Predictive maintenance without failure data

The problem

Supervised predictive-maintenance models learn from failures. To forecast how a bearing fails, they need bearings that have already failed, recorded with the same sensors under similar conditions. Few plants have that history, because machines are serviced before they fail, and failures recorded on one machine may say little about the next.

Simple indicators, one number per measurement, need no history. But a number that moves up and down without a clear trend does not tell a planner when to act.

We wanted an indicator that needs no such history and still trends steadily toward failure.

What we built

Tetracta PdM is a health indicator learned from a machine's healthy period. It gives one number per machine: low while the machine behaves as it did when healthy, rising as it moves away from that state.

It works from vibration, sound, current or thermal signals, and it needs the machine recorded in good condition first.

The method itself is not published.

What you need

Rotating machines such as pumps, compressors, fans, motors and gearboxes, with a period of recordings from normal operation.

The usual input is vibration recordings from an accelerometer on the machine. What the service can do depends on the data your sensors give. Raw waveforms or spectra support the full indicator and its trend. Sensors that report only single values, such as RMS, support a trend only. With low-bandwidth data it is unreliable for early bearing faults. In the first step we check your data, including whether the sensors sample fast enough, and label every channel with what it can support.

Sensors and gateways that give you your raw data work fully. Closed cloud platforms that do not export their data cannot be used.

How it runs

It runs as a web service, live at tetracta.ai/pdm. For each machine, it needs a healthy reference period and then the ongoing recordings; what comes back is the health indicator over time. We host it today. For a pilot we can also analyse recordings you send; an install on your own site is scoped with you.

The full validation is published on our predictive maintenance results page.

Results

We validated the indicator on 32 bearings that were run until they failed, from two public benchmarks: FEMTO (17 bearings) and XJTU (15 bearings). Both are accelerated lab tests, and the datasets are public.

Monotonicity measures how steadily an indicator moves in one direction as a bearing wears out: +1 means it rises at every step toward failure, 0 means no trend. The medians in the table mean that for the typical bearing the indicator rose almost steadily toward failure. Individual bearings vary: the hardest runs score much lower than the median, and the results page shows every run.

Against the raw FFT spectrum. On these benchmarks, an indicator built directly from the raw spectrum is at least as good as ours. If your team already tracks full spectra, these results do not show that we would add accuracy.

What a maintenance planner gains. One number to watch per machine. In the example run on the results page, the alarm came with 92% of the bearing's life still ahead.

On a phone: Check It

Check It is our Android app for a first look at rotating machines with the phone's own sensors: microphone, accelerometer and magnetometer. It is a screening tool for low-frequency faults such as imbalance, misalignment and looseness, and its analysis runs on the phone. It also has an acoustic lab and a visual check.

Access is by invitation: ask for one, or read the guide first. A phone is not a calibrated vibration sensor. Use it to decide where to look; do not use it to sign off a machine.

How a pilot starts

You send recordings from one machine, made while it was in good condition, and its maintenance log if you have one. We answer within one working day and say whether those recordings can support the indicator. The first job is a review of that one machine, one to three weeks long. We agree its scope and price in writing before any work begins, and no payment is due until you accept the review. Work on more machines is quoted after the review.

What you receive

• A written review of that one machine.
• The health indicator over the recorded period.
• A label on each channel for what it can support: the full indicator and its trend, or a trend only.
• The result of the data check, including whether the sensors sample fast enough.

See a sample report on public data

If the machine has no sensor or no recordings yet, we can say what to fit and what to record, starting with a period in good condition.

A pilot on several machines follows the same path:

1. Pick a few machines. Machines that have a vibration, sound, current or thermal sensor, or room to fit one, and a maintenance log.
2. Record a healthy baseline. Recordings made while the machines are in good condition.
3. Run it alongside what you have. The indicator runs next to your current monitoring. At the end we compare its trend with your maintenance log, machine by machine, and report what it caught, what it missed and where it raised false alarms.