What is Edge computing?

Edge computing and local data processing

Put computing where the data becomes useful

Edge computing places processing and, where needed, storage close to the source of the data. On a factory floor, that might mean analyzing machine measurements on a nearby computer before sending selected information to a remote service. The “edge” describes where a workload runs in relation to devices and the network, rather than a particular brand of hardware.

Cloud computing supplies shared computing resources; edge deployments handle selected tasks on devices, gateways or nearby servers. The Internet of Things concerns connected devices, while industrial IoT applies that connectivity to industrial operations. A connected machine does not necessarily perform local analytics. Connectivity and the placement of computing are separate design choices.

How does an edge device fit into an automation system?

An edge node could be an industrial PC, an embedded computer or a gateway with application processing. Some PLCs support additional analytics, but machine control and a broader edge application serve different purposes. A data logger captures measurements; its ability to analyze them depends on its hardware and software. SCADA provides supervision, alarms and operational visibility. These functions can share hardware without becoming interchangeable concepts.

A practical example: monitoring motor vibration

Suppose a sensor produces 10,000 vibration samples per second. At 16 bits per sample, one channel generates 20,000 bytes per second, or 72 MB per hour, before timestamps, protocol overhead and other channels are included. Here, one MB means 1,000,000 bytes.

A local application can calculate RMS and other indicators from the vibration signal, track changes and raise an alarm when a configured condition is met. It can send summaries and event records upstream while retaining raw samples around an event for investigation. A summary is not a lossless substitute for the original signal. Sampling rate, retention time and storage capacity need to be planned together.

Process data locally and send selected information to the cloud Site / local network Internet connection Machine / sensor Raw measurements Edge device Analyze + store Cloud Long-term analysis Raw data Summary / event Local display Alarm / status
Local processing can be designed to keep working without internet access.

Divide the work according to its requirements

An example division of responsibilities
Workload At the edge In the cloud
Immediate analysis Evaluate local measurements. Combine incoming results.
Data retention Local records and an upload queue. Long-term records across sites.
Machine learning Run a deployed model. Train models and analyze larger datasets.

This is a design example, rather than a mandatory allocation of tasks. A hybrid architecture combines local and cloud resources. Fog computing commonly describes multiple computing tiers between devices and the cloud; its meaning can overlap with edge computing, depending on the terminology used.

Design offline operation instead of assuming it

Local execution does not automatically remove cloud dependencies. An application can continue through an internet outage only if the software, model, configuration and authentication resources it needs remain available locally. A required remote service can still prevent an otherwise local workload from completing.

With store-and-forward, records wait in a local queue until connectivity returns. Available storage capacity and message retention settings limit how much can be preserved. Define what happens when storage fills, how records are ordered and how duplicate deliveries are handled. Test loss of connectivity and device restarts before relying on the arrangement.

Lower latency does not guarantee a control deadline

Moving a workload closer can reduce network delay, but the operating system, application execution, queues and local network still affect response time. Real-time operation requires meeting a timing deadline, rather than merely being fast on average. Real-time operating systems address part of that problem. Critical loops and safety functions belong in a control architecture that meets their requirements; a general-purpose edge analytics computer does not establish that suitability by itself.

Can older machines participate?

Often, with suitable interfaces and software. RS-485 describes an electrical interface; a protocol such as Modbus defines how data is exchanged. An Ethernet connection alone does not tell an application what a measurement means. Register addresses, data types, scaling, units and timestamps must be interpreted correctly. A successful connection is not proof of a valid measurement.

Does edge computing require AI?

No. Filtering, averages and threshold checks are useful edge workloads without machine learning. When a model is involved, distinguish training, which builds the model, from inference, which applies it to new input. An industrial analytics deployment should track model versions, performance on actual site data and execution time. Neither local processing nor an AI label guarantees a correct diagnosis.

Security and the measures that matter

Keeping data on site does not automatically secure it. Manage device identities, access permissions, updates and logs. Restrict communication between the machine network and external networks, and protect remote access and data transfers appropriately. Industrial cybersecurity also includes the physical access available to an edge computer installed near machinery.

Evaluate a deployment through response time, upstream traffic, outage tolerance and maintenance effort. In live monitoring, check how old the displayed measurement is. During record analysis, look for gaps and duplicates. In an Industry 4.0 project, the useful outcome is a measurable improvement in operations, supported by a system the team can maintain.

Technical references: IRTF RFC 9556 — IoT Edge · NIST — Fog Computing · AWS — Local processing · Microsoft — Offline operation · NIST — OT security