TFD-Net: Towards Intelligent Time-Frequency Mode Decomposition with Practical Applications

Time-frequency analysis (TFA) and mode decomposition for non-stationary signals are research hotspots in the field of signal processing. Current optimization-based decomposition methods require a good initial IF estimate. However, due to the Heisenberg uncertainty, achieving accurate ridge extraction from the time-frequency representation (TFR) necessitates empirical parameter adjustments. In this paper, we propose the TFD-Net framework, which takes time-series signals as inputs and adaptively conducts TFR construction and mode decomposition. Specifically, the framework integrates a physically interpretable TFA encoder and a query-based mode decomposition decoder. The highlights of this study include exploring the mathematical equivalence between deep convolutional operators and classical TFA methods. This enables the extraction of multi-scale features for TFR construction and mode separation in a data-driven manner, eliminating the need for signal-specific manual tuning and enhancing adaptability. Finally, simulated and real-world experiments demonstrate TFD-Net's superior performance over several state-of-the-art methods in complex signal processing.