Strong motion recording baseline drift recognition based on CNN-LSTM

Abstract

Strong motion records, as one of the important means to obtain earthquake information and understand the nature of earthquakes, provide a scientific basis for earthquake prediction and disaster prevention and mitigation. However, records that contain baseline drift can degrade the quality of the data and affect subsequent studies. In this paper, a baseline drift identification model based on convolutional neural networks and long short-term memory algorithms is proposed for identifying records containing baseline drift from strong motion records. To improve the accuracy of the model, Bayesian optimization is used to optimize the hyperparameters of the model. Using the strong motion records from the 1999 Taiwan Chi-Chi earthquake, we constructed a dataset and divided the data into two categories: high-quality records and low-quality records. The experimental results show that the proposed baseline drift recognition model can effectively identify baseline drift records, with an accuracy of 83 % and an AUC value of 0.847. It also demonstrates good generalization performance on cross-domain test sets composed of data from the Japan KiK-net and European ESM databases. Compared to other models, the recognition performance of the model in this paper is superior.