Industrial machine learning by country

For 45 countries: share of manufacturers using AI, IFR robot density, labour cost and sector mix, with the machine-learning use case most likely to pay back first in each country's dominant sector, every figure sourced, plus a downloadable workbook.

Readiness data and the first use case

Click a column header to sort. Adoption and robot density as published (Eurostat 2024–2025; IFR World Robotics 2025, 2024 installations, revised methodology). Labour cost in the source currency and year. The leading sector comes from official statistics; the use-case column is editorial and explained below.

The eight first use cases, by sector type

Changeover and minor-stop prediction

Automotive (Tier-1 assembly) (typical OEE 70–85%, world-class 85%, dominant loss: changeover & minor stops). Discrete assembly loses most OEE to changeovers and micro-stops; sequence models on PLC state + takt data predict them 10–30 min ahead. Needs per-station cycle timestamps.

Inline defect classification

Electronics & semiconductors (typical OEE 75–90%, world-class 88%, dominant loss: reduced speed & defects). SMT/semiconductor lines already produce AOI/SPI images and test logs; vision and tabular models cut false calls and escape rate. Needs labelled defects and traceability per board.

Startup-loss and giveaway reduction

Food & beverage (typical OEE 60–80%, world-class 82%, dominant loss: changeover & startup losses). Food lines lose OEE at startup and changeover and margin on overfill; regression on recipe, temperature and speed data targets both. Needs batch IDs joined to line data.

Batch deviation early warning

Pharmaceuticals (typical OEE 35–45%, world-class 70%, dominant loss: validation & changeover). Validated environments favour monitoring over control: anomaly detection on CPP/CQA trends flags deviations before release. Needs 21 CFR Part 11-compliant data capture.

Soft sensors and anomaly detection

Continuous process (chemical, pulp) (typical OEE 80–92%, world-class 92%, dominant loss: unplanned stops). Continuous plants have dense historian data; soft sensors replace lab delays and anomaly models catch drift. Needs a historian with tag metadata.

Tool wear and unplanned-stop prediction

Metal fabrication / machine shop (typical OEE 45–65%, world-class 78%, dominant loss: setup & idling). Machining cells lose time to setup, idling and tool breakage; spindle-load and vibration models predict wear. Needs CNC data (MTConnect/OPC UA) per machine.

Cycle-time and reject prediction

Plastics injection moulding (typical OEE 55–70%, world-class 80%, dominant loss: minor stops). Injection moulding exposes shot data (pressure, cushion, cycle); models predict short shots and warpage. Needs Euromap 63/77 or OPC UA from the press.

Speed-loss and stoppage analytics

Textile / apparel (typical OEE 40–60%, world-class 73%, dominant loss: reduced speed & quality). Textile lines run slow rather than stop; simple analytics on stop codes and speed come before ML. Needs stop-reason capture on each machine.

Three readiness patterns

Instrumented and automated (Korea, Germany, Denmark, Sweden, Belgium, Netherlands, Austria): both columns high, data exists, the constraint is labelling and ownership; start with prediction models on existing OEE data. Automated, not instrumented (Italy, Spain, Czechia, Slovakia, Hungary, France, Mexico): robot density above the EU average but adoption under 20%; the first project is data capture on the bottleneck line, the data-readiness scorecard is the test. Low-cost, low-density (India, Vietnam, Indonesia, Morocco, Tunisia): labour under 10 USD/hour makes downtime cheap per hour but frequent; the case is throughput and yield, not labour, so start with stop-reason capture and speed-loss analytics before any model.

Questions

Sources