A fast identification method for near fault ground motion based on LSTM

Traditional methods based on velocity pulse extraction struggle to accurately and efficiently identify the high-energy and changeable waveforms of near-fault ground motions. This paper explores the identification of near-fault ground motions using long and short-term memory (LSTM) neural networks. By utilising memory elements to process the ground motion time course, the approach offers improved accuracy and efficiency in identification. The 5356 non-near-fault ground motions and 154 near-fault ground motions in the PEER ground motion database were used as samples. Recognition is performed based on different neural network structures and preprocessed using different signal processing methods. In turn, the effects of various neural network structures and signal processing methods on the recognition results are compared. The results indicate that the prediction is significantly more accurate with one single hidden layer than with multiple hidden layers when using ground motion velocity time course as the neural network input. The training accuracy reaches a maximum of 93.07% with 95 neurons, and the test accuracy reaches a maximum of 92.65% with 100 neurons. When using the short-time Fourier transform to obtain the instantaneous frequency and spectral entropy of the signal as input to the neural network, the test accuracy reaches a maximum of 88.36% with 60 neurons. Further increasing the number of neurons does not improve the prediction effect. The ground motion is analyzed using the continuous wavelet transform with the 'db4' mother wavelet. The resulting wavelet coefficients are then used as inputs for the neural network. The neural network achieves the best prediction accuracy of 96.07% when the scale is set to 10 and the single hidden layer contains 400 neurons. The neural network for LSTM can effectively identify complex signals, such as those near fault ground motion, with appropriate preprocessing of input ground motion and neural network design.