Fast Fourier Transform (FFT)

The FFT algorithm decomposes a complex, noisy vibration signal—a waveform of acceleration over time—into its constituent sinusoidal frequencies. This transformation reveals the amplitude and phase of each frequency component present in the original signal. In predictive maintenance, this is critical because specific mechanical faults generate energy at characteristic frequencies. For example, a bearing defect will produce a distinct frequency peak related to its Ball Pass Frequency Inner Race (BPFI) , which is invisible in the raw time waveform.

The 'Fast' in FFT is its defining advantage over the Discrete Fourier Transform (DFT). A DFT computes frequency components with a computational complexity of O(N²), making it prohibitively slow for real-time, high-sample-rate sensor data. The Cooley-Tukey FFT algorithm reduces this to O(N log N) by recursively breaking down a DFT into smaller DFTs. This efficiency enables edge devices to perform continuous spectral analysis on streaming vibration data without overwhelming limited processing resources.

Once a time-domain signal is converted to a frequency spectrum, distinct failure modes become identifiable as peaks at specific frequencies:

• Imbalance: High amplitude peak at the shaft's running speed (1x RPM).
• Misalignment: Peaks at 1x, 2x, and sometimes 3x RPM.
• Bearing Defects: Non-synchronous peaks at characteristic frequencies (BPFO, BPFI, BSF, FTF) calculated from bearing geometry.
• Gear Mesh Faults: Peaks at the gear mesh frequency and sidebands spaced at the shaft speed.

A practical challenge in applying the FFT to finite, real-world signals is spectral leakage. When the sampled signal does not contain an integer number of cycles, energy 'leaks' into adjacent frequency bins, smearing the spectrum. To mitigate this, a window function (e.g., Hanning, Hamming, Blackman) is applied to the time-domain data before the FFT. The window tapers the signal to zero at its boundaries, reducing leakage and improving the accuracy of amplitude measurements at the cost of slightly reduced frequency resolution.

The diagnostic power of an FFT is governed by its parameters:

• Frequency Resolution (Δf): The spacing between frequency bins, determined by Δf = Sample Rate / N. A finer resolution is needed to separate closely spaced fault frequencies.
• Bandwidth (Fmax): The maximum frequency that can be analyzed, defined by the Nyquist Theorem as half the sample rate. To detect high-frequency bearing tones, a high sample rate is required.
• Lines of Resolution (LOR): The number of data points in the spectrum (N/2), dictating the detail of the analysis.

For early-stage bearing and gear faults, the low-energy impact signatures are often buried in high-energy, low-frequency noise from shafts and blades. Envelope analysis uses the FFT in a two-stage process: first, a band-pass filter isolates the high-frequency structural resonance excited by the repetitive impacts. The signal is then demodulated (enveloped), and a second FFT is performed on this envelope. This reveals the low-frequency repetition rate of the impacts, making the fault frequency clearly visible.

The measurement and interpretation of machine oscillations to detect imbalances, misalignments, and bearing faults. FFT converts raw time-domain vibration waveforms into the frequency domain, allowing analysts to match specific frequency peaks to known fault signatures such as inner race or outer race defects.

• Identifies 1x RPM peaks for unbalance
• Detects harmonics for looseness
• Reveals bearing defect frequencies (BPFO, BPFI)

A non-intrusive monitoring technique that analyzes the electrical current supply to a motor. FFT is applied to the current waveform to detect broken rotor bars, eccentricity, and winding faults by identifying sideband frequencies around the line frequency.

• Detects pole-pass frequency sidebands
• Non-invasive, no sensor mounting required
• Effective for submerged or inaccessible motors

The process of extracting statistical attributes from FFT spectra to train machine learning models. Key features include spectral centroid, kurtosis, skewness, and energy in specific frequency bands. These engineered features serve as inputs to Random Forest and Gradient Boosting classifiers for automated fault diagnosis.

• Spectral band energy ratios
• Cepstral coefficients for gearbox analysis
• Envelope spectrum features for bearing faults

Advanced techniques like Short-Time Fourier Transform (STFT) and Wavelet Transform that extend FFT to handle non-stationary signals. Unlike standard FFT, which assumes a steady-state signal, these methods reveal how frequency content evolves over time, critical for analyzing transient events and varying-speed machinery.

• STFT uses sliding windows for spectrograms
• Wavelets provide multi-resolution analysis
• Essential for startup/shutdown transient analysis

A two-stage signal processing technique that combines band-pass filtering, rectification, and FFT to extract repetitive impacts from noisy vibration signals. It is the gold standard for early detection of rolling-element bearing faults by isolating high-frequency resonance excited by metal-to-metal impacts.

• Demodulates the bearing resonance frequency
• Reveals fault frequencies hidden in noise
• Detects faults months before catastrophic failure

Defined as the inverse FFT of the logarithm of a signal's spectrum, the cepstrum separates periodic spectral patterns from the excitation source. In gearbox diagnostics, it decouples the gear mesh frequency from sideband families, enabling precise identification of gear wear and tooth damage.

• Detects harmonic families and sideband spacing
• Robust to transmission path effects
• Quantifies gear fault severity via quefrency peaks