CN-122024233-A - Decoupling phase segmentation method and system for microstructure of composite gel system

Abstract

The invention belongs to the technical field of phase segmentation, and discloses a decoupling phase segmentation method and a decoupling phase segmentation system for a microstructure of a composite gelation system, wherein (1) a reusable particle database is formed by particle images corresponding to each target gelation particle, (2) particles in the reusable particle database are inserted into a cement matrix area to obtain a phase 1 model training data set, a phase 1 general segmentation model is trained, 3 particles in the reusable particle database are adopted to train a phase 2 particle classification model, (4) the BSE images to be split are input into the phase 1 general segmentation model to obtain a matrix, pores and particle masks, and the phase 2 classification model is subjected to joint classification based on the particle masks and all the particles in the BSE images to be split to obtain specific phase types of each particle. The invention improves the universality.

Inventors

• ZHANG SHISHUN
• GAO YINGHAO
• Gan Yidong

Assignees

• 华中科技大学

Dates

Publication Date

20260512

Application Date

20260327

Claims (10)

1. 1. A decoupling phase segmentation method facing to a microstructure of a composite gel system is characterized by comprising the following steps: (1) Preparing a single component standard sample for each target gelled particle phase, and extracting, screening and storing standardized particle images by using an Otsu threshold method and a connected domain marking algorithm, wherein the obtained particle images form a reusable particle database; (2) Preparing an artificially marked pure cement paste BSE image, taking the artificially marked pure cement paste BSE image as a background image, and inserting particles in a reusable particle database into a cement matrix area to generate stage 1 model training data, so as to form a stage 1 model training data set; (3) Training the phase 1 general segmentation model by adopting a phase 1 model training data set, and training the phase 2 particle classification model by adopting particles in a reusable particle database; (4) And cutting particle slices based on the particle masks of the BSE image to be split, and inputting all particle slices in the same BSE image to be split as particle sets into the stage 2 particle classification model for joint classification to obtain the specific phase class of each particle.
2. 2. The method of phase-splitting with decoupling oriented composite gelling system microstructures as set forth in claim 1, wherein the stage 2 particle classification model is trained after applying edge erosion enhancement to particles in the reusable particle database.
3. 3. A method for phase-splitting decoupling of microstructure of a complex gel system as defined in claim 2, wherein the phase-splitting is a phase-splitting phase Hydration level at target age of Pixel ratio of erosion applied to the phase particles during training From the interval And (5) uniformly and randomly sampling.
4. 4. The method of claim 1, wherein the phase 1 universal segmentation model is a U-Net architecture, the preprocessing phase is to read the synthesized image and the corresponding three types of masks according to batch, and perform data enhancement, the training phase is to extract multi-scale features through the encoder-decoder structure, and multi-type cross entropy loss optimization parameters are adopted.
5. 5. The method of phase-separation decoupling for composite gel system microstructure according to claim 1, wherein the phase 2 particle classification model uses a context-aware aggregate attention framework, which performs joint processing on all particles extracted from the same BSE image as an unordered aggregate, so that the classification process of each particle can be based on morphological information of other particles in the particle aggregate.
6. 6. The method for phase-division decoupling of composite gel system microstructure according to claim 5, wherein the phase 2 particle classification model includes a gray scale feature encoder that calculates an average gray scale value for each particle image The average gray value is obtained by a multi-layer perceptron Embedded as gray feature vectors The gray feature encoder provides a direct gray level signal for classification using the physical characteristics of gray values of different phases in the BSE image in relation to the average atomic number.
7. 7. The method of phase-division decoupling for a composite cementitious microstructure according to claim 1, wherein the phase 2 particle classification model includes an available class constraint encoder that encodes available class constraints when the cementitious mix is known to obtain class constraint embedded vectors, and concatenates the class constraint embedded vectors to the characteristic tensor for each particle.
8. 8. The method of phase-separation decoupling of composite gel system microstructure according to claim 1, wherein the phase 2 particle classification model comprises a set transformer encoder that processes the particle feature set using an induced set attention block Each inducing set attention block comprises two multi-head attention sub-layers, wherein the first multi-head attention sub-layer enables inducing points to focus on all particle characteristics to aggregate global information, and the second multi-head attention sub-layer enables each particle to focus on the inducing points to acquire context information.
9. 9. A decoupling phase separation system facing a composite gelling system microstructure is characterized by comprising a memory and a processor, wherein the memory stores a computer program, and the processor executes the decoupling phase separation method facing the composite gelling system microstructure according to any one of claims 1-8 when executing the computer program.
10. 10. A computer readable storage medium, characterized in that it stores machine executable instructions that, when invoked and executed by a processor, cause the processor to implement the composite gel architecture microstructure oriented decoupled phase segmentation method according to any one of claims 1-8.

Description

Decoupling phase segmentation method and system for microstructure of composite gel system Technical Field The invention belongs to the technical field of phase segmentation, and particularly relates to a decoupling phase segmentation method and system for a microstructure of a composite gelation system. Background Phase segmentation and quantitative analysis aiming at the microstructure of a composite cement material system are important for scientific research and engineering quality control of the material. Phase segmentation, i.e. the accurate identification and differentiation of the type, number, morphology and spatial distribution of different phases (e.g. unhydrated particles, hydration products, pores, etc.) in microscopic images (usually back-scattered electron microscopy BSE images), is the basis for quantitative analysis. With the development of low-carbon cement technology, mineral admixtures such as fly ash, blast furnace slag and the like are widely applied to a composite gelling system, so that the microstructure of the mineral admixtures is more complex, and great challenges are brought to accurate phase segmentation. The current mainstream phase separation technology is mainly divided into two types: The segmentation method based on the BSE-EDS combines gray information of the BSE image and point/plane scanning element information of the EDS. The pixel brightness of the BSE image is related to the average atomic number of the point, and can be initially distinguished between heavy (bright) and light (dark) element-rich regions. EDS can provide specific points of chemical composition to accurately determine the phase. The method has the advantages of high accuracy, but has the disadvantages of low EDS analysis speed, high cost and high analysis difficulty. Segmentation methods based on image processing and machine learning such methods attempt to complete segmentation with BSE images only. Early methods relied on gray threshold methods, but the segmentation accuracy was inadequate due to the overlapping of gray scale intervals for different types of supplementary cementitious materials (e.g., fly ash, slag). In recent years, with the development of deep learning, semantic segmentation models typified by U-Net and the like have been applied to this field. The method can divide according to complex features such as the appearance, the texture and the like of the phase by learning a large number of artificially marked BSE images. In the prior art, machine vision and deep learning/machine learning are commonly adopted to carry out phase segmentation on microscopic images, namely, firstly, sample images (such as BSE and X-CT) are acquired, and a segmentation model (such as U) is trained based on pixel-level labeling dataNet and variants thereof), or to use random forest/UNet for extracted features of texture, edge, contrast, etc. Relevant representative disclosures include: CN118067577A (cement clinker mineral quantification method capable of automatically segmenting mineral phases) is characterized in that a clinker mineral (such as C3S, C S, holes and intermediate phases) is segmented and quantified by adopting UNet, CN118691825A (composite cement system phase segmentation method and electronic device based on machine vision) is characterized in that ROI labeling is carried out on a BSE image, multiple characteristics are extracted, a training random forest is trained to complete multi-phase segmentation and calculate volume fractions, and application number 202510092362.6 (concrete phase segmentation method based on light deep learning and SE attention mechanism) is applied to U Light weight and SE attention are introduced in Net, and the segmentation precision and efficiency are improved. Although the above prior art solutions achieve good results in specific scenarios, their inherent "end-to-end" and "full supervision" modes bring about the following core drawbacks: the model lacks versatility and has high expansion cost, and the model trained by the existing scheme is highly dependent on the phase combination in the training data. For example, a model trained on a "cement+fly ash" system cannot identify "slag". When a new auxiliary cementing material is required to be introduced into the system, the complete flow of preparing a new sample, collecting a new image, manually marking a property phase and retraining a model is required to be repeated, so that the phase-separating model is difficult to adapt to a material system which is continuously changed in research, and the expansion cost is extremely high. The data marking burden is heavy, the types of phases to be distinguished are more and more along with the increase of the complexity of the gel system, and the difficulty and the workload of manual marking are rapidly increased. The expert needs to carefully distinguish between various phases with similar morphology and grey scale, which is highly error prone and inefficient. This is a