Adaptive Signal Decomposition for Non-Stationary Time Series

This direction is merged into another direction: adaptive signal decomposition. Its last version stays readable and citable here.

Version 1 of 177 papers94 conceptsfrozen 2026-10-04

Most traditional signal processing tools rely on a fundamental assumption: that complex signals can be understood as a sum of simple, predefined building blocks. In Fourier analysis, these blocks are sine waves; in Wavelet analysis, they are "mother wavelets" of a specific shape.

However, real-world data—from the electrical firing of a muscle to the flickering of a distant star—is rarely so cooperative. These signals are often non-stationary (their statistical properties change over time) and non-linear. When we force a non-stationary signal into a fixed basis like a sine wave, we lose the very temporal nuances we are trying to study.

Adaptive signal decomposition, primarily through Empirical Mode Decomposition (EMD) and the Hilbert-Huang Transform (HHT), solves this by letting the data define its own building blocks.

The Problem with Fixed Bases

If you use a Fourier Transform to analyze a signal where the frequency jumps halfway through, the resulting spectrum will show two peaks, but it won't tell you when those frequencies occurred. While windowed approaches like the Short-Time Fourier Transform (STFT) attempt to fix this, they are limited by the uncertainty principle: a narrow window gives good time resolution but poor frequency resolution, and vice versa.

As Bo Peng, Yong Min, and Xianchuang Su argue in their study of muscle fatigue, traditional methods lack the resolution needed for signals with time-varying properties. When a muscle tires, its electrical signature shifts dynamically. Fixed-basis transforms "smear" these transitions, whereas adaptive methods can track the instantaneous drop in frequency with high precision.

Sifting for Intrinsic Rhythms

The core mechanism of this research direction is Empirical Mode Decomposition (EMD). Instead of projecting a signal onto a mathematical template, EMD "sifts" the signal into a set of Intrinsic Mode Functions (IMFs).

The intuition behind sifting is purely geometric:

  1. Identify all local maxima and minima (peaks and valleys) in the signal.
  2. Connect the peaks to create an "upper envelope" and the valleys for a "lower envelope."
  3. Calculate the mean of these envelopes.
  4. Subtract this mean from the original signal.

This process is repeated until the remaining component satisfies two conditions: it has the same number of extrema and zero-crossings, and its mean is zero. This component is an IMF—a natural, oscillatory mode inherent to the data. By repeating this on the residual signal, researchers like Kuo-Hao Wu have shown that we can separate noise from speech by isolating high-oscillation noise into the first few IMFs and keeping the meaningful signal in the later ones.

From Modes to Instantaneous Frequency

Once a signal is decomposed into IMFs, we can apply the Hilbert Transform to each mode. This creates the Hilbert-Huang Transform (HHT).

Unlike Fourier analysis, which gives you a global frequency for the entire signal, HHT provides instantaneous frequency and amplitude. This allows researchers to see how the energy of a signal "flows" across different frequencies over time. Yu Zhang, Shen Wang, and Songling Huang used this to untangle overlapping wave packets in metal plates, identifying distinct wave modes that would be indistinguishable using standard time-domain analysis.

Why It Works: The Power of Adaptivity

The reason this approach is so effective across diverse fields—from detecting black hole oscillations to stabilizing handheld video—is its unsupervised, data-driven nature.

In video stabilization, Jun Yu 0001 and colleagues demonstrate that EMD can separate erratic, high-frequency "jitter" from the smooth, low-frequency "intended" camera path. Because the decomposition is based on the signal's own local time scales, the filter automatically adjusts to the specific intensity and speed of the camera shake without needing a human to set a "cutoff frequency."

Similarly, in astrophysics, A. McQuillan and S. Aigrain use non-parametric decomposition to remove instrumental noise from Kepler satellite data. By not assuming a specific mathematical form for the noise, they can clean the data while preserving the delicate, non-linear fluctuations of the stars themselves.

Robustness and Hybrid Models

While EMD is powerful, it can suffer from "mode mixing," where a single IMF contains widely different scales. Recent work has focused on making these decompositions more robust.

For instance, Mengdan Feng and Yonghui Duan utilize CEEMDAN (Complete Ensemble Empirical Mode Decomposition with Adaptive Noise) to predict carbon prices. By adding controlled noise during the decomposition process, they force the model to find more stable, physically meaningful components, which are then fed into machine learning models like XGBoost.

This "decompose-then-process" pipeline is becoming a standard for complex systems. Whether it is Akshansh Gupta classifying mental tasks from EEG or Constantino Alvarez Casado extracting biometric data from digital stethoscopes, the intuition remains the same: to understand a complex, changing world, we must let the data speak in its own rhythms.

Papers in this version

Showing 60 of 77 papers.

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_adaptive_signal_decomposition_for_non_st_c3cfb31b,
  title = {Adaptive Signal Decomposition for Non-Stationary Time Series},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/adaptive-signal-decomposition-for-non-stationary-time-series/rdr_c3cfb31b64de4287b646f0e1b38f6d41},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_adaptive_signal_decomposition_for_non_st_c3cfb31b_v1,
  title = {Adaptive Signal Decomposition for Non-Stationary Time Series},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/adaptive-signal-decomposition-for-non-stationary-time-series/rdr_c3cfb31b64de4287b646f0e1b38f6d41/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}