Intsoft Intsoft Visual AI SOP System

AI Visual Monitoring of Workstation SOP Execution

The workstation is the smallest unit of SOP execution. We turn the video from the workstation camera into a structured sequence of actions, check each one against the standard process, and judge compliance with work standards such as dress, cleanliness and ESD control. Recognition runs locally at the edge and does not consume line takt time.

Isometric diagram of a single monitored workstation with camera and status panel

Key takeaways

  • Action granularity is selectable, from whether a process step was executed down to whether a specific action occurred.
  • One class, many poses: posture variation for the same action is adapted workstation by workstation.
  • Compliance checks cover personnel standards such as dress code, cleanliness and ESD control.
  • Inference is local, so a network outage does not affect decisions and production images are not sent to the public internet.
  • Action data can be published to a dashboard so team leaders see execution status across stations in real time.
Isometric diagram of an industrial camera and an edge computing box connected by a single cable
The station-level hardware: one industrial camera plus one edge computing box

The hard part of workstation monitoring is not whether an action can be recognised, it is how to turn recognition output into a process judgement your process engineers will accept. Two operators performing the same action look very different, and the same process step looks different on different products. The system has to normalise that variation into a single conclusion: the step was completed. Those judgements are what drive error prevention at the station.

Recognition Granularity: Four Levels

Action recognition from coarse to fine
LevelWhat it decidesTypical use
Process step levelWhether a given process step was carried outSOPs with many steps, where you only need to know that each one was done
Action levelWhether specific actions such as picking, assembling, fastening or inspecting occurredKey steps where the action itself has to be confirmed
State levelWhether a part is present, correctly positioned and correctly orientedDirect detection of missing or wrongly fitted parts
Work standard levelWhether personnel standards for dress, cleanliness and ESD control are metFood safety, cleanroom and ESD-sensitive operations

Engineering Practices That Reduce False Alarms

  • Consecutive-frame confirmation: a single frame never triggers an alarm. Several consistent frames in a row are required, which filters momentary occlusion and misclassification.
  • Per-station thresholds: confidence thresholds are calibrated for each station on its own instead of with one site-wide parameter set. Camera placement is the other half of that work - see how to choose visual SOP cameras.
  • Whitelist for non-standard actions: known legitimate variations are registered and passed through, so 3C electronics fastening stations do not alarm on each individual working style.
  • Time-window policy: during changeovers, line clearance and equipment commissioning, judgement can be temporarily downgraded or suspended.
  • Alarm tiering: prompts and blocks are separated, and only high-risk steps stop the line.

What This Capability Will Not Do

The edges are worth stating explicitly. The system reports on actions and parts; it does not identify people by default, does not judge ergonomics or safety behaviour, and cannot see through occlusion or beyond the imaging resolution.

  • A part smaller than a few pixels at the working distance cannot be judged reliably, so the field of view has to be chosen with that in mind.
  • Where the operator has a legitimately non-standard working style, the whitelist is what keeps the alarm count tolerable - see operator resistance for why that matters on the floor.
  • It is not a replacement for the driver torque record or the scanner part number; those confirm different facts.
  • Preconditions: the action list for each station, and a short tuning period to calibrate thresholds.

Common questions about workstation monitoring

How many cameras does one workstation need?

One is enough for most stations. Where the field of view is heavily occluded, or both hands and several material positions have to be watched at the same time, two cameras covering different angles may be needed.

How long does it take to adapt action recognition?

Standard actions are usually adapted within a few days. Fine judgements that have to distinguish specification, such as screw size, take longer and are normally resolved during the pilot.

Can it identify the operator?

Technically the system can link to badge scanning or face association, but this is off by default. Whether records are tied to individuals is up to you and depends on your management requirements.

Related content

Other products & capabilities

View all

Applicable industry solutions

View all

Want to see how this capability fits your line?

We provide on-site feasibility assessment, pilot deployment and complete MES integration.

Request a free trial plan