In the explosion experiment, the explosion vibration signal is the important information to evaluate the explosion equivalent or to locate the explosion point. However, the environment at the scene of the explosion is very complicated. The explosion vibration signal collected by the Distributed fiber Acoustic Sensor (DAS) system not only contains optical noise, but also contains a lot of non-stationary and non-Gaussian environmental background noise. In order to solve this issue, feed-forward denoising convolutional neural network (DnCNN) are used to denoise the signals during the noise suppression of explosion vibration signals. The initial application of this network was to purge images of additive Gaussian white noise. In order to make DnCNN adapt to the de-noising of explosion vibration data, we have performed numerous optimization tasks. Firstly, the input data is processed by an reversible downsampling factor in this paper, expanding the network’s perceptual field while also making training easier. Secondly, rather than using Gaussian white noise, the training set of DnCNN is rebuilt using background noise collected in the actual environment. Finally, the DnCNN parameters that impact network performance are modified and improved. From the experimental results in this paper, the DnCNN can effectively suppress the noise in the explosion vibration signal and preserve the effective signal.
|