I can't tell you how many times I've walked onto a manufacturing floor and seen the same thing: a half-finished IoT pilot with three disconnected sensor platforms, data going nowhere useful, and an engineering team that's lost confidence in the entire concept.
Here's the thing — Industrial IoT works. It works incredibly well. But the gap between a successful proof of concept on a lab bench and a reliable production deployment across 200 machines is enormous. And most of the guidance out there is vendor marketing dressed up as architecture advice.
Let me walk you through how we actually design IIoT systems for manufacturing clients — the sensor selection, the edge computing decisions, the protocol choices, and the data pipeline patterns that hold up when the shop floor gets messy.
Start With the Problem, Not the Sensors
Before you buy a single sensor, you need to answer one question: what decision will this data help you make?
I've seen companies deploy vibration sensors on every motor in the building because a vendor told them "predictive maintenance will save millions." Six months later, they have terabytes of vibration data and no idea what to do with it.
The right approach is working backward:
- Identify the business outcome — Reduce unplanned downtime by 30%? Improve OEE from 65% to 80%? Catch quality defects before they reach packaging?
- Define the decisions — What does an operator or engineer need to know, and when?
- Determine the measurements — What physical parameters indicate the condition you care about?
- Then select the sensors — Now you know what to measure, how often, and how precisely.
This sounds obvious, but I'd estimate 70% of failed IIoT projects skipped straight to step 4.
Sensor Selection: What Actually Matters
Sensor selection for manufacturing isn't the same as consumer IoT. Your sensors are going to live in environments with metal shavings, coolant mist, temperature swings, and electromagnetic interference from VFDs and welders. Here's what to evaluate:
| Factor | Consumer/Office IoT | Industrial IoT |
|---|---|---|
| Operating Temperature | 0-40C | -20 to 85C+ |
| Ingress Protection | IP20-IP44 | IP65-IP69K |
| EMI Tolerance | Minimal | Critical (CE/UL rated) |
| Vibration Rating | N/A | IEC 60068-2-6 |
| Power | Battery/USB | 24V DC, 4-20mA loop, PoE, or battery |
| Communication | WiFi, BLE | Wired (Ethernet/4-20mA), industrial wireless |
| Lifespan | 2-3 years | 10-15 years |
| Calibration | Rarely | Scheduled, traceable |
The Sensor Types We Deploy Most Often
For the manufacturing environments we work in, these are the workhorses:
- Vibration sensors (accelerometers) — Bearing wear, imbalance detection, misalignment. We typically use MEMS accelerometers for broadband monitoring and piezoelectric sensors when we need high-frequency resolution above 10kHz.
- Temperature sensors — RTDs for precision (0.1C accuracy), thermocouples for high-temperature processes (up to 1300C), IR sensors for non-contact measurement on moving parts.
- Current sensors — Split-core CTs on motor feeds. A motor drawing 15% more current than baseline is often the first sign of mechanical binding.
- Pressure/flow sensors — Hydraulic systems, coolant loops, compressed air. Pressure decay testing catches leaks that waste 20-30% of compressed air in a typical plant.
- Vision systems — Machine vision for quality inspection. This has gotten dramatically more accessible with edge AI inference — a $500 camera and an NVIDIA Jetson can replace a $50K inspection station for many use cases.
Edge Computing vs Cloud: The Real Trade-offs
This is the decision I spend the most time on with clients, because it's where the architecture either works or falls apart.
The Architecture Spectrum
Think of it as a spectrum, not a binary choice: