Normal-Data Feature Learning for Memory-Bank Anomaly Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Anomaly detection techniques using deep learning classification models face challenges due to the scarcity and exhaustion of anomaly data, particularly in industrial settings where anomaly data is difficult to obtain, and existing solutions like data augmentation and self-supervised learning are inadequate.
Innovation Solution
A machine learning-based anomaly detection method that utilizes normal data to train a feature extraction network and a dimensionality reduction network, generating a memory bank for anomaly detection, allowing accurate detection without requiring labeled anomaly data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep learning classification models are trained using conventional methods requiring large amounts of anomaly data, then detection accuracy may improve, but data acquisition difficulty and training cost increase significantly
Solution Approach 1:
The patent inverts the conventional anomaly detection approach by training the model exclusively on normal data rather than requiring anomaly data. The model learns what normal data looks like and identifies anomalies as deviations from this learned normal pattern, eliminating the need to collect and label scarce anomaly data while maintaining detection accuracy
Solution Approach 2:
The system uses only readily available normal data to train the model, making the training process self-sufficient without requiring external anomaly data resources. The model serves itself by learning from the abundant normal data that naturally exists in the system
2Quantity of substance
If data augmentation techniques are applied to generate synthetic anomaly data, then available training data increases, but data construction cost and time consumption increase
Solution Approach 1:
The patent extracts and eliminates the unnecessary step of generating synthetic anomaly data through data augmentation. By recognizing that anomaly detection can be achieved by learning from normal data alone, the method removes the time-consuming data augmentation process entirely while maintaining effective anomaly detection capability
3Measurement precision
If feature extraction and dimensionality reduction are performed using complex networks, then detection precision improves, but computational complexity and training difficulty increase
Solution Approach 1:
The patent segments the feature extraction process into two distinct neural network components: a feature extraction network that captures relevant characteristics from input data, and a dimensionality reduction network that compresses these features while preserving anomaly-discriminative information. This segmentation allows each network to be optimized independently, reducing overall complexity while maintaining high detection precision
Data Source
AI summary
A computing device for performing anomaly detection according to the present disclosure includes a memory including at least one memory bank that is a logical area, and a processor which: is configured to train a neural network by using at least one function module; generating a reduced feature extraction module; generating the memory bank on the basis of reduced feature data generated from the reduced feature extraction module; and determining whether inspection data is normal by using the generated reduced feature extraction module and the memory bank.