1
Tracks the trend
Horizon filters the noise of the health indicator and tracks its smoothed level and rate with a Kalman filter. The noise level is tuned from the data itself.
Horizon
When will a machine's health indicator reach its threshold? Horizon shows the answer as an uncertainty band compared with your planning lead time, not as one date. Your data stays in this browser.
Read the transcript
Narration: synthetic voice (Piper, voice “Joe”, CC0 public domain). Captions are burned into the video.
In three steps.
2
Projects with uncertainty
Thousands of possible futures are simulated, each carrying noise and rate uncertainty. The moment each path crosses the threshold gives a distribution of crossing times.
3
Reads a range, not a date
The result is not one date: P50 is the most likely time and the P10–P90 range shows plausible uncertainty. That range is compared with your planning lead time.
1
Data
Pick a scenario. All are mathematically generated examples, not real machine data.
My own health indicator
Drop a CSV file here
One numeric column: the row number is the time. With a date column, elapsed days are used.
2
Settings
Direction
Simulated paths
Forecast: measurements and future band
The pale line is measured, the dark line is the smoothed level. The shaded band is the likely path (P10–P90). Where the band meets the dashed threshold is the failure window.
When does it cross?
Each bar counts simulated paths that cross in that time window. Colours are zones against the lead time: red before it, amber near it, green later.
The needle compares the most likely crossing (P50) with the planning lead time.
Is it speeding up?
The rate is how much the indicator changes per unit of time; a rising line means degradation is speeding up. The shade is ±2 standard deviations.
Do the models agree?
Bars show the error (RMSE, lower is better) on the last 20% of the data, which each model did not see. The table lists each model's crossing estimate.
Export
How do I read the result?
- 1Forecast chart. The shaded band shows where the indicator may go. The range where it meets the dashed threshold is the failure window.
- 2Three numbers. P50 is the most likely crossing. P10 and P90 are the ends of the uncertainty. Do not look at P50 alone; P10 is compared with the lead time.
- 3Rate and dial. The rate chart shows whether degradation is speeding up. The dial shows where P50 sits against the planning lead time.
- 4Model agreement. If the models agree, you can trust the forecast more. If they diverge, read the range as wider than shown.
Horizon is a model that assumes the degradation continues. Maintenance resets, changes in load or regime, and sudden damage break the forecast. The threshold must be defined as loss of function. The result is not a guarantee, a diagnosis or a life table.
Frequently asked
What should I use as a health indicator?
Pick a measurable quantity that moves in one direction as the machine degrades: vibration RMS, filter differential pressure, a wear measurement. It should reflect the degradation mechanism, and the threshold should be the value where the function is lost, on the same indicator.
Why a range and not a date?
The future is uncertain because of load, maintenance, environment and noise. A single date hides that uncertainty. The P10–P90 range carries the model's own uncertainty and the noise honestly; it does not cover effects the model does not contain, such as sudden damage.
How much history is needed?
There is no fixed number. In practice a few hundred points and a clear trend help; a short record that is only noise gives a weak forecast. Earlier failures and the trend before them are very valuable. At least 12 points are needed; the model comparison runs from 20 points.
What does the planning lead time do?
Horizon compares the most likely crossing and the P10 value with your planning lead time. If P10 is shorter than that time, it raises an alert: a crossing is likely to come before the team can prepare. Changing the lead time changes the decision, not the event.
Is my data uploaded?
No. The calculation runs in your browser in a WebAssembly module written in C. Nothing is sent anywhere and it is gone when you close the page. Files you export stay with you.
Go deeper.
Full library
Reliability
Remaining useful life (RUL): report an interval, not one number
Learn how RUL is defined, what each of three common approaches needs as data, and how to report P10, P50 and P90 values with a worked trend calculation.
Reliability
Kalman filtering: a smoothed health indicator from a noisy sensor
Learn the predict and update steps, what the gain K means, and how a level-and-slope state yields a smoothed health indicator together with its rate of change.
Reliability
MTBF, MTTR and Weibull: reading reliability metrics correctly
Compute MTBF, MTTR and availability from a hypothetical record, then see why MTBF is not a lifetime and how the Weibull shape parameter β bears on scheduled replacement.
ML
Build a predictive maintenance pilot around a decision
Connect sensor behavior to operating context, maintenance events and a false-alarm budget.
Related questions
All questions
Start with a production problem.
Let’s review your data source, your existing system and the decision you want to improve.