Implementation of a neural network based system for estimating the strength of a board using mixed signals

An intelligent lumber grading system was developed to provide a new way for estimating the strength of a board by posing the estimation problem as an empirical learning problem. This system processed the X-ray image, extracted geometric features (of 1000 boards that eventually underwent destructive strength testing), and predicted the strength of the lumber by using a neural network. The X-ray image was passed through a threshold filter to separate the knots based on the fact that a denser knot produces a local maximum (as a rounded protrusion in an otherwise almost flat density surface) of the X-ray image. Each knot was modeled by a three-dimensional-cone with seven parameters. Information on all the detected knots such as volume, and knot-area-ratio were fed to a processor to generate 16 geometrical features (such as; the average of knot area ratio, and the number of knots detected in each board), which characterize each board. Then by using back-propagation as the training method, cross correlation as the measure of accuracy, and actual strength of a thousand boards as the empirical data set, a neural network was trained to estimate the strength of each board. The learning system consisted of three layers, with 1, 5, 16 neurons in output, hidden and input layer respectively. Ten-fold cross validation was used to produce an unbiased accuracy of the estimation problem. The learning and testing sets comprised of 900 and 100 boards respectively. By repeating the learning and testing for ten times and averaging the results, a coefficient of determination of 0.4059 was reached in this study for using X-ray images alone. The same methodology was applied to MOE (modulus of elasticity) and a coefficient of determination of 0.56 was reached. The results were improved by fusing the X-ray image and MOE using a learning system consisting of three layers, with 1, 5, 40 neurons in output, hidden and input layer respectively. Ten-Fold cross validation resulted in a coefficient of determination of 0.6101. By using the same data set of X-ray images, MOE and a mechanics based system methodology (consist of geometrical feature extraction, physical modeling, finite element analysis, and maximum stress failure theory for strength estimation) similar results of 0.4158, 0- .5805, and 0.6417 were reached for X-ray, MOE, and fusion of X-ray and MOE respectively. The results show that by fusing the MOE signal with X-ray images the estimation accuracy was improved by 10%. It shows a way to improve existing commercial lumber grading machines (such as the CLT), which are based on MOE alone. It also shows the way for future fusion of other signals to MOE signal in order to improve the grading accuracy.