Microseismic Event Identification Method Based on Lightweight U-Net Network

Microseismic event identification is the primary task of microseismic monitoring in hydraulic fracturing. Due to U-Net network will produce high time and space complexity in the training process, the learning efficiency of the model would be affected. Furthermore, the traditional cross entropy loss function is prone to interference from noisy samples when disposing microseismic data with a low signal-to-noise ratio (SNR), leading to a decrease in identification accuracy. In order to solve these problems, this paper proposes a microseismic event identification method using a lightweight U-Net network based on optimized cross entropy loss function. In this method, the label smoothing technology is used to optimize the cross entropy loss function, and the real label is adjusted by the smoothing coefficient, so that the model can better distinguish the effective signal from the noise features, and effectively suppress the oversensitivity of the cross entropy loss function to the noise features. The convolution layer in U-Net network is replaced by depthwise separable convolution (DSC), then the time and space complexity of the model can be reduced effectively through the synergistic effect of depthwise convolution and pointwise convolution, and the network model can be lightweight. After DSC, squeeze-and-excitation attention mechanism module is introduced to make the model focus on the channels with greater weight, which can improve the extraction ability and identification accuracy of effective features of the model. The experimental results show that, compared with the traditional U-Net network, the proposed model reduces the time and the space complexity by 64.6% and 65.2%. Also, the proposed model presents high identification accuracy and anti-noise ability when identifying true data with low SNR, and has good generalization ability.