What to expect

What kind of images make a good training data set? 

The training data should be as close as possible to the data on which predictions are to be made. Consider providing multiple angles, resolutions, and backgrounds for your training images if that is the case with your images during prediction. For example, if your use case requires low resolution and blurry images, then your training data should also contain low resolution and blurry images. 

The skill cannot assign labels to the images that even the humans cannot label when looking at it for a couple of seconds.

Do not adjust the camera (such as Pan, Tilt, and Zoom) after training, as this invalidates any training performed using the training data set.

What are the minimum conditions on the dataset?

Dimensions – You can modify the dimensions of the image dataset, or use the default settings. If you specify dimensions for the image dataset, then the new dimensions are used during training and prediction. The training dataset with the new dimensions are not stored.

When no dimensions are specified, then by default the image dimensions are readjusted to 400 pixels x 400 pixels. If the input image’s size exceeds 400 pixels x 400 pixels, it is downscaled. Images smaller than these are not upscaled, but resized to 400 pixels x 400 pixels. It is better to provide images in this dimension, otherwise for much smaller dimensions the image quality is lost during training and prediction process. 

Number of images – For Discrete State Detection, we recommend you provide 500 or more training images per class. For Anomaly Detection, we recommend you provide 500 or more good images.

How should the dataset look like? 

Complexity and diversity – Adding more images provides better predictions. More unique cases require more images per class. Just like humans, it is easier for model to learn difference between a circle and a line. It requires more time to learn 5–6 shapes and the training data requires different orientations, sizes, placement, and so on. 

What are the common issues related to training images?

Imbalanced Data – There might be cases where the number of samples for each class is not the same. Minor imbalance is not very problematic, but major imbalance can cause the model to be biased and the prediction results are very useful. We recommend you provide equal images that represents all classes. 

Data Leakage – Data leakage means that the model can use information while training that it is not supposed to. While trying to collect images for different classes, you might have one angle for positive images and different for negative images, and the model learns this information instead of learning the actual features. These angles might not hold during prediction.  

Overfitting and underfitting – With deep learning, there is a common phenomenon known as overfitting where the model memorizes the training data set (because of fewer samples in dataset) and gives really good results but cannot perform well on the test dataset. One way to detect this is when the training accuracy reaches ~100% while the validation accuracy remains lower consistently (~10% difference). In such cases, it is required that you add more diverse samples to your classes.

What are the performance parameters?

The performance of the skill depends on the following:

• The computer hardware (CPU, GPU, RAM)

• The type of skill selected

• The number of images

• The number of skills running at a time

• Consider the network bandwidth when posting data to AVEVA Insight.

How do I isolate the part or area in the plant?

There are various considerations to get the appropriate image that is most useful for the skill to identify any anomaly. Each camera comes with some pre-built settings that can be configured to change the raw image captured. Some items to consider are the physical placement of the camera, the aperture, focus, panning, and motion. Lighting is critical to capturing a good image, especially when the subject of the image is in motion. Also consider the alignment of the object in the image and overall position within the frame.

Test different settings and compositions to determine the best input to the skill.

What can the system not do?

AVEVA Vision AI Assistant is not suitable for scenarios where isolating or photographing an asset or Work In Progress (WIP) item is difficult, due to its position, access, or security concerns.