Continuous Anomaly Detection Model Training in Latent Space
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional image inspection systems for manufacturing applications face challenges such as high costs for labeling image data, lack of updatability to detect new types of defects, and lack of scalability, making them unsustainable in practice.
Innovation Solution
A computer-implemented continuous training method for systems identifying anomalies in images, which involves training the system using sets of images corresponding to different anomaly types, updating the model in a latent space, and calculating log-likelihood losses to adapt to new defects without downtime.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional imaging solutions use huge amounts of labeled data for training, then detection accuracy may be improved, but costs increase significantly and scalability decreases
Solution Approach 1:
The system performs preliminary action by pre-training on synthetic defect data generated before deployment. This allows the model to learn defect patterns in advance without requiring expensive labeled real defect data during operational training, thereby improving detection accuracy while avoiding the high costs associated with collecting and labeling large datasets.
Solution Approach 2:
The patent employs copying by generating synthetic defect images that replicate real defect patterns. Instead of using expensive real labeled defect data, the system creates copies of defect patterns through synthesis, maintaining detection accuracy while dramatically reducing data labeling costs and improving scalability.
2Adaptability or versatility
If conventional models are replaced with updated models to detect new defect types, then adaptability improves, but system downtime increases and scalability decreases
Solution Approach 1:
The system implements dynamics by enabling continuous adaptation of the detection model during operational phases. The model can dynamically update its parameters and learn new defect patterns on-the-fly without requiring complete model replacement, thereby maintaining adaptability to new defect types while avoiding system downtime associated with model retraining.
Solution Approach 2:
The patent applies continuity of useful action by allowing the anomaly detection system to continue operating while adapting to new defect types. The model updates its knowledge base continuously during normal operation rather than requiring shutdowns for retraining, ensuring both adaptability and uninterrupted service.
3Measurement precision
If conventional imaging solutions are designed for high detection accuracy, then measurement precision improves, but device complexity and cost increase making them unsustainable
Solution Approach 1:
The system applies taking out by extracting and isolating the critical training phase from the operational phase. The complex model training and synthetic data generation are performed separately before deployment, while the operational system remains relatively simple and lightweight, achieving high detection accuracy without maintaining excessive complexity in the deployed system.
Data Source
AI summary
A system identifying anomalies in an image of an object is first trained using first sets of images corresponding to first anomaly types for the object. A model of the object is formed in a latent space. A label for each anomalous image is used to calculate vectors containing means and standard deviations for each first anomaly types. The means and standard deviations are used to calculate a log-likelihood loss for each first anomaly type. The system is retrained using second sets of images corresponding to second anomaly types for the object. The vectors are supplemented using labels for each second anomaly types. A statistically sufficient sample of information in the means and standard deviations vectors is supplied to the latent space. A log-likelihood loss for each of the first and second anomaly types is calculated based on their respective mean and standard deviation.