TL;DR
This paper introduces a new data-driven method called Nonstationary Fourier Mode Decomposition (NFMD) for analyzing complex time-series data. NFMD excels at extracting instantaneous frequencies and amplitudes from signals that change abruptly or nonlinearly, overcoming limitations of traditional techniques like the Hilbert-Huang Transform (HHT). This advancement has significant implications for understanding dynamic systems across various scientific fields.
What Problem Does This Paper Address?
Time-series analysis is fundamental to many scientific and engineering disciplines, but analyzing signals that are nonstationary (their statistical properties change over time) and nonlinear presents a significant challenge. Traditional methods often assume stationarity, leading to inaccurate results when applied to real-world data exhibiting sudden shifts in frequency or amplitude. Existing advanced techniques like the Hilbert-Huang Transform (HHT) can struggle with signals that have discontinuities or nonlinear phase evolution, and often lack a strong theoretical foundation. This paper aims to bridge this gap by providing a robust and interpretable method for time-frequency analysis of such complex signals.
Methods
The core of the proposed method is the Nonstationary Fourier Mode Decomposition (NFMD). It builds upon several key ideas:
- Gradient Descent Algorithms: These are used to optimize the fitting of Fourier modes to the data.
- Fourier Transform: The fundamental building block for representing signals in terms of frequencies.
- Multi-resolution Analysis: The signal is broken down into smaller segments for analysis.
- Bayesian Spectral Analysis: Provides a framework for parameter estimation.
NFMD works by first decomposing the entire time series into segments. For each segment, it applies a Fourier Mode Decomposition (FMD) algorithm. FMD frames the problem as fitting a finite number of Fourier modes (sums of sines and cosines with flexible frequencies) to the data. This fitting is achieved through gradient descent optimization, aiming to minimize the error between the model and the data.
A key innovation is how NFMD handles the time-frequency aspect. Instead of analyzing the entire signal at once, it uses a segment-by-segment approach, similar to the Short-Time Fourier Transform (STFT). However, unlike STFT, NFMD fits Fourier modes to each segment, leading to improved frequency resolution and interpretability. Crucially, the learned frequencies from one segment are used as initial guesses for the next, significantly speeding up the process.
The method also explicitly models the instantaneous mean of the signal, which can capture non-periodic trends or drifts. This instantaneous mean is then fit to a model, providing insights into underlying forcing functions or system dynamics.
Datasets
The paper demonstrates the NFMD method on several types of data:
- Synthetic Test Signals: These are generated signals designed to mimic specific nonstationary and nonlinear behaviors, including:
- Signals with abrupt changes in instantaneous frequency and amplitude.
- Signals with additive noise at various signal-to-noise ratios (SNRs).
- Signals with discontinuities in instantaneous frequency and mean.
- Experimental Data: The method is applied to data from cantilever-based electrostatic force microscopy (trEFM). This data captures the time-dependent evolution of charging dynamics at the nanoscale, specifically analyzing the response of a metallic cantilever to applied forces. The experimental setup involves controlled perturbation forces with known relaxation times.
Key Insights and Ideas
- Bridging the Gap: NFMD effectively bridges the gap between traditional Fourier analysis and more advanced, yet sometimes less interpretable, methods like HHT, offering a more robust and theoretically grounded approach for nonstationary signals.
- Interpretable Decompositions: The method produces interpretable signal decompositions, allowing researchers to identify individual periodic components, their instantaneous frequencies, and amplitudes.
- Handling Discontinuities: A major advantage is NFMD's ability to accurately capture signals with sharp changes in frequency or amplitude, which often pose problems for other methods.
- Instantaneous Mean as a Forcing Indicator: The ability to accurately estimate the instantaneous mean of a signal is a powerful tool for inferring non-periodic forcing functions applied to oscillatory systems. This is demonstrated in the oscillator examples.
- Improved Resolution: NFMD offers superior temporal and spatial resolution compared to STFT, providing a more detailed view of signal dynamics.
Key Results
- Superior Decomposition: NFMD consistently outperforms the Hilbert-Huang Transform (HHT) in decomposing nonstationary signals, especially in the presence of noise. It more accurately identifies the correct number of modes and their instantaneous parameters (frequency and amplitude).
- Noise Robustness: Even with added noise (e.g., SNR=35), NFMD maintains good performance, whereas HHT shows erratic behavior.
- Accurate Mean Estimation: The instantaneous mean estimated by NFMD closely matches the true mean in synthetic examples and provides valuable insights into applied forcing functions in oscillator experiments.
- Experimental Validation: In trEFM experiments, NFMD successfully decomposes the cantilever signal into a periodic mode and an instantaneous mean, which is then used to estimate the relaxation time of applied perturbation forces. This method shows promise for identifying characteristic time constants in complex systems.
- Speed Improvement: Using learned frequencies from previous segments as initial guesses significantly speeds up the decomposition process.
Limitations and Future Work
- Window Size Limitation: The paper notes that the window size used for decomposition is limited by the frequency of the periodic components in the signal. For very fast periodic components, a smaller window is needed, which can impact the accuracy of fitting the instantaneous mean.
- Mode Identification for Mean: Identifying which mode represents the instantaneous mean can sometimes be challenging and may require inspection. The paper states that in their work, the lowest-frequency mode is generally assumed to be the mean.
- Future Work: The authors suggest that NFMD's ability to provide interpretable decompositions and accurately estimate instantaneous means makes it a viable architecture for integration into a diverse number of scientific processes. The method's application to experimental data from trEFM highlights its potential for real-world scientific discovery. Further exploration into optimizing the model fitting for the instantaneous mean and its application to even more complex systems is implied.
Summary updated