Machine Learning-Driven Adaptive Time-Frequency Analysis Using Wasserstein Distance

Traditional algorithms based on the synchrosqueezing transform (SST) struggle to adapt to the features of the signal's different segments, and their reliance on manual tuning restricts the applicability in automated scenarios. This study proposes an algorithm with a time-varying window length sequence (WLS) integrated with a sparse Bayesian optimizer (SBO) through a cost function using the Wasserstein distance (WD). First, a time-varying WLS is introduced into the signal to adapt to different local characteristics. Second, after performing an iterative reassignment operation, an SBO determines the high-dimensional WLS, avoiding manual tuning. Furthermore, this work employs WD for the first time to evaluate the ridge formation, enabling intelligent variation selection. Simulation and experimental results indicate that the proposed algorithm reduces the frequency estimation error by 49.36% compared with similar algorithms, while achieving higher energy regressions. This study also analyzes the relationship between the original signal and WLS, verifying the effectiveness of adaptive framework. The results indicate that the proposed algorithm can provide a reliable tool for the automated analysis of complex nonstationary signals.