Smart Super-Sampling-Frequency Vibration Signal Estimation for Skyscrapers under Typhoons Using Augmented Data Set

Recently, super-sampling-frequency (SSF) estimation for structural vibration signals has emerged as a state-of-the-art topic. By converting high-frequency signals using low-frequency samples, it significantly reduces the required data volume. However, a critical gap in previous studies is the lack of attention to SSF estimation for structures under typhoon events. In such scenarios, the existing SSF method based on compressed sensing (CS), i.e., CS with frequency-domain evaluation index (CS-FEI), is incapable of learning underlying data patterns of typhoon-induced structural vibrations. Although neural networks offer powerful learning capability that is well suited for such tasks, their applicability is often constrained by insufficient data volume. To mitigate these limitations, a combined SSF signal estimation framework is proposed that fuses CS with neural networks, leveraging their complementary advantages. A key novelty lies in SSF signal reconstruction, where high-frequency signals are reconstructed using low-frequency uniform signals to augment the data volume for neural networks. To implement the proposed framework, a method based on orthogonal matching pursuit (OMP) and long short-term memory network (LSTM), or OMP-LSTM, is proposed. Moreover, a frequency-domain assessment strategy is designed to evaluate the data quality of the original values of estimated signals. The theoretical foundation of CS-based SSF signal reconstruction is established, and its high performance is validated. Results on synthetic simulations and real-world typhoon recordings confirm that OMP-LSTM substantially enriches the data fed into neural networks and outperforms the existing CS-FEI method. This paper provides an efficient pathway toward resolving the data volume constraints in neural network–based signal estimation for structures under typhoon events.