FactoryThread

Guide

Data science in manufacturing, minus the hype

Where models genuinely pay, where good rules beat machine learning, and what has to exist before either works.

9 min read

Rules first, models second

A large share of the value attributed to machine learning in plants comes from consistent thresholds, good context and timely alerting. If a simple rule captures most of the benefit, ship the rule — it is explainable, cheap and maintainable.

Reserve models for problems where the relationship is genuinely multivariate and non-obvious: yield drivers, energy optimisation, quality prediction from process parameters, remaining useful life on well-instrumented assets.

The prerequisites nobody skips successfully

Labelled outcomes: you need reliable records of failures, defects and downtime reasons. Aligned time bases across sources. Sufficient history at the right resolution. Context that ties a signal to a part, order and asset state.

Most failed pilots failed here, not in the algorithm.

Deployment is the hard half

A model in a notebook changes nothing. It has to run on live data, on schedule, with monitoring for drift, and its output has to reach an operator or a system that acts. Plan for the Flow and the write-back before the modelling.

Trust is earned by exposure

Show the inputs, the confidence and the past accuracy alongside every prediction. Run it in shadow mode next to the current decision for a few weeks. Operators adopt models that have been right in front of them, not models that arrive certified.

Start here

Ask one question. Get one answer. Today.

Start on the free tier with a single connection, or bring a real question to a working session and we will build the first answer with you.

Free tier · No credit card · Cloud or self-hosted