Beyond the Spectrum: Why Time-Series Feature Extraction Outperforms FFT-Based Approaches in Fault Classification

There is a comfortable orthodoxy in the machine condition monitoring field: transform your vibration signal into the frequency domain, extract spectral features, feed them to a classifier. The approach has intuitive appeal. Rotating machinery generates periodic forces at predictable frequencies. The FFT makes those frequencies visible. The logic seems airtight.

It is not airtight. It is a convention that has calcified into an assumption, and that assumption is costing engineers diagnostic accuracy they do not know they are losing.

How the FFT Became the Default—and Why That Default Persists

The FFT's dominance in machinery diagnostics traces back to the 1970s and 1980s, when digital signal analyzers first made real-time spectral analysis practical on the shop floor. Vibration analysts learned to read spectra the way cardiologists read electrocardiograms—as a rich, interpretable representation of internal machine state. Defect frequencies for common fault modes were catalogued. Sidebands, harmonics, and sub-harmonics became diagnostic vocabulary.

This tradition predates machine learning by decades. When data scientists began applying classification algorithms to condition monitoring in the 2010s, they inherited the frequency-domain paradigm wholesale. Feature engineering pipelines extracted RMS values in spectral bands, peak amplitudes at defect frequencies, and cepstral coefficients—all computed from the FFT. The assumption was that if experienced vibration analysts read spectra, machine learning models should be trained on spectra.

The problem with inheriting a paradigm is that you also inherit its blind spots.

What the FFT Structurally Cannot Tell You

The FFT is a global transform. It integrates energy contributions across the entire analysis window to produce a single spectral estimate. This is simultaneously its power and its limitation. For stationary signals—signals whose statistical properties do not change over time—the FFT is an efficient and complete representation.

Fault signatures in rotating machinery are frequently non-stationary. A bearing with a developing spall does not produce a constant-amplitude sinusoid at the ball pass frequency. It produces a series of impulses, irregularly spaced in early degradation stages, each impulse exciting a broadband transient that decays over microseconds. The FFT of such a signal produces a spectrum, but that spectrum conflates the impulsive transient energy with the background vibration in a way that obscures the fault's actual character.

Consider what information the FFT discards by construction. It discards the temporal ordering of events within the analysis window. It discards the shape of individual impulses. It discards the variation in inter-event timing that distinguishes a developing fault from a fully developed one. It discards the relationship between the amplitude of one event and the next. All of this information exists in the time-domain waveform. None of it survives the transform.

Time-Domain Features That Carry Diagnostic Weight

The statistical characterization of a time-domain waveform is considerably richer than it might initially appear. Beyond the elementary metrics—RMS, peak, crest factor—there is a family of higher-order statistics that are specifically sensitive to the kinds of non-Gaussian, impulsive behavior that fault conditions produce.

Kurtosis is the most widely recognized of these. A healthy bearing vibration signal approximates a Gaussian distribution; its kurtosis is close to 3. As surface damage develops and impulses appear, the distribution develops heavy tails and kurtosis rises, sometimes dramatically. Kurtosis is sensitive to impulsive content in a way that RMS is not, because it weights the fourth power of deviations from the mean, making it disproportionately responsive to high-amplitude, short-duration events.

Skewness captures asymmetry in the amplitude distribution, which can indicate directional loading effects or asymmetric fault geometry. Permutation entropy quantifies the temporal complexity of the signal—its tendency toward ordered versus disordered amplitude sequences—and is sensitive to changes in signal dynamics that precede statistically detectable amplitude shifts. Hjorth parameters, originally developed for EEG analysis, characterize signal activity, mobility, and complexity in ways that map onto machinery degradation progression.

In power quality monitoring, the case is equally compelling. Voltage sag events, transient overvoltages, and harmonic injection from nonlinear loads all have characteristic time-domain morphologies. A voltage sag has a specific onset slope, depth, and recovery profile. These morphological details are preserved in the time-domain waveform and are largely destroyed when the signal is reduced to a spectral representation.

Empirical Evidence from Bearing Diagnostics

The Case Western Reserve University bearing dataset has become a standard benchmark for condition monitoring research, and it offers instructive lessons about feature domain performance. Studies comparing FFT-derived spectral features against time-domain statistical features on this dataset consistently find that the time-domain features—particularly kurtosis, crest factor, and higher-order statistics—achieve comparable or superior classification accuracy for early-stage fault detection, while requiring substantially less computational overhead.

More revealing are the failure cases. Spectral feature sets tend to fail gracefully for faults at expected frequencies but fail catastrophically when fault signatures do not align with predicted defect frequencies—which occurs more often in practice than textbook treatments suggest. Load variation, speed fluctuation, and multi-component interaction all shift defect frequencies away from their nominal values. Time-domain statistical features are largely insensitive to these shifts because they characterize the waveform's distributional properties rather than its spectral content.

For variable-speed machinery—a category that includes most modern drives with variable frequency control—this distinction is decisive. The FFT assumes a fixed relationship between time and frequency. Variable-speed operation violates that assumption continuously.

Combining Domains Intelligently

This is not an argument for abandoning the FFT. It is an argument for treating the FFT as one tool among several rather than as the self-evident starting point. The most robust condition monitoring models combine features from multiple representational domains: time-domain statistics, spectral features where the stationarity assumption is defensible, time-frequency representations such as the short-time Fourier transform or wavelet packet decomposition for non-stationary signals, and derived features such as envelope statistics computed after bandpass filtering.

The key discipline is intentionality. Each feature should be selected because it carries specific diagnostic information about a specific fault mode, not because it is what the FFT pipeline produces by default. Feature selection should be guided by an understanding of the physics of the fault being detected, not by the convenience of the transform being applied.

A Practical Recommendation for Model Development

Engineers building or auditing predictive maintenance models should conduct a systematic feature importance analysis that includes both spectral and time-domain candidates. In many cases, the results will be surprising. Features that are rarely discussed in vibration analysis textbooks—permutation entropy, sample entropy, Hjorth complexity—frequently rank among the most predictive for early-stage fault detection.

The FFT earned its place in the diagnostic toolkit through decades of practical utility. But the machinery of the 21st century—variable-speed drives, composite structures, tightly integrated multi-physics systems—generates fault signatures that the FFT was not designed to characterize. The engineers who recognize this limitation and expand their feature vocabulary accordingly will build models that detect failures their competitors' systems miss entirely.