TechnologyPredictive Intelligence
Can we understand waste before it becomes waste?
Patterns, anomalies and historical context to investigate where material is being lost, without confusing a signal with a cause.
CENTRAL QUESTIONCan we detect a loss before it becomes normal?
01CONTEXT
Scrap is usually discovered late: in a monthly close, in an audit, or when the disposal cost has already shown up. By then, what can be done is to explain it, not to prevent it.
A loss of material is rarely an isolated event. It is usually a gradual change in the behavior of a line, a shift or a process, and it becomes invisible precisely because it is gradual: each week looks like the last, until the new becomes the normal.
02WHY LOOKING AT THE LATEST DATA POINT IS NOT ENOUGH
The latest data point rarely says anything on its own. A high number can be an anomaly, a product change or just a longer shift. Comparing against a global average does not help either: what is usual for one line is not usual for another.
To tell a deviation from normal variation you need a baseline per process and per shift, and events that arrive with consistent context. That context is what Physical Intelligence aims to provide.
An anomaly is not a cause.
03OUR HYPOTHESIS
If material events are observed over time, it is possible to define what is usual, detect when something departs from it and open an investigation with context. But the signal is only the first step: what follows is carefully separating what was observed, what might explain it and what would have to be checked.
- 01BASELINEWhat is usual per process.
- 02EVENTSWhat gets recorded.
- 03DEVIATIONSomething departs.
- 04SIGNALWorth a look.
- 05INVESTIGATIONWe ask why.
- 06INTERVENTIONWe act.
- 07MEASUREMENTWe check the effect.
04HOW AN ANOMALY IS SEPARATED FROM A CAUSE Illustrative example
Davin is designed to present each finding in four parts that are not mixed. What was observed is a recorded fact. A possible factor is a hypothesis, shown with evidence for and against. The evidence says where each data point comes from. And the next check says what would have to be verified to confirm or rule out the factor.
A possible factor is never presented as a finding. That discipline avoids the most common mistake in scrap analysis: giving a correlation the tone of a conclusion.
- OBSERVED ANOMALY
A line records scrap above its usual behavior.
OBSERVED- POSSIBLE FACTOR
A product changeover coincides with the increase.
HYPOTHESIS- EVIDENCE
For: it matches in date. Against: another line running the same product did not change.
EVIDENCE- NEXT CHECK
Compare this line’s scrap before and after the product change.
PENDING
05CURRENT BOUNDARY
We do not claim causal prediction or validated automatic alerts. Today the analysis depends on the information available: without a baseline per process, the signal comes down to comparing trends between lots, and the process is sometimes recorded by hand. More precise detection is a direction of work.
06WHAT WE’RE LEARNING
Without consistent context, a deviation cannot be interpreted: it is unclear whether it is a real change or a difference in how it was recorded. Baselines need to be per process and per shift. And the value is not in the alert, but in the question it prompts in whoever knows the operation.
07WHAT COMES NEXT
Longer event series, better comparison between lines and processes, and a more direct connection to the decisions of Davin. Useful prediction depends on accumulating context; it cannot be improvised.