1 Introduction
1.1 Definition and distinguishing features
Non-stationary noise is noise whose statistical properties vary with time. The variation may involve the signal’s average level (mean), its spread (variance or amplitude), its frequency content, or the way samples relate to one another across time. In practice, these changes can appear as gradual drifts, abrupt bursts, short transients, evolving periodic components, or time-dependent correlation patterns. Because the underlying statistics are not constant, treating the noise as if it were stationary can lead to biased estimation and misleading uncertainty.
1.2 Stationary vs. non-stationary noise (conceptual comparison)
In stationary noise, descriptors such as mean and autocorrelation are invariant under time shifts, so the noise has a consistent “statistical character” throughout the record. Non-stationary noise violates this assumption: the same method applied across the entire duration may work in one interval and fail in another. Conceptually, stationary noise can be summarized by a single set of global statistics (e.g., a fixed power spectral density), whereas non-stationary noise often requires a time-indexed description (e.g., spectra that change with time or correlation functions that evolve).
1.3 Common sources and examples
Non-stationarity arises in many settings:
- Instrumentation artifacts: sensor gains may drift, electronic interference may vary, or environmental electromagnetic conditions can fluctuate.
- Changing environmental conditions: wind and temperature can alter measurement noise in mechanical or atmospheric monitoring.
- Network traffic fluctuations: packet arrival processes can show evolving load and contention patterns, affecting measured signals.
- Mechanical systems under time-varying loads: changing operating regimes (speed, torque, or contact conditions) can produce noise with time-dependent frequency content.
- Biological or biomedical measurements: motion, posture changes, and physiological state transitions can cause evolving artifacts superimposed on measured signals.
1.4 Why non-stationarity matters in analysis and inference
Many inference and processing pipelines implicitly assume stationarity to simplify modeling and to justify particular performance guarantees. When non-stationarity is present, those simplifications break down. Consequences include: (i) degraded denoising or detection performance, (ii) incorrect confidence intervals due to mismatched noise models, (iii) poor parameter estimates in system identification, and (iv) misleading interpretations when changes in noise are mistaken for structured signal events. Time-aware representations and adaptive methods are therefore central to reliable analysis.
2 Mathematical Characterization
2.1 Time-dependent statistical descriptors
2.1.1 Mean and variance as functions of time
A common starting point models the observed process \(x(t)\) as the sum of a signal component and a noise component, with the noise part having a time-varying first and second moments. The mean level can be written as \(m(t)=\mathbb{E}[n(t)]\). Similarly, the instantaneous variance can be expressed as \(\sigma^2(t)=\mathbb{E}[(n(t)-m(t))^2]\). In many applications, a varying mean corresponds to drift or bias, while time-varying variance corresponds to amplitude modulation, bursts, or changes in operating conditions.
2.1.2 Autocorrelation and cross-correlation vs. time lag
Beyond marginal moments, dependence structure can change over time. A time-varying autocorrelation can be formulated as \(R_n(t_1,t_2)=\mathbb{E}[n(t_1)n(t_2)]\), which depends on both time instants, not only on the lag \(t_2-t_1\). In multichannel settings, cross-correlation between different sensors or modalities may also vary with time, reflecting time-dependent coupling or changing measurement geometry. Non-stationarity often manifests as a failure of correlation functions to remain shift-invariant.
2.2 Spectral representations for time-varying signals
2.2.1 Instantaneous spectrum intuition
For stationary processes, a spectrum provides a global view of how energy is distributed across frequencies. For non-stationary noise, the “instantaneous spectrum” is an intuitive but careful notion: at each time, the local frequency content may differ. While a literal instantaneous spectrum is not always well-defined in strict stochastic terms, time-localized spectral descriptions (obtained via windows or transforms) provide a practical approximation to evolving frequency structure.
2.2.2 Time-dependent power spectral density
A widely used representation is a time-indexed power spectral density (often estimated indirectly). One approach computes spectra on sliding windows, yielding \(S_n(t,f)\) or related forms that express how power at frequency \(f\) changes with time \(t\). Such representations are valuable for identifying when certain bands become dominant, when noise bandwidth expands or contracts, or when spectral peaks drift.
2.3 Random-process viewpoints
2.3.1 Non-stationary stochastic processes
Non-stationary noise can be described as a stochastic process with statistical properties that change in time. Depending on the formulation, one may allow non-stationarity in mean and variance alone, or in higher-order statistics and dependence structure as well. The level of generality influences what can be proven and what can be estimated: full non-stationarity may be too complex for simple closed-form models, encouraging structured or approximate assumptions.
2.3.2 Locally stationary approximations
A common compromise assumes that the process behaves approximately stationary over short time intervals. Under local stationarity, time-varying spectra and correlations are treated as slowly changing, enabling estimation with methods designed for stationary segments. This assumption motivates short-time transforms and adaptive procedures: the analysis window is chosen so that the noise is nearly stationary within the window but allowed to change across windows.
3 Time–Frequency Analysis Methods
3.1 Short-time Fourier transform (STFT)
3.1.1 Windowing and resolution trade-offs
The STFT computes a spectrogram by applying a window function to short segments of data and taking the Fourier transform of each segment. The window length governs a trade-off: longer windows improve frequency resolution but smear time changes; shorter windows improve temporal localization but reduce spectral sharpness. For non-stationary noise, the window should be chosen to balance the timescale over which noise statistics remain approximately constant against the need to detect temporal evolution.
3.1.2 Spectrogram interpretation for noise
A spectrogram plots estimated power versus time and frequency. Non-stationary noise often produces patterns such as time-varying broadband energy, drifting narrowband components, or intermittent bursts that appear as bright regions. For detection or denoising, one can interpret these patterns as evidence that frequency content is time dependent, guiding mask design and adaptive filtering.
3.2 Wavelet transforms
3.2.1 Multi-resolution analysis
Wavelet methods represent signals using basis functions with varying resolution across time and frequency. Typically, they provide finer time resolution at high frequencies and finer frequency resolution at low frequencies. This multi-resolution property aligns well with many non-stationary phenomena, including transients and scale-dependent structures that are not well captured by fixed-window Fourier analysis.
3.2.2 Scalograms and transient detection
The wavelet transform produces scalograms (time-scale or time-frequency-like plots) that highlight localized events. Transients in non-stationary noise may show up as brief localized ridges or clusters across scales, helping separate short-lived disturbances from longer background fluctuations.
3.3 Choi–Williams and other smoothed time–frequency distributions
Quadratic time–frequency distributions can offer high sensitivity to structure, but they may suffer from cross-term interference. The Choi–Williams distribution and related smoothed distributions reduce these artifacts through kernel smoothing, improving interpretability in some cases. For non-stationary noise, these methods can reveal evolving spectral features when conventional spectrograms are blurred, while still requiring care to avoid over-interpreting distribution artifacts.
3.4 Empirical and data-driven spectral methods
When explicit models are difficult, data-driven techniques estimate time-varying spectra from the observed data. Examples include empirical mode decomposition variants paired with Hilbert spectral analysis, or adaptive time–frequency representations that tailor basis functions to the signal. These approaches aim to follow non-stationary structure more flexibly, though they can be sensitive to parameter choices and preprocessing.
4 Modeling Approaches
4.1 Parametric models
4.1.1 ARMA with time-varying parameters
Autoregressive moving-average (ARMA) models characterize a process using finite memory and linear dynamics. To handle non-stationarity, the model coefficients can be allowed to vary with time, producing a time-varying ARMA structure. This approach can capture evolving spectral shape and correlation length, provided the parameter evolution is not too rapid and can be estimated reliably.
4.1.2 State-space and stochastic state models
State-space models represent the observation as a function of hidden (latent) states that evolve over time. Non-stationary noise can be treated as part of the evolving state dynamics, for instance with time-varying system gains or noise variances. Such formulations are compatible with filtering and smoothing methods, making them useful for online estimation and for quantifying uncertainty.
4.2 Non-parametric and semi-parametric models
4.2.1 Kernel-based time-varying estimation
Non-parametric strategies estimate time-dependent statistics directly from data without fixed functional forms. Kernel smoothing can be applied in time to estimate local mean, variance, or spectra. The bandwidth controls the amount of smoothing: too much yields lagged or overly smooth estimates; too little increases variance in the estimator.
4.2.2 Adaptive smoothing and regularization
Adaptive regularization incorporates prior assumptions such as “statistics change gradually” or “bursts are sparse.” For example, one may use penalties that discourage rapid parameter fluctuations while still allowing sharp changes when supported by the data. This can improve stability in estimation when the signal-to-noise ratio is low.
4.3 Piecewise-stationary and regime-switching models
4.3.1 Change-point detection frameworks
Piecewise-stationary models assume that the noise is stationary within segments but changes at unknown times. Change-point detection methods estimate those boundaries and assign different statistical parameters to each segment. This framework is particularly effective when non-stationarity appears as discrete regime shifts rather than smooth drift.
4.3.2 Hidden Markov model style formulations
Regime-switching models can be represented using hidden discrete states where each state corresponds to a distinct noise regime with its own parameters. Transitions between states capture time-varying behavior such as switching between low- and high-variance noise. Hidden Markov model-style approaches provide a probabilistic mechanism for segmenting and estimating regimes jointly.
5 Detection and Estimation in Non-stationary Noise
5.1 Noise characterization from observed data
5.1.1 Baseline estimation with temporal drift
When the noise mean drifts, denoising and detection often require a baseline estimate that evolves over time. Approaches include detrending using smooth functions, robust local averaging, or model-based bias estimation in a state-space framework. The goal is to remove slowly varying components without erasing genuine signal structure.
5.1.2 Outlier and burst handling strategies
Non-stationarity frequently appears as bursts or heavy-tailed disturbances. Robust statistics and burst-aware estimation help distinguish transient noise from sustained patterns. Methods may incorporate outlier-resistant estimators, time-varying thresholds, or mixture models that explicitly represent “normal” and “burst” noise components.
5.2 Filtering and denoising strategies
5.2.1 Adaptive filters (conceptual overview)
Adaptive filtering adjusts its parameters based on incoming data to track changing noise properties. In non-stationary settings, adaptation aims to improve performance when spectral shape, variance, or correlation structure changes. Conceptually, such filters rely on measuring error or residual statistics to update the noise model or filter coefficients.
5.2.2 Time–frequency masking
Time–frequency masking modifies the spectrogram or wavelet coefficients to suppress noise-dominated regions while retaining components associated with the desired signal. Because non-stationary noise is time-localized in its frequency contributions, masks can vary across time, enabling selective attenuation. Mask design often depends on estimates of local signal-to-noise ratio or on learned or adaptive thresholds.
5.3 Performance metrics
5.3.1 SNR vs. perceptual/functional quality
A common metric is signal-to-noise ratio (SNR), but it may not correlate with perceived quality or task performance. In audio applications, perceptual measures and intelligibility metrics can better reflect usability. In other domains, downstream functional metrics (e.g., detection rate, tracking accuracy) quantify whether improvements in noise suppression translate to better interpretation.
5.3.2 Robustness to model mismatch
Performance can deteriorate if the assumed noise model does not reflect the actual non-stationarity. Robustness is therefore assessed by testing under different noise regimes, varying degrees of drift, or mismatch in burst statistics. Techniques that avoid overconfident parameter assumptions, or that adapt online, often exhibit better generalization.
6 Practical Considerations
6.1 Sampling, aliasing, and discretization effects
Discrete-time measurements approximate continuous processes. Sampling rate can cause aliasing, distorting observed frequency content and complicating time–frequency interpretation. Additionally, discretization effects influence variance estimates and autocorrelation calculations, particularly when noise varies rapidly relative to the sampling interval.
6.2 Boundary effects in time–frequency methods
Windowed transforms such as STFT suffer from edge artifacts because windows extend beyond the available data. Boundary handling—through padding, tapering, or modified windowing strategies—reduces spurious spectral leakage and prevents misleading interpretations near the start and end of records.
6.3 Computational cost and real-time constraints
High-resolution time–frequency analysis can be computationally intensive, especially for large window sizes, fine frequency grids, or multiple channels. Real-time systems must balance latency, throughput, and estimator accuracy. Often, the practical solution is to limit resolution, adopt efficient algorithms, or use approximate methods that meet timing requirements.
6.4 Experimental design for identifying non-stationarity
Good practice includes collecting data under diverse operating conditions, using repeated measurements to separate random variation from systematic time changes, and validating results with held-out segments. For diagnostic purposes, one typically compares statistics across intervals, checks whether local stationarity assumptions hold, and tests sensitivity to analysis parameters such as window length or smoothing bandwidth.
7 Applications
7.1 Signal processing and communications
In communications, non-stationary noise can arise from interference that changes with network activity, channel conditions that vary over time, or hardware behavior such as time-varying amplifier noise. Equalization, detection thresholds, and decoding performance may require time-adaptive models to maintain reliability.
7.2 Audio and speech enhancement
Speech enhancement systems often assume that noise characteristics are roughly stable, but real-world backgrounds can fluctuate due to traffic, crowd movement, or ventilation systems. Non-stationary modeling supports adaptive noise estimation, time-varying filters, and masks that track changing spectral patterns, improving intelligibility and reducing artifacts.
7.3 Vibration and mechanical monitoring
Monitoring systems face noise shaped by changing operating regimes, varying contact conditions, and load-dependent excitation. Time-varying spectra and regime-aware detection can help distinguish true mechanical events from changing baseline noise, supporting fault diagnosis and condition monitoring.
7.4 Biomedical signal analysis (time-varying artifacts)
Biomedical recordings can exhibit time-varying artifacts from motion, electrode impedance changes, or varying physiological state. Treating these disturbances as non-stationary noise can improve artifact suppression and preserve clinically relevant signal components.
8 Common Pitfalls and Diagnostics
8.1 Misidentifying non-stationarity as signal structure
A frequent error is interpreting noise evolution as part of the signal of interest. Without careful controls or baseline comparison, evolving variance or drifting spectral components may be mistaken for structured events, leading to incorrect conclusions.
8.2 Overfitting to time-varying noise characteristics
Flexible models can capture noise patterns too well, inadvertently removing or distorting the signal. Overfitting is especially likely when model complexity is high relative to data length, or when training and testing conditions differ.
8.3 Diagnosing stationarity assumptions with tests
Diagnostics aim to determine whether a stationarity assumption is reasonable. One can compare statistics across segments, test constancy of variance and correlation, or evaluate whether time-local spectral estimates are consistent with a single global spectrum. When deviations are significant, time-aware methods become necessary.
8.4 Choosing window sizes and model complexity
In time–frequency analysis, selecting window length affects both interpretability and estimator bias. In modeling, choosing how quickly parameters are allowed to change affects stability. A practical approach involves sweeping candidate parameters, examining sensitivity, and selecting values that reflect the expected timescale of non-stationarity rather than arbitrary defaults.
9 Further Reading and Resources
9.1 Key textbooks and surveys
Foundational resources include standard texts on stochastic processes, time–frequency analysis, and adaptive signal processing, along with survey literature on non-stationary spectral estimation. Readers typically benefit from first mastering stationary theory and then studying extensions such as local stationarity, wavelet methods, and time-varying models.
9.2 Benchmark datasets and tooling (high level)
Benchmarking often uses public recordings in audio enhancement, vibration monitoring, and sensor datasets where noise varies across time. Tooling commonly includes time–frequency transform libraries, wavelet toolkits, and experimental pipelines that support parameter sweeps and metric evaluation for denoising and detection tasks.
9.3 Suggested starter experiments and exercises
Useful exercises include: (i) generating synthetic non-stationary noise with controlled drift or bursts and observing how spectrograms change, (ii) comparing STFT and wavelet representations on the same data to evaluate resolution trade-offs, (iii) testing change-point detection on piecewise-stationary sequences, and (iv) implementing adaptive filtering or masking with tunable smoothing to measure robustness across different non-stationary regimes.