Quantitative analysis method for sintered neodymium-iron-boron magnet structure

Abstract

The invention discloses a quantitative analysis method for a sintered neodymium-iron-boron magnet structure, and belongs to the technical field of material microstructure analysis. The method is characterized by comprising the following steps: defining microscopic structure descriptors (such as a size factor, a shape factor, a main phase area fraction, Nd-rich phase coating property and the like) according to domain knowledge; constructing a backscattered electron image data set of the sintered neodymium-iron-boron magnet, performing fine marking, and performing fine tuning on the model on the data set to adapt to the characteristics of sintered neodymium-iron-boron image segmentation; performing high-precision segmentation on the sintered NdFeB back scattering electron image by using the trained MagSAM model, and preprocessing a segmentation mask; and carrying out feature extraction and quantitative analysis on the segmentation result based on the defined microscopic structure descriptor. According to the method, full-automatic and high-precision segmentation and standardized quantitative analysis of the complex microscopic structure in the sintered neodymium iron boron backscattered electron image are realized, and a set of universal structure evaluation system is established.

Description

Technical Field

[0001] This invention relates to the field of materials microstructure analysis technology, and more specifically, to a quantitative analysis method for the microstructure of sintered NdFeB magnets, which is used to achieve high-precision identification and standardized quantitative characterization of complex microstructures. Background Technology

[0002] With the rapid development of computer vision technology, image processing technology is increasingly widely used in the field of material microstructure analysis. In April 2023, Meta AI released a large-scale segmentation model, SegmentAnything Model (SAM), with zero-sample transfer capability. This model has attracted widespread attention in the field of image processing due to its millisecond-level segmentation speed and interactive segmentation features. Subsequently, in June 2023, Ke et al. improved SAM, enhancing its ability to segment local details in images, demonstrating great application potential in the field of material microstructure image segmentation (Ke, L., Ye, M., Danelljan, M., Liu, Y., Tai, Y., Tang, C., & Yu, F. (2023). Segment Anything in High Quality). ArXiv, abs / 2306.01567 . ).

[0003] Since their introduction in 1983, sintered NdFeB magnets have significantly contributed to the rapid development of modern science and technology and the information industry. Their high remanence and coercivity have led to their widespread application in new energy vehicles, wind power generation, and consumer electronics. The coercivity and remanence of sintered NdFeB magnets are highly dependent on their microstructure characteristics. However, existing technologies for rapid evaluation of the microstructure of sintered NdFeB magnets have significant shortcomings: firstly, the industry lacks a systematic quantitative evaluation system for microstructure, and existing evaluation methods are too simplistic, failing to extract key descriptors affecting performance from complex structures; secondly, the SAM model suffers from insufficient segmentation accuracy and numerous grain boundary burrs in segmentation tasks on backscattered electron images of sintered NdFeB magnets; furthermore, magnet structure evaluation heavily relies on expert visual interpretation of backscattered electron images. This method is highly subjective, time-consuming, and lacks repeatability, making it difficult to meet the needs of industrial-scale batch testing. Therefore, establishing a rapid quantitative evaluation system for microstructure is of great significance for the development and design of high-performance magnets. Summary of the Invention

[0004] This invention aims to establish a systematic rapid evaluation system for the microstructure of sintered NdFeB magnets. First, key microstructure descriptors are extracted. Based on an improved SAM model, fine-tuning and training are performed on a sintered NdFeB backscattered image dataset to obtain high-precision segmentation results. Then, the microstructure descriptors are automatically extracted and calculated based on the high-precision segmentation results, achieving quantitative correlation analysis between microstructure and magnetic properties.

[0005] The technical solution adopted in this invention is: a method for quantitative analysis of the microstructure of sintered NdFeB magnets, comprising the following steps: Step 1: Extract microstructure descriptors for sintered NdFeB magnets based on domain knowledge, including but not limited to: size factor, shape factor, main phase area fraction, and Nd-rich phase coverage.

[0006] Step 2: Construction of backscattered electron image dataset of sintered NdFeB magnets and fine-tuning of SAM model.

[0007] The specific steps of step 2 are as follows: Step 2.1: Acquire backscattered electron images of sintered NdFeB magnets. To make the microstructure clearly visible, adjust the brightness and contrast to enhance the visual effect. Each image is labeled with pixel-level mask labels for the main phase grains, Nd-rich phase, and other phases. The dataset is randomly divided into training, validation, and test sets in a 3:1:1 ratio.

[0008] Step 2.2: Import the dataset into the model for fine-tuning training to adapt to the segmentation characteristics of sintered NdFeB images, and obtain a dedicated MagSAM model adapted for sintered NdFeB image segmentation.

[0009] Step 3: Perform automatic high-precision segmentation of the backscattered electron image of the sintered NdFeB magnet based on the trained MagSAM model and preprocess the segmentation results. Step 4: Based on the processed segmentation mask, extract and calculate the microstructure descriptor defined in Step 1, and generate an analysis report.

[0010] Further, step 2.1 specifically includes: labeling each image in the sintered NdFeB magnet backscattered electron image dataset with pixel-level mask labels for the main phase grains, Nd-rich phase, and other phases.

[0011] Further, step 2.2 specifically includes: adding an additional ViT encoder to the feature vector generated by the Image Encoder to extract early ViT features and final ViT features (64×64). These are then upsampled to 256×256 through transposed convolution to align with the features of the Mask Decoder. The three sets of features are fused globally and locally through independent convolutional layers. The resulting HQ-Features enhance the model's learning of fine features such as grain boundaries. During training, the sintered NdFeB backscattered electron image dataset is randomly divided into training, validation, and test sets in a 3:1:1 ratio. The Encoder and Decoder parts of the original SAM architecture are frozen, and only the feature fusion module is fine-tuned. The model evaluates the segmentation results using mIoU as the metric. Training terminates when the model reaches the set maximum number of iterations. The relevant training methods can be obtained directly from https: / / github.com / SysCV / sam-hq.

[0012] Further, step 4 specifically includes: generating a distribution histogram of microstructure descriptors, drawing a visualization chart of the distribution of microstructure descriptors, and performing correlation analysis of magnetic properties.

[0013] Furthermore, step 4 specifically includes performing erosion and dilation operations and median filtering operations on the obtained high-quality mask to smooth the boundaries and eliminate image noise.

[0014] This invention provides a quantitative analysis method for the microstructure of sintered NdFeB magnets, with the following advantages: 1. Based on domain knowledge, a multi-dimensional microstructure feature descriptor covering size, shape, main phase area fraction, and Nd-rich phase coverage is extracted; 2. Through multi-scale feature fusion technology, a dedicated MagSAM model suitable for sintered NdFeB magnet segmentation is developed, reducing grain boundary burrs and improving segmentation accuracy; 3. A rapid quantitative evaluation system for the microstructure of sintered NdFeB magnets is established, significantly improving testing efficiency compared to manual methods and providing a reliable digital analysis tool for magnet design optimization. Attached Figure Description

[0015] Figure 1 This is the overall flowchart of automatic segmentation and quantitative analysis of backscattered electron images of sintered NdFeB magnets according to the present invention. Figure 2 This is a schematic diagram of the identification results of the sintered NdFeB test set in an example of the present invention. Detailed Implementation

[0016] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0017] A quantitative analysis method for the microstructure of sintered NdFeB magnets, such as Figure 1 As shown, it includes the following steps: Step 1: Extract microstructure descriptors for sintered NdFeB magnets based on domain knowledge, including but not limited to: size factor, shape factor, main phase area fraction, and Nd-rich phase coverage. Step 2: Construction of backscattered electron image dataset of sintered NdFeB magnets and fine-tuning of SAM model; Step 2.1: Acquire backscattered electron images of sintered NdFeB magnets. To make the microstructure clearly visible, adjust the brightness and contrast to enhance the visual effect. Each image is labeled with pixel-level mask labels for the main phase grains, Nd-rich phase, and other phases. The dataset is randomly divided into training, validation, and test sets in a 3:1:1 ratio.

[0018] Step 2.2: Import the collected dataset of 842 images into the model for fine-tuning training to adapt to the segmentation characteristics of sintered NdFeB images, and obtain a dedicated MagSAM model adapted to sintered NdFeB image segmentation.

[0019] To address the feature vectors generated by the Image Encoder, an additional ViT encoder is added to extract early and final ViT features (64×64). These are then upsampled to 256×256 via transposed convolutions to align with the Mask Decoder's features. The three sets of features are fused through independent convolutional layers to achieve global and local feature fusion. The resulting HQ-Features enhance the model's learning of fine features such as grain boundaries. During training, the sintered NdFeB backscattered electron image dataset is randomly divided into training, validation, and test sets in a 3:1:1 ratio. The Encoder and Decoder parts of the original SAM architecture are frozen, and only the feature fusion module is fine-tuned. The model is evaluated using mIoU as the metric for segmentation results, and training terminates when the model reaches the set maximum number of iterations. The relevant training methods can be obtained directly from https: / / github.com / SysCV / sam-hq.

[0020] This example compares the segmentation performance of SAM and MagSAM on the test set. MagSAM provides more accurate and refined segmentation of grain boundary contours, reducing grain boundary "burr" phenomena and improving the evaluation metric mIoU from 0.89 to 0.92. This improvement significantly enhances the accuracy of descriptor extraction and calculation. Specific results are shown below. Figure 2 As shown.

[0021] Step 3: Perform automatic high-precision segmentation of the backscattered electron image of the sintered NdFeB magnet based on the trained model and preprocess the segmentation results; Step 3.1: Perform erosion, dilation, and median filtering operations on the obtained high-quality mask to smooth the boundaries and eliminate image noise. Step 4: Based on the processed segmentation mask, extract and calculate the microstructure descriptors defined in Step 1, draw a visualization chart of the distribution of microstructure descriptors, and perform magnetic property correlation analysis.

Claims

1. A method for quantitative analysis of the microstructure of sintered NdFeB magnets, characterized in that: Includes the following steps: Step 1: Extract microstructure descriptors for sintered NdFeB magnets based on domain knowledge, including but not limited to: size factor, shape factor, main phase area fraction, and Nd-rich phase encapsulation. Step 2: Construction and fine-tuning of the large segmentation model Segment AnythingModel (SAM) for the backscattered electron image dataset of sintered NdFeB magnets; Step 2.1: Acquire backscattered electron images of sintered NdFeB magnets, adjust brightness and contrast to enhance image visual effects; then perform fine annotation on each image; The dataset is divided into training, validation, and test sets according to a certain ratio. Step 2.2: Import the dataset into the model for fine-tuning training to adapt to the segmentation characteristics of sintered NdFeB microstructure, and obtain a dedicated MagSAM model adapted to the segmentation of sintered NdFeB microstructure; Step 3: Perform automatic high-precision segmentation of the backscattered electron image of the sintered NdFeB magnet based on the trained model and preprocess the segmentation results; Step 4: Based on the processed segmentation mask, extract and calculate the microstructure descriptor defined in Step 1, and generate an analysis report.

2. The method according to claim 1, characterized in that, Step 2.1 specifically includes: each image in the sintered NdFeB magnet backscattered electron image dataset is labeled with pixel-level mask tags for the main phase grains, Nd-rich phase, and other phases.

3. The method according to claim 1, characterized in that, Step 2.2 specifically includes: adding an additional ViT encoder to the feature vector generated by ImageEncoder to extract early ViT features and final ViT features (64×64). The two are upsampled to 256×256 through transposed convolution to align with the features of Mask Decoder; the three sets of features are fused into global and local features through independent convolutional layers, and the generated HQ-Features can enhance the model's learning of fine features such as grain boundaries; during training, the sintered NdFeB backscattered electron image dataset is randomly divided into training set, validation set and test set in a 3:1:1 ratio, the Encoder and Decoder parts in the original SAM architecture are frozen, and only the feature fusion module is fine-tuned. The model evaluates the segmentation results using mIoU as the indicator, and training is terminated when the model reaches the set maximum number of iterations.

4. The method according to claim 1, characterized in that, Step 4 specifically includes: generating a distribution histogram of microstructure descriptors, drawing a visualization chart of the distribution of microstructure descriptors, and performing correlation analysis of magnetic properties.

5. The method according to claim 1, characterized in that, Step 4 further includes performing erosion and dilation operations and median filtering operations on the obtained high-quality mask to smooth the boundaries and eliminate image noise.

Images