Noise Reduction for High-G Accelerometer Signals Using Deep Learning With Residual Dense Module

High-G accelerometers are essential for measuring high overloads during high-speed motion processes such as impact and penetration; however, noise can contaminate the calibration of the accelerometers, leading to signal distortion and quality degradation. We propose a denoising method that combines a convolutional neural network (CNN) and residual dense modules (RDMs) to remove the noise and ensure the accelerometer's accuracy effectively. First, short-time Fourier transform (STFT) is applied to the output of the high-G accelerometers to convert signals into images in the 2-D time-frequency domain. Then, the images are fed into the CNN model for time-frequency feature extraction. The network is designed with RDMs to realize the reuse and accumulation of features, enabling full use of the hierarchical components of the original signal. Finally, the output is converted back to the time domain using inverse STFT to obtain the noise-reduced acceleration signals. Experimental results indicate that the proposed method significantly enhances signal-to-noise ratio, effectively suppresses noise, and ensures signal fidelity. In contrast to conventional approaches, such as wavelet thresholding and ensemble empirical modal decomposition, our method demonstrates a more stable denoising effect.