Improving on human analysis with AI

NDT.ai is an image detection and classification project which is developed by DNV GL’s Artificial Intelligence research centre in Shanghai. The project aims to develop novel methodology to automatically measure indications in radiography testing images digitally, thereby reducing reliance on faulty human judgements where welding is concerned in shipyards.

In shipyards, radiography testing is widely and essentially adopted to detect weld imperfections -porosity, slag, and cracks. These X-ray images need to be reviewed by Non Destructive Testing (NDT) experts such that unqualified welds are discarded, requiring considerable manpower.

Furthermore, processing radiography testing with physical x-ray films is difficult, requiring special storage and darkroom processing. In order to detect tiny defects, analysis requires LED lamps, introducing the possibility of human error into the equation. Communication, meanwhile, is difficult, as there is normally only one copy of the x-ray film.

Automatic weld imperfection tool

The DNV GL artificial intelligence research center is using deep learning networks and has developed an automatic weld imperfection control tool.

The first step is image quality control, common to x-ray images, which requires optical density and IQI tweaks. This includes an optical density method to evaluate the image grayscale value, and a deep learning algorithm to detect IQI area and the values in the x-ray images.

This is followed by indication prediction, which uses the welding steam detection method to do the segmentation both on horizontal and vertical. Then, the indication property measurement phase is used for edge detection, contour fitting of the size of individual region(s); Then, media axis transform determines the size of connectivity of 2D shape/region.

The benefits

In order to improve weld quality control by overcoming the hit-and-miss nature of human analysis, DNV GL will accelerate the NDT review process in a bidirectional manner, namely, reducing manual efforts - detection of weld indication is replaced by AI algorithms – demoting human intervention to acceptation review; and speeding up the assessment process. Unacceptable parts are reported immediately, accelerating of the shipbuilding phase and enhancing communication.