Ultrasonic signal denoising based on autoencoder.

Area of Science:

• Signal Processing
• Machine Learning
• Ultrasonic Testing

Background:

• Traditional signal denoising methods require manual parameter tuning, which is susceptible to human error.
• Existing algorithms like Singular Value Decomposition (SVD), Principal Component Analysis (PCA), and wavelet transforms have limitations in noise suppression and signal feature preservation.

Purpose of the Study:

• To develop a signal adaptive denoising method for ultrasonic signals using a denoising autoencoder.
• To evaluate the effectiveness of the proposed method against established denoising techniques.

Main Methods:

• A novel denoising autoencoder was developed for adaptive signal denoising.
• The proposed method was compared with SVD, PCA, and wavelet algorithms using sample signals.
• Performance was evaluated using signal-to-noise ratio (SNR), root mean square error (RMSE), and autocorrelation coefficient.

Main Results:

• The proposed method effectively suppresses noise across various noise intensities.
• It outperforms PCA in overall denoising and surpasses wavelet and SVD algorithms at lower noise levels.
• The method demonstrates superior performance in retaining signal saltation information compared to wavelet algorithms.

Conclusions:

• The signal adaptive denoising autoencoder offers a robust solution for ultrasonic signal processing.
• This method provides improved noise suppression and signal feature preservation over conventional techniques.
• Validated effectiveness in both time and frequency domains for real-world ultrasonic signal applications.