You bolt a vibration sensor on a pump, set an alarm at 10 mm/s, and walk away. The alarm never goes off for two years. Then the bearing fails overnight. The sensor logged the whole thing, but nobody looked at the trend, and the threshold was set so high that the fault signal never crossed it until it was already too late.
Thresholds are not a strategy
A fixed alarm threshold makes sense for a machine that never changes speed, never changes load, and runs in a clean environment. Almost no machine does. A pump that starts and stops every hour has a vibration profile that swings with the operating state. Set one threshold and you either get false alarms on every startup, or you set it high enough that real faults hide below it.
The useful number isn’t the absolute velocity. It’s the change from that machine’s normal. A 5 mm/s reading on a good pump might be a 50 percent increase over baseline, which is the early warning. The same reading on a different pump might be normal for that bearing. Without the baseline, the number means nothing.
Build the baseline first
After installing the sensor, collect data for at least two weeks of normal operation. Include startups, stops, and the range of production conditions. The baseline isn’t a single number; it’s a distribution. You’re looking for the typical vibration level at each operating state, and the natural variation around it.
Then set the alarm as a percentage increase over that baseline, not an absolute value. ISO 10816 gives absolute guidance, but the trend alarm should come from your own data. A 30 percent increase over baseline is a reasonable early warning threshold. A 100 percent increase is urgent.
What the spectrum tells you that RMS doesn’t
Overall velocity (RMS) catches gross imbalance and looseness, but it misses early bearing faults. A bearing on its way out produces high-frequency bursts at specific frequencies: the ball pass frequency outer race (BPFO), ball pass frequency inner race (BPFI), and the cage frequency. These show up in the spectrum, not in the RMS trend.
An expensive sensor that only logs RMS is a waste of money. You need frequency spectrum data, at least periodically, to see the developing fault. A cheap sensor that logs RMS every minute will tell you the bearing is failing in month three; a spectrum-capable sensor tells you in month one, when you can plan the replacement.
Sampling rate and window
A vibration reading taken once an hour is good enough for trend monitoring. It’s not good enough for fault detection. A bearing fault develops over days to weeks, so hourly trends catch it. But if you want to see the high-frequency fault frequencies, you need a sample rate at least 2.5 times the highest frequency of interest, and you need to capture it under steady operating speed. A sample taken during a startup is garbage for spectrum analysis.
Most IIoT vibration sensors handle this automatically: they log a low-rate RMS value continuously, and burst a high-resolution spectrum on a schedule or on trigger. Make sure you’re paying for that capability, not just for a temperature sensor with an accelerometer bolted on.
The false-alarm tax
If the system alarms every time the machine starts up, operators mute the alarms. Once the alarms are muted, a real fault gets ignored. This is the most common failure mode of IIoT monitoring: not too few alarms, but too many. Every false alarm trains the maintenance team to distrust the system.
Filter the alarms by operating state. Don’t alarm on a transient during startup. Don’t alarm on a speed change. Send the alert to someone who can act on it, not to a generic inbox. And review the alarm history every quarter: if an alarm never results in a work order, either the threshold is wrong or the notification path is broken.
What it can’t do
Vibration monitoring catches mechanical faults: bearing wear, imbalance, misalignment, looseness, cavitation. It doesn’t catch electrical faults, like a failing motor winding or a cracked rotor bar. It doesn’t catch lubrication starvation until the bearing is already damaged. It doesn’t tell you the root cause; it tells you something is changing.
Use it as one input to a maintenance strategy, not as the whole strategy. Oil analysis, thermal imaging, and visual inspection still have a role. The sensor tells you where to look; it doesn’t replace looking.
Bottom line
Install the sensor, then wait two weeks before setting any alarm. The baseline is the whole point. A system with no baseline is a data logger that nobody reads, and a threshold set from a textbook will either cry wolf or stay silent until the machine stops.
