Hearing an engine fault long before an inspection would.
Today a jet engine is watched in two ways: gauges in the cockpit that alarm at a fixed limit, and inspections every few hundred flights where an engineer looks inside. Both only notice a fault once it's big. This report shows a third way: turn every sensor's repeating flight pattern into a Fourier series, use it to put every reading on the same scale, and let a predictive system spot the pattern shifting long before anything is visible.
- 01Every flight repeats a pattern
- 02A pattern is a handful of circles
- 03Standardise every sensor
- ·Watch the circles drift
- 04Spot the shift early
- 05Twin vs conventional
An engine lives in cycles. Each flight traces the same shape.
An airliner's engine does the same thing every flight: idle at the gate, roar at take-off, ease off through the climb, hold steady in cruise, throttle back to land. Every sensor on it follows that rhythm. Exhaust temperature spikes at take-off and settles in cruise. Oil pressure rises and falls with engine speed.
So a sensor's data isn't a random stream. It's one shape, repeated flight after flight, with a little natural wobble from weather, load and route. That repetition is the key to everything below: if you know the shape, you know what "normal" looks like.
Any repeating shape can be rebuilt from spinning circles.
This is the Fourier series. Take one repeating shape, here a single flight's exhaust temperature, and you can rebuild it from circles stacked tip to tip. Each circle spins at its own steady speed: the first once per flight, the next twice, and so on. The pen at the tip of the last circle draws the shape.
A few big circles get the rough outline; more circles add the detail. Try it. A handful of numbers, each circle's size and starting angle, now describe the whole flight. That compact description is what lets a system compare thousands of flights quickly. It's the same maths as the drawing toy.
Five sensors, five patterns, one method.
The twin does this for every sensor it reads, using that engine's own healthy flights. Vibration repeats every turn of the shaft, thousands of times a minute; the rest repeat once per flight. Different units, different speeds, the same kind of description: a set of circles. That's what makes them comparable.
Measure every reading in "normal wobbles" from expected.
Now each new flight can be laid over the healthy pattern. The twin first takes the weather out (a hot, high airport makes any engine look worse), then measures the gap between this flight and the pattern. Dividing that gap by how much this sensor normally wobbles turns it into a z-score: zero means perfectly normal, ±1 is everyday variation, and beyond ±3 almost never happens on a healthy engine.
That's the standardisation. Degrees, psi, kilograms an hour and g all become the same unit, so one rule can watch every sensor at once. Watch it happen below: the flights stream in, each one becomes a single z-score, and the score drops onto the chart. Partway through, the engine's blades start to wear.
Expected, from the twin's circlesThis flight, liveThe gap
Normal, |z| under 3Beyond 3
Every circle has a normal size. When one drifts, that's the signal.
The circles are more than a way to draw the pattern. Each one stands for something physical: the fan's once-per-turn shake, the core's hum, the take-off peak in exhaust heat. On a healthy engine every circle keeps roughly the same size, flight after flight.
So the twin can track the size of each circle over time. A circle that grows, shrinks or suddenly appears is a fault announcing itself, and its drift gives a clean detection point: the moment a circle leaves its normal range and stays out. A cracked bearing, for example, adds a brand-new circle spinning 4.3 times per shaft turn. It starts tiny and grows with the crack, and it shows up in the circles long before the overall shaking looks unusual.
Why gauges and inspections find faults late.
Conventional monitoring works on raw readings. A cockpit gauge alarms at a fixed limit, the exhaust temperature redline for example. But a raw reading swings by tens of degrees with weather and load, so the limit has to sit far above normal to avoid false alarms. A slow drift stays hidden in that swing until it's almost at the redline.
Inspections have the opposite problem: they look closely, but rarely. An engineer opens the engine every few hundred flights with a borescope, and only damage big enough to see gets caught.
The twin watches the same sensor, but standardised. With the weather and the normal pattern taken out, the drift stands clear of the noise after a fraction of the flights. Same data, very different picture.
Same engine, same five sensors. Which system notices first?
Pick a fault and play. At flight 100 the fault starts and slowly gets worse. Both systems receive exactly the same readings from the same sensors every flight; the only difference is what they do with them. On the left, conventional monitoring checks each raw reading against its cockpit limit and inspects every 100 flights. On the right, the twin standardises each sensor against its Fourier pattern and flags any sensor that stays beyond ±3 for five flights. The fidelity setting is how many readings the twin takes per cycle, and so how many circles it can keep.
Conventional monitoring
Each raw reading against a fixed cockpit limit, plus an inspection every 100 flights.
Digital twin
The same readings, weather removed, scored against the healthy pattern. Anomaly: beyond ±3 for five flights.
| Twin fidelity | Readings per cycle | Sensors per signal | Twin detects at | Sensors the twin flags | Conventional detects at | Head start |
|---|
Illustrative model: realistic methods and shapes, a made-up engine and made-up numbers. Not real engine data.
The method, step by step.
-
01 · Collect
Read every sensor, many times per cycle
Accelerometers on the fan and core casings, exhaust thermocouples, oil pressure and temperature, fuel flow, shaft speeds, and inlet temperature T₂ and pressure P₂.
-
02 · Transform
Fit a Fourier series to each pattern
x(θ) ≈ a₀ + Σk=1K Ak sin(kθ + φk)θ runs over one shaft turn for vibration (order tracking keeps a part on the same term at any speed) and over one flight for the slow signals. With M readings per cycle you can keep K = ⌊(M − 1) / 2⌋ circles. A bearing crack ringing at 4.3× shaft speed needs at least 10 readings per turn to show up at all.
-
03 · Correct
Take the weather out
θ = T₂ / 288.15 K, δ = P₂ / 101.325 kPa, Nc = N / √θ, EGTc = EGT / θxStandard gas-turbine corrections to a reference day, before comparing with the pattern.
-
04 · Standardise
Turn every gap into a z-score
z = ( x − x̂ ) / σhealthy, σ ∝ 1 / √Sx̂ is the twin's expected value from the pattern, σ the sensor's normal wobble. Averaging S sensors shrinks σ. Several z-scores can combine into one health index, for example D² = rᵀ Σ⁻¹ r.
-
05 · Detect
Flag a sustained shift, per sensor
anomaly when |z| > 3 for 5 flights runningA healthy reading lands beyond ±3 about 0.27% of the time, so five in a row is vanishingly rare by chance. Because each sensor is watched on its own, the twin also says where the problem is.
The circles in this report are the same ones in the toy.
Every pattern above was drawn by a Fourier series. The toy lets you do it with anything, from a word to a photo, using up to 2,048 circles. It's the quickest way to get a feel for why a few circles can describe a messy signal, and why a circle changing means something changed.