Anti-frequency long short-term memory model for stable estimation of structural response under noise conditions

Abstract

Deep learning models for structural response estimation exhibit degraded performance when the training and input data characteristics differ, particularly because of noise. This study proposes an anti-frequency long short-term memory (AF-LSTM) model designed to ensure a stable estimation regardless of noise conditions. The term “anti-frequency” is used to describe the process of suppressing predefined frequency components by setting them to zero in the frequency domain. The AF-LSTM model introduces an AF layer before the LSTM layer, which suppresses specific frequency components before learning. The AF layer transforms the input into the frequency domain, zeroes out the components within predefined noise-prone frequency bands, and converts the signal back to the time domain. This process enables LSTM to effectively learn and estimate structural responses with improved consistency, even under noise conditions. The proposed model was verified using a numerical three-degree-of-freedom system, demonstrating stable estimation performance under varying noise frequencies and amplitude ratios. Experimental validation on a three-story steel frame structure and acceleration data from a real 55-floor building with environmental noise confirmed the model's ability to estimate stable responses across non-stationary inputs.