An innovative method integrating two deep learning networks and hyperparameter optimization for identifying fiber optic temperature measurements in earth-rock dams

Abstract

Since one of the main threats to the safety of earth-rock dams is leakage, its timely and accurate identification is crucial. Distributed fiber optic sensing system (DFOS) is considered as one of the ideal methods for leakage monitoring in earth-rock dams. However, the working conditions of earth-rock dams are complex, and the identification of fiber optic temperature measurements has issues such as low efficiency and high misjudgment rate. For improving the identification efficiency and accuracy of fiber optic temperature measurements in earth-rock dams, a signal identification method integrating least squares generative adversarial network (LSGAN), one-dimensional convolutional neural network (1DCNN), and white shark optimization (WSO) algorithm is presented. Firstly, the LSGAN model is used to augment the signals of different categories to reduce the effect of data set unbalance on the identification result. According to the variation characteristics of fiber optic temperature measurement signals in earth-rock dams, a 1DCNN model is designed to extract signal features for classification. To reduce the blindness in hyperparameter setting of 1DCNN model, the WSO algorithm is introduced to optimize its key hyperparameters, which further enhances the identification accuracy of the model. The new method is applied to a data set specifically acquired with tests on a physical model of an earth-rock dam. The identification accuracy obtained with the new method reaches 99.76 %, which is better than the accuracy of other commonly used identification methods. Upon completion of the pre-training, the new method can fulfill the practical needs of fast identification and has promising applications.