Every part, checked.
An industrial camera and a vision model on an edge PC look at each part as it passes and signal the PLC to reject anything with a scratch, crack, dent, burr or missing hole. Try adding a defect.
See the failure coming.
Vibration, temperature and current from each critical asset are compared with what “normal” looks like for that machine. Pick an asset to see what maintenance would be told.
Seven use cases.
Start with one. Each is scoped as a pilot with a success metric agreed up front. Point at one to see what’s involved, the data it needs and what to aim for.
-
Visual quality inspection
Cameras and lighting at the line, a vision model on an edge PC, rejects signalled to the PLC and every image stored for traceability.
- Typical target
- 50–90% fewer escapes, a full image record per part
- Data needed
- 200–500 images of good and bad parts. We help collect them.
- Time to pilot
- 6–8 weeks, cameras included
- How it works
- We design the camera station first: lighting, optics and how the part is presented. Then an anomaly model trained mostly on good parts, plus a classifier for known defect types, runs on an edge PC in tens of milliseconds and rejects through digital I/O or OPC UA.
- What we measure
- Escape rate and false-reject rate, both, with the threshold tuned to what a missed defect and a scrapped good part each cost you.
- The problem
- Manual inspection is tiring and inconsistent, especially late in a shift, and escapes cost you in returns and claims.
-
Predictive maintenance
Retrofit vibration and temperature sensors, or tags you already log, feed an anomaly model per asset. Alerts land in Teams or your CMMS with the likely cause and a suggested window.
- Typical target
- 30–50% less unplanned downtime, 10–25% lower maintenance cost
- Data needed
- 3+ months of history, or start fresh: models learn “normal” in 2–4 weeks
- Time to pilot
- 6–10 weeks on 3–5 critical assets
- How it works
- Vibration spectra and envelope analysis pick up bearing wear, imbalance and misalignment. A model per asset learns its normal behaviour across loads and speeds and tracks the margin left to the alarm level.
- What we measure
- Unplanned downtime hours, the share of alerts that led to a real fault, and mean time between failures.
- The problem
- A spindle, bearing or compressor fails without warning and the line stops, while fixed schedules replace parts with months left in them.
-
Production scheduling
A constraint optimiser that knows your orders, routings, changeovers, materials and staff. When something changes it re-plans in seconds and shows the planner the trade-offs.
- Typical target
- +5–15% OEE, planning time from hours to minutes
- Data needed
- Orders, routings, machine calendar, changeover times
- Time to pilot
- 6–8 weeks
- How it works
- Constraint programming over routings, changeover times, due dates, material and staff. Planners lock what must not move, and compare candidate plans side by side before releasing one.
- What we measure
- OEE, on-time delivery, and changeover hours per week.
- The problem
- The plan lives in one planner’s head and a spreadsheet. A rush order or a broken machine means hours of re-planning.
-
Demand & inventory forecasting
Forecasts per item from order history, seasonality and customer behaviour, turned into reorder points and safety stock purchasing can act on.
- Typical target
- 15–30% less inventory with the same or better service level
- Data needed
- 2+ years of order history per item
- Time to pilot
- 4–6 weeks
- How it works
- Gradient-boosted and hierarchical time-series models per item and product family, using open orders, price changes and seasonality as signals. Slow, intermittent items get their own method instead of being forced into the same model.
- What we measure
- Forecast error (WAPE) against your current method, stock value and service level, item by item.
- The problem
- Too much of what doesn’t sell, too little of what does, and cash tied up in the warehouse.
-
Energy optimisation
Meter data, the production plan and spot prices combined: flexible loads move to cheaper hours, and machines that use more than they should are flagged.
- Typical target
- 8–20% lower energy cost, plus data ready for sustainability reporting
- Data needed
- 15-minute meter data, production plan, tariff
- Time to pilot
- 4–6 weeks
- How it works
- Load profiles per machine from your meters, a day-ahead price feed and the production plan. Flexible jobs move within limits you set, and machines drawing more than their baseline are flagged for maintenance.
- What we measure
- Euros and kWh per unit produced, and peak demand charges.
- The problem
- Energy is a top cost line, yet you see it once a month on a bill, with no idea which machine or shift caused it.
-
Shop-floor assistant
Operators and technicians ask on a tablet or by voice, in their own language, and get answers from your manuals, work instructions and past repair tickets, with the source page shown.
- Typical target
- Up to 40% faster troubleshooting, onboarding shortened by weeks
- Data needed
- Manuals, SOPs and maintenance logs
- Time to pilot
- 4–6 weeks
- How it works
- Retrieval over your manuals, work instructions and repair history, with every answer linked to its source page. When the answer isn’t in your documents it says so instead of guessing, and it can run fully offline on a local model.
- What we measure
- Mean time to repair and the share of questions answered without calling a senior colleague.
- The problem
- Know-how sits with a few senior people. New hires take months to ramp up, and the night shift waits for a phone call.
-
RFQ & quoting
Reads incoming requests and drawings, finds similar past jobs, estimates material, machine time and cost, and prepares a draft quote for your estimator to check.
- Typical target
- Quote turnaround from days to hours, steadier margins
- Data needed
- Past quotes, drawings and actual job costs
- Time to pilot
- 6–8 weeks
- How it works
- Drawings and specs are read, similar past jobs found by geometry and material, and a cost estimate built from your routings and actual times. The estimator reviews and approves every quote.
- What we measure
- Quote turnaround, win rate, and quoted versus actual margin.
- The problem
- Estimators dig through old jobs by hand, quotes take days, and faster competitors win the order.
Shop-floor lessons.
Things we design around on every production project, because they decide whether a system is still running a year later.
Lighting before models.
Most vision projects that fail, fail on lighting and how the part is presented, not on the model. We build and test the camera station before we train anything.
A false reject costs money too.
A system tuned to catch every defect also scraps good parts. We set the threshold on the real cost of each mistake and report both rates, every week.
Train on good parts.
Defects are rare, so we don’t wait for hundreds of examples of each. Anomaly models learn what a good part looks like and flag anything that isn’t.
Read first, write later.
We start by reading PLC tags over OPC UA or MQTT. Anything that writes back to control happens only after validation, together with your automation engineer.
An alert needs a work order.
An alarm nobody acts on is noise. Every maintenance alert arrives in your CMMS with the likely cause, the evidence and a suggested window.
Walk us through your line.
A 30-minute call or a half-day on site. We’ll tell you which of these fits your plant, what data you already have and what a pilot would cost. Or write to hello@aisensus.eu.