Skip to content

Services · Digital & automation

Machine learning on production data

We build models on data from PLCs, quality control, the ERP system, test rigs and photos of parts. A model indicates the causes of scrap, predicts the actual lead time of an order and power demand, or assesses a part in a photo.

What we do

  • Scrap cause analysis and risk assessment for the current batch
  • Starting settings for a new product
  • Operation time and order due date forecasts
  • Forecasts of power demand and contracted capacity exceedance
  • Analysis of test rig data
  • Vision inspection: surface defects and assembly completeness
  • Part recognition and reading of markings
  • Detecting unusual batches and runs
  • Data suitability assessment before a model is built

Example applications

The scope of data and the type of model depend on the application. Five of the six below use data a plant usually already collects; photos for vision inspection most often have to be taken first.

Quality

Causes of scrap

We analyse

  • Process parameters logged by the PLC: pressure, temperature, cycle time
  • Material batch, mould or tool, machine, shift
  • Inspection and measurement results
  • Scrap reasons from inspection records and shift reports

The model indicates

  • Parameters and parameter combinations associated with scrap
  • Scrap risk for the current batch before final inspection
  • The range of settings within which the part meets requirements

What you gain

You correct the settings during the batch, not after it has been rejected. Scrap cause analysis rests on data from several hundred batches, not on observations from the last shift.

Process

Settings for a new product

We analyse

  • Setting sheets and recipes from previous orders
  • Product features: material, wall thickness, weight, overall size
  • Number of trial parts and corrections at each start-up
  • Start-up result: dimensions, scrap, cycle time

The model indicates

  • Starting settings for a product that has not been made before
  • The settings with the greatest influence on the result
  • The range of settings worth covering in trials

What you gain

Start-up of a new product takes fewer trial parts and corrections. The choice of settings no longer depends on the experience of one setter.

Planning

Operation time and order due date

We analyse

  • Order history from the ERP: routings, planned and actual times
  • Part features: material, size, number of operations, batch size
  • Machine load and order queue on the day the order came in

The model indicates

  • The actual operation time for a new part compared with the standard time
  • The lead time as a range, not a single date
  • Orders in progress at risk of delay

What you gain

The customer receives a date that can be kept, and the quote is based on the actual time similar parts took.

Energy

Power demand and contracted capacity

We analyse

  • Load profile from the meter: readings every 15 minutes or every hour
  • Production plan and shift schedule
  • Operating time of compressors, furnaces and the largest loads
  • Outdoor temperature

The model indicates

  • Expected demand for the next day and week
  • The hours in which contracted capacity may be exceeded
  • The loads with the largest share in the peak
  • Energy use per order or per part

What you gain

You move the start of a furnace or compressor outside the peak and avoid the charge for exceeding contracted capacity. You set contracted capacity on the basis of data, without an excessive margin.

Testing

Test rig and end-of-line test

We analyse

  • Test traces: pressure, leakage, force, temperature
  • Test reports and results from spreadsheets, including archived ones
  • Product data: type, dimensions, component batch
  • Complaints linked to a serial number

The model indicates

  • The expected test result after the first stage of the test
  • Units that passed, but whose trace differs from the typical one
  • Test stages that add no further information

What you gain

You shorten the test where the data justifies it and keep the full scope for unusual units. The test rig no longer limits assembly output.

Vision

Vision inspection of the part

We analyse

  • Photos of conforming parts — a few dozen to begin with
  • Photos of defects, if the model is to distinguish their types
  • Photos taken under fixed lighting and with a fixed part position
  • The inspector's assessment of the same parts

The model indicates

  • A deviation from the reference: scratch, dent, crack, pore
  • A missing component or its incorrect position
  • The type of defect and its location in the photo

What you gain

Every part is assessed against the same criteria, regardless of the shift and the inspector's workload. The photo stays in the archive and documents the condition of the part if a complaint is raised.

Image recognition applications

Surface defects

Scratches, dents, pores, chipped coating and discolouration on machined, cast, painted and chrome-plated surfaces.

Assembly completeness

Presence and position of seals, bolts, plugs and labels before the part is passed on.

Part and variant

Distinguishing similar parts and variants during sorting, kitting and goods receipt.

Markings and nameplates

Reading stamped, engraved and printed numbers and nameplates, with the result recorded in a register or the ERP.

Gauges on older machines

A pressure gauge or display read by a camera where the reading cannot be taken from the PLC. Deployment is preceded by a trial on one gauge.

Part condition in a photo

Assessment of the degree of wear of a tool edge, a seal or a piston rod surface. Deployment is preceded by a trial on photos from your plant.

Data needed to start

Scope of data (indicative)Source
Scrap causes Several hundred batches or orders with a recorded inspection result, including batches with scrapMachine PLC, inspection records, quality department spreadsheets
Process settings Setting sheets from a few dozen start-ups, each with an assessment of the resultSetting sheets, recipes in the PLC, the process engineer's spreadsheets
Order time and due date Orders from one to two years with planned and actual timesERP, time recording, job cards
Power demand Meter readings from one year and the production plan for the same periodGrid operator's portal, meter, energy management system, ERP
Test rig Several hundred recorded tests with their resultsTest rig data logger, test reports, spreadsheets
Vision inspection A few dozen photos of conforming parts; to distinguish defects — photos of each defect as wellPhotos taken at a workstation fitted with a camera and fixed lighting

Which models we use

The type of model depends on the data and on the question. Every model is verified on data that was not used for training.

Models on tabular data — scrap, settings, operation time. Data from the ERP, inspection records and the PLC takes the form of a table in which a row corresponds to a batch, an order or a cycle. We use models based on decision trees, which are effective from a few hundred rows. Besides the prediction, the model reports the variables with the greatest influence on it — hence the list of parameters associated with scrap.

Forecast with a range — order due date and power demand. The result is a range, not a single value. The width of the range shows how far the forecast can be relied on for a given part or day.

Anomaly detection — unusual batches and test traces. The model learns the typical course of a batch, cycle or test and indicates those that differ from it. It needs no examples of defects.

Image recognition — vision inspection. We use a neural network pre-trained on a large set of images and fine-tuned on photos of your part. In the basic version the model learns from conforming parts only and indicates deviations from the reference. To distinguish types of defect it needs examples of each.

Machine learning and machine monitoring

Machine monitoring concerns technical condition: wear of a pump, bearing or seal and the time remaining before a repair. The models described on this page concern the product, the order and the cost — the machine can be in working order and the batch still fails to meet tolerance.

The two services often use the same signals. If the PLC does not record the parameters a model needs, we start by logging them as part of monitoring.

Quoting from a drawing is described separately: AI for manufacturers. The operation time model supplies it with the actual time similar parts took.

Vision inspection station

The result of vision inspection depends equally on the model, the lighting and repeatable part positioning. We fit the camera, lighting and guard at the existing workstation and bring the result signal to the line PLC or the operator’s screen.

We make the camera mount, guard and wiring in-house, as we do control cabinets.