Proposing a time-frequency analysis method using nonlinear wave modulation for machine learning-based detection of bolt looseness in non-gaussian noise environment

Abstract

Detecting bolt loosening in bolted structures under non-Gaussian noisy environments is challenging with existing time-frequency analysis methods. This study proposes a novel fractional lower-order fractional synchrosqueezing-extracting transform (FLOFSSET) to effectively process linear chirp signals in such environments. Additionally, a data fusion approach based on the Dezert-Smarandache Theory is introduced to enhance machine learning model performance. A nonlinear wave modulation (NWM) technique combined with chirp signal excitation (chirp-NWM) is also implemented to improve early damage detection in bolted structures. The FLOFSSET method ensures high accuracy in analyzing close and far instantaneous frequencies, with superior energy concentration compared to conventional methods. Furthermore, decision-level data fusion is employed using machine learning models, including, Newton method for sparse support vector machines, sparse artificial neural network, and adaptive neuro-fuzzy Inference System. This fusion utilizes features extracted by the FLOFSSET method and a multi-sensor fusion method. The FLOFSSET method outperformed conventional approaches in analyzing close and far instantaneous frequencies, achieving higher energy concentration. For bolt-loosening detection, integrating FLOFSSET features with data fusion improved the average accuracy of machine learning models from 85 % to 99.5 %. In contrast, using conventional features reduced the accuracy of the combined models to 92.5 %. In scenarios with close instantaneous frequencies, only the FLOFSSET method accurately identified bolt loosening, with the fused model achieving 99.1 % accuracy. Conventional time-frequency analysis features were unable to accurately differentiate these frequencies, making them ineffective and resulting in misclassifications. The proposed FLOFSSET method, combined with Dezert-Smarandache Theory-based data fusion, significantly enhances the performance of machine learning models in detecting bolt loosening under non-Gaussian noise. FLOFSSET features provide higher discrimination capability and accuracy compared to conventional time-frequency methods, making it a robust tool for early damage detection in bolted structures.