SKF 6205 deep-groove ball bearing, 9 rolling elements. Raise the outer-race defect and watch where it shows up: barely in the raw spectrum, unmistakably in the envelope spectrum.
Normal
Alert
Fault
Bearing — outer race defect at 12 o'clock
Rolling elementDefect strike
Operating conditions
Shaft fr
29.9 Hz
BPFO
107.4 Hz
BPFI
162.2 Hz
Cage FTF
11.9 Hz
RMS
0.00 g
Kurtosis
3.0
Crest factor
3.0
Envelope BPFO index
0.00
1 — Time waveform, raw accelerometer (fs = 20 480 Hz)
AccelerationDefect impact instants
2 — Raw spectrum (magnitude FFT, Hann window)
SpectrumShaft 1× and harmonicsStructural resonance band
What to notice. The defect energy is a train of tiny impacts. In plot 2 it is buried: the shaft imbalance at 1× is a hundred times larger, and the impacts smear across a wide resonance band around 3 kHz rather than appearing at 107 Hz. Plot 3 band-passes that resonance, takes the analytic envelope, and transforms again — the impact rate falls out as a clean line at BPFO with harmonics. Nothing here is machine learning yet. This is deterministic signal processing driven by bearing geometry, and it is the feature the model consumes.