Quality inspection applications in industry are required to move towards a zero-defect manufacturing scenario, with non-destructive inspection and traceability of 100% of produced parts. Developing robust fault detection and classification models from the start-up of the lines is challenging due to the difficulty in getting enough representative samples of the faulty patterns and the need to manually label them.
Automatic Quality inspection of industrial parts is crucial. However, the proposed solutions to this day are not as generalizable to different scenarios as one might wish as they heavily rely on case and situation specific features in order to design the inspection system.
This building block provides access to a Deep Learning based methodology for the development of robust and flexible industrial inspection systems.
This methodology has the following steps:
Stage 1 – Anomaly detection: In the first stage of the set-up of a new production line, an anomaly detection approach would be used. By training a network using only non-defective parts, the products that are out of normality, i.e. anomalous or defective, would be identified. In addition to classifying them, the defective areas within the product would also be marked for inspection, i.e. defect segmentation. In this way, it would not be necessary to wait for having enough defective samples for training, and it would be possible to start detecting defects from the beginning.
Stage 2 – Supervised training: During the lifetime of the production line, it will generate defective products which will be automatically identified and annotated at pixel level by the model from the first stage. Once a handful of these defective samples are accumulated, they will be used to train a second model in a supervised manner and using methods that require little data for training. This second model will be trained specifically to search for concrete defects reaching higher accuracy rates than the previous model in detection.
Stage 3 – Model adaptation: In a production system, there may be features that are not common and rarely appear during production. It could also happen, that similar inspection scenarios could take advantage of existing models. For this case, the methodology proposes two DL based techniques called Few-shot learning and Transfer Learning, which take advantage of the previously trained models to obtain new models that will be adapted to work on the new line. Using already trained models as the starting point alleviates the need for defective data which accelerates the deployment of the new line.
For more information, please contact:
- Luka Eciolaza – email@example.com