On the factory floor, AI is a sensor-reading colleague with a maintenance rota
AI in manufacturing, logistics & supply chain
The question: How do predictive maintenance, quality vision and routing actually earn their keep in operations?
Industrial AI is unglamorous and lucrative: machines stream sensor data, cameras see every unit, routes burn diesel. You have already trained the maintenance classifier in the lab — this lesson puts it in its industrial context.
What the lesson covers
Predictive maintenance is the flagship: sensors (vibration, temperature, torque, wear) feed a failure classifier, and the alert threshold is an economics dial — your lab found the optimum accepts many false alarms because a missed failure costs 25× an inspection. The industrial reality adds drift (seasons, product mix), sensor decay, and the workflow question: who gets paged, who schedules, who overrides.
Quality inspection is computer vision's home turf: cameras catch defects at line speed that tired humans miss. The gotchas are representativeness (a model trained on day-shift lighting fails at night) and the appeal path — a unit wrongly scrapped is money, a defect wrongly passed is reputation. Convolutional networks (the vision architecture) learn visual features layer by layer, the same representation-learning story as Lesson 3.
Logistics optimisation is often classical AI (Lesson 2's "not everything needs an LLM"): routing, load-building and scheduling are constraint problems where solvers beat neural nets — with ML forecasting demand and traffic as inputs. The blend is composite AI in work clothes.
Supply-chain resilience is the executive layer: risk scoring of suppliers, disruption simulation, and document AI on the paperwork tide (customs forms, bills of lading — your multimodal extraction from Lesson 6 with a hard hat on). Every deployment inherits the operations discipline: monitoring, drift alerts, and humans owning exceptions.
The constraint that makes factory AI different from every other deployment in this course is that the plant does not stop for your model. A line running three shifts cannot pause while somebody investigates an alert, which means every prediction has to arrive with a default action attached: what happens if nobody looks at this within an hour? An alert with no default is not a control, it is a notification — and the failure mode is not a dramatic incident but a queue of unread alerts that everybody learned to ignore in the second week. Design the default first, then the model. It is the same lesson as the autonomy ladder, arriving from the direction of shift patterns rather than of risk.
Key points
- Maintenance thresholds are economics dials (your lab: recall 0.93 at the cost optimum); drift and paging workflows are the real work.
- Vision QC fails on unrepresentative data (shifts, lighting); design the appeal path for false scrap.
- Routing/scheduling are constraint-solver territory — ML forecasts feed them; composite AI in practice.
- Document AI digests the supply-chain paperwork tide with Lesson 6's validation rules.
- A plant does not stop for your model. Every prediction needs a **default action** for when nobody looks within the hour — an alert without one is a notification, and notifications get ignored by week two.
Framework — Sense → Predict → Schedule → Verify → Learn
The industrial AI loop. Sensors feed predictions; predictions feed schedules only through economic thresholds; humans verify exceptions; every outcome retrains the loop.
The lab
Extend your maintenance-lab result into a full industrial deployment design.
Open this lesson, its lab and its quiz
Sources and further reading
- Predictive maintenance lab (your trained model) — this academy
- NIST AI RMF — reliability in deployment — NIST