Back to Projects

Predictive Maintenance AI

Simulated Demonstration
All telemetry on this page is synthetic. Waveforms, spectra, health scores and work orders are generated in your browser by a physics-flavoured signal model — no real plant, customer, equipment or serial number is represented. The FFT, RMS/kurtosis statistics and the trend regression behind the confidence interval are, however, computed for real from the synthetic buffer.
Fleet — Simulated Site day 0.0
0Healthy
0Watch
0Critical
Simulation Controls
Playback speed
Inject fault into selected asset
Injected severity 60%

—

—
Health—
Runtime—
Shaft speed—
Sampling—
—
Vibration waveform — time domain —
RMS
—
g, overall
Peak
—
g, |max|
Crest factor
—
peak / RMS
Kurtosis
—
3.0 = gaussian
Novelty
—
unexplained energy
Frequency spectrum — 1024-pt FFT, Hann window
—
spectrum 1×/2×/3× shaft BPFO BPFI BSF FTF unmodelled peak
Condition trend — 90 days history + projection —
health index RMS velocity 80% RUL projection band functional-failure threshold
Prognosis — remaining useful life 80% CI
—
—
0 d——
Outside training distribution
Fault classification —
Alerts & work orders 0 open
No open alerts.
Inject a fault, or select a degrading asset.
Event history
How this demo works. A synthetic accelerometer signal is generated at 1280 Hz for the selected asset from its shaft speed and bearing geometry. A radix-2 Cooley–Tukey FFT is run over a 1024-sample Hann-windowed buffer; the peaks you see at BPFO / BPFI / BSF / FTF are the actual harmonic content of that buffer, not painted-on markers. Classification scores come from signal-to-noise at those geometry-derived frequencies and their sidebands; the RUL interval is an ordinary-least-squares fit on the health trend with the slope's standard error propagated to an 80% prediction interval. The out-of-distribution detector measures the fraction of above-floor spectral energy that cannot be attributed to any modelled frequency — when that fraction is sustained above threshold, the model refuses to output an RUL and asks for a human inspection instead. A real deployment would substitute measured data and a trained model; the honesty properties — calibrated intervals and an explicit “I don't know” state — are the point of the demonstration.