The Cloud Ingress Trap
Industrial IoT projects often begin by forwarding every available signal upstream. That approach may be appropriate for a bounded diagnostic capture, but continuous operation can create avoidable network, storage, and query load.
The engineering question is not whether cloud or central processing is useful. It is which detail must cross the site boundary, which evidence must remain available locally, and which events require immediate delivery.
Signal volume depends on tag count, sampling interval, change frequency, payload shape, and retention policy. Before filtering, teams should measure these characteristics and classify the evidence needed for operations, troubleshooting, compliance, and model development.
This is why modern industrial architectures rely on Smart Filtering at the edge.
- Measured: Post-filter data volume
- Policy-based: Retained evidence detail
Data Volume: Before vs After Edge Filtering
| Item | Data points per hour |
|---|---|
| Raw (no filter) | 3,600,000 |
| Deadbanding | 720,000 |
| Aggregation | 60,000 |
| Threshold-only | 1,000 |
Illustrative figures for one fast signal, to show the order of magnitude each strategy works at. They are not a measurement; your own reduction depends on the signal.
Observed performance depends on workload shape, node capacity, and deployment design.
What is Smart Filtering?
Smart Filtering (often coupled with Edge Analytics) is the process of examining, reducing, and aggregating raw machine data before it leaves the physical factory building.
Instead of treating the edge gateway as a transparent relay, smart filtering applies an explicit policy to each signal: transmit the sample, aggregate it, retain it locally, or create an event window.
The policy must also define maximum reporting intervals, quality changes, exceptions, audit retention, and recovery behavior. Filtering without these controls can hide the evidence needed for diagnosis.
10ms Sampling → Smart Filter & Deadband: Data Storm
Smart Filter & Deadband → Optimized Data: Policy-selected data
How smart filtering changes data volume
A proper edge computing platform employs three primary filtering strategies to convert a massive "Data Storm" into a highly optimized "Smart Stream":
Deadbanding (Change-of-State)
In many industrial processes, a value doesn’t change for hours. Why send a temperature reading of 72°C every second for a straight hour? With deadbanding, the edge gateway sends a new payload when a value changes by a configured percentage or absolute amount, or when a maximum reporting interval expires. The reduction depends on signal variability, polling frequency, threshold choice, and the events that must remain observable; it must be measured on representative process data.
Time-Based Aggregation
Sometimes, you don't care about the microsecond fluctuations; you just want the trend. For example, an edge gateway may collect readings over a defined window, calculate average, minimum, maximum, and sample count, then send the aggregate with its time boundary and quality. Aggregation reduces resolution; it does not preserve every transient.
Event-Driven Transmission
The most advanced form of filtering. The Edge Gateway runs continuous localized logic (such as a local Rule Engine). It monitors a high-speed vibration sensor on the local network and can avoid routine cloud transmission until the vibration crosses a warning threshold. Then it streams high-resolution data surrounding the anomaly to the cloud for deeper analysis.
The Proxus Approach to Edge Filtering
At Proxus, we recognize that Edge Computing is fundamentally an exercise in bandwidth and cost optimization.
When you deploy a Proxus Edge Gateway, it isn't just a protocol converter: rules and pipelines run on the gateway itself. The gateway keeps no configuration of its own; everything it runs is defined on the hub and delivered to it. Without writing code, you can use pipelines to:
- Normalize tags: convert units, scale raw PLC counts, rename tags and add context.
- Filter, drop or route values before they are stored and published.
- Raise alerts at the edge with expression and anomaly rules, which keep running when the network is down.
- Buffer delivery during outages: 48 hours by default, up to 30 days per gateway, replayed with original timestamps (the edge historian guide covers sizing and replay).
Proxus does this shaping at the source; it does not aggregate at the edge. Where a consumer needs averages or window summaries, compute them in the historian or in the consuming system, so the raw values that reached the hub stay available for investigation. When a C# function is the better tool, it is built and deployed from the hub like any other configuration.
Conclusion
Smart filtering can control upstream load by processing data close to its source. Every filter is also an information-loss decision, so raw retention, aggregation, event windows, and delivery policies must be designed together and verified. In Proxus, this capability belongs to the edge data pipeline of the Industrial Data Platform.
When this may not be suitable
- Lower-frequency telemetry may not justify full distributed complexity.
- Small single-line plants may prefer simpler architectures first.
- Strict legacy constraints may require phased adoption.
- Safety-critical closed-loop control should remain in PLC/Safety PLC layers.
Outcomes depend on workload profile, hardware capacity, and deployment topology.
Frequently Asked Questions
Won't filtering discard data I might need later?
It can. A deadband or aggregate intentionally removes resolution. Configure exception reporting, maximum intervals, quality-change handling, and raw retention according to each signal's operational and audit requirements. Validate the result with known events before production rollout.
What is a deadband and how do I set it?
A deadband defines the minimum change threshold before a new value is transmitted. For a temperature sensor with ±0.5°C accuracy, a 1°C deadband is sensible - changes smaller than sensor accuracy are noise. For vibration monitoring used in predictive maintenance, use much tighter deadbands (or none) to capture high-frequency patterns.
How does Smart Filtering interact with Store and Forward?
They work in sequence: Smart Filtering reduces the volume of data generated, Store and Forward ensures the delivery of the remaining data during outages. Together they minimize both bandwidth costs and data loss risk.
References
- OPC UA Part 8 (Data Access) - Defines deadband filtering for analog items, the standard model adapted by many edge platforms.
- ISA-95 / IEC 62264 - Provides the data categorization framework for determining which data to filter vs. preserve.
Last updated August 6, 2026