Neural Network Modeling for Paper Property Predictions

 

Sandra Lenz	Renae Koerbitz	John Rudd
Appleton Coated	Appleton Coated	NuSoft Technologies
Locks Mill	Locks Mill	10505 Talleyran Drive
540 Prospect St.	540 Prospect St.	Austin, TX 78750
Combine Locks, WI  54113	Combine Locks, WI  54113

 

---

Appleton Coated Locks Mill became interested in utilizing neural network based empirical modeling technology to predict and optimize process parameters such as formation, porosity, opacity and strength properties.  These models can be used for off-line troubleshooting and for on-line operation.

This paper describes the development and deployment of several of these types of models on multiple machines at the mill.  Also discussed will be examples of the use of these models for off-line troubleshooting analysis.

---

Background

Appleton Coated purchased a neural network software package called Insights from Pavilion Technologies, Inc. in December 1998.  The software is designed for process engineers to solve process problems.  It builds empirical models for the process by utilizing stored real-time historical data.  The software used handles large data sets, linear and non-linear processes, and many inputs and outputs.  The software performs off-line analyses for process characterization.  It ranks the effects of process inputs on the outputs and performs “what-if” simulations for process predictions.  Finally, the off-line models can be implemented on-line in real time.

For the past two years, the software tool has been used many times for off-line analyses and trouble-shooting purposes.  From October – December 1999 and again in February 2000, on-line models were run as trials providing soft sensors for several paper machine properties.

The trial models included formation on the mill’s No. 7 and No. 5 machines along with porosity on the No. 6 machine.  The models ran on-line during the trial period predicting the outputs every five minutes, sending the information to the mill’s process data historian.  The results were trended and made available to operations.

Discussion of Modeling Methodology Used

The neural network software is designed to take process historical data (inputs) which consist of all the variables that relate to the property being predicted (output) and develop a prediction model.  Historical data were analyzed from the mill’s process data historian consisting of six months of data (June 1 through December 1, 1999) in five-minute intervals.  False data were eliminated.  With this data, models were trained using all the possible inputs corresponding to a particular property.  When a model was trained, a sensitivity rating was produced (Table 1) which listed the variables from the most important to the least important.  This information allows the possible elimination of the variables that have the least amount of impact on the property.  On the same list, there is a column, which gives either a negative or positive value.  This shows if the variable has a positive or a negative effect on the property being predicted.  A predicted versus actual graph (Figure 1) displays the slope, standard deviation, and R-squared values for the trained model.  Models are trained until the R-squared and slope values are acceptable by eliminating bad data and dropping variables that are very insensitive to the property.  After a model is trained and developed off-line, the model can be used as an on-line tool to accurately predict the property.

In addition to developing on-line models, the software has been used to predict various outputs for trouble-shooting purposes.  These include models on No.1, No. 5, No.6 and No. 7 machines.  Some of the issues dealt with are loss of machine speed, drying limitations, porosity, formation, bleed through and runnability at the converting plants.

No. 7 On-Line Formation Model

Multiple on-line models for formation on No. 7 were developed to predict M/K lab formation.  The final model settled on after elimination of low sensitivity variables had thirty-two inputs.

The on-line formation model with 32 inputs had an R-squared value of 0.937.  Figure 2 is the predicted online model over a period of 1.5 days without any breaks.  Figure 3 is the predicted formation over a ten-day period.

After a model is trained and put on-line, it doesn’t mean that the model is finalized.  There is still refining that can and must be done with more data.  Figure 4 shows one of the preliminary models, which has spikes (breaks) and plateaus.  The plateaus were due to OBA flow and K-system NHPD/T values being outside the range of the data stored in the historian for the period over which the model was trained.  By training the model with data over the entire operating range and having the model hold the last good output during a sheet break, these issues are eliminated.

No. 6 On-Line Porosity Model

The same method was used to develop a porosity model for the No. 6 machine that had an R-squared value of 0.921.  When the model was put on-line for the trial period, a definite cycling in the predicted output was seen.  In trouble-shooting the problem, two variables with erratic behaviors were the source of the problem.

Headbox stock flow and the vertical slice position were fluctuating during this period due to the addition of a trial filler at the ‘pre-accepts’ position disturbing the signal from the headbox stock flow element (Figure 5).  Normally, ground calcium carbonate is added at the ‘pre-fan pump’ position.  After collecting more data, a new model was trained with the slice position variable eliminated.  With this model, the porosity trended well.

No. 5 On-Line Formation Model

The task of developing a model for the No. 5 machine was more difficult because there are not many on-line measurements available.  When the model was trained, it was found that coat weight was the most sensitive variable.  In fact, its sensitivity was over four times the next most sensitive variable in magnitude and the model had an R-squared value of 0.792.  This indicates that the coating applied can cover the underlying formation and give higher numbers.

With this in mind, a second model was trained that included grades that applied starch only at the coater (R-squared = 0.826).  For these grades, the model does a better job of indicating the sheet formation without being biased by the coat weight.

The next step was to combine both models, applying a transform that would interchange the models based on coat weight.  One model would run if the coat weight went above 2 lbs/t and the other would run if the coat weight went below 2 lbs/t.  Figure 6 shows that this configuration predicts the M/K lab formation test results very well.

Using Models as On-Line Trouble-Shooting Tools

One additional advantage of on-line models is that their behavior can provide insight into problems that are occurring within the process. 

In early February of the second trial period, the No. 7 formation model indicated that there was problem with one of the input variables.  Note: With neural network technology, predictions are made within the trained data range for each input variable.  It does not attempt to extrapolate. 

In about 10 minutes, it was discovered that the headbox temperature was out of the model’s range.  Upon investigation, the temperature was checked and found to be low.  In addition, the white water silo temperature was low.  The machine personnel had not yet discovered the low white water temperature and were appreciative of the information.  The instrument technician checked the system and found the temperature control loop was not responding.  The steam valve to the silo had to be repaired.

Also in February, a shift in the predicted porosity on No. 6 occurred.  For a thirteen-hour period, the porosity prediction showed a 20-point decrease in porosity.  Then the prediction came back to normal.  During trouble-shooting, the problem found was related to the A-system on-line drainage unit being out of service. This is not a normal condition.  The on-line drainage input was a key variable to the model prediction.  The stock prep operator said that there was an electrical problem and the unit went down.  Once the drainage unit was repaired, the model prediction was back to normal.

Conclusions

Neural network models are good process-engineering tools providing insight into the process.  They have been used numerous times at the Locks Mill as a paper machine trouble-shooting tool and for providing understanding about runnability issues at the converting plants.

On-line models for three applications, with R-squared values ranging from 0.792 to 0.937, have been run to demonstrate that soft sensor applications are viable and valuable.  Prediction of key properties every five minutes gives valuable information about the process quickly without having to wait for reel turn-up data, which can be delayed by as much as two hours.  Operators can be made aware of the situation immediately on a trend screen, and can then fix the problem, eliminating the possibility of two hours of rejected paper.

The Locks Mill has plans to move forward with a program to make the on-line predictions permanent in 2001.  Also, operations are now suggesting other possibilities for modeling opportunities.  Some of these include internal bond, steam usage, opacity, brightness, charge and break detection.

Neural network technology can supply valuable information to engineers and operators about key process variables in real time that traditionally is obtained through lab results.

 

TABLE 1: Sensitivity Analysis         Output Q7:FMINDX.QA

Rank	Input Name	Avg. Abs.	Avg.
1	HB Cons.	0.344	0.344
2	HW Blend Flow	0.284	0.284
3	Total Head	0.273	-0.273
4	Wire Speed	0.190	-0.190
5	Thick Stock Flow	0.186	0.186
6	Couch Vacuum	0.176	-0.176
7	% F System Softwood	0.173	-0.173
8	AKD Flow	0.156	-0.075
9	BMA Flow	0.139	0.139
10	K System HPD/T	0.128	0.128
11	HB Temp.	0.110	-0.110
12	OBA Flow	0.101	0.081
13	J Drainage	0.087	-0.087
14	Floc Index	0.086	-0.058
15	HB % Ash	0.074	-0.0291
16	PCC Flow	0.066	-0.051
17	Tray Ash %	0.063	-0.063
18	% Northern Hardwood	0.062	0.032
19	BC Drainage	0.039	-0.038
20	H Slice Position	0.034	-0.034
21	D Drainage	0.030	-0.002
22	Skimmer Flow	0.030	0.030
23	V Slice Position	0.030	0.030
24	%BCTMP	0.027	0.026
25	HB Zeta Potential	0.025	0.007
26	% Aspen	0.024	0.024
27	% D System Softwood	0.024	-0.012
28	% Broke	0.019	0.013
29	BC Conductivity	0.019	0.016
30	Silo pH	0.018	0.018
31	F Drainage	0.018	0.002
32	HB Conductivity	0.009	-0.007