Edge computing has been discussed heavily over the past few years, but down on the shop floor the questions are practical ones: what equipment do I actually put on site? What does it compute? And who consumes the result?
Start with a clear definition: edge computing means placing data storage, processing and analysis close to where the data is generated, rather than sending everything to a remote data center or the cloud. In a factory, that "edge" is usually an industrial computer, an edge gateway or an edge server sitting in a workshop cabinet.
Why Factories Need Edge Computing
This is not about following a trend. Four classes of problems can only be solved at the edge:
- Latency: equipment interlocking, vision inspection and safety interlocks demand millisecond response. The physical round trip to the cloud, measured in hundreds of milliseconds, is simply unacceptable.
- Bandwidth: one machine generating thousands of points per second, a line running dozens of cameras — pushing all of that to the cloud is neither economical nor realistic.
- Network outages: external connectivity loss is routine in plants. Critical control and local alarming cannot depend on the availability of an outside link.
- Compliance: process parameters and production data are commercially sensitive, and many companies require raw data to never leave the plant.
Four-Layer Architecture: What Each Layer Solves
Layer 1: Device Access
PLCs, instruments, drives, sensors, cameras and power meters. The core job of this layer is turning heterogeneous devices into readable data points, and the difficulty lies in the mix — protocols, brands and equipment vintages all vary.
Common practice: serial devices are aggregated over an RS485 bus, Ethernet devices go into a local switch, and sensors without a communication port are converted through IO modules or smart transmitters.
Layer 2: Edge Acquisition
Handled by an edge gateway or an industrial computer, responsible for protocol parsing, data buffering and uplink. The capabilities that matter here are offline buffering and time synchronization — data may arrive late, but it must not go missing or arrive out of order.
Engineering note: time synchronization is chronically underestimated. If data from several workshops carries inconsistent timestamps, energy curves and fault-sequence analysis are both distorted. Use a single time source on site and configure periodic clock correction on the gateway side.
Layer 3: Edge Computing
This is the layer that separates a true edge solution from a pure acquisition project. What can be done at the edge includes:
- Data cleansing: removing spikes, filling gaps, converting units, so dirty data never reaches the systems above.
- Threshold and rule evaluation: over-temperature, out-of-range, timeout and abnormal state, judged locally and triggering audible/visual alarms or SMS notification.
- Statistical aggregation: output, energy consumption and OEE by minute, hour or shift, computed at the edge before uplink, which can cut bandwidth usage by an order of magnitude.
- Edge inference: deploying trained models onto edge devices for vibration anomaly detection, visual defect inspection, energy forecasting and more — fast response with raw data staying inside the plant.
This layer demands compute, storage and stability in equal measure, and is typically carried by a fanless embedded industrial computer or a rack-mount industrial computer rated for continuous long-term operation.
Layer 4: Platform and Applications
SCADA, MES, energy management systems, visualization dashboards and reporting on a cloud or local server. The edge layer uploads cleansed, structured data, leaving the platform free to focus on business analytics, cross-plant comparison and long-term trends.
How to Specify Hardware: Tier by Compute Requirement
- Lightweight acquisition (a few hundred tags, forwarding only): an edge gateway is enough — low power, DIN-rail mounting, fanless and wide-temperature.
- Moderate compute (a thousand-plus tags, rule evaluation, local storage): a fanless embedded industrial computer, multiple serial and Ethernet ports, industrial-grade solid-state storage.
- Higher compute (vision inspection, model inference, multi-channel video): a high-performance industrial computer or edge server with a discrete GPU or AI acceleration module, where heat dissipation and power supply are the problems to solve.
Do not judge a selection by CPU model alone. What matters more is sustained performance of the complete unit at the target temperature. Fanless devices shed heat through the chassis and will throttle under high ambient heat, so leave thermal headroom.
A Phased Rollout Path
- 1Phase one: take stock. Complete the device inventory, tag table and network topology, and establish which devices can communicate and which need retrofitting. This step cannot be skipped, and should not be skipped.
- 2Phase two: single-point validation. Pick one production line or one distribution room as a pilot, run the full chain of acquisition, buffering, uplink and visualization, and verify that protocols and polling intervals meet requirements.
- 3Phase three: replicate horizontally. Once the pilot is stable, replicate by workshop, establishing tag naming conventions, alarm rule templates and O and M documentation in parallel.
- 4Phase four: add edge intelligence. Only after data has accumulated reliably for a while should analytics and model inference be introduced, by which point the data foundation is trustworthy.
A common mistake: buying the most expensive edge server first, only to find the tag table was never sorted out and most equipment has no communication port, leaving the project stalled at step one. The right order for edge computing is data first, analytics second, intelligence last.
Four Pitfalls That Are Easy to Hit During Implementation
- No network planning: mixing the field device network with the office network invites broadcast storms and security risk. Segment the networks from day one.
- Careless tag naming: without a unified naming convention, later additions become unmaintainable and systems fail to match up with each other.
- Neglecting power and grounding: a considerable share of communication faults trace back to this, and troubleshooting later costs far more than doing the installation properly upfront.
- No operations design: equipment spread across multiple workshops with no remote status monitoring means failures surface only when someone on site reports them.
Frequently Asked Questions
Does edge computing make the system more complex and harder to maintain? It adds equipment in the short term and reduces complexity in the long term. Because data is cleansed and aggregated locally, the platform only has to handle structured results, which shortens the chain. The precondition is a solid remote-operations design: device status viewable remotely, configuration pushed remotely, firmware upgradable in batches.
We already have SCADA. Is an edge layer still necessary? It depends on how SCADA is deployed and how extensible it is. If SCADA already covers the critical equipment and runs stably, the edge layer mainly serves to bring in legacy equipment that cannot be connected, to add offline buffering, and to take on new compute workloads such as vision inspection. The two are collaborative rather than mutually exclusive.
What environmental criteria should edge devices be specified against? Specify against the measured extreme temperature at the installation location, not against the average. Cabinets at ceiling level or near a heat source can exceed 50 degrees Celsius in summer, so the specification should cover a wide-temperature range of -20 to 60 degrees Celsius or wider, with a wide-range power input and proper protection.