Picking the radio is usually the first irreversible decision in an industrial sensing project. Everything else — battery size, enclosure, gateway count, even where you are allowed to mount the sensor — follows from it.
Start from the message, not the technology
Write down three numbers before you look at any radio: how many bytes per message, how often, and how far. Most industrial sensing turns out to be tens of bytes every few minutes over a few hundred metres. That profile rules things in and out immediately.
LoRaWAN
Good for small payloads, long range and multi-year battery life. Bad for anything that resembles a stream. The duty cycle limits in most regions are real constraints, not fine print — if your device needs to talk more often than the regulations allow, no amount of firmware cleverness will fix it. Best fit: distributed sensors across a large site where you control the gateway placement.
4G / LTE Cat-M and NB-IoT
Good when there is no site infrastructure to borrow and the devices are scattered across places you do not own. You get coverage without building anything. You also get a SIM per device, a bill per device, and a dependency on a carrier’s network plans, which change. Best fit: assets in the field, vehicles, remote pump stations.
Wired Ethernet
Unfashionable and frequently correct. If a cable tray already runs past the measurement point, wired is more reliable, needs no battery, and removes an entire class of intermittent fault that is genuinely miserable to diagnose. Best fit: inside a plant, on a line, anywhere the infrastructure already exists.
Wi-Fi
Convenient for a pilot, awkward for a rollout. It shares a network someone else administers, and it will be reconfigured without telling you. If you use it, plan for credential rotation from day one.
Most real sites use two
A site that has been running for twenty years does not have one uniform environment. Wired where the trays reach, LoRa across the yard, one 4G gateway for the outbuilding nobody wanted to trench to. Choose gateway hardware that accepts more than one uplink, and specify store-and-forward — an outage should leave a gap that fills in later, not data that never existed.
Before you commit
Do a site survey with the actual hardware, at the actual mounting height, with the machinery running. Range figures on a datasheet are measured in conditions your plant does not have. One afternoon of measurement saves a re-spec later.