Juha Kortelainen UPM RD Paper and Pulp Finland

Juha Kortelainen UPM R&D, Paper and Pulp Finland Avogadro Scale Engineering November 18 -19, 2003 The Bartos Theater, MIT

Contents ● UPM overview ● Jämsänkoski Paper Mill ● Paper quality and data analysis

UPM Key Figures, 2002 ● One of the world's largest paper producers ● Yearly production corresponds to 170, 000 km 2 area covered by paper! (land area of Massachutes is 20, 000 km 2) ● Mills mainly in Europe, North America and China

From the Forest to the Customer

Jämsänkoski – Finland, year 2002 Founded: Capacity: Personnel: 1888 815. 000 t/a 940 Products: - PM 5&6: uncoated magazine - PM 4: coated magazine - PM 3: label paper 570 000 t/a 125 000 t/a 110 000 t/a

Jämsänkoski SC PM 6 ● 325 000 t/a, 39 … 56 g/m², 9. 30 m width, 25 m/s speed

Automation Hierarchy, open systems

Paper Formation ● micrometer range variations, fibre level − paper surface structure, small defects − optical and printing properties ● several meters range, CD and MD profiles − paper web brakes ~ up to 100 km range

Paper Web Break Camera Monitoring

Image analysis ● Microscopic image analysis for fiber dimensions − fiber length ~2 mm, width ~40 um, cell wall ~ 2 um − automatic fibre analysers with 1, 5 um pixel resolution − paper structure with SEM using 0, 2 um pixel resolution ● Real-time image analysis for web defects and brakes − on-line camera scanner defects down to 0, 5 mm size ● Real-time microscopic scale? − 20 um pixel resolution − 10 meter web width − 25 m/s speed 12500 images / second with 1 MPix image size

On-line control ● Distributed Controls − thousands positions ● Supervisory Controls: Paper quality data with web scanner − e. g. cross-direction profile control − basis weight − moisture − caliper − colour. :

Time series data – Multivariate Auto. Regressive analysis ● Time dependent cross-correlation disturbance sources ● Numerically efficient method needed (FFT) − e. g. 1000 channels, 10 s sample period, 8. 6 E 6 samples/day ● Problems: − not efficient enough for long process delays − assumes stationary process state during analysis period − assumes linearity needs data prehandling, about 80 % of manual work!

Data Clustering ● Automatic clustering often ends up to distinct time periods, which are (more) stationary − product grades, process states ● Principal Components, k-means ● Neural networks: Self Organised Maps by T. Kohonen − visualization! ● Problems: − poor numerical efficiency − does not practically help in data prehandling

Modelling of paper quality ● Paper strength ● Optical properties ● PM control variables dominate ● some correlation from raw material disturbances

Neural Networks: Self Organised Maps (T. Kohonen)

Clustering of SOM by k-means

Summary for data-amounts / hour ● DCS data − 5 Hz rate, 10, 000 channels − multichannel: vibration, NIR spectra 2 E 8 samples / hour ● Paper web scanner − six channels, 1000 Hz 2 E 7 samples / hour − typically 5 scanners for one production line ● Camera systems − many fast speed camera applications in use ● off-line image analysis applications real time needs − in future 20 um resolution? 5 E 13 pixels / hour