0065] Among them, the attention weight of the low-magnification image channel:

[0066]

[0067] The attention weight of the high-magnification image channel:

[0068]

[0069] Multiply the self-attention weight matrices of the low-magnification channel and the high-magnification channel respectively with their corresponding Value matrices to obtain the output features after self-attention:

[0070] Output of the low-magnification image channel:

[0071] F l out =A l ·V l

[0072] Output of the high-magnification image channel:

[0073]

[0074] Finally, concatenate the outputs of the low-magnification and high-magnification channels along the channel dimension to obtain the fused feature map after fusion:

[0075]

[0076] Finally, use a transposed convolutional layer or an upsampling layer to restore the fused feature map to the size of the input image, perform semantic segmentation to generate a semantic segmentation map, and use the softmax function as the activation function.

[0077] The training process adopts forward propagation, calculates the loss, and then performs backpropagation optimization.

[0078] Finally, obtain the trained dual-channel microscopic component segmentation network, which can realize multi-scale information extraction and component segmentation of SEM images.

[0079] Step 3: Construct a process classification model and train it

[0080] Specifically, a dynamic optimization random forest is used as the process classification model for classification. As the cement hydration process progresses, the proportion of unhydrated particles gradually decreases, the proportion of hydration products increases, and the pore volume also gradually decreases. Therefore, first, multi-feature data such as the proportion of unhydrated particle area, the proportion of hydration product area, and the average pore diameter are extracted from the multi-scale component segmentation image, and the above data are standardized or normalized to ensure that each feature is on the same scale and enhance the model performance. The dataset is divided, with 70% as the training set and 30% as the test set. The hydration stage data is labeled for the training set for model training. The hydration stage is divided into five stages: pre-induction period, induction period, acceleration period, deceleration period, and stable period.

[0081] Furthermore, in different hydration stages, the contribution degrees of features to stage classification are different: the proportion of unhydrated particle area is more important in the early stage and gradually decreases as hydration progresses; the proportion of hydration product area is the opposite; the average pore diameter is the most important in the acceleration period, reflecting the change in hydration rate. Based on the above analysis, a time series weight allocation algorithm is designed to dynamically adjust the decision path of the random forest, enabling the model to adjust the feature importance in real time according to the hydration process stage, thereby improving the classification performance.

[0082] The process of constructing a dynamic weight random forest model is as follows:

[0083] Hydration stage time annotation: A pseudo-timestamp is deduced and assigned to the sample data to represent its relative time position in the hydration process. A reasonable time range is assigned to each hydration stage: pre-induction period S1 Induction period S2 Acceleration period S3(4, 12], deceleration period S4(12, 24], stable period S5(24, 48]. For the i-th sample belonging to a certain stage Sk, according to the time range (t k-1 , t k of this stage, a pseudo-timestamp is generated: t i = t k-1 + (t k - t k-1 ) · r, where r is a random number in the range [0, 1].

[0084] Definition of time-related weight function: For the case where the importance of features changes with time, a time-related weight function is defined for each feature. For the proportion of unhydrated particle area, exponential decay is adopted: For the proportion of hydration product area, linear increase is adopted: ω 2 (t) = k 2 · t + b 2 ; for the average pore diameter, a Gaussian function is adopted: where τ 1 , k2 , b 2 , μ 3 , σ 3 are key parameters and are reasonably set based on prior knowledge and experimental experience.

[0085] Random sampling: Assume there are N samples in the training set. Randomly draw N times to form a sub-dataset. The size of the new sub-dataset is also N, but the samples may be repeated. Repeat the sampling process to generate multiple different sub-datasets. Each sub-dataset is used to train a decision tree, that is, the trees in the random forest. The number of trees in the random forest needs to be set within a reasonable range.

[0086] Decision tree generation: For the sub-dataset, in the training stage, the decision tree starts from the root node, that is, all features in the sub-dataset, and randomly selects a feature X i and a split point s to perform node splitting. The splitting operation is as follows: For the data samples on each node, the decision tree divides the samples into two parts according to the selected feature X i and the split point s. If the feature value of the sample satisfies X i ≤ s, it enters the left child node; if the feature value of the sample satisfies X i > s, it enters the right child node. Traverse all possible split points on the child node for splitting until the stopping requirement is met, until a split point is found that minimizes the weighted Gini coefficient after splitting.

[0087] Among them, the number of selected features is generally the square root of the total number of features. The node splitting stopping requirements include the depth of the tree, the number of nodes, etc., and need to be set within a reasonable range.

[0088] Introduce time-related weight to adjust the Gini coefficient: Defined as the mixing degree of samples in node D, used to measure the purity of the node, and the formula is:

[0089]

[0090] Among them, ω i (t) is the time-related weight of the feature, and p i is the proportion of samples in the node belonging to class i.

[0091] The process of generating a random forest is as follows: For each tree, perform node splitting according to the sampled features until the stopping requirement is met. All the trees together form a random forest.

[0092] Ensemble prediction training process: In the training and testing stage, when a new sample X iWhen classifying a sample, each decision tree in the random forest will independently classify it. That is, when the sample reaches the last node, i.e., the leaf node, along the tree structure, a class label will be output. Assuming that the splitting rule of a certain tree can be represented by T as a set of decision paths, then the predicted class of the input sample x can be expressed as:

[0093]

[0094] where T(x) represents the path from the root node to the leaf node according to the splitting rule of decision tree T, and the finally output class is the majority class label included in the leaf node.

[0095] In the random forest, each decision tree T j (where j = 1, 2, …, M) will independently classify the sample X i and obtain a class prediction Since each tree uses different samples and subsets of features during training, their prediction results may be different. The final predicted class of the random forest is determined by majority voting, and the formula is:

[0096]

[0097] where c represents the class label, i.e., the hydration stage class; M is the total number of decision trees; is the indicator function, which takes 1 when and 0 otherwise; argmax c means finding the class c with the most votes.

[0098] After training is completed, the classification and judgment of the hydration stage of the cement-based material are realized.

[0099] As an embodiment, as Figure 2 shown, the application process of a method for discriminating cement microscopic components and hydration process in the present disclosure is as follows:

[0100] Step 1) Obtain low-magnification and high-magnification SEM scanning images of cement at different hydration stages;

[0101] Step 2) Input them into a dual-channel microscopic component segmentation network respectively. Among them, the global features of the low-magnification SEM image are extracted through the low-magnification image channel, and the local features of the high-magnification SEM image are extracted through the high-magnification image channel;

[0102] Step 3) Realize multi-scale feature fusion of the global features and local features in the multi-head self-attention mechanism to obtain a fused feature map;

[0103] Step 4) After the fused feature map restores the image size through the upsampling operation, semantic segmentation is performed to obtain a multi-scale component segmentation image;

[0104] Step 5) Extract the proportion of unhydrated particle area, the proportion of hydrated product area, and the average pore diameter in the multi-scale component segmentation image, and input them into the random forest model for classification to realize the judgment and recognition of the cement hydration stage.

[0105] Example 2

[0106] In an embodiment of the present disclosure, a system for discriminating cement microscopic components and hydration process is provided, including:

[0107] An image acquisition module, configured to acquire low-magnification SEM images and high-magnification SEM images of cement-based materials and perform preprocessing;

[0108] A feature extraction module, configured to input the preprocessed low-magnification SEM image and high-magnification SEM image into a dual-channel microscopic component segmentation network respectively, extract the global features of the low-magnification SEM image through the low-channel network branch, and extract the local features of the high-magnification SEM image through the high-channel network branch;

[0109] A feature fusion module, configured to fuse the global features and local features in the multi-head self-attention mechanism to obtain a fused feature map;

[0110] A component segmentation module, configured to perform semantic segmentation using the fused feature map to obtain a multi-scale component segmentation image of the cement-based material;

[0111] A process discrimination module, configured to extract multi-feature data such as the proportion of unhydrated particle area, the proportion of hydrated product area, and the average pore diameter from the multi-scale component segmentation image, input the multi-feature data into the process classification model, and output the hydration process category label to realize the classification and judgment of the hydration stage of the cement-based material.

[0112] Example 3

[0113] In an embodiment of the present disclosure, a computer program product is provided, including a computer program, and when the computer program is executed by a processor, it implements the method for discriminating cement microscopic components and hydration process as described above.

[0114] Example 4

[0115] In an embodiment of the present disclosure, a non-transitory computer-readable storage medium is provided, and the non-transitory computer-readable storage medium is used to store computer instructions, and when the computer instructions are executed by a processor, the method for discriminating cement microscopic components and hydration process as described above is implemented.

[0116] Example 5

[0117] In one embodiment of the present disclosure, an electronic device is provided, including: a processor, a memory, and a computer program; wherein, the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device runs, the processor executes the computer program stored in the memory so that the electronic device executes a method for discriminating cement microcomponents and hydration processes as described above.

[0118] The present disclosure is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to embodiments of the present disclosure. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, as well as the combination of flows and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, such that the instructions executed by the processor of the computer or other programmable data processing devices generate means for implementing the functions specified in one process Figure 1 one process or multiple processes and / or blocks Figure 1 or multiple blocks.

[0119] These computer program instructions can also be loaded onto a computer or other programmable data processing devices, such that a series of operation steps are executed on the computer or other programmable devices to generate computer-implemented processing, so that the instructions executed on the computer or other programmable devices provide steps for implementing the functions specified in one process Figure 1 one process or multiple processes and / or blocks Figure 1 or multiple blocks.

[0120] Although the specific embodiments of the present disclosure have been described above in conjunction with the accompanying drawings, this is not a limitation on the protection scope of the present disclosure. Those skilled in the art should understand that, based on the technical solutions of the present disclosure, various modifications or deformations that can be made by those skilled in the art without creative efforts are still within the protection scope of the present disclosure.

Claims

1. A method for distinguishing cement microscopic components and hydration progress, characterized in that: include: Obtain low-magnification SEM images and high-magnification SEM images of cement-based materials and pre-process them; The preprocessed low-magnification SEM image and high-magnification SEM image are respectively input into the dual-channel micro-component segmentation network, the global features of the low-magnification SEM image are extracted through the low-channel network branch, and the local features of the high-magnification SEM image are extracted through the high-channel network branch; The global features and local features are fused in a multi-head self-attention mechanism to obtain a fused feature map; The fused feature map is used for semantic segmentation to obtain a multi-scale component segmentation image of cement-based materials; Multi-feature data such as the area ratio of unhydrated particles, the area ratio of hydration products and the average pore diameter are extracted from the multi-scale component segmentation image, and the multi-feature data are input into the process classification model to output the hydration process category label to realize the classification and judgment of the hydration stage of cement-based materials.

2. A method for distinguishing cement microscopic components and hydration progress as claimed in claim 1, characterized in that: The low magnification range of the low magnification SEM image is 100-1000 times, and the high magnification range of the high magnification SEM image is 1000-10000 times. The preprocessing of the low magnification SEM image and the high magnification SEM image includes data enhancement, size adjustment and brightness adjustment.

3. A method for distinguishing cement micro-components and hydration progress as claimed in claim 1, characterized in that: The dual-channel micro-component segmentation network includes a low-channel network branch and a high-channel network branch. The low-channel network branch is suitable for deep convolution layers and large pooling kernels to extract large-scale global features, and the high-channel network branch is suitable for shallow convolution layers and small pooling kernels to extract detailed local features. Two residual blocks are designed for each channel network branch.

4. A method for distinguishing cement microscopic components and hydration progress as claimed in claim 1, characterized in that: After extracting global features and local features, the feature maps of global features and local features are adjusted to the same size through interpolation operations and aligned in the spatial dimension. A multi-head attention mechanism is introduced after the two channel network branches. The two aligned feature maps are converted into sequence form and input into the multi-head attention mechanism. The self-attention weight matrices of the feature maps of global features and local features are multiplied by their corresponding Value matrices to obtain the output features after self-attention. Finally, the output features after self-attention are spliced ​​along the channel dimension to obtain the fused feature map. The deconvolution layer or upsampling layer is used to restore the fused feature map to the size of the input image and perform semantic segmentation to generate a multi-scale component segmentation image.

5. A method for distinguishing cement microscopic components and hydration progress as claimed in claim 1, characterized in that: A dynamic weighted random forest model was constructed and used as a process classification model. The area ratio of unhydrated particles, the area ratio of hydration products and the average pore diameter were extracted from the multi-scale component segmentation image and input into the dynamic weighted random forest model to classify and judge the hydration stage. The hydration process was divided into five stages: pre-induction period, induction period, acceleration period, deceleration period and stable period.

6. A method for distinguishing cement microscopic components and hydration progress as claimed in claim 5, characterized in that: When the sample data of the area ratio of unhydrated particles, the area ratio of hydration products and the average pore diameter are input, each decision tree of the random forest will independently classify the sample. When the sample walks along the tree structure to the last node, the leaf node, it will output a category label. The final output category is the majority category label contained in the leaf node. In the dynamic weight random forest, each decision tree will independently classify the sample x and obtain a category prediction. The final prediction category of the dynamic weight random forest is determined by majority voting.

7. A system for distinguishing cement micro-components and hydration progress, characterized in that: include: An image acquisition module, used for acquiring low-magnification SEM images and high-magnification SEM images of cement-based materials and preprocessing them; A feature extraction module is used to input the preprocessed low-magnification SEM image and high-magnification SEM image into the dual-channel micro-component segmentation network respectively, extract the global features of the low-magnification SEM image through the low-channel network branch, and extract the local features of the high-magnification SEM image through the high-channel network branch; The feature fusion module is used to fuse global features and local features in a multi-head self-attention mechanism to obtain a fused feature map; The component segmentation module is used to perform semantic segmentation using the fused feature map to obtain a multi-scale component segmentation image of the cement-based material; The process discrimination module is used to extract multi-feature data such as the area ratio of unhydrated particles, the area ratio of hydration products and the average pore diameter from the multi-scale component segmentation image, input the multi-feature data into the process classification model, output the hydration process category label, and realize the classification and judgment of the hydration stage of cement-based materials.

8. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, a method for distinguishing cement microscopic components and hydration progress as described in any one of claims 1 to 6 is implemented.

9. A non-transitory computer-readable storage medium, characterized in that: The non-transitory computer-readable storage medium is used to store computer instructions. When the computer instructions are executed by the processor, a method for distinguishing cement microscopic components and hydration progress as described in any one of claims 1 to 6 is implemented.

10. An electronic device, characterized in that: include: A processor, a memory and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory so that the electronic device executes a method for distinguishing cement microscopic components and hydration process as described in any one of claims 1-6.

Images