Empirical Mode Decomposition (EMD)

The defining characteristic of EMD is its complete adaptivity. The decomposition basis is derived directly from the signal itself, not from a fixed mathematical function like a sine wave or a Morlet wavelet. This makes it exceptionally well-suited for analyzing non-stationary and nonlinear data where predefined harmonic bases fail to capture the underlying physics. The algorithm acts as a dyadic filter bank on white noise but adapts its shape to the data for deterministic signals.

The output of EMD is a finite set of Intrinsic Mode Functions (IMFs). Each IMF must satisfy two conditions:

• The number of extrema and zero-crossings must differ by at most one.
• The mean value of the envelope defined by the local maxima and the envelope defined by the local minima is zero at any point. This ensures each IMF is a mono-component signal with a physically meaningful instantaneous frequency, avoiding the cross-term interference that plagues quadratic time-frequency distributions.

The core algorithm is an iterative procedure called sifting. For a signal x(t), the process is:

1. Identify all local maxima and minima.
2. Interpolate between maxima (upper envelope) and minima (lower envelope) using cubic splines.
3. Calculate the local mean m(t) of the two envelopes.
4. Extract the proto-IMF: h(t) = x(t) - m(t).
5. Repeat steps 1-4 on h(t) until it satisfies the IMF conditions. This computationally intensive loop separates the finest oscillatory mode from the residual trend.

EMD is a complete decomposition. The original signal can be perfectly reconstructed by summing all extracted IMFs and the final residual trend: x(t) = Σ IMFᵢ(t) + r(t). While not theoretically guaranteed, the decomposition is practically orthogonal, meaning the IMFs carry distinct frequency information with minimal energy leakage between components. This allows for targeted filtering by selectively removing specific IMFs, such as high-frequency noise or low-frequency baseline wander, without affecting other signal structures.

EMD excels where traditional methods fail: analyzing signals with time-varying frequency content. Because the IMFs are not constrained to constant amplitude and frequency, they can represent natural physical processes like a chirp signal or a seismic wave with changing frequency. The subsequent application of the Hilbert transform to each IMF yields a sharp, physically meaningful instantaneous frequency, enabling the construction of the Hilbert spectrum—a high-resolution time-frequency-energy distribution free from the smearing artifacts of the Short-Time Fourier Transform.

A primary vulnerability of the original EMD is mode mixing, where a single IMF contains signals of widely disparate scales, or a signal of a similar scale appears in multiple IMFs. This often occurs when the signal is intermittent or contains noise. To address this, the Ensemble Empirical Mode Decomposition (EEMD) was developed, which adds finite-amplitude white noise to the signal, performs EMD on multiple noisy trials, and averages the resulting IMF sets. The noise-assisted approach forces the dyadic filter bank behavior, dramatically improving scale separation.

The fundamental oscillatory components extracted by EMD. An IMF must satisfy two strict conditions: the number of extrema and zero-crossings must differ by at most one, and the local mean envelope (defined by the average of the upper and lower envelopes) must be zero at any point. This ensures the IMF represents a simple oscillatory mode with a well-defined instantaneous frequency, free from riding waves.

The iterative, data-driven algorithm at the heart of EMD. It works by repeatedly extracting the local mean from a signal until an IMF is isolated:

• Step 1: Identify all local maxima and minima of the signal.
• Step 2: Interpolate a cubic spline through the maxima to form an upper envelope, and through the minima for a lower envelope.
• Step 3: Calculate the local mean envelope.
• Step 4: Subtract the mean from the signal to produce a proto-IMF.
• Step 5: Repeat until the stopping criterion is met.

A two-step adaptive data analysis framework that combines EMD with the Hilbert spectral analysis. First, EMD decomposes a non-stationary signal into a finite set of IMFs. Then, the Hilbert transform is applied to each IMF to calculate its instantaneous amplitude and frequency. The result is a high-resolution time-frequency-energy distribution called the Hilbert spectrum, which is particularly effective for analyzing nonlinear and non-stationary phenomena.

A noise-assisted enhancement designed to solve the mode mixing problem of standard EMD, where a single IMF contains signals of disparate scales. EEMD adds finite-amplitude white Gaussian noise to the original signal, performs EMD on the noisy ensemble, and averages the resulting IMFs across multiple trials. The added noise populates the time-frequency space uniformly, providing a natural dyadic filter bank structure and forcing the decomposition to separate scales more robustly.

A non-recursive, mathematically rigorous alternative to EMD that extracts modes concurrently by solving a variational optimization problem. VMD defines each mode as an amplitude-modulated-frequency-modulated (AM-FM) signal with a specific bandwidth and sparsity prior. It minimizes the sum of the bandwidths of all modes, constrained by the requirement that their sum equals the original signal. This approach is highly robust to noise and sampling artifacts, avoiding the recursive error propagation of the sifting process.

Mode mixing is a primary challenge in EMD, defined as either a single IMF containing widely different frequency scales or similar scales appearing across different IMFs. It is often triggered by signal intermittency. To control the sifting iteration, a stopping criterion is essential. The standard Cauchy-type criterion terminates sifting when the normalized squared difference between two consecutive proto-IMFs falls below a threshold (e.g., 0.2-0.3), preventing over-decomposition and the extraction of physically meaningless amplitude-only modulated components.