DEV Community

Ajay Oloruntoba
Ajay Oloruntoba

Posted on

Closing the Feedback Loop in Industrial AI

A well designed industrial AI system should not end here.
Within connected industrial systems there should be methods of checking what has happened after an authorised action.

Difference between Decision Process and Decision and Action Process

Let’s say you have an industrial application that detects the need for action.

It takes the observations and makes a recommendation. It then goes through an authorisation process so that an action can be taken.

We now have a number of questions:

Has any action been taken?

Has the correct equipment responded?

Has the expected condition been observed?

Is the feedback source trusted?

Do we need to make another decision?
Without this we could be making incorrect assumptions about the outcome.

Basic Architecture

Asset Identification
|
Sensor Observations
|
Data Validation
|
AI Decision
|
Authorisation
|
Authorised Action
|
Outcome Observation
|
Result Verification
|

Next Decision / Review

The last few steps of this process bring us back to the physical world.

*Considerations to make when implementing
*

  1. Command or Confirmation?

A command will indicate that a command has been sent however it will not confirm that the physical action has been taken.

The system should take into account what needs to be observed to confirm it has been completed.

  1. Delayed or No Feedback

There will often be delays or disruptions to communications and sensors within an industrial system.

The process should have a method to identify if it has been confirmed, failed to take action or if it has not been confirmed yet.

These are separate processes and should not be treated as one.

  1. What happens if it can’t be confirmed?

If the result doesn’t match the expected result we need a defined process for how the system will respond.

Depending on the application this could be reviewing the situation, alerting, escalating to an authorised individual or using an approved recovery process.

Review

The whole process

Testing just the AI will not show that the whole process works as intended.

It should take into account the data quality, the authorisation process, command delivery, outcome feedback and exception handling.

Also consider how the system will react when there is incomplete information or communication issues.

Final Thoughts

When we start linking the decisions made by the AI to observable outcomes it adds real meaning to the process.

Rather than asking ‘what did the model recommend?’.

Ask "What happened after the authorized action, and how do we know?"

Find out more about our Physical AI and AIoT research here:

https://apertureventurestudio.com/physical-AI-aiot-research/

IndustrialAI #AIoT #SoftwareEngineering #IndustrialAutomation #Manufacturing

Top comments (2)

Collapse
 
ajay_oloruntoba_b8a722b0f profile image
Ajay Oloruntoba •

discuss

Collapse
 
ajay_oloruntoba_b8a722b0f profile image
Ajay Oloruntoba •

Ask "What happened after the authorized action, and how do we know?"