How can I automate the processing of my industrial inspection data?

It will not have escaped you that in recent yearswe have been collecting data, a lot of data. Drones are one of the drivers of this massive collection. More than 300 images, for example, are collected from a fifty metre very high voltage tower in order to detect and characterise the various potential structural defects. Thousands of pylons will need to be inspected in the coming years, so there will be hundreds of thousands of images to be analysed!

In the last few years,artificial intelligence algorithms (neural networks) have been brought to the forefront, thanks in particular to the increase in the power of computers and the creation of huge labelled databases that have allowed the neural networks to be pre-trained. By training on thousands of qualified images, these algorithms are able recognise complex patterns such as animals, cars, bridges, etc.

This family of algorithms is fully deployable in the case of industrial inspections. A few principles and rules should be considered and understood.

This type of algorithm is not just a "pre-existing algorithmic mill" that is ingesting thousands of pieces of data to produce a relevant result. IThese neural networks must very often be developed to adapt them for specific industrial fields. Like a human cognitive process, it is also often necessary to nest multiple neural networks, each dedicated to a specific task: recognition, classification, segmentation of an image. The detection of corrosion levels on an infrastructure is a good example of this complexity.

Without qualified data, no results!The data must be acquired and labelled according to a rigorous and constant process: taking images at a distance and in a similar orientation for the same type of item, the same sufficiently high resolution, the same optical or infrared sensor, etc. This aspect should not be underestimated in terms of guaranteeing the performance of the solutions.

It is often said that it takes millions of images or pieces of data to achieve a good level of automation of detection/characterisation on images. Each industrial case is unique but a few thousand qualified images are often sufficient to initiate automation in the detection process. Consider the inspection of wind turbine blades (for impacts, abrasion, dirt, cracks, etc.) where more than 200 images are acquired by inspection; detection capabilities/characterisations similar to those made by a human are achievable after about forty wind turbines have been inspected.

The industrialist wishing to implement such an approach will have guaranteed results (time gained and quality of detection) within the framework of a medium-term industrial approach. This approach is a virtuous circle that will also open up access to new possibilities such as predictive maintenance.

Sterblue has now developed a dozen algorithms dedicated to the automatic detection/characterisation of complex shapes/defects for the industry.