Adaptive Time-Frequency Analysis for Non-Stationary Signals

142 concepts136 papers1 proposal

In an ideal world, sensor data is clean, predictable, and stationary—meaning its statistical properties do not change over time. In the real world, however, the signals that matter most are inherently non-stationary and non-linear. Whether it is the transient shudder of a failing bearing, the turbulent cavitation inside an industrial valve, or the erratic voltage of a power grid, critical physical events manifest as fleeting, time-varying disruptions.

Traditional signal processing tools are fundamentally unsuited for these environments. To understand why, and to see how modern adaptive techniques overcome these limitations, we must look at how traditional tools handle time and frequency, and why data-driven decomposition changes the paradigm.

The Failure of Stationary Assumptions

The cornerstone of classical signal analysis is the Fourier Transform. It operates by mapping a signal from the time domain to the frequency domain, decomposing it into a sum of infinite sine and cosine waves. This approach works beautifully if the frequencies present in the signal last forever.

But when a mechanical component degrades—such as a rolling bearing developing a microscopic crack—the resulting physical signature is an instantaneous, localized impact. As Songling Wang et al. point out in their work on fan diagnostics, traditional techniques rely heavily on the assumption of signal stationarity. When this assumption collapses, classical frequency analysis fails.

Because the Fourier Transform integrates across all time, a brief, high-frequency snap gets smeared across the entire frequency spectrum. Its energy is diluted, rendering it invisible beneath the continuous, low-frequency hum of normal operations. Furthermore, real-world industrial environments introduce severe background noise and shifting operational loads. As Thomas Decker et al. note, complex mechanical faults produce signal modulations and overlapping frequencies that obscure underlying physical patterns.

Linear tools cannot untangle these overlapping components. For example, Xinfang Ge demonstrates that standard linear frequency analysis fails to characterize engine-induced vibrations because the non-stationary structural resonances directly overlap with the engine's firing frequencies. To isolate these components, we need a way to let the signal dictate its own coordinates.

Data-Driven Signal Decomposition

To analyze a non-stationary signal without smearing its features across time, researchers use adaptive decomposition. Instead of forcing the signal onto rigid, predefined basis functions (like the fixed sinusoids of Fourier or the static mother wavelets of standard wavelet transforms), adaptive algorithms break a complex signal down into a set of data-driven sub-signals based on its own local characteristic time scales.

Two primary methodologies dominate this space: recursive sifting and variational optimization.

Recursive Sifting: EMD and CEEMDAN

Empirical Mode Decomposition (EMD) and its advanced variant, Complete Ensemble EMD with Adaptive Noise (CEEMDAN), treat the signal as a mixture of fast and slow oscillations. The algorithm identifies all local maxima and minima, connects them with cubic splines to form upper and lower envelopes, and subtracts the mean of these envelopes from the signal.

Raw Signal ──> [Identify Local Extrema] ──> [Generate Envelopes] ──> [Subtract Mean Envelope] ──> IMF 1 (High Freq)
                                                                             │
                                              [Repeat on Residual] <─────────┘ ──> IMF 2 (Medium Freq)

This process is repeated recursively, peeling away high-frequency oscillations (called Intrinsic Mode Functions, or IMFs) until only a low-frequency trend remains. Chengwei Li demonstrates that this adaptive approach allows for highly effective denoising of chaotic friction signals, separating underlying physical trends from stochastic noise without altering the fundamental waveform.

Variational Optimization: VMD

While recursive sifting is highly intuitive, it suffers from "modal aliasing"—where disparate frequencies leak into the same IMF due to intermittent noise. Variational Mode Decomposition (VMD) solves this by abandoning recursion. It frames decomposition as a non-recursive, calculus-of-variations optimization problem.

VMD looks for a predefined number of sub-bands simultaneously, optimizing for the narrowest possible bandwidth around an evolving center frequency. Tongguang Yang et al. leverage this mathematical rigor to isolate subtle bearing fault signatures under highly variable speeds and loads. Because VMD determines the frequency bands dynamically, it adapts to shifting operational conditions. To maximize its precision, single-author work by Yuan Chai shows that bio-inspired optimization algorithms can be used to automatically tune VMD’s parameters, ensuring that the decomposition perfectly isolates the fault frequencies from intense industrial background noise.

Other multi-resolution techniques, such as the Wavelet Packet Transform (WPT) and Empirical Wavelet Transform (EWT), offer a middle ground by adaptively partitioning the Fourier spectrum. As the authors of a study on mining gearboxes explain, these methods provide the structural flexibility required to capture complex mechanical impacts where rigid, single-basis functions fail.

Quantifying Chaos via Complexity Metrics

Decomposing a signal into clean sub-bands is only half the battle. Once we have isolated the relevant IMFs or sub-bands, we must extract features that indicate whether the system is degrading. Simple statistical metrics like root-mean-square (RMS) or peak amplitude are often insufficient because a fault might change the structure of the signal's randomness without significantly changing its total energy.

To capture these structural shifts, researchers turn to entropy and complexity metrics, which measure the predictability and information dynamics of a time series.

  • Permutation Entropy (PE): This metric quantifies the predictability of a signal by analyzing the order of neighboring data points. A completely predictable signal has low permutation entropy, while a highly chaotic or noisy signal has high entropy. Jing Guo et al. introduced Composite Multiscale Transition Permutation Entropy to prevent the information loss that often occurs during standard signal processing, allowing models to detect subtle mechanical degradation across multiple time scales simultaneously.
  • Lempel-Ziv Complexity (LZC): LZC evaluates the rate at which new patterns arise along a digitized sequence of the signal. Tongguang Yang et al. paired optimized VMD with LZC to provide a quantitative metric for bearing wear, showing that as a component degrades, the emergence of periodic fault impacts systematically alters the signal's structural complexity.

By combining adaptive decomposition with complexity metrics, engineers can transform a messy, non-stationary raw signal into a highly descriptive, low-dimensional feature space that isolates the underlying physics of the system.

Synergizing with Deep Learning

While adaptive features provide a clean representation of non-stationary data, mapping those features to specific operational states or multi-fault classifications requires powerful pattern recognition. The modern consensus relies on feeding these adaptively extracted features into deep learning architectures.

Linear models and fixed thresholds struggle with real-time industrial data because combustion environments, fluid dynamics, and mechanical loads introduce chaotic, non-linear fluctuations (Yuanhao Shi et al.). Deep learning models excel at mapping these complex, non-linear relationships, but they perform poorly when fed raw, noise-corrupted non-stationary signals directly.

The combination of adaptive signal decomposition and deep learning creates a powerful synergy:

Raw Non-Stationary Signal ──> [Adaptive Decomposition (VMD/CEEMDAN)] ──> [Clean Sub-bands / IMFs] ──> [Deep Neural Network (CNN/LSTM)] ──> Precise Diagnosis

For instance, George Ioannou et al. demonstrate that combining 1D Convolutional Neural Networks (CNNs) with Long Short-Term Memory (LSTM) networks allows models to isolate highly complex, stochastic mechanical vibrations in nuclear reactors. Similarly, Dengxue Cao et al. utilize an optimized Wavelet Packet Transform alongside a Temporal Convolutional Network (TCN) to capture early-stage component failures hidden beneath industrial noise.

By using adaptive decomposition to handle the non-stationarity of the physical world, and deep learning to handle the non-linearity of the data, modern diagnostic frameworks can reliably detect structural anomalies long before they lead to catastrophic system failures.

History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_non_stationary_signal_analysis_0e48cca8,
  title = {non-stationary signal analysis},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/non-stationary-signal-analysis/rdr_0e48cca8db0a4abf85747a762443c64b},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_non_stationary_signal_analysis_0e48cca8_v1,
  title = {non-stationary signal analysis},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/non-stationary-signal-analysis/rdr_0e48cca8db0a4abf85747a762443c64b/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake