Search

CN-121999941-A - Multi-scale strain-microstructure analysis method and system for material with self-adaptive grid, multi-mode fusion and life prediction

CN121999941ACN 121999941 ACN121999941 ACN 121999941ACN-121999941-A

Abstract

The invention discloses a material multi-scale strain-microstructure analysis method and system with self-adaptive grids, multi-mode fusion and life prediction. The method comprises the steps of acquiring multi-angle projection data of a material under loading conditions by a switchable multi-mode internal imaging probe, fusing multi-source sensing information to form unified volume reconstruction, realizing three-dimensional noise reduction and detail enhancement by utilizing a generated countermeasure network, calculating and screening high-reliability data based on voxel confidence coefficient, inputting the data into a self-adaptive grid refinement module, realizing the improvement of global and local matching precision, establishing quantitative association between microstructure characteristics and a strain field input statistics/depth modeling module after the microstructure characteristics are segmented by a convolutional neural network, and outputting fatigue life or failure probability results through a life prediction interface. Compared with the prior art, the method improves the suitability of the multi-scale strain measurement method in an imaging mode, solves the problem that the fixed grid division precision and the calculation efficiency are difficult to consider, and obviously reduces the matching error in a low signal-to-noise ratio scene.

Inventors

  • XI LI
  • WANG YAN

Assignees

  • 北京理工大学

Dates

Publication Date
20260508
Application Date
20260130

Claims (10)

  1. 1. A material multi-scale strain-microstructure analysis method with self-adaptive grid, multi-mode fusion and life prediction is characterized by comprising the following steps: Firstly, collecting multi-angle projection data of a material sample under loading conditions through an internal three-dimensional imaging probe with a switchable imaging mode, wherein the imaging mode comprises an X-ray tomography mode, an optical coherence tomography mode or a scanning electron microscope mode, and multi-source sensing data comprising thermal imaging, infrared imaging, ultrasonic detection or spectral imaging can be accessed; secondly, performing spatial registration and gray scale normalization processing on the multi-angle projection data, and forming a unified volume data set by using mode normalization factors determined by mutual information maximization; thirdly, performing global digital volume related calculation on the unified volume data set through a coarse grid unit to obtain a preliminary result of a macroscopic displacement field and a strain field; Step four, according to the strain gradient distribution of the macroscopic strain field, generating fine grid units by the self-adaptive grid controller, executing local fine scale iterative optimization on a strain concentration area, and calling and generating an countermeasure network to perform noise suppression and texture detail enhancement on the volume data before calculation; Fifthly, in the fine-scale iterative optimization, calculating voxel confidence through a voxel confidence calculation unit based on local gray stability, local gradient size and matching residual errors respectively, and participating in matching update only for voxels higher than a preset threshold; sixthly, fusing displacement and strain data subjected to coarse scale and fine scale processing to form a multi-scale strain field, and extracting microstructure features through a convolutional neural network segmentation module; And seventhly, inputting the strain field and the microstructure characteristics into a statistical regression, long-short-term memory network or Bayesian network modeling module to establish a quantitative association model, outputting a fatigue life or failure probability prediction result through a life prediction interface module, and visually displaying prediction distribution in a three-dimensional model.
  2. 2. The method of claim 1, wherein the switchable imaging mode internal three-dimensional imaging probe of the first step comprises an imaging source, a detector and a multi-mode switching member, the multi-mode switching member being an electronically controlled switching structure for adjusting the different imaging modes optical axis to coincide with the geometric center.
  3. 3. The method of claim 1, wherein the gray scale normalization process of the second step determines normalized weighting factors for each mode by maximizing mutual information between the fusion result and the reference mode volume data while maintaining gradient consistency of each mode at the edge structure during the fusion process.
  4. 4. The method of claim 1, wherein the adaptive mesh controller in the fourth step automatically adjusts the spatial arrangement density of fine mesh cells according to the second-order strain gradient index η, refines the cell size in the region where the strain gradient is higher than the set threshold, and sparsely distributes the cells in the region where the strain is uniform, so as to optimize the overall calculation efficiency and the local matching accuracy.
  5. 5. The method of claim 1, wherein the voxel confidence in the fifth step The method is characterized by comprising the steps of obtaining weight parameters through standard sample training, wherein the weight parameters are used for realizing stability and universality under different noise conditions, and the weight parameters comprise the ratio of the standard deviation of local gray scale to the mean value, the normalized local gradient size and the weighted sum of the inverse of the matching residual error.
  6. 6. The method of claim 1, wherein the generating the countermeasure network in the fourth step uses a three-dimensional convolution generator and multi-discriminant structure, and the loss function includes a countermeasure loss, a reconstruction error, and a high frequency component fidelity loss to achieve both noise suppression and detail preservation.
  7. 7. The method of claim 1, wherein the microstructure features in the sixth step include grain boundaries, two-phase interfaces, hole distribution, fiber orientation, and cell alignment, and the convolutional neural network segmentation module is a three-dimensional U-Net structure and has a bi-directional data interface with an associated modeling module or a graph neural network to achieve online feature optimization.
  8. 8. The method according to claim 1, wherein the lifetime prediction interface module in the seventh step calculates a fatigue lifetime or failure probability distribution according to the strain field parameters, the microstructure evolution parameters, the load sequence, and the material fatigue constitutive model, and the prediction result can be mapped to the material three-dimensional model and displayed in a multidimensional manner through the visual interface.
  9. 9. A material multi-scale strain-microstructure analysis system with adaptive mesh, multi-modal fusion and life prediction, comprising: The data acquisition module is used for acquiring multi-angle projection data of the material sample under the loading condition through an internal three-dimensional imaging probe capable of switching imaging modes and can be connected with a multi-source sensor signal; the data processing module is used for executing multi-mode data registration fusion, three-dimensional reconstruction and multi-scale digital volume correlation calculation with generation of counternetwork noise reduction and voxel confidence weighting; The microstructure modeling module is used for extracting microstructure features through convolutional neural network segmentation, inputting the microstructure features and a multi-scale strain field into a statistical modeling or deep learning network together, and establishing a quantitative association relation; The life prediction module is used for outputting a fatigue life or failure probability prediction result according to the correlation model; And the visual interface is used for superposing and displaying the strain field, the microstructure characteristics and the life prediction result in the three-dimensional geometric model.
  10. 10. The system of claim 9, wherein the data acquisition module comprises an imaging source, a detector, and a multimode switching probe having an optical or ray path adjustment function adapted to X-ray tomography, optical coherence tomography, and scanning electron microscope modes; The multi-scale digital volume correlation calculation unit of the data processing module comprises a coarse grid unit, a fine grid unit and a self-adaptive grid controller, wherein the self-adaptive grid controller is used for automatically refining or sparsely refining the distribution of the fine grid unit according to a coarse-scale strain gradient and calling a three-dimensional generation countermeasure network noise reduction module to improve the data quality before fine-scale calculation, and the fine-scale strain calculation can be replaced by a mixed finite element-volume correlation calculation method FE-DVC and a three-dimensional reconstruction noise reduction method based on a diffusion probability model so as to adapt to analysis requirements of different material systems and failure mechanisms.

Description

Multi-scale strain-microstructure analysis method and system for material with self-adaptive grid, multi-mode fusion and life prediction Technical Field The invention relates to the technical fields of X-ray computer tomography, three-dimensional image processing and neural network, in particular to a material multi-scale strain-microstructure analysis method and system with self-adaptive grids, multi-mode fusion and life prediction. Background In the service process of the material, the mechanical property change and the failure behavior of the material often relate to the coupling evolution process from macroscopic to microscopic, multi-scale and multi-physical field. The distribution and evolution of the strain fields is not only affected by external loading, but is also closely related to internal microstructure features (e.g., grain boundary orientation, two-phase interface morphology, pore distribution, fiber alignment and orientation, etc.). Therefore, the method realizes the measurement of a trans-scale strain field, establishes the quantitative association relationship between the strain and the microstructure, and has important significance for understanding the failure mechanism, the optimal design and the life prediction of the material. However, despite the rapid development of digital volume correlation (Digital Volume Correlation, DVC), three-dimensional imaging, and machine learning techniques in recent years, the prior art has various bottlenecks. The earliest volume-related measurements relied primarily on single imaging modalities, such as DVC methods based on X-ray computed tomography (X-ray CT). Such methods derive displacement fields and strain fields by analyzing changes in voxel gray fields in the volumetric images before and after loading. Subsequently, the development of the synchrotron radiation X-ray tomography (SR-CT) technology remarkably improves the spatial resolution and the signal to noise ratio, and is applied to multi-scale deformation researches such as biological materials, metal fatigue crack initiation, composite material interface peeling and the like. However, such high-end imaging devices are expensive and demanding in application conditions, with high demands on sample shape, size, and density. In particular, when processing a region containing local high gradient strain (such as a crack tip, a heterogeneous interface, etc.), the fixed grid division method either sacrifices local precision and trades global calculation speed, or greatly increases calculation amount to maintain precision, which makes efficiency and precision difficult to be compatible. According to the multi-scale DVC method, fine-scale iterative optimization is performed on the local strain concentration area after coarse-scale grid matching, so that the calculation efficiency is improved compared with that of the traditional global fine-scale grid DVC. However, the refinement strategy adopted by the method is based on a fixed gradient threshold, cannot dynamically adapt to spatial distribution of strain characteristics under different material types and loading states, and meanwhile, in low-dose imaging data, the matching precision is still limited due to obvious noise influence. In addition, the existing multi-scale DVC is universally and equally weighted in voxel selection, namely all matching points are considered to be equally reliable, and a voxel confidence assessment mechanism is not introduced to reject low-confidence data. This results in regions with more noise or artifacts negatively affecting the global matching. In terms of noise suppression, conventional three-dimensional image preprocessing often uses linear or nonlinear filters such as gaussian smoothing and median filtering. These methods inevitably attenuate or obscure the microstructure features while reducing noise, especially the lack of fidelity to critical defect information such as internal microcracks, voids, etc. of the material. In recent years, generating a countermeasure Network (GAN) has achieved significant effects in two-dimensional image denoising and super-resolution reconstruction, and has gradually expanded to the field of three-dimensional medical image reconstruction. However, a special GAN noise reduction architecture for three-dimensional imaging data characteristics of materials is not available, and it is difficult for the existing scheme to combine texture detail fidelity and noise suppression performance required by strain measurement. In the link of strain and microstructure association analysis, the traditional method mainly extracts specific microstructure characteristics through image segmentation or manual labeling, and then carries out statistical association analysis with a strain field. This "step-wise" flow has significant delays in data interaction and on-line optimization, and most models do not integrate life prediction functionality. For example, some studies have utilized multiple regressi