An industrial AI prototype can produce useful predictions and yet still not be part of the operational process.
This is a very important engineering issue when you go from the experimental phase to actual deployment.
A prototype may well demonstrate a particular capability. A solution that is going to be used in production must also take into account how the information is received, understood, reviewed and used.
The issue isn't just about the performance of the model. It's the design of the entire workflow around it.
Start with the operational event
Consider a manufacturing environment in which an AI system spots a potential equipment problem.
The model's output may be of use, but it's not the whole of the operational process.
A simplified information flow might look like this:
Equipment activity
|
v
Data collection
|
v
Context and validation
|
v
AI-generated insight
|
v
Human review
|
v
Existing operational workflow
|
v
Outcome evaluation
Each stage has a function.
Data collection provides the raw data that enables the analysis. Context and validation are used to determine whether the data is fit for interpretation. The AI then produces an insight. The human judgement as to how that insight should be treated is provided by human review and existing operational processes.
- Separate data collection from interpretation
A fairly common design error is to treat the data you've collected as if it's already established the operational conclusion.
A sensor reading, an equipment event or a production update is an observation. What it means is entirely dependent on the context.
Systems should therefore distinguish between:
The original observation
The source and time of the observation
Relevant equipment or process context
Any validation applied
The resulting interpretation
This separation makes it far easier to see how the insight was generated.
- Make the output relevant to its intended user
Different operational roles require different information.
A maintenance fella might be after the detail on a particular piece of equipment event, whereas a production manager would be after the wider picture of how that event relates to production activity.
The same AI output may well need to be presented in different ways depending on the intended use.
A technically accurate result is no guarantee of a useful operational message.
- Keep recommendations separate from execution
An AI generated recommendation should not be confused with an action which has already been approved or which has already been completed.
A system should make it clear whether information represents:
An observed event
A detected pattern
A prediction
A recommendation
An approved operational decision
A completed action
Maintaining those distinctions facilitates clearer communication and helps to avoid any confusion between analysis and execution.
- Include a feedback and evaluation stage
Deployment shouldn't end when the AI system has produced an output.
Organisations need some way of judging whether the output was useful and whether the intended operational process in fact achieved the desired outcome.
This could involve a review of the original event, the insight that was generated, the decision that was taken and the eventual outcome.
Such feedback can be used by the team to identify weaknesses in data quality, in the way the data is interpreted, in the workflow design or in the performance of the system.
- Design for operational variation
A workflow that works in one production environment may well need tweaking in another.
Equipment configurations, processes, data available and user responsibilities will all vary.
A scalable design should allow you to work out which bits can be re-used and which will need to be adapted.
Final consideration
Industrial AI engineering should take into account the full journey from observation to operational outcome.
A good model is important, but so are data context, output design, human review, workflow integration and evaluation.
When you look at all of those elements together then organisations have a much clearer basis on which to go beyond a successful demonstration and assess whether the AI solution is in fact suitable for day to day industrial use.
Explore industrial technology and venture development perspectives at https://apertureventurestudio.com/
Top comments (0)