Multidimensional Iterative Filtering method for the decomposition of high-dimensional non-stationary signals

Antonio Cicone, Haomin Zhou

Abstract

Iterative Filtering (IF) is an alternative technique to the Empirical Mode Decomposition (EMD) algorithm for the decomposition of non-stationary and non-linear signals. Recently, IF has been proved to be convergent for any $L^2$ signal and its stability has been demonstrated through examples. Furthermore, Fokker-Planck (FP) filters have been introduced, which are smooth at every point and have compact supports. Based on those results, this paper introduces the Multidimensional Iterative Filtering (MIF) technique for the decomposition and time-frequency analysis of non-stationary high-dimensional signals. We present the extension of FP filters to higher dimensions and illustrate the promising performance of the MIF algorithm when applied to the decomposition of 2D signals.

Introduction

In fields ranging from structural engineering to medical imaging, researchers frequently encounter "non-stationary" signals—data where frequency and amplitude vary over time or space. Traditional analytical tools like Fourier Transforms or Wavelets often struggle with such data because they rely on fixed, predetermined bases that do not adapt to the specific signal characteristics.

The Empirical Mode Decomposition (EMD) algorithm was a significant breakthrough due to its fully data-adaptive nature. However, EMD's reliance on cubic splines to create "envelopes" lacks a rigorous mathematical proof of convergence and can be highly sensitive to noise. Furthermore, extending EMD to multidimensional signals (2D or 3D) is computationally intensive and theoretically complex. This work bridges these gaps by providing a stable, convergent, and multidimensional version of Iterative Filtering.

Methodology: The MIF Algorithm

The core contribution is the Multidimensional Iterative Filtering (MIF) algorithm. It decomposes a signal into a set of Intrinsic Mode Functions (IMFs) through two nested loops.

The Sifting Process

The inner loop, or "sifting" process, extracts a single IMF. Unlike EMD, which uses splines for moving averages, MIF employs convolution with a filter function. It iteratively subtracts the moving average from the signal until a stopping criterion is met.

The Decomposition Loop

The outer loop subtracts the extracted IMF from the original signal and repeats the process on the residual to find subsequent IMFs. This continues until only a trend remains.

The Multidimensional Iterative Filtering (MIF) algorithm pseudocode showing the nested loop structure.

A critical component is the Generalized Fokker-Planck (GFP) Filter. By rotating or scaling 1D Fokker-Planck filters—which are infinitely differentiable and have compact support—the authors create smooth multidimensional filters without solving high-dimensional partial differential equations.

Results and Discussion

The authors evaluated MIF using artificial 2D signals and real-world hyperspectral data. A key advantage of MIF is its mathematical stability; unlike EMD, it is proven to converge for any $L^2$ signal.

Accuracy and Separation

In artificial tests involving mixtures of non-stationary signals, MIF demonstrated remarkable accuracy. It successfully separated components of fundamentally different natures, such as smooth waves from jagged, non-smooth signals. As shown in the comparison plots, the extracted IMFs align almost perfectly with the ground truth.

Comparison between the extracted IMF2 and the ground truth for an artificial signal slice, demonstrating the high decomposition accuracy of the MIF method.

Real-World Application

In hyperspectral chemical detection, MIF served as an effective pre-processing step. By identifying and removing high-frequency noise (typically captured in the first IMF), the algorithm significantly reduced false alarms in classification maps compared to raw data analysis.

Conclusion and Future Work

MIF provides a mathematically rigorous, data-adaptive alternative to EMD for high-dimensional signal analysis. While promising, several areas for future research remain, including the automatic identification of optimal filter supports in multiple dimensions and the extension of instantaneous frequency concepts (like the Hilbert Transform) to multidimensional IMFs.

Summary updated

Abstract

Iterative Filtering (IF) is an alternative technique to the Empirical Mode Decomposition (EMD) algorithm for the decomposition of non-stationary and non-linear signals. Recently in [1] IF has been proved to be convergent for any $L^2$ signal and its stability has been also showed through examples. Furthermore in [1] the so called Fokker-Planck (FP) filters have been introduced. They are smooth at every point and have compact supports. Based on those results, in this paper we introduce the Multidimensional Iterative Filtering (MIF) technique for the decomposition and time-frequency analysis of non-stationary high-dimensional signals. And we present the extension of FP filters to higher dimensions. We illustrate the promising performance of MIF algorithm, equipped with high-dimensional FP filters, when applied to the decomposition of 2D signals. [1] A. Cicone, J. Liu, and H. Zhou, Adaptive local iterative filtering for signal decomposition and instantaneous frequency analysis, arXiv:1411.6051, 2014.
BibTeX
@inproceedings{cicone2017multidimensional,
  author = {Antonio Cicone and Haomin Zhou},
  title = {{Multidimensional Iterative Filtering method for the decomposition of high-dimensional non-stationary signals}},
  booktitle = {Numerical},
  year = {2017},
  doi = {10.4208/nmtma.2017.s05},
}
Figure