- Anomaly detection
Seamless anomaly detection technology from unsupervised to supervised learning
Further increase the accuracy of anomaly detection models using anomaly data obtained during operations.
- Using anomaly data obtained during operations for learning enables the seamless transition from an initial unsupervised model based on learning from normal data only to a highly accurate supervised model.
- Because anomaly data does not need to be gathered in advance, simple anomaly detection can be introduced first, and accuracy can be increased through operations.
- After the transition, the deep neural network structures and anomaly score distribution do not change dramatically, so the burden of operations resulting from the transition can be minimized.
Applications
- Product inspections at manufacturing sites
- Monitoring of anomalies in equipment and facilities
Benchmarks, strengths, and track record
- Consistency (Histogram Intersection) between detection performance (AUROC) evaluated using a variety of public image dataset and before model update (AE)
| Conventional method | Proposed method | |||
|---|---|---|---|---|
| Database name | Evaluation index | Autoencoder (AE) | Autoencording Binary Classier (ABC) |
AEAL |
| MNIST | Detection accuracy | 0.94 | 0.99 | 1.00 |
| Consistency with before model update (AE) | - | 0.27 | 0.84 | |
| F-MNIST | Detection accuracy | 0.93 | 0.99 | 0.99 |
| Consistency with before model update (AE) | - | 0.13 | 0.86 | |
| CIFAR-10 | Detection accuracy | 0.66 | 0.93 | 0.88 |
| Consistency with before model update (AE) | - | 0.03 | 0.86 | |
Inquiries
Contact the Toshiba Corporate Research & Development Center
Please include the title “Toshiba AI Technology Catalog: Seamless anomaly detection technology from unsupervised to supervised learning” or the URL in the inquiry text.
Please note that because this technology is currently the subject of R&D activities, immediate responses to inquiries may not be possible.
