Context
Compressors and other rotating equipment are usually protected by built-in threshold switches and by periodic manual inspection rounds. Both tend to catch problems late. A threshold only trips once a fault has progressed far enough to push a single measured value out of range, and some developing faults never push any single value far enough to trip anything at all. The obvious alternative, training a model on past failures, runs into a practical wall: failures are rare, and when they do occur they are often logged loosely or not at all. Many sites have no usable failure history to learn from.
What we tested
We built a monitor that learns the normal relationships between signals the machine already produces. Each signal is estimated from the others, and when the machine drifts from its learned behaviour the estimates stop matching the measurements. That gap is condensed into a single anomaly indicator. The model uses a class of lightweight machine-learning techniques (reservoir computing) suited to time-series signals and edge deployment. It was validated on a publicly available dataset covering months of continuous operation, including several real failures. No failure information was used in training or in setting the alarm threshold.
What it shows
The monitor identified every failure present in the data, including one that the unit's own pressure switch never registered, and raised the alarm earlier on another. Because each signal is estimated separately, the indicator also points to which signal is behaving abnormally.
This may apply to your equipment
If you run compressors, pumps, or other rotating equipment protected mainly by threshold switches and periodic inspection rounds, this approach may be worth exploring, particularly if you have plenty of normal-operation data but few failure records. If you do hold reliable failure records, they can be used to tune the alarm point further. A short conversation is usually enough to tell whether it fits.