Study on denoising multilayer simulated impact signals based on DnCNN

Abstract

During the penetration of projectiles through multi-layer targets, the signals are significantly affected by complex environmental noise and measurement interference, which directly impact the acquisition of impact signals and subsequent mechanical analysis. To address the limitations of traditional low-pass filtering, wavelet transform, and empirical mode decomposition (EMD/EEMD) methods, such as limited denoising accuracy and strong parameter dependency, this study proposes a multi-layer impact signal denoising framework based on short-time Fourier transform (STFT) and DnCNN. First, high-fidelity clean signals are obtained using a self-developed multi-impact equivalent simulation test rig, and noisy datasets are constructed by superimposing Gaussian and uniform noise. Then, the signals are transformed into complex matrices via STFT and input into the DnCNN-S-15 network for training and prediction, followed by inverse STFT to reconstruct the time-domain signals. Experimental results demonstrate that the denoised signals from DnCNN-S-15 achieve an average SNR of 17.31 dB and a Pearson correlation coefficient of 99.24%. Compared with low-pass filtering, wavelet transform, and EEMD, the proposed method outperforms them in SNR, correlation coefficient, impact peak error, and impact pulse width error, with SNR improvements of approximately 7.55, 9.56, and 9.20 dB, and correlation coefficient improvements of approximately 7.26%, 12.38%, and 11.08%, respectively. This parameter-free method can achieve high-precision denoising of multi-layer impact signals, providing an effective technical basis for precise projectile control and dynamic response analysis.