Synthetic Component Image Generation for AI Diagnostics
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The manual generation and identification of large amounts of image data for training artificial intelligence systems in industrial settings is inefficient and impractical, especially in complex systems where components need to be photographed from multiple perspectives and states, requiring significant human effort and limiting the ability to capture data from critical or dangerous operating states.
Innovation Solution
A computer-implemented method and apparatus that automatically generate identified image data by calculating a visual representation of components using design data and surface property specifications, allowing for the simulation of various states and conditions, and assigning identifiers to these representations, which can be used for diagnostics and analysis.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual identification and capture of image data is performed, then image data of existing real systems can be acquired, but enormous human effort and time are required making it unmanageable for complex technical systems
Solution Approach 1:
The patent creates virtual copies of real components through 3D models and simulates their appearance under various conditions. Instead of manually photographing real components in different states, the system generates synthetic images from digital models, eliminating the need for physical manipulation and manual identification while maintaining identification accuracy.
Solution Approach 2:
The patent replaces the mechanical process of physically capturing images of real components with a computational system that generates images algorithmically. The visual representation calculation unit computes what components would look like in various states without requiring physical access to actual components, substituting manual photographic processes with automated image synthesis.
2Adaptability or versatility
If components are photographed from multiple perspectives and in different operating states to train AI systems, then comprehensive training data can be obtained, but significant human effort is required to demount and reposition components
Solution Approach 1:
The patent dynamically generates images from static 3D models by virtually changing viewing angles, lighting conditions, and component states. Instead of physically repositioning components to capture different perspectives, the system rotates and renders the digital model from any desired angle, providing comprehensive data coverage without physical manipulation.
Solution Approach 2:
The patent changes parameters of the virtual component model (such as surface properties, operating states, environmental conditions) to generate diverse training data. By modifying digital parameters rather than physical component states, the system achieves high adaptability in data coverage while eliminating the ease-of-manufacture issues associated with physical data collection.
3Loss of information
If real components are exposed to capture image data of dangerous or critical operating states, then data from critical states can be obtained, but safety risks and unmanageable effort increase
Solution Approach 1:
The patent uses virtual copies of components to represent dangerous or critical states that would be unsafe to capture with real components. The simulation generates images of components in critical operating conditions without requiring actual exposure to hazardous environments, eliminating safety risks while preserving access to critical state data.
Solution Approach 2:
The patent performs preliminary generation of critical state data through simulation before any real component would need to be exposed to dangerous conditions. By pre-computing what critical states would look like from safe 3D models, the system obtains necessary information without subjecting real components to harmful conditions.
Data Source
AI summary
The present disclosure relates to the automatic generation of characterized image data. For this purpose, a visual representation of a component is computed on the basis of existent structure data for the component, wherein the surface properties of the visual representation of the component may be varied based on predefined characteristic properties. Because, in the computation of such visual representations, both the component itself and the underlying characteristic surface properties are known, this information may be used for characterizing the corresponding parts in the visual representation in order to achieve automatic characterization of the computed visual representation.