A feature extraction method for distributed optical fiber strain monitoring data of tunnel structures based on an LDSA-TCN

By using distributed optical fiber-based tunnel deformation monitoring technology, high-density strain data can be obtained from numerous measurement points. However, characterizing the load strain response relationship of a tunnel on the basis of the features of such monitoring data remains challenging because of the strong noise, nonstationarity, and complex spatiotemporal coupling in the strain response. To address this issue, a feature extraction method for the distributed optical fiber strain monitoring data of tunnel structures based on an LDSA-TCN is proposed. First, relying on the high-density strain monitoring data of a tunnel, the local dynamic spatial autocorrelation (LDSA) method is employed to identify abnormal points along the distributed sensing fiber. A temporal convolutional network (TCN) is subsequently introduced, taking the time series strain data of the abnormal points as its inputs. Through multilayer dilated causal convolutions that are used to predict the response at the next time step, the method achieves time series modeling and feature extraction for the strain data. An analysis of field-measured data reveals that the proposed method outperforms the traditional three-standard-deviation method in terms of spatial anomaly detection. With respect to time series modeling, the proposed method demonstrates better feature stability and higher sensitivity to external load variations than LSTM and attention-based networks do. Furthermore, the extracted features can be applied to accurately assess the structural conditions of tunnels. • Proposes an LDSA-TCN framework for spatiotemporal feature extraction in distributed optical fiber tunnel strain monitoring. • Develops a temporal convolutional network with dilated causal convolutions and self-supervised feature learning. • Proposed LDSA-TCN framework suitable for abnormal detection and condition assessment of tunnel.