SBIR-STTR Award

Situation or Problem: The U.S. forest products industry contributed $406 billion to the U.S. economy and employed 2,140,399 people in 2002. Currently the forest products industry is facing unprecedented competition from international imports and high wood costs. In 2003, the engineered wood panel sector produced 64.3 billion square feet of panels of which wood waste ranged from 3% to 9%. Reducing wood waste by 1% can translate into annual savings of $500,000 to $700,000 per producer and save 1.9 to 5.9 billion square feet of wood. Two of the largest contributors to wood waste in engineered wood manufacture are rejected panels and high density targets. Rejected panels lead to rework and high density targets result from excessive process variation. High levels of wood waste lead to poor wood yield, and subsequently higher resin and energy use. Reducing wood waste and improving wood yield can help this important economic sector improve and sustain competitiveness. Indirect benefits to society from wiser use of the forest resource are immeasurable. Purpose: This Phase I project will address the problems of wood waste and poor wood yield in engineered wood manufacture by developing a real-time prediction system for physical properties using a hybrid Genetic Algorithm/Neural Network (GANN) with distributed data fusion. The proposed system will lower the rate of rejected panels, optimize throughput, identify key process parameters, optimize wood usage, promote lower resin use, lower energy use and improve wood yield. OBJECTIVES: The goal of Phase I is to provide proof-of-principle for the concept of real-time prediction of the physical properties of engineered wood using a GANN with distributed data fusion. The proof-of-principle will demonstrate implementation of this technique for medium density fiberboard (MDF) manufacture, by accurately predicting the real-time internal bond (IB) of MDF. The system will also be designed to be transportable to all types of engineered wood and other wood properties. There are three technical questions related to the research objectives. 1. Development of an automated real-time distributed data fusion system. -- The automated fusion of diverse types of real-time sensor data with event-based destructive tests will advance information technology, automated data fusion and associated issues. 2. Development of a real-time Genetic Algorithm to predict the physical properties of engineered wood. This objective will advance the mathematical and wood sciences by determining the real-time sources of process variation that influence physical properties. Critical technical questions will be addressed in defining NN-based genetic representation and genetic operations to predict the value of a single material property from parameter values acquired during real-time data acquisition from a GA-optimized number of sensors. 3. Development of an information software platform for industrial use. Ease-of-use of an object oriented program and efficiency in processing information across an industrial LAN will be the critical computational science question addressed by this objective. APPROACH: 1. Development of an automated real-time distributed data fusion system: Real-time process data will be aligned with the internal bond strength of medium density fiberboard (MDF) at one manufacturing test site. Lag times, corresponding to the period of time required for the furnish to travel through the process from the point where a given parameter has an influence, to the point where the panel is extracted for destructive testing will be estimated and included in the system. Statistical estimates will be derived for the real-time data and exclude null fields, data type error, string data type, time overlap data and statistical outliers. A unique number will be generated when the panel is extracted from the process, and will be later used to match process data with lab results. When the lab results are matched with the process data, this combined data will be recorded in an automated relational database. 2. Development of a real-time Genetic Algorithm/Neural Network system for prediction of the physical properties of engineered wood: Three principal tasks comprise the GANN portion of the proposed work: a) Incorporate Automated Input Parameter Selection: A typical composite production line may include sensors for monitoring the values of several hundred to more than one thousand process variables. Not all of these are of relevance to the prediction of a given material property. The selection task is one especially well suited to solution by GA methods. b) Add Automated Network Pruning: One of the hazards in almost any application of neural network methods is inadvertent production of networks of excessive size. In our GANN systems, it is possible to adjust fitness measures to favor networks that are no larger than necessary to perform the predictive tasks required of them. A scheme in which an estimate of generalizing capability (as opposed to mere mimicry) will be incorporated directly into one of the fitness measures. c) Perform Extensive Network Validation Studies: Although our preliminary statistical validation results appear most promising, the GANN method will be accepted in industry only to the extent that it can be demonstrated to be statistically valid and robust. A believable demonstration of predictive performance can only be performed in the context of extensive and broad data spanning a plausible range of process parameter values, material property values, and measurement noise. It is our intention that a final report on the Phase I work will include extensive measures of GANN performance over a sufficiently broad spectrum of conditions such that others contemplating the use of the method will be able to make informed and careful judgments concerning potential applicability 3. Development of an information software platform for industrial use The Phase I GANN software platform for industrial will have three Visual Net components: a) Genetic Algorithm/Neural Network (GANN) Trainer 1.0, b) Genetic Algorithm/Neural Network (GANN) Processor 1.0, c) Genetic Algorithm/Neural Network (GANN) Client 1.0. The Visual Net software will have point and click functionality

The Phase II project will advance innovation in manufacturing of forest products and improve the overall competitiveness of this industry by lowering costs, maximizing throughput and improving quality of product. The importance of the work lies in manufacturing innovation and economic efficiencies. This sector contributed $240 billion to the U.S. economy in 2002 and employed 1,009,555 Americans in that same year. Wood costs represent the largest cost component of total costs for most forest products manufacturers. Wood waste is the largest avoidable cost. Reducing wood waste by 1% can lead to annual savings of almost $700,000. Two of the largest contributors of wood waste are unusable panels due to poor strength and over-specifying density. Rejected panels lead to costly rework operations. High-density targets result from excessive process variation. High levels of wood waste also result in high resin and energy use which result in higher operating costs and greater risk to the environment. "Purpose": To develop a commercial software system for the real-time prediction of the mechanical properties of an engineered wood panel and other process parameters using the GANN method and build upon the proof-of-concept system developed in Phase I. The commercial software system will model multiple types of engineered wood panels and associated broader groups of mechanical properties. The system will include features of real-time prediction that lead to maximum production throughput, energy reduction and lower pollutant emission. OBJECTIVES: There are eight technical objectives for the proposed Phase II project. These objectives will produce a commercially ready genetic algorithm/neural network (GANN) system (which will be called MIGANN) of sufficient robustness that it can be applied commercially to a broad range of wood manufacturing facilities. Phase II objectives are: (1) Develop data fusion software; (2) Convert Phase I GANN C code to MATLAB code; (3) Provide for predictions of several mechanical properties; (4) Incorporate cross-validation features in the new GANN; (5) Archive models for retrieval; (6) Expand GANN system validation from the two original Phase I plants (medium density fiberboard and oriented strand board) to three new plants (particleboard, laminated-veneer-lumber and wood-plastic composites); (7) Conduct a cost/benefit analysis; and (8) Create a GANN system for commercial use (called MIGANN). The Phase II project will lead to improved plant operations by increasing throughput, reducing waste, reducing energy use and reducing pollutant emission. This project and the resulting commercial software directly supports Executive Order 13329 (69 FR 9181) entitled Encouraging Innovation in Manufacturing, by improving the business competitiveness of the forest products industry, a key economic sector of the U.S. economy. The Phase II project will build upon the successful results of the Phase I project. Phase I results were: 1) Real-time data fusion system of destructive test lab data and real-time process sensor data; 2) GANN (Genetic Algorithm/Neural Network) system that allows automatic creation of models for engineered wood plants which detect influences of process parameters on mechanical properties and makes real-time predictions; and 3) GUI software of the system LAN capable. APPROACH: The methodology will include eight high-level tasks: Develop user-friendly commercial GUI software for data fusion. Phase II will develop a user-friendly GUI software module in Visual Studio.Net that will interface with any SQL-based data warehouse. An important feature of this system will allow users to enter the time-lagging and median sample size for process sensor tags. Core GANN code. The new system will include multilayer feedforward networks using tan-sigmoid transfer functions within the hidden layers and a single linear transfer function for the output layer. The Scaled Conjugate Gradient (SCG) algorithm will be used to train the neural networks (NN)s. The genetic algorithms (GA)s will be used to select the combination of process variables that will serve as inputs in the neural networks. The GAs will control the architecture of the neural networks. Predictions of several mechanical properties. The core code of the Phase II GANN will allow for simultaneous modeling and prediction of several mechanical properties. We can run two to three applications on a dual processor computer while keeping enough CPU power for the other programs to operate freely. It will be possible to run two to three GANN optimizations at a time and select the best performer from those parallel jobs to run a single GANN optimization. Incorporate cross-validation features. We will use a segmented cross-validation sequence that will allow optimizing the process variables selection and the neural networks architecture based on prediction capabilities as well as calibration performances. A loop will be created to define subsets (75% for a calibration subset and 25% for a validation subset) from the initial dataset selected for the optimization. Implement Model Archiving and Retrieval. We will rank models based on their cross-validation and real-time validation performances, e.g., root mean square error of cross-validation (RMSECV) and real-time sum of squares error (SSE), and the number of records for which the model was built and validated. These models and their associated performances will be archived by product type, quality parameter and date-time. Validation at Three Additional Plants. The intent of validation will be to ensure that the acquired data represent all principal modes of plant operation, e.g., medium density fiberboard, particleboard, oriented strand board, laminated-veneer-lumber and wood plastic composites. Cost/Benefit analysis. The real-time data warehouse will be used to retrieve line speed and raw material set points for time periods when the GANN was used for production operations. An analysis of line speed improvement and raw material input reductions will be conducted to determine cost savings and improved productivity. Estimates of improved profitability will be made by working with company accountants. Generalized GANN system for commercial use. The MIGANN commercial software platform will be install-ready and available on CD, DVD and from web-link downloads.