Enhanced concrete crack detection using nonlinear ultrasonic signal denoising and feature enhancement with CEEMDAN, VMD and GRU

Abstract

The nonlinear ultrasonic detection signals of concrete are highly sensitive to micro-damage changes and are thus susceptible to noise interference. Consequently, denoising these detection signals is crucial. This study proposes a comprehensive automated denoising method for nonlinear ultrasonic detection signals of concrete damage. The method combines the complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN), the variational mode decomposition (VMD), and the gated recurrent unit (GRU). The original detection signal is first decomposed into intrinsic mode functions (IMFs) obtained from CEEMDAN to reduce the error of ensemble averaging. These IMFs are then categorized into three groups of integrated IMFs (IN-IMFs) based on the sample entropy. And the VMD results of each IN-IMF are input into the GRU model for training and denoising. The denoising results of each IN-IMF type are combined by superimposing or the GRU model to obtain the final comprehensive denoising result. Analysis of the denoising metrics indicates that the proposed method achieves slightly improved performance compared to the VMD-gray wolf optimizer combined with the EMD algorithm, CEEMDAN combined with fuzzy rough sets, and Dominant noise-aided EMD. The proposed method yields a signal-to-noise ratio (SNR) of 11-12, a root mean square error (RMSE) of 0.12-0.13, and a normalized cross-correlation (NCC) of 0.90-0.95. Unlike existing methods, the proposed denoising method enhances the nonlinear characteristics of the signal, increasing the range of nonlinear indicator changes (5.5 times or 27.9 times that before denoising). It also reduces the volatility of nonlinear coefficients with damage changes. In engineering applications, this method effectively eliminates interference information from nonlinear ultrasonic detection signals, enhancing the features of the target information.