Klimasauskas Group^TM
Leading the Edge in Emerging Technologies

Various of Klimasauskas Group staff have worked extensively in applying Neural Networks to developing empirical models of non-linear dynamic physical processes. These models are used for inferential sensing, control, and optimization. The four applications described here were selected to illustrate Klimasauskas Group Staff's innovative approach to crafting application specific solutions, and the integration of multiple technologies into high-performance hybrid solutions.

Hybrid Process Model (Klimasauskas & Guiver., Patent 5,877,954, assigned to Aspen Technology). One of the issues that the most common Neural Network algorithms face is the issue of extrapolation beyond the training set. Whereas other modeling approaches, notably linear regression, extrapolate in predictive ways that mirror the processes they are modeling, Neural Networks often produce unpredictable results. In particular, one of the most popular methods, Back-propagation, will saturate, or output will move toward the average as you extrapolate beyond the training data. This prevents the direct application in any form of closed loop control situation. A solution to this is to use a primary linear model and use a non-linear neural network model to predict the error. Another detector determines the degree to which the non-linear model is extrapolating and smoothly reduces it's contribution to the overall model prediction.

Data Selection for Improved Model Performance (Guiver & Klimasauskas, Patent, 5,809,490, Assigned to Aspen Technology). In process control data as well as many other applications in Data Mining, even large data sets may have concentrations in a very small portion of the space. The most popular algorithm for fitting this data, Back-propagation, uses sum squared error as a metric to minimize. This metric is dependent on the distribution of the data. Very large concentrations of data at a few points in the process space (such as during steady-state operation) can skew model performance, particularly in less dense areas (such as during process transitions). One approach to addressing this problem is to use a Self-Organizing Map (SOM) to separate the data into equal area regions, and sub-select examples from those. This patented approach substantially improved model accuracy, particularly during process transitions.

Dynamic Inferential Sensors from Laboratory Sampled Data. The most common approach to developing inferential sensors is to make the assumption that a manually taken sample corresponds to the steady-state properties of the system, and to average the continuous process data over a long period of time (1-2 hours) to approximate steady-state for the inputs. This approach works well for predicting properties during steady-state operation. However, there are substantial benefits from an inferential sensor that predicts well through transitions. However, during transitions, process dynamics often play a substantial role in product properties. The standard mathematical techniques for developing dynamic models require product properties at fairly frequently. This is often not feasible from either a cost or laboratory through-put standpoint. A novel approach suggested by Casey Klimasauskas was to use a Genetic Algorithm to identify the parameters of a first-order state-space model. This was actually quite successful, and has been deployed by Aspen Technology in their Aspen IQ product.

Process Diagnostics. The manufacture of polymer film involves a number of process elements all of which can impact ultimate product performance. In one particular application, several roles of film failed to meet customer coatability requirements. This was particularly troublesome, because all of the process data from those runs were within normal process tolerances. Linear modeling failed to identify the cause of the fault. One approach that worked was to use a fuzzy-encoding of all of the continuous input variables. This had the effect of dividing the input space into a number of discrete bins. This transformed data set was used to identify three colluding variables (all within normal operating specifications) that were characteristics of the fault.

Scheduling. The actual schedule of work-flow through a job-shop or chemical plant can have a substantial impact on profitability. Klimasauskas Group has worked with a variety of algorithms for route scheduling based on greedy algorithms, Genetic Algorithms, Simulated Annealing, and hybrid approaches. If you have a special scheduling need, please contact us.

Klimasauskas Group has the capabilities to work with you to address the specific requirements of your applications. We have the capability to develop innovative solutions which the performance of your process modeling applications.

Key Benefits

• Hybrid models combining the best of linear and non-linear models for closed loop process control.
• Improved overall model accuracy through selection of training data.
• Dynamic inferential models with improved accuracy, particularly through process transitions.
• Fault identification in a "normally" operating process.

Capabilities

Klimasauskas Group staff have extensive experience in developing hybrid solutions using Neural Networks, Genetic Algorithms, and Fuzzy Logic. In these projects, we were able to leverage our understanding of these technologies to leverage the benefits while mitigating the weaknesses of existing solutions.

• Application specific solutions for improving neural network prediction performance
• Hybrid integration of Genetic Algorithms and Neural Networks
• Novel hybrid integration of statistical and Neural Network models for improved application dependent characteristics.
• Integration of Fuzzy Logic and Neural Networks to solve complex non-linear problems.

Selected Publications

Klimasauskas Group staff have extensive experience in developing Neural Network technology, and working with the integration of existing products such as NeuralWorks Professional II/Plus. In this project, we were able to leverage our understanding of Neural Network learning mechanisms to improve overall consistency. (Klimasauskas Group Staff in bold.)

• Klimasauskas; Casimir C. (Sewickley, PA); Guiver; John P. (Pittsburgh, PA), March 2, 1999, Hybrid linear-neural network process control, Patent 5,877,954. Assigned to Aspen Technology.
• Guiver; John P. (Pittsburgh, PA); Klimasauskas; Casimir C. (Sewickley, PA), September 15, 1998, Apparatus and method for selecting a working data set for model development, Patent 5,809,490. Assigned to Aspen Technology.
• Klimasauskas, C. (1997).Perspective on Neural Networks. Control Engineering. July 1997, p. 30.
• Zhao., H., Klimasauskas, C., Guiver, J. (1997). Nonlinear Industrial Model Predictive Controller Using Neural Networks. Submitted to: 1997 American Control Conference, June 4-6, 1997, New Mexico.
• Klimasauskas, C. (1995).Using Fuzzy Pre-processing with Neural Networks for Chemical Process Diagnostic Problems. Intelligent Hybrid Systems, Goonatilake, Suran (ed.); Khebbal, Sukhdev (ed.). Chapter 8. pp. 143-151. New York: John Wiley & Sons.
• Klimasauskas, C. (1993). The Incipient Revolution in the Process Industries. Advanced Technology for Developers. March, 1993. High-Tech Communications: Sewickley, PA.
• Klimasauskas, C. (1993b). Simulated Annealing and the Traveling Salesperson Problem. Advanced Technology for Developers. June, 1993. High-Tech Communications: Sewickley, PA.
• Klimasauskas, C. (1992). Neural Network's: An Engineering Perspective. IEEE Communications Magazine. September 1992. pp. 50-53.
• Klimasauskas, C. (1991). An Introduction to Neural Networks with Applications to an Adaptive PID Controller. AUTOFACT'91, November 10-14, 1995. Dearborn Michigan: Society of Manufacturing Engineers. Technical Paper MS91-458.