Predictive maintenance with vibration and temperature sensors: how it works

A motor that is about to fail usually says so weeks in advance. It runs a little rougher, a bearing warms up, a frequency appears in its vibration that was not there before. Predictive maintenance is the practice of listening for those signs with sensors, so that a bearing is replaced in a planned window rather than after it seizes at two in the morning. This article explains how the two most common sensor types do that job.

What predictive maintenance is

Maintenance strategies sit on a scale. Reactive maintenance fixes things after they break. Preventive maintenance replaces parts on a schedule whether they need it or not. Predictive maintenance measures the condition of the equipment and intervenes when the measurements say a failure is developing.

The measurement is the hard part, and for rotating equipment (motors, pumps, fans, gearboxes, compressors) two quantities carry most of the information: how it vibrates and how warm it gets. Both can be measured from the outside of the machine, both change gradually as a fault develops, and both are cheap to sense continuously. That combination is why condition monitoring retrofit projects almost always start here.

The output is a trend and a set of thresholds. Nobody needs to interpret a vibration spectrum in real time; the system watches the numbers, compares them with what is normal for that machine, and raises an alert when they drift. People then decide what to do.

What vibration tells you about a machine

Every rotating machine vibrates, and a healthy one vibrates in a way that stays stable over time. A vibration sensor, in practice an accelerometer bolted or glued to the machine casing, measures that motion many times a second. The signal is processed into two kinds of number.

The simplest is an overall level: a single figure for how much the machine is shaking, often expressed as a velocity. Standards bodies publish guidance on what levels are acceptable for machines of different sizes and mountings (the ISO 20816 series, Mechanical vibration: measurement and evaluation of machine vibration, is the one most charts derive from), and an overall level that rises steadily over weeks is the most basic warning sign there is.

The richer number is the spectrum: the vibration broken down by frequency. Different faults show up at different frequencies. An imbalance in the rotor appears at the speed the shaft turns. Misalignment between a motor and the pump it drives shows at multiples of that speed. A damaged rolling-element bearing produces characteristic frequencies that depend on the bearing’s geometry and are published by the bearing manufacturer; when energy appears at those frequencies, a specific bearing is telling you it has a defect. Gear problems show at the gear-mesh frequency. Looseness shows as a spread of harmonics.

Two practical consequences follow. First, a sensor that only reports an overall level is enough to know that something is changing, but a sensor or gateway that reports a spectrum, or at least the energy in a few frequency bands, is what tells you what is changing. Second, the machine’s speed must be known, because every diagnostic frequency is relative to it; a variable-speed drive that changes the running speed changes every frequency with it, and the system must account for that.

Vibration sensors vary in what they can measure. The upper frequency a sensor can see depends on its design and on how it is mounted, and bearing defect frequencies can be high, so sensor selection and mounting are not separable from what you hope to detect.

What temperature adds

Temperature is slower and simpler. A bearing that is failing generates more friction and gets warmer. A motor winding under electrical stress gets warmer. A pump running dry gets warmer. A temperature sensor on the casing does not tell you which fault is developing, but it confirms that something is producing heat that was not there before, and it does so with a sensor that costs little and needs no signal processing.

Temperature earns its place in two ways. It corroborates vibration: a rising vibration level with a rising bearing temperature is a much stronger signal than either alone, and an alert that requires both is far less likely to be false. And it catches faults that vibration misses: lubrication problems, cooling failures and electrical faults often show up as heat before they show up as motion.

The measurement is relative. What matters is the temperature of the bearing housing compared with the ambient temperature and compared with the same housing last week. A hot day raises everything; a failing bearing raises one thing. Systems that monitor temperature usually record ambient alongside and work with the difference.

Where to mount the sensors

Mounting decides whether the sensor sees the fault. For vibration, the rule is to get as close to the bearing as possible, on a rigid part of the housing, with the sensor’s sensitive axis pointing along the direction the fault will move the shaft. On a horizontal motor that usually means a sensor on the drive-end bearing housing, and often one on the non-drive end too. Sensors on flimsy covers, painted surfaces or thick gaskets measure the cover rather than the bearing.

Attachment matters. A stud screwed into a tapped hole gives the best coupling and the widest frequency range. A strong adhesive pad is the usual retrofit choice and works well for most faults. A magnet is convenient for a survey and poor for a permanent install, because it limits the upper frequency and can shift.

For temperature, the sensor goes on the housing near the bearing, with thermal contact ensured by a small amount of thermal compound, and away from external heat sources such as a nearby exhaust or direct sun.

Wireless sensor nodes have made this far easier, because the battery, radio and sensor are one sealed unit that sticks to the machine and reports every few minutes. They have a power budget to respect: a node that samples vibration at high rates all day will not last on a battery, so most sample briefly at intervals and send a summary. That is the right trade for trend monitoring, and the wrong one for catching a short transient.

From raw readings to an alert

Raw readings become useful in three stages. First they are processed into features: an overall vibration level, the energy in a few frequency bands, a peak temperature, a temperature difference from ambient. This can happen on the sensor node, on a gateway, or on a server, and the further down it happens the less data has to travel; the edge computing article is about exactly that placement.

Then each feature is trended against its own history. The first weeks of monitoring establish what normal looks like for that machine under its usual load. Every machine is different, and a level that is alarming on one is routine on another.

Finally, thresholds turn a trend into an alert. The most useful are set relative to the machine’s own baseline: a warning when a level rises by some fraction above normal, and an alarm when it rises further or when two independent features (vibration and temperature, say) move together. Published vibration severity charts, such as those in the ISO 20816 series, give a starting point when there is no baseline yet. The alert itself usually travels as an MQTT message to whatever people watch. Some systems add automatic anomaly detection on top, which flags any pattern that differs from the baseline without naming the fault; that is useful for catching the unexpected and noisy if the baseline was collected on a bad week.

Every alert should carry the evidence: which machine, which feature, the trend, and the threshold it crossed. A maintenance planner then decides whether to inspect, to schedule a replacement, or to keep watching. The system predicts; people plan.

What this approach cannot see

Vibration and temperature are outside-in measurements of rotating machinery, and they have edges. They say little about a fault that produces neither motion nor heat until it fails suddenly, such as some electrical insulation faults or a structural crack. They do not see the state of the lubricant, which is what oil analysis is for. They struggle on very slow machines, where the frequencies of interest are low and the energy is small. And they cannot tell you about a machine you have not put a sensor on.

They are also only as good as the baseline. A machine that runs under varying loads and speeds produces varying readings, and a monitoring system that ignores load will alert on every busy shift. Where a machine’s duty changes, the load or speed needs to be measured alongside, or the thresholds need to be set per operating state.

Within those limits, a vibration sensor and a temperature sensor on each critical bearing, reporting to a system that trends them and alerts on drift, is the single most productive place to start with condition monitoring. It is retrofit-friendly, the physics is well understood, and the first bearing it catches before it fails pays for the rest.