NEURAL NETWORK METHODS FOR ULTRASOUND SIGNAL EXTRACTION FROM NOISE

Abstract

This paper addresses neural network–based methods for extracting ultrasonic signals from noise in applications of time-of-flight (TOF) distance measurement, non-destructive testing (NDT), and ultrasonic imaging. The study is motivated by the limited robustness of conventional filtering and correlation techniques under non-stationary noise, reverberation, and multipath propagation, where even small distortions of the echo waveform may lead to significant TOF estimation errors. Mathematical models of ultrasonic pulse observations are considered, including additive, correlated, impulsive, and reverberation noise components, with particular emphasis on preserving the temporal structure of reflections. Quality criteria relevant to practical ultrasonic measurements are formulated, such as signal-to-noise ratio improvement, mean absolute and squared errors, and, most importantly, the time-of-arrival estimation error directly affecting distance accuracy.

The paper systematizes modern neural denoising architectures applicable to ultrasonic signals. These include one-dimensional convolutional neural networks with residual learning for real-time processing, denoising autoencoders for structured noise suppression, and U-Net–based models operating on time–frequency representations obtained via short-time Fourier transform. Attention-based models and transformers are discussed in the context of long reverberation tails and complex interference patterns. Special attention is given to training strategies in scenarios where clean reference signals are unavailable, including Noise2Noise and blind-spot self-supervised approaches, which enable learning directly from field measurements.

It is shown that optimizing neural networks solely with energy-based losses may lead to excessive smoothing and temporal bias; therefore, loss functions should be aligned with the physical measurement objective, namely minimizing TOF estimation error. Practical recommendations for integrating neural denoising modules into ultrasonic signal processing chains are provided. The results demonstrate that properly designed neural network denoisers can significantly enhance noise immunity while preserving echo timing, enabling more accurate and reliable ultrasonic measurements in challenging environments.