Measurement model · research
AI belongs at the measurement signal layer.
We research an AI model within the sensor to analyse signal patterns alongside physical calculations, aiming to produce readings with confidence assessments.
- Position
- Inside the sensor, before a reading is formed
- Stage
- Research: model design and simulation tests
- Next
- Collect real signals
- Target output
- Readings with confidence assessments
Our view
Reading confidence depends on measurement conditions
-
How readings are formed
Sensors calculate readings from measurement signals using physical formulas. Those formulas assume particular measurement conditions.
-
Conditions can change
Bubbles, particles, temperature or installation may depart from those conditions. A reading alone often cannot show whether the conditions hold.
-
Signals carry more information
A reading is one result calculated from a signal. We research learned models to analyse the richer signal patterns alongside physical calculations and assess reading confidence.
We therefore see greater scope for AI before a reading is formed; planned cloud analysis will build on that layer’s outputs.
Scope
What the model aims to do, and its limits
-
Research aim
Research aims
-
Analyse signal patterns alongside physical calculations.
-
Support reading assessment under complex conditions.
-
Produce readings with confidence assessments.
-
-
Boundaries
Limits
-
Do not replace physical formulas.
-
Not metrological certification; no replacement for calibration.
-
No accuracy figure is published before independent verification.
-
Validation
Research conclusions need real-data validation
Test on simulated data, then compare real signals against independent reference readings. Return to design if the results do not hold.
-
In progress
Model design and simulation tests
Design the model framework and test it on simulated data; a first round of tests is complete.
-
Next
Collection of real signals
Record raw signals on a test pipe with independent reference readings. Collection has not started.
-
Later
Validation on real data
Compare with conventional calculations against independent references; return to design if the results do not hold.
-
Later
Embedding in the sensor
Compress the validated model to run on the sensor processor.
Next steps