: Due to the strong non-stationarity of rotating machinery vibration signals under coupling effects of noise environments and time-varying conditions, traditional time-frequency analysis (TFA) methods and existing time-frequency networks struggle to dynamically characterize closely-spaced instantaneous frequencies (IFs) under noisy environments. Therefore, the time-frequency denoising and optimization network (TFDON) is proposed. In the TFDON, an attention-guided sparse denoising sub-network (SDSN) is first designed to eliminate noise aliasing interference and obtain the clean time-frequency representation (TFR). Then, the time-frequency optimization sub-network (TFOSN) with three-stage hybrid Transformer blocks (HTB) cascade is constructed. Within each HTB, an efficient grouped Swin-Transformer (EGST) is developed to compute the spatiotemporal character istics, and guided by a dual-layer attention mechanism, the time-frequency concentration is iteratively enhanced. Additionally, a weight-controllable joint loss function tailored for the TFR denoising and optimization is designed to achieve the optimal balance in two tasks. The performance of the TFDON in characterizing and noise suppression is verified by a simulated signal with closely-spaced IFs. Meanwhile, two bearing and a planetary gearbox vibration signal added noise are further analyzed, and the TFDON achieves the lowest Rényi entropy of 6.055, 6.387, and 6.077 at −5 dB, *Corresponding author. Downloaded for personal academic use. All rights reserved. https://papernode.online/
Download PDF: https://gemilang.eu.org/xLDGND