Crack Parameter Interpretation Research Based on a Forward-Reverse Interactive Long Short-Term Memory Model Using Distributed Optical Fiber Sensing Strain Data

Summary Distributed optical fiber sensing (DOFS) technology, with its high spatial resolution, long-distance monitoring capability, and resistance to electromagnetic interference, demonstrates significant application potential in hydraulic fracturing crack monitoring. However, due to the complexity of signals during crack propagation and their susceptibility to noise interference, accurately inverting crack geometric parameters remains a technical challenge. To address this issue, we propose an innovative forward-reverse interactive (FRI) long short-term memory (LSTM) (FRI-LSTM) model based on deep learning techniques. The model utilizes strain data from a forward model of crack propagation to invert crack parameters, including fracture length, inlet width, and average width. First, the study compares the performance of models such as recurrent neural network (RNN), gated recurrent unit (GRU), LSTM, and Transformer in inverting distributed optical fiber strain signals. The results show that traditional LSTM outperforms other models in capturing dynamic features and modeling nonlinear relationships. Building on this, the proposed FRI-LSTM model leverages forward and reverse feature interaction modeling and a joint optimization mechanism to significantly enhance the global modeling capability and prediction accuracy of the LSTM model. Compared with traditional LSTM, the proposed model achieves a 47% reduction in mean absolute error (MAE), a 47% reduction in root mean square error (RMSE), and an increase in R2 from 0.6179 to 0.7500. This study is the first to introduce neural networks into the field of distributed optical fiber strain data interpretation, providing an innovative solution for efficient crack parameter interpretation. The findings provide a new interpretation approach for hydraulic fracturing crack monitoring, while also offering a methodological reference for addressing complex time-series modeling challenges in broader applications such as pipeline surveillance and seismic signal analysis.