guest lecture · V. R. Siddhartha Engineering College
AI and Digital Twins in Instrumentation
What AI and digital twins actually do in instrumentation — and what they don't. Sixty minutes for Electronics & Instrumentation students, tracing a single bearing failure from the physics to the alarm nobody trusts.
The useful test for anything sold to you as a twin: what does it predict that a human could not, and how was that prediction validated? A 3D model with live sensor values painted on it is a visualization. It might be a good one. It is not a twin, because it cannot answer "what happens if I raise the feed rate by ten percent." The second simulator below is a twin by that test — it runs a physics model of a bearing forward in time and produces predictions you can check against what happens.
Everything below runs entirely in your browser. No build step, no CDN, no network calls, no analytics, no storage. They work offline, and they reflow on a phone — though the live monitor computes four FFTs a second, so an older phone will run warm.
Bearing Kinematics
where the frequencies come fromThe pedagogical companion to Part 2. A guided six-stage walkthrough that builds up the fault frequencies from geometry — anatomy, the rolling constraint, then FTF, BPFO and BPFI derived one at a time — plus an Explore tab with sliders for RPM, ball count, d/D ratio and contact angle.
- Watch the cage rotate at ~40% of shaft speed — the constraint is visible, not stated
- Turn on the outer spall and count impacts as balls pass 12 o'clock. Compare to n × FTF
- Presets for SKF 6205, SKF 6308, NU 220 — see how BPFO changes with ball count and geometry
Envelope Analysis, the intuition
no equations requiredThe gentlest possible companion to §2.4. Five plain-English scenes tell the same story — every bearing defect is a tiny hammer, each strike rings the housing, drawing a line over the loudness reveals the pattern, and an FFT of that line counts the strikes automatically. No jargon; the words Hilbert and analytic signal do not appear.
- Animated bearing shows a ball striking the defect in real time; each hit lights up the waveform
- The envelope is drawn on top of the waveform as an amber outline — visual proof that the pattern was there all along
- Ends on a clean 107 Hz peak. For students who want the actual math, links to the deep-dive walkthrough
Bearing Fault Lab
one measurementA single vibration snapshot from an SKF 6205 bearing at 1797 rpm. The defect is invisible in the raw FFT and obvious in the envelope spectrum. Four presets — start on Early defect and try to find the fault before switching views.
- BPFO derived from bearing geometry: 107.4 Hz, before anything is measured
- Raw spectrum vs. band-pass → Hilbert envelope → FFT, computed live in the browser
- Envelope BPFO index: 0.36 healthy → 3.2 early defect (Alert) → 7.7 advanced (Fault)
- Noisy plant keeps the same 22% defect but drops the index to ~1.7, back under the alert line — a missed detection
Cooling with RL
where the 40% came fromA 1 MW data centre with the three cooling levers a smart optimiser gets to touch — chilled-water setpoint, cooling-tower fan, free-cooling bypass. Watch a baseline policy and an RL policy run the same week on the same weather. The gap opens overnight and widens when the outside is cool.
- Chiller COP falls with lift, fan power scales cubically with speed, and free-cooling kills the compressor entirely — the three physics facts the savings live on
- Verified: 42% reduction over a synthetic week, PUE 1.19 → 1.10, no safety violations
- Stage 5 strips the safety margin so the RL agent hunts the limit and the safety layer clamps it — the reason you never let RL explore in production
Live Condition Monitor
the twin, running nowA level-one physics twin. Four identical motors on a simulated clock, one reading each per four simulated hours. MTR-402 develops an outer-race spall on day 18 and it grows. Every point on the trend is computed from the twin — synthesise, band-pass, Hilbert envelope, FFT — not replayed from a table.
- Health index = 30·√(BPFO band energy) ÷ RMS. Healthy ≈ 0.36, defect climbs to ~7.7
- False alarms come from random non-periodic knocks a healthy machine actually takes, not a faked small defect
- The twin makes a falsifiable prediction: days-to-fault. It's confidently wrong early on
Twin or Dashboard?
rapid-fire · 60 secThe vendor test as a room game. Five mock vendor screens — a live 3D turbine, a SCADA console, an MPC preview, an 'AI-powered enterprise digital twin platform', and your own bearing model. The room shouts twin or dashboard, then the reveal lands with the one-line reason.
- Full-screen mode, space bar to advance — manual pace so you can wait for the room to commit
- Two twins and three dashboards, in a mix that mirrors what actually gets sold under the name
- Closes on the bearing twin you demoed earlier — the smallest possible thing that passes the test
The twin, running now
MTR-402 fails on day 18It needs about forty simulated days before the trend chart says anything, so leave it running while you read. Once the defect takes hold, the threshold slider shows the trade-off between warning time and false alarm rate.
Where a twin stops working
Three things get sold under the name. A physics model — first principles, accurate where your equations are right, slow. A data-driven surrogate — trained to mimic the plant, fast, interpolates well, extrapolates terribly. And a hybrid, a physics model whose parameters are continuously re-estimated from live data. That last one is the interesting one and the hardest to build.
The trap is in the second. Faults are extrapolation — that's what makes them faults. A surrogate trained on a healthy plant will predict a normal Tuesday beautifully and tell you nothing about the one event you actually wanted warning of.
My own Raspberry Pi rig — light, motion, temperature and proximity sensors driving the home automation — has years of normal readings and essentially no failures. So does most industrial data. Without faults to train on, the useful approach is to model normal and measure distance from it.
Both simulators are single self-contained HTML files — bearing lab, live monitor. Save the page and they still work offline.