What Is Empirical Mode Decomposition?

Empirical mode decomposition (EMD) is an adaptive, data-driven technique for decomposing nonlinear and non-stationary time series signals into a finite set of oscillatory components called intrinsic mode functions. Unlike the Fourier transform, which decomposes a signal into fixed sinusoidal basis functions, or wavelet analysis, which uses a predetermined mother wavelet, EMD extracts basis functions directly from the data based on the local time scale of the signal itself. This adaptivity makes it particularly well suited for signals whose frequency content changes over time and whose oscillations do not conform to sinusoidal templates. The method was introduced by Norden E. Huang and colleagues at NASA Goddard Space Flight Center and published in the Proceedings of the Royal Society of London A in 1998.

EMD is the foundational step of the Hilbert-Huang transform (HHT), a complete time-frequency analysis framework. After decomposing the signal into intrinsic mode functions, the HHT applies the Hilbert transform to each component to obtain instantaneous frequency and amplitude as functions of time, producing a time-frequency-energy representation called the Hilbert spectrum. The combination of EMD and the Hilbert spectrum has been applied across geophysics, biomedical engineering, mechanical fault detection, and oceanography.

Intrinsic Mode Functions

An intrinsic mode function satisfies two conditions: the number of extrema and the number of zero crossings differ by at most one, and the mean value of the envelope defined by local maxima and the envelope defined by local minima is zero at every point. These conditions ensure that the function is locally symmetric about zero and that a Hilbert transform applied to it will yield a meaningful instantaneous frequency. The decomposition process, called sifting, identifies the signal's local extrema, interpolates upper and lower envelopes using cubic splines, subtracts the mean envelope from the signal, and repeats until the residual meets the IMF criteria. Successive applications of the sifting process extract components from the highest-frequency oscillations present in the signal down to any residual trend, which is typically a monotone function representing the signal's long-term behavior. The PMC article on EMD software in Python provides a computational description of the sifting algorithm alongside open-source implementations for signal analysis workflows.

Hilbert Spectrum and Time-Frequency Analysis

Once the signal has been decomposed into intrinsic mode functions, the Hilbert transform converts each real-valued IMF into an analytic signal whose instantaneous amplitude and frequency can be computed as continuous functions of time. Plotting instantaneous amplitude squared as a function of both time and frequency yields the Hilbert spectrum, which represents the signal's energy distribution in the time-frequency plane without the smearing introduced by the fixed windowing of short-time Fourier methods. This representation is especially informative for signals with intermittent components or with frequencies that drift continuously, both of which are common in real physical systems. The original Huang et al. 1998 paper in the Proceedings of the Royal Society demonstrates the Hilbert spectrum on ocean wave, earthquake, and bridge vibration data, establishing the range of physical systems where the approach offers advantages over Fourier methods.

Ensemble EMD and Variants

A recognized limitation of the original EMD algorithm is mode mixing, which occurs when a single intrinsic mode function contains oscillations at widely separated scales, typically triggered by signal intermittencies. Ensemble empirical mode decomposition (EEMD), proposed by Wu and Huang in 2009, addresses mode mixing by adding white noise to the signal multiple times, running EMD on each noise-added realization, and averaging the resulting IMFs across the ensemble. The added noise spreads intermittent structures uniformly across scales in individual trials, and the averaging removes the noise while preserving the separation of the underlying modes. Further variants, including complete EEMD with adaptive noise (CEEMDAN) and multivariate EMD (MEMD), extend the framework to multichannel and multivariate signals. The Nature Scientific Reports paper on best practices for EMD and iterative filtering evaluates the numerical behavior of these variants on a range of synthetic and real signals.

Applications

Empirical mode decomposition has applications in a wide range of fields, including:

• Biomedical signal processing, including EEG, ECG, and EMG analysis for disease detection and brain-computer interface development
• Mechanical fault diagnosis, where IMFs isolate bearing and gear defect signatures in vibration data
• Oceanographic and geophysical data analysis, including tide gauge records, seismic waveforms, and airborne gravity measurements
• Speech and audio processing, where instantaneous frequency tracks voice pitch and formant dynamics
• Financial time series analysis, separating trend from cyclical and noise components in economic data