Emd via mEMD: multivariate noise-Aided Computation of Standard EMD

Naveed ur Rehman, Cheolsoo Park, Norden E. Huang, Danilo P. MandicAdv. Data Sci. Adapt. Anal 2013 · DOI

TL;DR

This paper introduces NA-MEMD, a novel method that leverages multivariate empirical mode decomposition (MEMD) with added noise to improve the computation of standard empirical mode decomposition (EMD). NA-MEMD effectively addresses issues like mode-mixing and noise interference in EMD, offering better frequency localization and more robust analysis, particularly for multivariate signals like EEG.

What Problem Does This Paper Address?

Empirical Mode Decomposition (EMD) is a powerful data-driven technique for analyzing non-stationary signals by breaking them down into intrinsic mode functions (IMFs). However, standard EMD suffers from two major drawbacks: mode-mixing, where a single IMF contains signals of disparate scales or a single scale is spread across multiple IMFs, and its limitation to univariate data. Mode-mixing degrades the accuracy of time-frequency analysis and can lead to misinterpretations of signal components. While methods like Ensemble EMD (EEMD) attempt to mitigate these issues by adding noise, they can introduce residual noise into the decomposed signals. This paper aims to provide a noise-assisted approach that overcomes these limitations, particularly for multivariate signals, without introducing direct noise artifacts.

Methods

The core of this paper is the Noise-Assisted Multivariate Empirical Mode Decomposition (NA-MEMD) algorithm. It builds upon the Multivariate Empirical Mode Decomposition (MEMD) framework, which is designed to handle multi-channel data.

Here's how NA-MEMD works:

  • Multivariate Signal Creation: Instead of adding noise directly to the input signal (as in EEMD), NA-MEMD creates a multivariate signal by combining the original data channels with additional, uncorrelated Gaussian white noise channels. This effectively creates a "noise subspace" separate from the data subspace.
  • MEMD Processing: The combined multivariate signal (data + noise channels) is then processed using the MEMD algorithm. MEMD, due to its quasi-dyadic filter bank structure, aligns the components of the signal and the added noise.
  • IMF Reconstruction: After MEMD decomposition, the IMFs corresponding to the original data channels are extracted, while the IMFs from the noise channels are discarded.

This approach leverages the filter bank property of MEMD for the added noise to enforce a structured decomposition, thereby reducing mode-mixing and improving frequency localization. The paper also discusses an Ensemble NA-MEMD variant that averages results from multiple noise realizations to further enhance robustness, similar to EEMD but without direct noise addition to the signal.

Datasets

The paper evaluates NA-MEMD using:

  • Synthetic signals: Used to demonstrate the reduction of mode-mixing and compare performance with EMD and EEMD.
  • Real-world EEG data:
    • Steady-State Visual Evoked Potential (SSVEP) data: Collected from 7 EEG channels of a subject presented with a 15 Hz visual stimulus.
  • Motor Imagery EEG data: From the BCI Competition IV Dataset I, recorded from four healthy subjects using 59 electrodes, with 11 selected channels for analysis.

Key Insights and Ideas

  • Decoupling Noise and Data: NA-MEMD's key innovation is adding noise in separate channels, creating a distinct "noise subspace." This prevents direct noise artifacts in the final decomposition, unlike EEMD.
  • Leveraging MEMD's Structure: By using MEMD on a multivariate signal that includes noise channels, NA-MEMD exploits the quasi-dyadic filter bank property of MEMD to better align oscillatory modes and reduce mode-mixing.
  • Improved Frequency Localization: The method demonstrably leads to better separation of frequency components in the IMFs, which is crucial for analyzing complex, non-stationary signals like EEG.
  • Multivariate Signal Advantage: NA-MEMD is inherently designed for multivariate data, making it particularly well-suited for signals like EEG where channels are coupled and exhibit complex interactions.
  • Noise Power Control: The effectiveness of NA-MEMD is sensitive to the power of the added noise. While too little noise makes it similar to standard EMD, too much can degrade its data-driven nature. An optimal range of 2-10% of the input data power is suggested.

Key Results

  • Mode-Mixing Reduction: Simulations on synthetic and real EEG data show that NA-MEMD significantly reduces mode-mixing compared to standard EMD and EEMD, leading to cleaner and more localized IMFs.
  • Enhanced Frequency Localization: In EEG analysis (SSVEP and motor imagery), NA-MEMD achieved better separation of frequency bands (e.g., mu and beta rhythms) in the IMFs, facilitating more accurate analysis.
  • Robustness to Noise: NA-MEMD exhibits more stable residual noise levels compared to EEMD, even with increasing input noise power.
  • Multivariate Capability: NA-MEMD effectively processes multivariate EEG data, outperforming channel-wise application of univariate EMD/EEMD.

Limitations and Future Work

  • Sensitivity to Noise Power: While NA-MEMD is less sensitive than EEMD, excessive noise power (above 10% of input signal variance) can still degrade performance and lead to mode-mixing. The paper suggests an optimal range of 2-10%.
  • Pure Oscillatory/Rotational Signals: NA-MEMD may not offer significant improvement, or could even perform worse, on signals that are already pure oscillations or rotations (like sums of sinusoids), as these are already well-handled by standard EMD. However, most real-world signals, like EEG, are not purely oscillatory.
  • Future Work: The paper highlights that NA-MEMD is a promising tool for analyzing multivariate signals, particularly in BCI applications, and suggests further exploration of its capabilities in this domain. The authors also note that the work on improving classification of EEG-based motor imagery tasks using NA-MEMD is currently in progress.

Summary updated

BibTeX
@article{rehman2013emd,
  author = {Naveed ur Rehman and Cheolsoo Park and Norden E. Huang and Danilo P. Mandic},
  title = {{Emd via mEMD: multivariate noise-Aided Computation of Standard EMD}},
  journal = {Adv. Data Sci. Adapt. Anal},
  year = {2013},
  doi = {10.1142/s1793536913500076},
}
Figure