GlossaryKind of term: Manufacturing termMachine learning
What is machine learning?
Machine learning is a branch of AI in which a model derives a relationship from examples rather than from rules that someone writes down.
Nobody writes the relationship out as a formula; it lives in the model's parameters. Training sets the model's parameters so that its outputs fit the examples as closely as possible. For a cycle time, that would mean many parts with geometry, material, machine and measured time, and a model that predicts a time for a new part from them. Statistical regression, decision trees and neural networks all count as machine learning.
What does machine learning depend on?
A model can be no better than the examples it was built on. It can capture relationships nobody wrote down, but also ones that were pure chance; this is called overfitting, and it shows only on examples held out of training. Outside the range of the examples, such as a new material, a different machine or a much larger part, a prediction is uncertain. Whether an output can be explained depends on the type of model: a regression shows its coefficients, a large neural network does not directly.
How does machine learning work in practice?
A booked time with a tool breakage or a breakdown in it is not an ordinary time. Use such times unflagged and the exception becomes part of the model; leave them all out and the model can predict too optimistically. In one-off and high-mix work the first run is often the normal case, not the exception. The booking itself matters too: whether setup, machining and waiting are booked separately, and against the right part. A model from one shop does not carry over to another with different machines, tooling and programming style. A predicted time remains an estimate and belongs beside the measured time, not in its place.
What is machine learning often confused with?
Machine learning is one branch of AI, alongside rule-based systems for example, so the two words are not interchangeable. Adjusting standard times to actuals is something else again: there you change an existing standard per operation yourself and can see what moved, whereas machine learning builds a model whose relationships come from the data.
Frequently asked questions
Can I use my actuals to train a model?
Only if you know what is in them. A time that includes a breakdown or tool breakage is an exception. Record per order the conditions under which the time was booked, so exceptions are flagged rather than silently kept or left out.
How many examples does a model need?
That is not fixed. It depends on how many characteristics count, the spread in your work and how precise the prediction has to be. Test the model on examples held out of training.
What is the difference between machine learning and adjusting standard times?
Adjusting standard times means changing an existing standard per operation to match measured times, so you can see which standard moved and why. Machine learning builds a model whose relationships come from the examples; whether an output can be explained depends on the model type.