Ultrasonic signal noise reduction based on convolutional autoencoders for NDT applications

Abstract

One of the most challenging problems of ultrasonic non-destructive testing is the signal distortion caused by the presence of noise, yielding the sound wave corruption and thus degrading the ultrasonic imaging technology performance due to Time of flight methods’ loss of precision. Deep learning algorithms have proven their effectiveness in reducing noise on several types of signals in different domains. In this paper, we propose a one-dimensional convolutional autoencoder for ultrasonic signal denoising. The efficiency of the proposed architecture is compared to the wavelet decomposition method, collating the peak signal-to-noise ratio values on the denoised signals. Our method proved its potential for NDT applications in recovering temporal information even on very noisy signals, and improving the PSNR by about 30 dB.