Engineers from Moscow Polytechnic University have created a neural network that automatically detects cracks and other critical defects in cast metal parts directly during production. The system promises to replace exhausting manual inspection, improve quality control accuracy, and minimize the risk of accidents in aviation, automotive manufacturing, and the energy sector.

At modern factories, cast parts (housings, brackets, turbine blades, etc.) typically undergo visual quality inspection by human inspectors. A person examines the hot or cooled surface and evaluates cracks, pores, cavities, and oxidation marks. The method is simple, but extremely unreliable the human eye becomes tired after 1–2 hours of intensive work, lighting conditions change and glare interferes, on complex textures or oxidized surfaces, a thin hot crack can easily go unnoticed.

A missed crack in a critical component can lead to consequences: structural failure in flight, a road accident, or a shutdown of a power unit. Meanwhile, production volumes are so large that checking every part with microscopic precision manually is physically impossible.

Classical computer vision algorithms also struggle. They work well in ideal conditions (clean surfaces, uniform lighting) but fail when dealing with real-world issues such as material irregularities, oxidation, blurred defect boundaries, or cracks blending into the natural relief of the casting. The result is either false rejections (production losses) or missed defects (safety risks).

The developers combined two technologies: a convolutional neural network (CNN) analyzes images of the part, identifies suspicious areas, and classifies defects, and fuzzy logic processes uncertainty — such as oxidation levels, surface type, material properties, and image context. Instead of a strict “defect/no defect” decision, the system produces a weighted assessment of danger.

The AI combined a convolutional neural network that analyzes images of the part with fuzzy logic that can handle uncertainty. The system doesn’t just detect a crack — it evaluates its level of danger while considering the context: the nature of the surface, the degree of oxidation, and the type of material. This represents a fundamentally different level of diagnostics compared to current approaches.

The neural network is trained using a large labeled dataset. Researchers collect thousands of photographs of real defective castings, manually marking every defect (cracks, pores, cavities) and specifying their shape, size, and location. The more diverse the dataset — including different alloys, lighting conditions, and shooting angles — the better the model can generalize to new parts.

Fuzzy logic becomes especially useful in borderline cases, where traditional algorithms would produce an incorrect binary decision.

Source: News.am

Image: News.am