Planning real-time industrial data collection begins with how quickly a decision needs trustworthy information. Sensor response, PLC update rate, polling interval and screen refresh are different measurements. Combining them under a single “live” label can give users confidence that the system does not support.
Create a signal inventory
For each signal, record equipment identity, meaning, unit, data type, source interface, permitted polling rate and quality state. Use a physical description such as “north cold-room ambient temperature” rather than an ambiguous label. Document whether a meter value represents instantaneous power or cumulative energy.
Verify the register map against the correct device and firmware documentation, then compare it with a known field condition. Scaling, signedness and word order can produce a plausible but incorrect number. During commissioning, compare the raw read, the normalised record and the displayed value.
Distinguish sampling from presentation
Fast sampling does not require sending every point to a browser at the same rate. A collector may retain events while a screen presents a slower summary. Conversely, refreshing a page frequently does not make the source measurement newer. Users need the last source timestamp.
| Time | Meaning | Common mistake |
|---|---|---|
| Measurement | When the physical value was produced | Replacing it with server arrival |
| Collection | When the collector obtained it | Treating it as exact source time |
| Processing | When a service handled the record | Hiding network delay |
| Presentation | When the screen rendered it | Treating a new screen as new data |
Clock synchronisation can fail. If source time is unreliable, carry that uncertainty in the quality information rather than claiming an exact event order. Storage should use a consistent time standard while the interface makes its display time zone clear.
Illustrative volume calculation
Assume 20 measurement points each produce one record every 10 seconds. That gives 20 × 8,640 = 172,800 records per day. This is a record count, not a storage estimate. Actual byte size depends on payload structure, packaging, indexes and operational overhead; measure representative records before estimating capacity.
A short machine event can occur between periodic reads and disappear. An event record or a device-maintained counter may be more suitable for that signal. Slowly changing ambient temperature may need a different approach. Collection frequency should preserve the required information without adding unnecessary device load.
Missing data is a distinct state
A valid zero, sensor failure, unreachable device and absent record are not equivalent. A trend should not silently connect across a missing interval. If the last good value remains visible, show its age. Alarm rules must not repeatedly treat stale data as a new measurement.
Store-and-forward records belong at their event time when they arrive late. Preserve a stable identity so replay does not create duplicate observations. Decide whether late data changes historical reports and how users learn that a previously incomplete period has been updated.
Commission the failure cases
Test normal production, network loss, clock drift, sensor faults and reconnection separately. Report delay distributions and data loss rather than relying only on an average. Remote monitoring is not the same requirement as a deterministic machine control loop.
The completed pilot should include the signal dictionary, reasons for each sampling policy, measured volume and documented failure behaviour. Those records prevent the assumptions of the first installation from being lost when new equipment is added.
Turn a delay requirement into a test
Suppose a maintenance team wants a status change to become visible within a defined operating interval. Agree on where the interval begins and ends: a physical transition, the PLC updating its state, the collector reading that state, or the browser receiving it. Measuring only the final network request ignores everything before it. If the source cannot timestamp a transition, report the uncertainty introduced by polling rather than presenting the detection time as an exact occurrence time.
Use a controlled sequence of events with an independent reference during commissioning. Include two transitions closer together than the usual polling interval and an event during a temporary disconnection. Observe whether the source retains them, whether the collector sees them and how the report presents them. A successful slow test does not establish that brief production events will be captured. Conversely, a slow environmental measurement does not automatically need the same collection rate as a machine-state transition.
Record a distribution of observed delays under representative load. The typical result, occasional long delays and periods with no valid result answer different questions. Identify the workload and network conditions used in the test so a later comparison is meaningful. A desktop demonstration on an unloaded network is useful for development, but it is not a timing guarantee for the installed system.
Define recovery without rewriting history
Consider a gateway that reconnects with fifteen minutes of buffered observations. Sending all old records before collecting new ones may increase the age of the current view. Prioritising current measurements without a bounded recovery plan can leave history incomplete indefinitely. Specify both responsibilities and measure whether they compete for device access, network capacity or server processing.
Late records need a clear reporting policy. A shift report can display a provisional state until the recovery window closes, or publish a later revision with the corrected coverage. The choice follows the operating process; it should not be hidden in an undocumented database job. Keep a count of rejected records and the reason category, while avoiding unnecessary raw values in diagnostic logs. During discovery, bring examples of the shortest event that matters, the longest outage to cover and the maximum useful age of a displayed reading. These requirements make collection planning more concrete than a request for a generally live dashboard.