Training Data Generation with Synthetic Anomalies for Automated Inspection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for detecting anomalies in content, such as video or audio, require manual selection of anomaly candidates by users, which is time-consuming and costly, and can vary based on threshold settings.
Innovation Solution
A method and device for generating a learned model using training data that includes normal and artificially generated anomalous content, allowing for automated anomaly detection by machine learning, reducing the need for manual selection and threshold adjustments.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual inspection by specialized personnel is used to detect anomalies in content images, then detection accuracy can be maintained through careful monitoring, but the method imposes large mental and physical burden on inspectors and introduces individual differences in anomaly detection
Solution Approach 1:
The patent replaces manual mechanical inspection by specialized personnel with an automated machine learning system. The system uses trained models to automatically detect anomalies in content images, eliminating the need for human inspectors to manually examine each image while maintaining or improving detection accuracy through consistent application of learned patterns.
Solution Approach 2:
The system enables self-service anomaly detection where the machine learning model automatically identifies and flags anomalous content without requiring human intervention for each inspection case. The model serves itself by continuously learning from training data and applying its knowledge to new content, reducing operational burden on human inspectors.
2Ease of operation
If mechanical detection using detection software with threshold values is used to detect anomalies, then the method reduces reliance on human inspectors, but it requires setting threshold values for multiple parameters which generates differences in anomaly detection
Solution Approach 1:
The patent transitions from using multiple parameters with manually set threshold values to a machine learning approach where the system learns optimal detection parameters automatically from training data. The model processes input features and outputs anomaly probabilities without requiring manual threshold configuration, ensuring consistent detection across different users and scenarios.
Solution Approach 2:
The system replaces the mechanical threshold-based detection method with an intelligent machine learning system. Instead of relying on pre-set thresholds for multiple parameters, the model learns complex patterns and relationships from training data, providing more consistent and adaptable anomaly detection that does not depend on manual parameter tuning.
3Extent of automation
If machine learning algorithms are used for detecting anomalies in images with user-selected anomaly candidates, then automated detection capability is introduced, but it takes a large amount of time and cost for users to select anomaly candidates
Solution Approach 1:
The patent applies preliminary action by pre-training machine learning models using extensively curated training data that includes both normal and anomalous content. This pre-training phase captures diverse anomaly patterns in advance, so when the model is deployed for actual anomaly detection, it can automatically identify anomalies without requiring users to manually select or annotate candidate anomalies, significantly reducing time and cost.
Solution Approach 2:
The system enables self-service anomaly detection where the pre-trained model automatically identifies and flags anomalies in new content without requiring human users to manually select anomaly candidates. The model independently processes input images and outputs anomaly detections based on its learned knowledge, eliminating the time-consuming manual selection process.
Data Source
AI summary
To generate training data based on normal content and anomalous content generated from the normal content. A training data generation method for generating training data used for generating a learned model for determining whether there is an anomaly in an inspection target, the training data generation method including: receiving normal content regarding the inspection target and anomalous content generated from the normal content; and generating training data based on a set of the normal content and one or more pieces of the anomalous content.