Stopping a bad run before the product is lost
A model that predicts the yield of a production run while it is still running — accurate enough to act on, and transparent enough to trust.
A run that goes wrong is only discovered once it has finished — and at this product value, every wasted unit is expensive.
A hybrid model that reads the machine’s existing sensors and predicts the eventual yield mid-run, with its reasoning visible so an engineer can act on it.
Yield predicted at 98% accuracy, early enough to intervene — and with the likely cause attached, so maintenance knows where to look.
This client makes high-value products on one specialised machine, and asked not to be named, so the method and the numbers are here and the identity is not. The machine is already densely instrumented: sensors record what happens throughout every run. What that data was not doing was warning anyone early enough.
What the model does
Predicts mid-run
It reads the sensors already in the machine and projects where the run will end up, long before the product is finished.
98% accurate
Accurate enough that an operator can act on a warning rather than wait for certainty that arrives too late to matter.
Shows its reasoning
Every prediction carries the factors behind it. An engineer sees why the model expects a poor run, not only that it does.
Points at the cause
The signals that predict a poor yield also indicate what is driving it, so mechanics start where the problem is instead of searching for it.
The machine was already producing plenty of data. The problem was timing: nothing in it told anyone a run was going wrong until the run was over — by which point the product, and the hours that went into it, were already lost.
How we worked
We started with the simplest thing that could work, and added complexity only where it earned its place.
-
Simple methods first
Statistical baselines on the existing sensor data, to establish how much could be predicted without any machine learning at all.
-
Machine learning where it paid
Where the simpler approaches ran out we moved up to machine learning, using a neural network — and measured the gain rather than assuming it.
-
Explainability as a requirement
Every model was judged on whether an engineer could see why it said what it said, not only on how often it was right.
What changed on the floor
Two things — and the second turned out to matter as much as the first.
-
Intervene during the run
A predicted poor yield now arrives while there is still something to do about it: correct the process, or stop it, rather than finish and scrap.
-
Less time spent diagnosing
The model points at the conditions behind a bad prediction, so mechanics know where to look. Faster fixes mean less downtime, and more of the machine’s time spent producing.
Why hybrid, and not just machine learning
A machine-learning model on its own — in our case a neural network — would have been marginally more accurate on paper and far less useful in the control room: nobody stops an expensive run on a number they cannot interrogate. Pairing it with traditional, interpretable methods keeps the accuracy where it matters and the reasoning where people can see it, which is what makes a prediction something to act on rather than something to argue with.
Key outcomes
-
Yield predicted at 98% accuracy while a run is still in progress
-
Poor runs caught early enough to intervene, rather than scrapped at the end
-
The likely cause surfaced alongside the prediction, so maintenance starts in the right place
-
Less downtime spent diagnosing, and more of the machine’s time spent producing
Ready to unlock value from your data?
Let’s connect for a no-obligation conversation about your goals and how we might support them.