Edge Computing in Industrial IoT: What to Process Where
Industrial IoT systems have three places to compute: on or beside the machine, on an edge gateway or server at the site, and in the cloud. Each has different latency, connectivity and maintenance characteristics. Putting a function in the wrong tier creates either a fragile system that stops when the internet does, or an expensive one that ships raw data nobody uses. This article gives a way to decide.
At a glance
- 1.7 GB
- Raw data per day from one vibration channel sampled at 10 kHz with 16-bit samples.
- Milliseconds
- The timescale of machine control loops, which belong on the controller, not across a network.
- Autonomy
- Site functions that must survive an internet outage need to run locally.
- Fleet view
- Comparisons across sites and long-term trends are where the cloud adds most.
The problem
Sending everything to the cloud is the simplest architecture to start and the costliest to run. High-rate signals such as vibration or current waveforms generate far more data than a site uplink or cellular plan can carry, and most of it is never examined.
Pushing everything to the edge has the opposite problem. Every site becomes a small data centre that needs patching, monitoring and spares, and analytics improved in one place must be rolled out to many.
Process data where the decision is needed. Ship the conclusions, and keep the raw data close in case you need it later.
Reducing data at the edge
The first figure shows why edge processing matters for high-rate signals. One channel at 10 kHz and 16 bits is about 1.7 GB per day. Computing eight features such as RMS, peak and band energies once per second, stored as 32-bit values, brings that under 3 MB.
Keep raw data for a short rolling window at the edge, so when a feature crosses a threshold the waveform around the event can be uploaded for diagnosis. Report by exception and deadbands do the same job for slow signals such as temperature and pressure.
What good looks like
List each function with its latency requirement, its data rate and whether it must keep working when the site loses its internet connection.
Place control and protection on the controller, local alarms and data reduction at the edge, and cross-site analytics and long-term storage in the cloud.
Plan how edge software is deployed, updated and monitored remotely, since it runs on hardware spread across many sites.
| MB per day | |
|---|---|
| Raw, 10 kHz, 16-bit | 1728 |
| Features every second | 2.76 |
| Features every minute | 0.046 |
| Tier | Latency | Typical functions |
|---|---|---|
| Controller (PLC, drive) | Milliseconds or less | Control loops, interlocks, protection |
| Edge gateway or server | Milliseconds to seconds | Protocol conversion, feature extraction, local alarms, buffering |
| Cloud | Seconds and above | Fleet analytics, model training, long-term storage, reporting |
What it changes
Lower Data Cost
Only features, events and selected raw data leave the site.
Site Autonomy
Local alarms and logging continue through internet outages.
Fast Response
Time-critical decisions are made close to the machine.
Better Analytics
Cloud models see consistent, pre-processed data from every site.
Where to go next
Keep safety functions in dedicated safety systems. Edge and cloud software can report on them but should never be the protection itself.
Containerised edge software makes deployment and rollback across sites more manageable, provided the gateway has the resources to run it.
Problems this prevents
- Uplinks saturated by raw waveforms
- Local alarms that need the cloud to fire
- Edge devices nobody can update remotely
- Protection logic running on IT hardware
Discuss your project
Email support@aavico.com or use Get in Touch when the website finishes loading.