Anomaly Detection for Machine Vision
For anomaly detection you need a Deep Learning HW (VSC523xxx.xxP-000) and the technology package - mappVison 5.28 and all together is already available for mass production. The training of the AI-Networks, the so-called DeepLearning models or DL models, can be created by the mappVision “Technology Packages Add-ons” - mappVision_DeepLearningTool_5.28.
What is Anomaly Detection?
With anomaly detection, the camera learns what “good” looks like, allowing it to quickly recognize any deviations from that and you combined with rules-based algorithms.
To make it easier for our customers to select and apply the correct DL model, we offer a user-friendly tool with a simplified workflow. With predefined parameters, they can quickly set up the desired function and execute it directly on the deep learning smart camera. The advantage of this approach is that you get the benefits of deep learning, all what you need is our smart camera with an AI processor.
How?
Take the existing images (from our SmartCamera) or collect some images, it is not a huge effort to get an idea, 10 to 30 images should be enough and for all this no special HW is needed, you can simulate it offline in mappVision!
Workflow
-
collect good images
- put all images into the folder OK
- put all images with a defect (for validation reason) into the folder NOK
-
add the Anomaly Detection Vision Function (VF) in your Vision Application (VA)
-
create a Deep Learning (DL) config in the Automation Studio (AS)
-
open the DL Tool (button in the read box)
-
Start the Review of you dataset (have all images the right label?)
- all images are auto labeled via the folder structure
-
create a split (pre defined)
-
start the training
- training parameter are pre defined
-
start the evaluation, to make sure that the training was successful before you export the calibrated DL model for the AI processor
-
start the offline UI in the AS (image above, button in the yellow box) and parametrize the VF.
VSC523xxx.xxxP-000 Hardware is needed
To execute a DL (DeepLearning) Model a special hardware with an AI processor is required. This available HW (VSC523xxx.xxP-000) is all what you need to use any DeepLearning feature like anomaly detection and in the future classification, object detection and segmentation.
Use Cases
You can use it for various tasks like inline inspection
- detecting deformations e.g. during the punching process of vial caps
- find unfitted brushes
- to locate undefined faults in the cap seal
- detecting faults during label shrinking in a hot foil process
- recognize deformations, thread defects e.g. in screw production
end-of-line quality inspection
- vial inspection
- surface inspection to find defects
- find unexpected and undefined parts to sort them out
and these are just a couple of examples!
QnA
- How long does it take to train the DL network?
- That depends on the GPU performance, a few numbers: NVIDA A500 () ~2h and with the NVIDA RTX4070 ~10min.
- How many images i need for the Training?
- Depends on the complexity of failure cases, but a rough idea a number between 30 and 100 images
- What should I do if the results don’t make sense?
- please check the label classes, in most of the cases there is wrong labeled image
- consider the local and global anomalies separately (default is combined) often anomalies in the background are stronger than the actual error
- How important is the image sensor resolution?
- not really the input resolution from the DL Model is 256x256 Pixel and you don’t have to be nervous, that’s typically more than enough!
- BUT if you need a larger FOV and only want to crop a small area for anomaly detection … then a 5M pixel sensor could be a needed solution.