Vibration Recognition Based on Feature Extraction by Deep Autoencoder

Vibrational response identification of high-rise building structures under excitation of varying ambient conditions is of great significance for structural vibration control and health monitoring. Traditional strategies allocating time-frequency domain transformation or graph neural network for feature extraction cast great loss on computational efficacy and engineering applicability. Herein, this study proposes a novel vibration recognition method based on feature extraction by deep autoencoder. In this method, a deep autoencoder (AUE) is established for vibration signal reconstruction in frequency domain, on the basis of which the vibration sensitive parameters are defined and quantified based on the discrepancy between original signals and the reconstructed ones. The extracted indicators are then performed as input feature engineering of downstream linear classifiers in a supervised pattern, thereby achieving rapid identification of vibration response. To validate the capacity of the proposed method, a task of vibration response identification of a super high-rise building under typical ambient excitations was conducted. The performance results achieved a satisfactory accuracy exceeding 0.95, suggesting a powerful capability and potential applicability.