How to efficiently automate data processing with drones? The corrosion use case

For the last years, you have surely noticed that we have been collecting a lot of data. Drones can help in this massive data collection. For instance, more than 300 pictures are collected during a structural defect detection inspection of an extra high voltage electrical pylon of 50m high. Thousands of pylons have to be inspected in the next years, we can then expect millions of pictures to be analyzed!

These previous years, artificial intelligence algorithms (neural networks) have been brought to the forefront thanks to a hardware computation power increase and the labelling of huge data bases that made possible the pre-training of these algorithms. Thanks to a suitable training on thousands of labelled data, these algorithms can recognize some complex patterns such as animals, cars, bridges, etc.

This family of algorithms can perfectly be used on some industrial inspection applications. A few golden rules and principles need to be respected.

  • This kind of algorithms are not just a pre-existing “magic algorithmic cauldron” that you feed with data expecting a pertinent result. It is necessary to develop and adapt these neural networks to specific industrial activities. Following the example of a human cognitive process, it is most of the time necessary to mix different neural networks. Each one of them is dedicated to a specific task : recognition, classification, segmentation, number of classes in a pictures, etc. Corrosion detection is a good example of this complexity.

  • No quality labelled data, no result! Data must be collected and labelled following a rigorous and constant process: same distance and orientation acquisition, similar resolution, same exposure, same optical or infrared sensor, etc. This topic does not have to be under estimated in order to reach a good performance level

  • We often hear that millions of images or data are necessary to reach a high automation level for detection/characterization on images. Each industrial case is unique but usually a few thousands of labelled data are enough to kickstart the automation. Sterblue and its industrial partner Omexom have been working together for more than 3 months on a neural network, dedicated to corrosion levels & bended bars detection on extra high voltage electrical towers. Training has been done on only 638 images (equivalent of 3 towers). Results are quite encouraging, with a false negative rate (defects not detected) of 2% and a false positive rate (false alarms) of 25%. Once another dozen of towers will have been through the tool, outputs should reach human one's!

The industrial that wants to implement such a project will have some guaranteed results (detection quality and saved time) over a mid term period. This kind of project enhances the industry and sets it in a virtuous circle enabling new opportunities such as predictive maintenance.