VECTORALAB EDGE
DSP Pipeline & Feature Extraction
2048-point FFT — 32 statistical & spectral features in under 8 ms
DSP Pipeline
On-Device DSP Pipeline
All signal processing executes on the STM32H750's Cortex-M7 FPU using CMSIS-DSP routines. The pipeline transforms raw vibration samples into a compact 32-element feature vector in under 8 ms per frame.
01Hanning Window2048-point window applied to reduce spectral leakage. CMSIS-DSP arm_mult_f32.
02FFT (2048-pt)arm_rfft_fast_f32 — real FFT in ~2.1 ms. Output: 1025 complex bins.
03Magnitudearm_cmplx_mag_f32 converts complex bins to magnitude spectrum.
04Band EnergySpectral energy computed in 8 sub-bands covering 0–6 kHz range.
05Time-Domain StatsRMS, Kurtosis, Crest Factor, Peak-to-Peak, Skewness per axis.
06Feature Assembly32 features assembled into float32[32] vector — input to TinyML model.
Feature Extraction
32-Dimensional Feature Vector
Each 2048-sample frame produces a 32-element float32 feature vector combining time-domain statistics and spectral sub-band energies. This vector is the sole input to the TinyML inference engine.
×3 axes
RMS
Root Mean Square — overall vibration energy indicator.
×3 axes
Kurtosis
Fourth-order statistical moment — impulsive fault sensitivity.
×3 axes
Crest Factor
Peak / RMS ratio — early bearing defect indicator.
×3 axes
Skewness
Third-order moment — signal asymmetry indicator.
×3 axes
Peak-to-Peak
Max amplitude range per frame.
Energy
Band Energy 0–750 Hz
Low-frequency structural resonance band.
Energy
Band Energy 750 Hz–1.5 kHz
Gear mesh fundamental band.
Energy
Band Energy 1.5–3 kHz
Bearing outer-race fault frequency band.
Energy
Band Energy 3–6 kHz
High-frequency impulsive fault band.
Hz
Spectral Centroid
Frequency center-of-mass of the spectrum.
Hz
Dominant Frequency
Peak magnitude frequency in 0–6 kHz range.
Scalar
Spectral Entropy
Frequency-domain disorder indicator.