Distributed fiber optic sensing data noise suppression based on U-net

Distributed fiber optic acoustic sensing (DAS) is a developing vibration observation technology recently. DAS has attracted widespread attention in the fields of structural monitoring, leakage detection, transportation, oil and gas exploration, and natural seismicity. Compared with conventional geophones, on the one hand, DAS has the advantages of low cost, high density, high sensitivity, efficient construction, and long-term monitoring. On the other hand, the signal-to-noise ratio of the DAS data is relatively low, so it is of great importance to suppress the DAS noise.

Most of traditional noise suppression methods rely on a prior information, which affects the final denoising effect. It also reduces the processing efficiency especially the amount of data is large. In recent years, the application of artificial intelligence methods in seismic data processing and interpretation has widespread gradually. Deep learning methods can dig deeper features of the data through multi-layer structure, so as to suppress the noise. To build the training dataset, we use fractional order Fourier Transform (FrFT) to construct a median filter to suppress the high (low) frequency noise. The soft-threshold curvelet transform is used to suppress random noise. The amplitude equalization f-k filtering is used to suppress the linear noise. In this way the denoised seismic record is obtained using three improved mathematical transform methods. In our U-net, the patching technique is used to generate many small-scale patches from the input data, together with their labels. The denoised data are reconstructed from the patches using the unpatching technique. This is helpful in reducing the computational cost and improve the ability to extract essential features from large-scale seismic data. And help to keep the same matrix dimension of input and output of the U-net. The Mish activation function is used instead of the traditional activation function (Sigmoid, ReLU or Tanh) in the U-net. The upper unbounded property of the Mish avoids the sharp drop of training speed. The lower bounded produces a strong regularization effect and can smooth the training model to get a better generalization ability. The non-monotonic property not only helps to keep little negative values that contribute to stabilizing the gradient of the network, but also avoiding the risk of gradient vanishing like the ReLU activation function.

After the calculation based on a real seismic data, three common noises mentioned above are suppressed by the U-net. The weakly hidden effective signals can be recovered from raw DAS data. Furthermore, our method does not involve the multiple waves suppressing. However, the curvelet transform can also achieve suppression of multiple waves. It can help form the training set for U-net. This is an aspect that needs to be further improved in the future.