Prediction of Overlying Rock Deformation based on LSTM in Optical Fiber Sensor Monitoring

The deformation of overlying rock layer is one of the key issues related to occurrence of rock burst and gas explosion, and this will aggravate the threat to electric equipment and miner's life during the mining process underground coal mine. Therefore, the prediction of deformation of overlying rock layer is of great significance to safety of coal mine. In this paper, a model is developed by combining the long short-term memory neural network (LSTM) with Synthetic Minority Oversampling Technique (SMOTE) and First difference transformation (FDT) for forecasting the frequency shift value of sensing fiber on distributed optical fiber sensor monitoring. Accurately predicting the frequency shift value of fiber sensor is used to infer the deformation state of the rock. Then, verification experiments were performed on dataset generated by 6 sensors on Fyn and Fv12 optical fiber. The average absolute percentage error (MAE), maximum absolute percentage error (MaxAPE) and root mean square error (RMSE) are the evaluation indicators of model performance. The experimental results show that the average of MaxAPE, MAPE and RMSE are 16.68%, 4.22%, and 8.70 on 6 points, which are lower than RNN and ES method, respectively. The results demonstrate that prediction of SMOTE-FDT-LSTM is accurate and robust, and the model can improve prediction the deformation of overlying rock layer.