A Deep Learning Approach for Denoising Air-Coupled Ultrasonic Responds Data

Ensuring material quality is a central objective in production and manufacturing. Non-contact nondestructive testing methods without the use of coupling media are of particular interest with regard to mechanical or biochemical properties of the material. For this purpose, air-coupled ultrasonic is a useful method for quality control. The challenge is the poor signal-to-noise ratio, which makes it difficult to apply the classical approaches. This makes it impossible to distinguish between defect structures and noise. We are developing a method for denoising air-coupled ultrasonic data by applying deep neural networks by using a geometry-analytical component that detects defect structures. During the evaluation we show that we are able to obtain the data almost free of noise, so that incorrectly classified noisy pixels are mainly located at the edges of the defect structures, which cannot be clearly delimited. It is shown that the quality of the data is significantly improved for detection processes.