What a Football Stadium Taught Us About Factory Noise

In 2023 a Fraunhofer IDMT and IMMS team our founder worked in installed five wireless acoustic sensor nodes around the Veltins-Arena in Gelsenkirchen for the city's Open Innovation Lab. The nodes measured noise levels and classified the source on the device: traffic, trains, music, construction. Each drew about 8.5 W. The public write-up is here. A stadium car park is not a factory, but four of the problems were the same ones we meet on a shop floor.

Mixed sources, one microphone

At the arena a single node heard a motorway, a tram line, a car park filling up, a public-address system and, on match days, sixty thousand people, often at once. Nothing waited its turn. The system had to answer a question about one source while the others were still running.

That is the ordinary condition inside a plant, and it is the reason so many acoustic pilots die. A model trained on a clean recording of a bearing learns the bearing and the room it was recorded in. Put it next to a press line and the press is louder than the fault. The techniques that survive are the ones that learn a representation in which the target is separable rather than merely loud: time-frequency features that preserve the temporal structure of the event, training data that contains the interference rather than excluding it, and augmentation that mixes sources deliberately so the model never sees a target in isolation.

The instinct to record the machine on a quiet Sunday is the single most expensive mistake in an industrial audio project. The quiet recording produces a model that has never met the factory.

Sensor placement decides the dataset

Each of the five nodes in Gelsenkirchen was placed for a reason, and the placement determined what the node could ever learn. A node sheltered behind a building hears a different city from one on an open approach road. No amount of model capacity recovers a source that the microphone is acoustically screened from.

In a plant this is the same decision with more constraints: mounting points, cable runs, wash-down zones, hot areas, the machine operator's line of sight and the safety officer's opinion. We treat placement as part of the measurement plan and not as an installation detail, because the data from a badly placed sensor is not slightly worse, it is a different problem. Where it is genuinely uncertain, the cheapest answer is a short campaign with a portable unit in three candidate positions before anything is mounted permanently.

The corollary is that a plant's existing recordings are more valuable than they look, and less transferable than they look. Useful because they come from the real position; limited because they only describe that position.

A power budget is a design constraint, not an afterthought

About 8.5 W per node is not much. It is enough for continuous audio capture, a front-end transform and a small network, and it is not enough for anything casual. The design had to fit the classification into that envelope, which pushes every decision in the same direction: a compact spectral front-end computed once and reused, a model sized to the processor rather than to the benchmark, and inference scheduled rather than free-running.

Industrial deployments meet the same envelope for different reasons: PoE budgets, intrinsically safe areas, battery-backed cabinets, or simply a unit cost that has to survive multiplication by two hundred machines. A model that needs a GPU per asset does not scale to a plant, and a plant-wide monitoring programme that starts with a GPU per asset usually ends as a three-machine pilot.

Designing for the power budget first is also what makes the sovereign architecture affordable. If inference fits on the node, no raw audio has to leave the site, and the data-governance conversation that stops so many projects never starts.

Maintenance-free means drift-tolerant

Nobody was going to climb to those nodes to re-tune them. They sat outdoors through weather, seasons, traffic pattern changes and whatever the city did next to them. A system that is deployed and then left alone has to survive the slow drift of its own input distribution.

Factories drift too, just differently: a new product mix, a replaced bearing, a resurfaced floor, a neighbouring machine installed six months later, summer heat in an un-air-conditioned hall. The engineering answer is the same in both places and it is unglamorous. Monitor the input as well as the output, so drift is visible before accuracy falls. Keep a versioned dataset so retraining is a scheduled operation rather than a rescue. Design the alarm so its sensitivity can be adjusted in the field without a model rebuild. And decide, before deployment, who owns the retraining and on what trigger.

What transfers to a plant, and what does not

Four things transfer directly: training in the presence of interference, treating placement as a dataset decision, designing to a power budget from the start, and planning for drift. Those are not stadium lessons, they are deployment lessons that a stadium makes unusually visible because everything happens at once and nobody can intervene afterwards.

Two things do not transfer. A city soundscape has an enormous, freely observable variety of sources, so a general model of urban sound is a reasonable target; a plant has a small number of highly specific sources and a general model of factory sound is not a useful object. And a noise-monitoring system reports; it does not stop a line or reject a part. The moment an acoustic decision touches production, the evaluation standard changes completely, and the false-alarm budget shrinks by an order of magnitude.

Everything difficult about a field deployment is visible in a car park in Gelsenkirchen. Everything difficult about a production decision only appears once the line stops.

If any of this matches a problem on your line, the fastest way to find out what is possible is a free discovery call followed, where it makes sense, by a feasibility study of two to ten days.