By Matt Taylor
Iggesund Paperboard (www.iggesund.com) in Workington, Cumbria, U.K., has five main-line refiners that each require 15 megawatts to operate, so saving energy is essential to reduce our operating costs. However, while we needed a solution to optimize our energy use, we also had to make certain that the refiners produce the required pulp for our final paperboard products.
Consequently, the model we recently implemented to predict paper stiffness is based on an artificial neural network (ANN) with a feed-forward, single-layer perception (FFSLP) structure. We chose this model over a multilinear regression (MLR) model because the process needs were nonlinear, and the ANN is more accurate at making predictions over a wide range of grades. In addition, using just one model for all paper grades is convenient and makes the system easy to use for refinery operators.
We also coupled the ANN with a multivariate data analysis (MVDA) approach for online monitoring because it allows for more adaptability to operating conditions, such as paper grade changes. The resulting nonlinear model accurately predicts the machine directional (MD) variation of the cross direction (CD) bending stiffness.
Our facility in Workington employs approximately 380 people and has a capacity of 200,000 tonnes of Incada per year. Incada is a multilayered folding box board (FBB) that is manufactured from fiber of known and traceable origin. It's produced at a site where pulp production is integrated with the paperboard mill (Figure 1).
Next, we designed a predictive model with National Instruments' (www.ni.com) LabVIEW software and ANN technology, combined with advanced adaptive control algorithms for closed-loop adaptive refiner control. The solution incorporates an offline tool that enables simulation of refiner loads for different pulp types. This gives operators and development engineers the opportunity to run different refining configurations and compare the predicted final paperboard quality measurements. By simulating the effects of changing parameters with LabVIEW, we saved time and money by avoiding expensive full-scale trials.
In addition, LabVIEW's data-logging and supervisory control (DSC) module makes it easy to run several models in parallel and provides straightforward online performance monitoring. Using the DSC module, we bounded the model within the normal operating conditions for the pulp mill. After variables move outside of these conditions, an alarm activates to notify an operator that a problem has occurred, and the system communicates which variable is at fault, as well as the best action to resolve it.
Modeling Leads to Online Monitoring
At present, Iggesund Paperboard is successfully using this new model in its online operations. Using a score plot, the system notifies machine operators when variables deviate from specifications. The score plot displays the principle component for each of the variables and updates them in one-minute intervals. For example, the scores for two latent vectors are plotted against each other to form a two-dimensional monitoring chart (Figure 2). The 2-D score graphs are then used to show the relationship between the two latent vectors and are displayed along with a 95% to 99% Hotelling's ellipse, which defines the normal operating region for the process. This allows operators to easily identify outliers.
All similar data points are clustered together in the 2-D score graph, with groupings of points representing an operating point for the process. These graphs may show one group of points to indicate a single operating point, or more than one cluster of points to indicate more than one operating point. If the data points remain within the Hotelling's ellipse, this indicates that the operating point is within the normal region. Any abnormal shift in the process variables—whether the basic correlation between variables remains intact or not—is clearly indicated in the 2-D score graph. With the help of a loadings graph, users can then determine the root cause of the problem and isolate the fault.
Consequently, the model we built with the DSC module gives our board machine operators real-time process information and clearly indicates when deviation from the required pulp quality setpoint has occurred. The pulp mill operators can select the desired setpoint and refiner energy to achieve optimal results.
As a result, Iggesund Paperboard has achieved significant benefits from the LabVIEW DSC modeling solution. Once the model was implemented, the plant experienced a 1% increase in throughput, which is a significant amount of energy savings for a large plant. In fact, cost savings from the energy reduction are estimated at £720,000 (USD1,117,755) per year.
Another benefit is better quality control and less repulping of board that isn't suitable for sale. Production losses due to re-pulping are £12,000 (USD18,630) per hour. We estimate that the model will improve board machine uptime by 30 hours per year for an annual cost savings of about £360,000 (USD558,877).
Also, using LabVIEW for model-based multivariate optimization allows us to easily adapt the method and techniques we implemented for other areas of the process. This provides a path for future process optimization without needing numerous software packages.
Predictive pH Control
While using model-predictive control and ANN technology combined with advanced adaptive control algorithms for closed-loop adaptive refiner control is very useful, the model is bound within the normal operating conditions for the pulp-mill. If conditions outside normal occur, an alarm sounds that not only indicates that there is a problem, but also provides details of what variable is at fault and the best action to resolve that fault. Advances in computer power, combined with the development of object process control (OPC) and improvements in LabVIEW, have enabled us to use a desktop PC to control large manufacturing plants such as our paperboard factory.
However, one problem with conventional process control is variables are monitored individually, which indicates the state of only the individual variables and ignores the interaction of variables. So normal process monitoring gives the operators warning alarms when only that single process is out of specification.
On the other hand, multivariate techniques take into account the internal interactions among variables. Therefore, we can model the process on detecting univariance with individual variables. This can then be compared with the univariance for combined variables to identify the true process performance, which can then be used to identify when the process is unstable.
In the first stage of this project, we're finding the primary contributing chemicals that drive the pH in our pulp process. We find these chemicals through previous experiments and observing the effects of the different chemicals using multi-linear regression (MLR). A predictive control method using an ANN can then be trained and developed using the same data used for the MLR model.
In general, an ANN is an information processing paradigm inspired by the way biological nervous systems, such as the brain, process information. The key element of this paradigm is the novel structure of the information processing system. It is composed of a large number of highly interconnected processing elements (neurones) working in unison to solve specific problems. Like people, ANNs learn by example. They're typically configured for specific applications, such as pattern recognition or data classification, through a learning process. Also, an ANN is a nonlinear model, so it can predict the nonlinear process much better. An indication of how well it matches the actual output response is given by the root mean square error (RMSE), which improved from 0.13 to 0.045, providing far more accurate results.
Because of these capabilities, we used LabVIEW to develop a pH model that demonstrates that it's possible to model a nonlinear complex process. It confirms the pH meter readings and provides a basic online diagnostic tool for operators. Because the model is only taking in the key variables that affect pH, if there is a change in the incoming flows of one of these, then it's highlighted on the model. This change can be observed much sooner than normal due to process dwell time. The lower and upper control limits are set at three standard deviations, below and above the mean, respectively, so any difference between the actual and predicted pH values displayed on the screen indicates either a drift with the online pH meter or a problem with the incoming chemicals. All the chemical flows are displayed on the GUI under the pH trend. This means that the operators have an instant reading of all the chemical flows.
This model aided in monitoring and optimizing chemicals to control the pH more smoothly by displaying real-time information. This increased process stability is due to improvement of the control of a process-critical variable.