Vibration Analysis Support Apparatus for Unknown Defect Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing technologies for analyzing vibration data in manufacturing, such as those using machine learning techniques, struggle to provide clear interpretations and detect defect patterns not represented in the initial data, especially with reduced defect rates and varying manufacturing conditions.
Innovation Solution
An analysis support apparatus that includes a feature interpreting section to classify features into abnormality determination features, an abnormality determination range creating section to identify abnormal ranges, a similar model selecting section to find overlapping models, and an abnormality range difference calculating section to extract differences, enabling clear model interpretation and improvement considerations for uncollected data patterns.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If machine learning techniques are used to establish determination rules from previous data, then determination accuracy for known defect patterns is improved, but the ability to detect unknown defect patterns deteriorates
Solution Approach 1:
The patent applies preliminary action by pre-defining multiple candidate determination rules covering various defect patterns before actual inspection. These rules are prepared in advance based on historical data and expert knowledge, allowing the system to handle both known and unknown defect patterns effectively without requiring retraining for each new pattern.
Solution Approach 2:
The patent utilizes parameter changes by adjusting the weighting and selection of different determination rules based on the specific inspection context. The system dynamically changes parameters such as rule priority, threshold values, and feature importance to optimize detection performance for different defect patterns while maintaining high accuracy for known patterns and adaptability for unknown ones.
2Manufacturing precision
If the manufacturing technology is improved to lower defect rates, then product quality is improved, but the availability of defective products for collecting training data deteriorates
Solution Approach 1:
The patent addresses this contradiction by performing preliminary action through pre-collecting and storing determination rules and reference data during periods when defective products are more available. The system prepares comprehensive training datasets and rule sets in advance, creating a knowledge base that can be reused even when actual defective products become scarce due to improved manufacturing quality.
Solution Approach 2:
The patent applies copying by creating virtual representations of defective patterns through synthesized data and simulated defect scenarios. Instead of relying solely on physical defective products, the system generates copies of defect patterns through computational methods, allowing continuous model improvement without requiring actual defective units.
3Adaptability or versatility
If model interpretation is performed to extract features for undetectable defect patterns, then the model improvement capability is improved, but the complexity of the analysis process deteriorates
Solution Approach 1:
The patent introduces an intermediary component that automatically performs feature extraction and model interpretation tasks. This intermediary layer translates complex analysis requirements into automated processing steps, reducing the apparent complexity for users while maintaining strong model improvement capabilities through systematic feature analysis and pattern recognition.
Data Source
AI summary
An analysis support apparatus includes: a feature interpreting section extracting a feature from a target analysis model generated from vibration data and classifying the feature into an abnormality determination feature, or a feature representing a vibration and enabling determining whether abnormality occurs; an abnormality determination range creating section identifying a range being determined to be abnormal of the abnormality determination feature of the target analysis model as an abnormality determination range; a similar model selecting section calculating an overlapping degree indicating how much the abnormality determination ranges of the target analysis model and one or more predetermined reference analysis models overlap and selecting, based on the overlapping degree, a similar analysis model similar to the target analysis model from the reference analysis models; and an abnormality range difference calculating section extracting a difference between the abnormality determination ranges of the target analysis model and the similar analysis model.