FFT Spectrum Analysis: How to Read Vibration Spectra
Key Takeaways
- FFT transforms a time-domain vibration signal into a frequency-domain spectrum. Each peak in the spectrum corresponds to a specific mechanical event, not just overall vibration level.
- Sampling rate, time record length, and window function selection directly determine what faults are detectable and how accurately amplitudes are measured.
- Common fault signatures are well-defined: imbalance dominates at 1x running speed, misalignment at 2x, looseness across multiple harmonics, and bearing defects at non-synchronous calculated frequencies.
- Waterfall plots and spectrum averaging extend single-spectrum analysis to trend detection and noise reduction.
- FFT analysis works best on steady-state signals. Transient events, variable-speed machines, and very early-stage defects require supplemental analysis methods.
What Is FFT Spectrum Analysis?
Vibration signals from rotating machinery are time-domain waveforms: amplitude measured against time. These waveforms are difficult to interpret because every bearing, gear mesh, imbalance, and structural resonance in a machine produces vibration simultaneously, and all of them combine into a single complex waveform at the sensor.
FFT spectrum analysis solves this by decomposing that combined signal into its individual frequency components. The Fast Fourier Transform is a computational algorithm that efficiently calculates the Discrete Fourier Transform, expressing a time-domain signal as a sum of sine waves at different frequencies and amplitudes. The output is a frequency spectrum: a graph with frequency on the horizontal axis and vibration amplitude on the vertical axis. Each peak in the spectrum corresponds to a specific repeating mechanical event in the machine.
This frequency-domain view is the foundation of vibration analysis in industrial maintenance. When an analyst sees a dominant peak at exactly twice the running speed, they know misalignment is the likely cause, not because the waveform looks unusual, but because the physics of misalignment produces that specific forcing frequency. FFT analysis gives that level of diagnostic specificity.
The Mathematics: From Time Domain to Frequency Domain
The Fourier Transform is based on the principle that any periodic signal can be represented as a sum of sinusoidal components at different frequencies, amplitudes, and phases. For a discrete, digitally sampled signal, the Discrete Fourier Transform (DFT) performs this decomposition. The FFT is an optimized algorithm that computes the DFT far faster by exploiting symmetry in the calculation, making real-time spectrum analysis practical.
The key inputs that determine what the FFT can resolve are:
- Sampling rate (Fs): How many data points are captured per second. Determines the highest frequency visible in the spectrum.
- Number of lines (N): The number of frequency bins in the resulting spectrum, typically 400, 800, 1600, or 3200. More lines produce finer frequency resolution but require longer measurement time.
- Time record length (T): The duration of the signal captured. T = N / Fs. Longer records provide better low-frequency resolution.
- Frequency resolution (df): The spacing between adjacent spectrum lines. df = 1 / T = Fs / N. A 1-second record gives 1 Hz resolution; a 10-second record gives 0.1 Hz resolution.
Nyquist Theorem and Sampling Requirements
The Nyquist theorem states that to accurately reconstruct a frequency component, the sampling rate must be at least twice that frequency. If the highest fault frequency of interest is 10,000 Hz, the minimum required sampling rate is 20,000 samples per second. Frequencies above half the sampling rate (the Nyquist frequency) cannot be represented and fold back into the spectrum as aliased artifacts.
In practice, analysts apply an anti-aliasing low-pass filter before the analog-to-digital converter to remove frequency content above the Nyquist limit before it can alias. Because real filters have a gradual rolloff rather than a sharp cutoff, a common convention is to use a sampling rate of 2.56 times the maximum frequency of interest. This provides clean, alias-free data across the full analysis range.
For vibration monitoring systems that measure high-frequency bearing defects, this means accelerometer sensors with flat response well above the target analysis range, digitizers sampling at high rates, and careful selection of the analysis bandwidth for each machine type.
Amplitude Units: Velocity, Acceleration, and Displacement
Vibration amplitude in a spectrum can be expressed in three different physical units, and the choice matters for what faults are most visible:
| Unit | What it measures | Best for detecting | Typical frequency range |
|---|---|---|---|
| Displacement (mils or microns) | Physical movement of the shaft or housing | Low-frequency imbalance, looseness, slow-speed bearing wear | 10 Hz to 1,000 Hz |
| Velocity (mm/s or in/s) | Rate of change of displacement | Mid-range faults: imbalance, misalignment, looseness across most machines | 10 Hz to 10,000 Hz |
| Acceleration (g or m/s²) | Rate of change of velocity | High-frequency bearing defects, gear mesh frequencies, cavitation | 1,000 Hz to 20,000 Hz+ |
ISO 10816 and ISO 20816 severity standards are expressed in velocity (mm/s RMS) because velocity provides a relatively flat sensitivity response across the frequency range most relevant to general machine health. For high-speed or high-frequency applications such as rolling element bearing diagnostics, acceleration spectra are preferred because bearing defect frequencies produce impact energy that integrates favorably in acceleration units.
Key Spectral Features and What They Mean
1x Running Speed (Synchronous Component)
The 1x component occurs at a frequency equal to one times the shaft rotational speed in Hertz. For a motor running at 1,800 RPM, the 1x frequency is 30 Hz. A dominant 1x peak is the primary indicator of mass imbalance. The amplitude of the 1x peak at the bearing housing is roughly proportional to the magnitude of the imbalance force.
Harmonics
Harmonics are integer multiples of a fundamental frequency: 2x, 3x, 4x, and so on. Each fault type produces a characteristic harmonic pattern. Misalignment typically produces elevated 2x alongside 1x, sometimes extending to 3x in severe cases. Mechanical looseness produces a spectrum rich in multiple harmonics, often extending to 5x, 10x, or higher. Gear mesh frequencies appear at the product of tooth count and shaft speed, with sidebands spaced at running speed on either side.
Sub-synchronous Components
Sub-synchronous peaks appear below the 1x running speed frequency. Oil whirl in journal bearings characteristically appears at approximately 0.43x to 0.48x running speed. Partial rubs, misaligned couplings, and rotor instability can also produce sub-synchronous content. Because these frequencies fall below the dominant 1x peak, they are easy to miss when only monitoring overall amplitude levels.
Non-synchronous Components
Non-synchronous peaks are frequencies that are not simple multiples or fractions of running speed. Rolling element bearing defect frequencies fall into this category. Each bearing generates four characteristic frequencies based on its geometry and the shaft speed: Ball Pass Frequency Outer race (BPFO), Ball Pass Frequency Inner race (BPFI), Ball Spin Frequency (BSF), and Fundamental Train Frequency (FTF). These are computed from the bearing's pitch diameter, ball diameter, number of balls, and contact angle.
Common Fault Signatures: Reference Table
| Fault Type | Dominant Frequencies | Supporting Indicators | Notes |
|---|---|---|---|
| Mass Imbalance | 1x dominant | High radial vibration, 90-degree phase shift between horizontal and vertical | Amplitude increases with speed squared |
| Angular Misalignment | 1x and 2x axial | High axial component, 180-degree phase across coupling axially | Pure angular misalignment shows mainly axial 1x and 2x |
| Parallel Misalignment | 2x dominant radially | High radial 2x, 180-degree radial phase across coupling | In practice, misalignment is usually a combination of both types |
| Mechanical Looseness | Multiple harmonics of 1x (2x, 3x, up to 10x+) | Sub-harmonics (0.5x) possible, noisy floor | Often triggered or amplified by another fault (imbalance, misalignment) |
| Rolling Element Bearing Defect | BPFO, BPFI, BSF, or FTF (non-synchronous) | Sidebands at running speed, raised noise floor, elevated high-frequency content | Early stages may only show in demodulated (envelope) spectrum |
| Gear Mesh Fault | Gear mesh frequency = teeth count x shaft speed | Sidebands at shaft speed, ghost frequencies from manufacturing | Tooth damage produces sidebands; wear produces elevated mesh amplitude |
| Oil Whirl (Journal Bearing) | 0.43x to 0.48x running speed | Sub-synchronous peak tracking with speed | Can progress to oil whip locked at rotor critical speed |
Window Functions
An FFT algorithm assumes the captured time record is one complete cycle of a periodic signal. In practice, the signal is not perfectly periodic within the measurement window, so the signal begins and ends at different amplitude values. This discontinuity at the edges of the record creates spectral leakage: energy from a true single-frequency peak spreads into adjacent frequency bins, raising the noise floor and masking nearby peaks.
Window functions address this by multiplying the time record by a weighting curve that smoothly tapers the signal to zero at both ends, eliminating the discontinuity. The two most common windows in machinery diagnostics are:
- Hanning window: Good frequency resolution and reasonable amplitude accuracy. The standard choice for most machinery fault diagnosis work. Reduces leakage significantly while maintaining the ability to distinguish peaks that are close in frequency.
- Flat-top window: Excellent amplitude accuracy (within 0.1 dB), at the cost of wider peaks that may merge if two frequencies are close together. Used when the primary goal is accurate amplitude measurement for alarm setting, trending, or calibration verification rather than frequency resolution.
A rectangular (uniform) window applies no weighting and is used only when the signal is known to be exactly periodic within the record, such as during precision tonal testing. For general rotating machinery diagnostics, the Hanning window is the correct default.
Spectrum Averaging
A single FFT spectrum contains random noise contributions on top of the deterministic signal from machine forces. Spectrum averaging computes multiple consecutive spectra and averages the amplitude at each frequency bin. Because deterministic signals (imbalance, bearing defects, gear mesh) are coherent and repeat each revolution, they reinforce through averaging. Random noise, being incoherent, averages toward zero at a rate proportional to the square root of the number of averages.
Common averaging types used in industrial vibration analysis include:
- Linear (synchronous) averaging: Simple mean of N spectra. Reduces random noise by factor sqrt(N). Most common for general diagnostics.
- Peak hold averaging: Retains the maximum amplitude seen at each frequency bin across all spectra. Useful for capturing intermittent events or transients that might be missed in a single spectrum.
- Exponential averaging: Weights recent spectra more heavily. Provides a continuously updated view that responds gradually to real changes in machine condition. Used in online monitoring systems.
For routine route-based data collection, 4 to 8 linear averages is a typical practical balance between measurement time and noise reduction.
Waterfall Plots
A waterfall plot stacks a series of FFT spectra along a third axis, creating a three-dimensional representation of how the frequency spectrum changes over time or over a range of machine speeds. Each individual spectrum forms one "slice" of the waterfall, displayed as a ridge in the plot.
Waterfall plots are particularly powerful for two diagnostic applications:
- Run-up and coast-down analysis: As shaft speed increases during startup, running-speed harmonics (1x, 2x, 3x) sweep upward in frequency. Any frequency that remains fixed as speed changes is a structural resonance, not a forced vibration. This separation is impossible to see in spectra taken at a single speed but immediately visible in a waterfall plot.
- Long-term trending: Waterfall plots from periodic route measurements or continuous online monitoring show visually when new frequency components first appear, when existing components grow, and whether growth is gradual or sudden. This is valuable input for failure prediction models and scheduling maintenance interventions at the right time.
Worked Example: Interpreting a Pump Spectrum
Consider a centrifugal pump running at 1,800 RPM (30 Hz) with a 6-vane impeller. The expected spectrum features are:
- 1x at 30 Hz: Always present. Amplitude determines whether imbalance is within tolerance.
- Vane pass frequency at 180 Hz (6 x 30): Normal operating component. An increase in this peak often indicates cavitation or recirculation.
- 2x at 60 Hz: Minor 2x below the 1x amplitude is acceptable. If 2x approaches or exceeds 1x, investigate misalignment.
- Bearing defect frequencies: Calculated from the bearing specification and shaft speed, typically falling between 100 Hz and 500 Hz for a bearing in this speed range. Any peaks at calculated BPFO, BPFI, BSF, or FTF frequencies warrant investigation.
Suppose the analyst observes: 1x at 30 Hz with 4.5 mm/s, 2x at 60 Hz with 4.1 mm/s (nearly equal to 1x), and a cluster of peaks near 220 Hz (non-synchronous, close to calculated BPFI). The pattern suggests combined misalignment (elevated 2x approaching 1x) and an early inner race bearing defect. The correct response is to schedule an alignment check and bearing inspection at the next planned maintenance window, not to wait for vibration severity to breach the overall alarm threshold, which may not trigger until the fault has progressed significantly.
This kind of early, specific diagnosis is the core value of predictive maintenance programs built on FFT spectrum analysis.
Limitations of FFT Spectrum Analysis
FFT spectrum analysis is a powerful tool, but it has real constraints that analysts must understand to avoid misdiagnosis:
- Assumes stationarity: The standard FFT assumes the signal statistics do not change during the measurement window. Variable-speed machines, machines accelerating through resonance, or signals with impacts and transients violate this assumption. Alternative methods such as the Short-Time Fourier Transform (STFT), wavelet analysis, or order tracking are better suited to these cases.
- Fixed time-frequency tradeoff: Better frequency resolution requires a longer time record. This means very low-frequency phenomena (sub-1 Hz components in slow-speed machinery) require long measurement windows that are impractical in some field conditions.
- Early-stage bearing faults: Very early bearing defects produce low-amplitude, high-frequency impacts that may not be visible in a standard velocity or acceleration spectrum. Envelope analysis (demodulation), spike energy, or high-frequency resonance excitation techniques are needed to detect faults before they become visible in the standard FFT spectrum.
- Requires analyst expertise: Reading a spectrum correctly requires knowledge of the machine's rotational speed, bearing geometry, gear ratios, number of impeller vanes, and normal baseline spectra. Without this context, peaks can be misidentified or overlooked entirely.
- Aliasing if setup is incorrect: An incorrectly set sampling rate or a missing anti-aliasing filter will fold high-frequency content back into the spectrum as ghost peaks, potentially leading to false diagnoses.
These limitations are why experienced reliability programs pair FFT spectrum analysis with complementary techniques: envelope analysis for early bearing detection, phase analysis for confirming misalignment and looseness, and oil analysis for gearbox and journal bearing health. No single technique covers every failure mode across every machine type.
FFT Analysis in Continuous Monitoring Systems
Condition monitoring platforms that capture data continuously rather than on periodic routes can compute FFT spectra on every measurement cycle, storing spectra for trend analysis and triggering alerts when new frequency components appear or when known fault frequencies exceed thresholds.
This continuous approach closes a critical gap in route-based monitoring. A bearing that develops a defect three days after a technician visits the machine will not be detected until the next route, which may be weeks away. Continuous spectral monitoring captures the onset of fault frequencies as they first develop, giving maintenance teams the lead time needed to plan repairs before failure occurs.
The practical application in modern asset condition monitoring systems is automated spectrum comparison against stored baselines and fault frequency templates. Rather than requiring an analyst to review every spectrum manually, the system flags spectra that deviate from baseline, presents the specific frequency components of concern, and ranks alerts by severity. The analyst's role shifts from routine spectrum inspection to investigating flagged exceptions, which is a far more efficient use of expertise.
Frequently Asked Questions
What is FFT spectrum analysis in vibration monitoring?
FFT spectrum analysis converts a time-domain vibration signal into a frequency-domain spectrum using the Fast Fourier Transform algorithm. The resulting spectrum shows how vibration energy is distributed across individual frequencies, allowing analysts to identify specific mechanical faults by their characteristic frequency patterns rather than reacting to an overall amplitude threshold.
What sampling rate is required for FFT analysis?
The Nyquist theorem requires a sampling rate at least twice the highest frequency of interest. To analyze bearing defect frequencies up to 10,000 Hz, for example, the sensor must sample at a minimum of 20,000 samples per second. In practice, analysts commonly use a sampling rate 2.56 times the maximum frequency of interest to include a small safety margin and account for anti-aliasing filter rolloff.
What does the 1x component in a vibration spectrum mean?
The 1x component, also called the synchronous or running speed frequency, represents vibration at exactly one times the rotational speed of the shaft. A dominant 1x peak is the classic signature of mass imbalance. It can also appear during misalignment or looseness, so analysts always examine the full spectrum and the phase relationship between measurement points before diagnosing based on a single spectral peak.
How do bearing fault frequencies appear in an FFT spectrum?
Rolling element bearing defects generate non-synchronous frequency components that are not simple multiples of running speed. The four characteristic frequencies are BPFO, BPFI, BSF, and FTF, all calculated from bearing geometry and shaft speed. Early bearing faults also produce raised noise floor and sidebands spaced at running speed around the bearing defect frequency.
What is the difference between a Hanning window and a flat-top window?
A Hanning window reduces spectral leakage and is the standard choice for machinery fault diagnosis because it provides good frequency resolution. A flat-top window provides better amplitude accuracy but wider peak shapes, making it the preferred choice when precise amplitude measurement matters more than resolving closely spaced peaks.
What is a waterfall plot and when is it used?
A waterfall plot stacks multiple FFT spectra collected at different times or machine speeds along a third axis. It is particularly useful during run-up or coast-down testing to separate resonance frequencies from forced vibration, and for trending to show when new frequency components first appeared over time.
What are the main limitations of FFT spectrum analysis?
FFT analysis assumes the signal is stationary during the measurement window. It performs poorly on transient events, variable-speed machines, and very early-stage bearing defects. Resolution is fixed at the inverse of the time record length. Very early bearing faults may not be visible in a standard spectrum and require envelope analysis or other supplemental techniques.
The Bottom Line
FFT spectrum analysis is the most powerful diagnostic tool available for rotating machinery. It converts a raw vibration waveform into a precise map of forces acting inside a machine, with each peak traceable to a specific physical event. Imbalance, misalignment, looseness, bearing defects, and gear faults each produce distinct frequency signatures that an analyst can identify before the machine reaches a failure-critical state.
The technique is not passive reading. It requires correct setup: adequate sampling rate to capture the fault frequencies of interest, appropriate window function, sufficient averaging to reduce noise, and baseline spectra to compare against. Analysts who understand the physics behind each fault signature, rather than simply watching amplitude thresholds, extract far more diagnostic value from spectrum data.
When deployed in a continuous monitoring platform, FFT analysis shifts from a periodic diagnostic activity to an ongoing early-warning system. Faults that would go undetected between monthly route visits become visible at the moment they begin generating their characteristic frequencies, giving maintenance teams days or weeks of advance notice to plan corrective action safely and cost-effectively.
See How Tractian Analyzes Machine Vibration
Tractian's condition monitoring platform captures continuous vibration data, computes FFT spectra automatically, and alerts your team when fault frequencies exceed thresholds. No manual spectrum review required.
See How Tractian Analyzes Machine VibrationRelated terms
Maintenance KPI: Definition
Maintenance KPIs are metrics used to measure maintenance program performance. Learn the most important KPIs, formulas, benchmarks, and how to set targets by maturity stage.
Maintenance Logs: Definition
Maintenance logs are chronological records of all maintenance activities performed on equipment. Learn what they contain, types, compliance value, and how a CMMS automates logging.
Asset Performance Management: Definition
Asset Performance Management (APM) is a strategy that combines real-time monitoring, data analytics, and predictive maintenance to maximize the reliabil...
Asset Inventory Management: Definition
Asset inventory management is the systematic process of tracking, recording, and maintaining an organization's physical and digital assets throughout th...
Capacity Utilization: Definition
Capacity utilization is the percentage of a facility's or machine's maximum possible output that is actually being produced over a given period, calcula...