Optimized Noise Suppression of Acceleration Records from Shaking-Table Tests Using Signal Processing and Deep Learning Techniques

Reliable displacement reconstruction from acceleration measurements is a persistent challenge in shaking-table testing because sensor noise and baseline drift can accumulate through numerical integration, biasing the inferred kinematic response. This study investigates and compares two complementary strategies for noise suppression and displacement estimation: (i) an optimized signal-processing workflow that combines band-limited integration with systematic parameter tuning, and (ii) a data-driven approach based on a Long Short-Term Memory (LSTM) neural network trained to map acceleration time histories to displacement. For the signal-processing workflow, filter parameters are selected through a multi-objective search using Latin Hypercube Sampling (LHS) to minimize reconstruction error against reference displacement measurements while limiting drift and spurious low-frequency content. For the LSTM approach, model hyperparameters are selected via Bayesian optimization to balance accuracy and generalization across excitation phases. The methods are assessed on shaking-table records from an instrumented experimental campaign, using displacement transducers as reference. Results indicate that both approaches substantially reduce integration drift and noise-induced artefacts compared with conventional fixed filtering and detrending. The optimized signal-processing pipeline provides a transparent, physically interpretable baseline with strong accuracy, whereas the LSTM model can achieve comparable performance with reduced need for manual tuning and improved robustness to non-stationary noise characteristics. The proposed framework offers a reproducible benchmark for computational-intelligence methods in vibration data post-processing and supports more reliable displacement estimation in experimental structural dynamics.