Abnormality Detection Model Training with Selected Feature Dimensions
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing abnormality detection models trained with only normal data using unsupervised learning face accuracy issues due to the lack of sufficient abnormality data, making efficient training challenging.
Innovation Solution
An information processing apparatus that extracts features of n dimensions from training images, selects k dimensions based on feature differences or random selection, and generates an abnormality detection model for accurate inference, allowing training with or without abnormal images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If unsupervised learning is used to train abnormality detection models with only normal data, then the cost for generating training data is low and introduction to the site is easy, but the accuracy of the abnormality detection models is low
Solution Approach 1:
The system performs preliminary action by collecting and storing abnormal images in advance during the model training phase. This allows the model to be trained with both normal and abnormal data before deployment, improving detection accuracy while maintaining the ease of unsupervised learning approach. The abnormal images are gathered proactively rather than waiting for them to occur naturally during operation.
Solution Approach 2:
The system implements feedback by using detected abnormal images as new training data for subsequent model iterations. When the model detects an abnormality, that image is fed back into the training dataset, allowing the model to continuously improve its accuracy by learning from real-world abnormal cases it encounters during operation.
2Measurement precision
If supervised learning is used to train abnormality detection models with abnormal data, then the detection accuracy is improved, but it is not efficient to execute supervised learning after collecting sufficient abnormality data
Solution Approach 1:
The system applies dynamics by making the training approach flexible and adaptive. The model can be trained using unsupervised learning with only normal data initially, then dynamically transition to supervised learning when abnormal data becomes available. This dynamic adjustment allows the system to optimize between ease of deployment and detection accuracy based on the availability of training data.
Solution Approach 2:
The system changes the training parameter composition by starting with only normal data for unsupervised learning, then gradually incorporating abnormal data as it becomes available. This parameter change in the training dataset composition allows the system to improve detection accuracy without requiring all abnormal data to be collected upfront, thereby maintaining training efficiency.
3Stability of the object's composition
If defects rarely occur in inspection targets at the site, then the inspection process is stable, but it is difficult to collect sufficient abnormality data for training
Solution Approach 1:
The system performs preliminary action by proactively collecting and storing abnormal images during the model training phase, rather than waiting for them to occur naturally during routine inspection. This advance collection ensures sufficient abnormal data is available for training while the inspection process remains stable with rare defects.
Solution Approach 2:
The system uses copying by generating or obtaining abnormal images that can be used as training data. These copied or synthesized abnormal images serve as substitutes for real defect occurrences, allowing the model to be trained on sufficient abnormal data without requiring actual defects to be present during normal inspection operations.
Data Source
AI summary
According to one embodiment, an information processing apparatus includes a processor. The processor is configured to acquire at least one training image including an inspection target from an image database that stores the training image, extract first features of n dimensions of the training image output from a feature extraction model by inputting the training image to the feature extraction model, select k dimensions from the n dimensions, and generate an abnormality detection model used to infer a state of the inspection target by executing training using the first features of the selected k dimensions among the first features of the n-dimensions of the training image.