GAN Anomaly Detection Using Synthetic Defect Signals
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In the field of defect detection using artificial intelligence, the imbalance between defect and normal samples makes it challenging to build a high-accuracy defect detection model, especially when only normal samples are available, as it is difficult to collect all possible defect samples for training.
Innovation Solution
An anomaly detection device and method based on a generative adversarial network architecture that uses single-type training data composed of normal signals to train an anomaly detection model, including an encoder, generator, discriminator, and random vector generator, where random latent vectors are used to generate synthesized signals, and the discriminator determines their validity based on a predetermined threshold.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If supervised learning method is used for defect detection, then detection accuracy can be improved with sufficient training data, but it becomes difficult to build a high-accuracy model when defect samples are insufficient or unavailable
Solution Approach 1:
The patent uses a generator network to create synthetic defect samples by transforming normal samples. The generator learns the distribution of normal samples and generates artificial defect instances that serve as training data, effectively copying and transforming existing data to overcome the scarcity of real defect samples.
Solution Approach 2:
The patent performs preliminary training of the generator and discriminator networks using only normal samples before using the generated defect samples for final defect detection modeling. This preliminary action creates a pool of synthetic defect data that can be used for subsequent supervised learning without requiring actual defect samples during the initial training phase.
2Ease of manufacture
If only normal samples are available for training, then data collection is simplified, but the model cannot learn to distinguish between normal and defective patterns
Solution Approach 1:
The patent inverts the traditional approach by first training the model to recognize normal patterns, then using a generator to create artificial defect patterns from those normal patterns. Instead of directly teaching the model what defects look like, it teaches what normal looks like and then generates the opposite (defects) for comparison and learning.
Solution Approach 2:
The generator network acts as an intermediary that transforms normal samples into synthetic defect samples. This intermediary component bridges the gap between available normal data and the needed defect detection capability, enabling the model to learn defect patterns without direct access to real defect examples during training.
3Measurement precision
If synthesized abnormal signals are generated and used for training, then the model can learn from imbalanced data, but the complexity of the training process increases with multiple training stages
Solution Approach 1:
The patent segments the training process into distinct stages: first training the discriminator and generator with normal samples, then generating synthetic defect samples, and finally training the defect detection model with both real and synthetic data. This segmentation makes the complex training process more manageable and systematic.
Solution Approach 2:
The patent maintains continuous useful action by having the generator continuously produce synthetic defect samples that feed into the training process. The generator-discriminator system runs continuously to provide the necessary training data, ensuring that the useful action of data generation and model learning proceeds without interruption.
Data Source
AI summary
An anomaly detection device based on a generative adversarial network architecture, which uses the single-type training data composed of multiple normal signals to train an anomaly detection model. The anomaly detection device includes an encoder, a generator, a discriminator, and a random vector generator. In the training phase of anomaly detection model, the random latent vectors generated by the random vector generator are sequentially input to a generator to generate the synthesized signals with the same dimension as the normal signals. The synthesized signals are sequentially input into a discriminator to output the corresponding discriminant values. When the corresponding discriminant values are under the predetermined threshold, the corresponding synthesized signals are selected as the anomalous class training data, and the real normal signals are selected as the normal class training data.