Perimeter security guide
Do MEMS fence sensors false-alarm in wind and rain?
Wind and rain are the dominant source of nuisance alarms in fence-mounted MEMS sensing, and a poorly tensioned fence is the hardest case of all. It is a real limitation of the technology rather than a fault in any one product — so the question worth asking a supplier is not whether it happens, but what their sensor does about it.
A MEMS accelerometer on a fence feels every vibration that reaches it. An intruder climbing the fence produces vibration. So does a gust of wind, a lorry passing on the far side of the road, rain striking the fabric, and a dog pushing along the line. The sensor cannot choose what reaches it; it can only choose what it reports.
This is a constraint of accelerometer-based sensing rather than a gap in one vendor's implementation, and it is well understood by anyone who has commissioned a fence sensor on an exposed site. Any supplier telling you their MEMS sensor is unaffected by weather is describing a product that does not exist.
Why wind is the hard case
A fence is a tensioned membrane. Wind loads it continuously, and the fabric responds by flexing and, if the tension is low, resonating. That resonance can be larger in amplitude and longer in duration than the vibration an actual climb produces.
A sensor that decides by amplitude alone therefore has no good threshold available. Set it low enough to catch a careful climber and the wind triggers it; set it high enough to ignore the wind and a careful climber walks through. Neither setting is wrong — the information needed to separate the two is simply not present in amplitude.
Fence tension is the variable that makes the difference, and it is an installation property rather than a sensor property. A well-tensioned fence on sound posts gives any vibration sensor a far better signal to work with than a slack one, and no amount of signal processing fully recovers what poor tensioning costs.
Rain is a different problem from wind
The two are often described together as weather, but they look nothing alike to a sensor. Wind is sustained, low in frequency and builds and decays over seconds. Rain is impulsive, broadband, and — importantly — arrives across the whole perimeter at once rather than at one point on it.
That difference is usable. An event appearing on every node along a fence simultaneously is almost certainly weather; an event on one node, or on two adjacent ones, is a point disturbance. Rejecting rain and distinguishing wind are separate problems with separate solutions, and a supplier who treats them as one has probably solved neither.
Threshold detection versus classification
The useful distinction between perimeter sensors is not how sensitive they are but how much they decide. A threshold sensor reports that vibration exceeded a level. A classifying sensor reports what the vibration resembled.
Classification works because intrusion methods have structure. Cutting a chain-link strand is a short, sharp, repeated impulse. Climbing is a slow sequence of loading and unloading as weight shifts. Drilling is sustained and periodic. A hammer strike is a single high-energy transient. These differ from each other, and they differ from wind, in ways that amplitude alone cannot express but a signature can.
Where that decision is made matters too. Classifying at the node means a nuisance event is discarded on the fence and never consumes bandwidth, operator attention or control-room screen space. Classifying at a server means every vibration is transmitted first and filtered afterwards.
What Prachir reports and what it discards
Prachir classifies on the N100 node rather than at a server. It reports four alarm events — climbing, cutting, drilling and hammer impact — each identified separately, so an operator is told which is happening rather than simply that something moved.
Three sources are identified as nuisance and discarded at the node before they reach the control room: rain, passing vehicle traffic, and animals crossing or crawling along the line.
Because every node is individually addressed on the CAN bus, an event resolves to a specific point on the fence rather than to a long zone — which also means a correlated event across many nodes is visible as exactly that.
Where this still goes wrong
Four things we would rather tell you now than after a purchase order.
- Rock walls are not supported. The vibration signature does not transfer through the material predictably enough to classify reliably. If your boundary is a rock wall, a fence-mounted MEMS sensor is the wrong technology and we will say so.
- A slack fence bounds what any sensor can do. If the fabric is poorly tensioned or the posts move, tension it before specifying detection. This is cheaper than the sensor and it improves every sensor you might buy, including a competitor's.
- We publish no false-alarm rate. A figure would depend on fence construction, tension, wind exposure, traffic proximity and local wildlife, so a number from our test site would not predict yours. We would rather describe the reject classes, which are checkable, than quote a percentage that is not.
- Sustained high wind on long, light-gauge chain link remains the hardest condition for this technology, ours included. On an exposed site, ask for a trial run on your own fence rather than taking any supplier's word — including ours.
Questions worth asking any PIDS supplier
These apply whoever you are buying from, and the answers are more revealing than a specification sheet.
- Which event types do you classify separately, and which do you report as one generic alarm?
- Which nuisance sources do you explicitly reject, and where is that decision made — on the sensor or at a server?
- What surfaces is the sensor qualified for, and which are excluded?
- What happens during a storm? Does the system flood the operator, suppress a zone, or degrade in some defined way?
- Can the sensor prove it is healthy, so that silence means something? A failed sensor and a quiet sensor should not look the same.
- Can we trial it on our own fence, in our own weather, before committing?