Interactive machine learning in industrial control

Abstract

In industrial control, digital radiography is used to detect defects in parts. However visual inspection by a trained expert can take up to a few minutes to identify anomalies on one image, which creates a bottleneck. That is why projects like ADRIC (automatic defect recognition in industrial control) try to use machine learning to help the operators speed up the process.

However the manual defects identification is not fully reliable, which can confuse the machine learning algorithm tasked to discriminate those defects. This is why we want to design a method that will use data efficiently, in order to achieve the best results with less data than before.

In this document, we show that, using suitable data selection in the framework of active learning in conjunction with Bayesian deep learning, we achieved better results than random, for a patch based method, while we did not achieve convincing results for a more realistic image based method. More work need to be done in this direction to get similar results on full images.

We also developed post-processing techniques on heat-maps in order to enhance the performance metric, as well as the accuracy. We also noticed that the heat-maps production was not fast enough, only marginally surpassing the human experts time-wise. We thus designed a heat-maps production algorithm that offers speed ups, but with lower accuracy.

Finally, we found out that using weights trained on a synthetic dataset as a starting point, we can achieve better performance, without almost any additional cost, on a real dataset.