BDU-Net Segmentation for γ′ Phase Distribution in Superalloys
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional methods for detecting and analyzing γ′ phases in metal materials are limited by small view fields, low statistical efficiency, and lack of representativeness, making it difficult to accurately quantify and analyze the large-scale distribution of γ′ phases in single crystal superalloys.
Innovation Solution
A full-view-field quantitative statistical distribution representation method using deep learning-based image segmentation, specifically optimizing the U-Net network to BDU-Net, for automatic and high-speed recognition and extraction of γ′ phases, enabling the collection and analysis of feature maps across a large area with high throughput and accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional image processing software (Image-Pro Plus, Photoshop) is used to manually process and analyze γ′ phase features, then measurement precision can be maintained for individual features, but productivity is severely limited to only hundreds to thousands of features per analysis
Solution Approach 1:
The patent replaces manual mechanical image processing operations with an automated computer-based system that uses trained models to automatically identify, segment, and measure γ′ phase features. This substitution of manual mechanical processing with automated computational methods enables the system to process hundreds of thousands of features while maintaining measurement precision through consistent algorithmic application.
Solution Approach 2:
The system performs self-service by automatically processing images without requiring manual intervention for each feature. The trained model autonomously identifies and measures all γ′ phase features in the images, eliminating the need for operators to manually process each individual feature while maintaining measurement quality.
2Measurement precision
If manual post-processing methods are used to meet quantitative statistics requirements, then measurement precision can be ensured, but loss of time increases significantly due to the labor-intensive nature of processing
Solution Approach 1:
The system performs preliminary action by pre-training models on labeled datasets before actual measurement tasks. This preliminary training phase enables the system to automatically perform measurements without requiring manual post-processing during the actual analysis, significantly reducing the time required while maintaining precision through the pre-established measurement protocols.
Solution Approach 2:
The patent replaces time-consuming manual post-processing operations with automated computational methods. The system automatically performs all measurement and analysis tasks that previously required manual intervention, eliminating the significant time loss associated with manual processing while maintaining measurement precision.
3Ease of operation
If only partial multiple view fields are observed for statistical analysis, then ease of operation is maintained, but reliability of statistical representativeness deteriorates because the sample size is insufficient to represent the global material properties
Solution Approach 1:
The system performs multi-functionality by automatically processing and analyzing multiple view fields simultaneously. Rather than requiring operators to selectively choose which fields to analyze, the system universally processes all provided images, ensuring comprehensive sampling and reliable statistical representativeness while maintaining ease of operation through automated workflows.
Solution Approach 2:
The system ensures continuity of useful action by automatically processing images in sequence without interruption or manual selection. This continuous automated processing ensures that all available view fields are analyzed, providing sufficient sample size for reliable statistical representation while maintaining operational simplicity.
4Measurement precision
If traditional SEM technology with limited field of view is used, then measurement precision for individual features can be achieved, but the area of stationary object that can be analyzed is severely constrained, preventing full-view-field statistical representation
Solution Approach 1:
The patent applies segmentation by dividing the large-area full-view-field analysis into multiple smaller image fields that can be processed individually with high precision. The system then automatically combines results from all segments to achieve comprehensive full-view-field statistical representation, resolving the contradiction between maintaining measurement precision and expanding the total area analyzed.
Solution Approach 2:
The system transitions from analyzing a single two-dimensional view field to analyzing multiple two-dimensional fields that collectively represent a larger three-dimensional sample space. This dimensional expansion allows the system to maintain measurement precision at the individual feature level while achieving full-view-field statistical representation across the entire material sample.
Data Source
AI summary
The present invention discloses, a full-view-field quantitative statistical distribution representation method for microstructures of γ′ phases in a metal material, comprising the following steps: step a: labeling γ′ phases, cloud clutters and γ matrixes by Labelme, and then making standard feature training samples; step b: building a deep learning-based feature recognition and extraction model by means of BDU-Net; step e: collecting γ′ feature maps in the metal material to be detected; step d: automatically recognizing and extracting the γ′ phases; and step e: performing in-situ quantitative statistical distribution representation on the γ phases in the full view field within a large range. The full-view-field quantitative statistical distribution representation method for microstructures of γ′ phases in a metal material provided by the present invention realizes automatic, high-speed and high-quality recognition and extraction of features of γ phases in the metal material