Separation of multi-source microseismic signals based on time-frequency dual path Transformer

Surrounding rock stability is the cornerstone for safe production in coal mines. Microseismic monitoring technology evaluates the stability of rock mass by monitoring the vibration signals generated during rock deformation and fracturing, thereby providing vital information for the prevention and control of coal or rock dynamic disasters. In the complex geological environment of underground mines, the rock mass fracturing often accompanies the occurrence of multiple concentrated sources, which leads to the overlapping of multiple sources in microseismic monitoring signals. Such mixing of microseismic signals severely affects the downstream tasks such as microseismic event analysis and accurate source localization, so that it is very necessary to separate the mixed signals. However, classical separation methods are usually based on certain assumptions, and the separated results are prone to deviating from the original characteristics of signals. Although those separation models based on deep learning have the ability to automatically learn the mapping relationship from mixed signals to separated signals, they primarily focus on learning temporal dependencies among the features. Therefore, a multi-source microseismic signal separation model named MS-SepNet based on the time-frequency dual-path Transformer is proposed, including an encoder, separator and decoder. And the core modules of them are also designed, in which the separator takes a time-frequency dual-path feature learning module as its crucial component to capture the temporal dependencies among features and fully learn the frequency domain features. To achieve the separation and avoid inconsistent signal amplitude before and after separation, a joint loss function is designed, which incorporats both frequency domain mean square error and time domain signal-to-noise ratio. In addition, the model adopts the transfer learning framework to deal with distribution discrepancy between training data and test samples. By training the model with an abundant seismic dataset, freezing its common feature extraction layers, and fine-tuning it with microseismic data, a microseismic signal separation model can be obtained. The experimental results show that MS-SepNet can effectively separate microseismic signals mixed in different degrees. Even when the overlap rate reaches 80%, the signal-to-noise ratio and waveform similarity before and after separation still reach 7.36 dB and 91.52%, respectively, significantly outperforming other separation methods including Conv-TasNe, DPRNN, Sepformer.