Transfer learning from physical scale information: A signal identification method for mine system security monitoring

Earthquake signals and microseismic signals are elastic waves excited by the sudden release of energy inside the geological medium, but they differ in the scale of their respective seismic sources. This makes it possible to learn from physical information at different scales. We proposed a one-dimensional convolutional neural network under transfer learning framework to learn how to identify microseismic signals using the discrimination of earthquake magnitude. We first trained the convolutional neural network using numerous labeled seismic signal data and then fine-tuned the partial weights using a small amount of microseismic data. Experimental results demonstrate the classification results on the microseismic database are as high as 100 %. When the number of samples is reduced to 100 signals, the model can still maintain a minimum accuracy of 96 % on the same dataset. When the signal-to-noise ratio gradually increases to 0dB, the minimum accuracy of the model reaches 95.35 %. Compared with traditional machine learning (SVM, Logistic Regression, Naive Bayes), the accuracy of the model increased by 6.19 %, 16.00 % and 40.30 % respectively. The research result above show the model has excellent classification accuracy and high robustness, providing better pre-technical support for the safety and stability of mine systems.