When a machine fails, it is natural to think of the failure as the moment it stopped. A pump was running and then it was not. In nearly every case that moment is the end of a story that began much earlier, and the story was written in physical signals that anyone with the right instruments could have read.
That idea, that failure has a signature, is the foundation of condition monitoring and of AFFERENT's work. This article explains what those signatures are, the order in which they tend to appear, and why so many of them still go unread.
A bearing, from beginning to end
Rolling-element bearings are among the most common components in industry and among the most common sources of failure, which has made their decline one of the best documented processes in engineering. Public run-to-failure datasets, such as the bearing test records in NASA's prognostics data repository, show the same broad arc again and again.
At first there is nothing to see or hear. Microscopic fatigue damage forms beneath the surface of a raceway. The earliest evidence appears at very high frequencies, far above human hearing, as rolling elements pass over the first imperfections and excite tiny resonances in the structure.
As the damage reaches the surface, each pass produces a small impact. The impacts repeat at a rate fixed by the bearing's geometry and the shaft speed, so they appear as a distinct component in the vibration signal. An analyst who knows the bearing can tell which part is damaged from the frequency alone. That calculation is the starting point of vibration analysis, still the most established form of machine condition sensing.
Later the damage spreads. Metal particles appear in the lubricant. The repeating pattern grows stronger, then blurs as the surfaces roughen. Noise becomes audible to a person standing beside the machine. Temperature rises as friction increases. By the time the bearing is hot to the touch, very little life remains.
Reliability engineers describe this with the P-F curve: the point P at which a developing failure becomes detectable, and the point F at which it becomes a functional failure. The interval between them is the window for action. Which signal you watch decides where P sits. High-frequency vibration and ultrasound move it early; a hand on the housing moves it late.
A breakdown is an event. Wear is a process. And processes leave evidence. We want to recognize that evidence early enough that someone still has time to act.
How much warning is there?
The length of the P-F interval is the most important number in condition monitoring, and it is also the hardest to pin down. It depends on the failure mode, the machine, its load, its lubrication and the signal being watched. A bearing in a lightly loaded fan may decline over months. The same bearing type in a heavily loaded gearbox, or one that has lost its lubricant, may go from detectable to failed in days. Electrical faults can be slower still, or abruptly fast.
The interval matters because it sets how often you have to look. A common rule of thumb in reliability engineering is to check a machine at intervals no longer than half the expected P-F interval, so at least one inspection falls inside the window with time left to plan the repair. For a fault that gives a month of warning, a monthly route is too slow. For a fault that gives a few days, only continuous measurement will reliably catch it.
Run-to-failure studies make the same point from the other direction. Bearings of the same type, run under the same conditions in the same laboratory, still fail at different times and decline at different rates. Any prediction of remaining life is therefore an estimate with a spread, and an honest system reports it that way.
The same pattern elsewhere
Bearings are the textbook case, but the general shape recurs across physical systems, each with its own signature.
- Electrical insulation deteriorates gradually. As it weakens, small partial discharges occur within it, producing ultrasonic and electromagnetic emissions long before a breakdown.
- Loose electrical connections have higher resistance, and resistance produces heat. The joint warms, sometimes for months, before it arcs or fails.
- Misalignment and imbalance in rotating equipment produce vibration at characteristic multiples of running speed, and they load bearings and seals in ways that shorten their life.
- Cavitation in a pump has a recognizable crackling sound and broadband vibration, and it is present before the impeller shows damage.
- Lubrication problems raise friction, which shows up as heat and as a change in acoustic character, especially in the ultrasonic range.
In each case the failure is preceded by a physical process, and the process emits energy in some measurable form: motion, sound, heat or electrical change. That emission is what we mean by a signature.
One fault, several signals
A useful consequence follows. Most failure modes leave evidence in more than one place, and different modes favor different signals. Vibration is the strongest general indicator for mechanical faults in rotating machines. Ultrasound is sensitive to friction, lubrication and electrical discharge. Temperature is late for bearings but early for electrical connections. Electrical measurements reveal problems inside motors and power equipment that a sensor on the outside may miss.
Combining signals also helps with the opposite problem, false alarms. A temperature rise on its own could be a sunny afternoon or a heavier load. The same rise alongside growing vibration at a bearing defect frequency is much harder to explain away. When two independent kinds of evidence point the same direction, a system can raise a finding earlier and with more confidence than either signal would justify alone.
This is why the international guidance on condition monitoring, ISO 17359, starts from the failure modes of a machine and works backward to the measurements that can detect them, rather than starting from whichever sensor is cheapest. It is also why AFFERENT is designing NODE around several sensing modalities at once.
A signature in the electrical current
Mechanical sensors are not the only way to read a machine. An induction motor's own supply current carries information about its condition, and motor current signature analysis has been used in industry for decades to read it.
A worked example shows how it works. A four pole motor on a 60 hertz supply has a synchronous speed of 1,800 revolutions per minute. Under load it runs a little slower, say 1,764, which is a slip of 2 percent. If one of the rotor bars cracks, the current spectrum develops small sidebands around the 60 hertz supply line at a spacing of twice the slip frequency. Twice the slip, 0.04, times 60 hertz is 2.4 hertz, so the sidebands appear near 57.6 and 62.4 hertz. Their height relative to the main supply peak, usually expressed in decibels, indicates how severe the damage is, and tracking that ratio over months shows whether it is growing.
The appeal of current analysis is access: the measurement can be taken in the motor control cabinet, away from the machine, which helps with submerged pumps and other equipment that is hard to reach. Its limits are real too. Load fluctuations, variable frequency drives and certain driven equipment can produce sidebands of their own, so readings need context about speed and load. Like every signature, it is most reliable when combined with others, such as vibration at the bearings and temperature at the windings.
Why signatures go unread
If the evidence is there, why do so many failures still arrive as surprises? Three reasons dominate.
Nobody is measuring. Most equipment is instrumented for control, not for condition. The signals exist, but nothing captures them. Portable route-based measurements, taken monthly by a technician, help, but a fault that develops in two weeks can slip between visits.
The data goes unread. Where data is captured, it often flows into a historian or a dashboard that nobody has time to study. Reading a spectrum is a specialist skill, and specialists are scarce. Simple thresholds are the usual substitute, and a fixed limit set for all machines fits none of them well.
Context is missing. A change in vibration might mean a developing fault, or it might mean the machine is running at a different speed or load. Without knowing the operating state, a system cannot tell the difference, and it either misses real problems or raises false alarms until people stop listening.
Reading the signature well
The practical lessons are straightforward to state and hard to execute. Measure the signals that appear early for the failure modes that matter. Measure continuously enough to catch faults that develop between inspections. Learn what normal looks like for each machine at each operating condition. Combine signals so that one can confirm or rule out another. And deliver findings in a form a planner can act on: which machine, what changed, how confident, and what to check.
None of that removes the need for experienced people. It gives them earlier warning and better evidence. Some failures will always be sudden, caused by an external shock or a defect with no detectable precursor. But a large share of costly failures announce themselves well in advance, and that is the opportunity behind machine failure signatures as a field of work.
