Formed Metal Part Defect Detection Using Physics-Informed Anomaly Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for detecting defects in metal parts manufactured through stamping processes are inefficient and unreliable, particularly due to the high speed of production lines and the complexity of defects, which often require extensive human labeling and are limited to specific defect types, making it difficult to identify all defects consistently.
Innovation Solution
A physics-informed anomaly detection method using deep learning models that integrate physics-based simulations to identify high-risk regions in synthetic images, overlaying these regions on real training images to train a deep learning model for defect detection without requiring defective data, focusing on sensitive areas for efficient defect identification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If sampling-based inspection by human operators is used, then the inspection process is simple and low-cost, but the detection reliability is insufficient due to high production speed and small defect size
Solution Approach 1:
The patent replaces the mechanical human inspection system with an automated optical inspection system using deep learning models. The system captures images of formed parts and uses trained neural networks to automatically detect defects, eliminating the limitations of human visual inspection while maintaining compatibility with high-speed production lines.
Solution Approach 2:
The patent applies preliminary action by pre-training deep learning models using physics-based simulations that predict high-risk regions for defects. These simulations are performed before actual production, allowing the model to be pre-equipped with knowledge of where defects are most likely to occur, improving detection reliability without adding runtime complexity.
2Adaptability or versatility
If extensive human labeling of defective data is performed for model training, then the model can be trained to detect specific defect types, but the process is time-consuming and limits the model to known defect types only
Solution Approach 1:
The patent uses physics-based simulations to generate synthetic images that copy the characteristics of real formed parts without requiring actual defective samples. These simulated images incorporate predicted defect locations and patterns, allowing the model to learn from virtual data that represents both normal and defective states, eliminating the need for extensive manual labeling of real defective parts.
Solution Approach 2:
The system performs self-service by using physics-based simulations to automatically generate training data without human intervention. The simulations autonomously predict high-risk regions and generate corresponding images for model training, eliminating the need for human operators to manually label defective data while maintaining adaptability to various defect types.
3Measurement precision
If the inspection system focuses on the entire part surface, then comprehensive coverage is achieved, but the detection sensitivity for small defects in specific high-risk regions is reduced
Solution Approach 1:
The patent applies local quality by directing the inspection system to focus computational resources and model attention on specific high-risk regions identified by physics-based simulations. Instead of uniformly processing the entire part surface, the system concentrates analysis on areas where defects are most likely to occur based on simulation predictions, improving detection sensitivity for small defects without requiring exhaustive scanning of the entire surface.
Data Source
AI summary
A method for detecting defects in a formed metal part includes locating one or more regions of interest in a synthetic image of a part manufactured by a forming process. The synthetic image is informed based on a physics-based simulation of the forming process. The regions of interest indicate a high risk of having a defect from the forming process. A set of training images including real images of actual manufactured parts are registered with the synthetic image. The regions of interest are overlaid on each training image, to extract patches from the training images that correspond to high-risk regions. An anomaly detection model is trained on the patches extracted from the training images to detect a defect in a formed metal part from an acquired image of the formed metal part, by detecting an anomaly in a patch extracted from the acquired image that corresponds to a high-risk region.