RESUMEN
This paper presents a theoretical approach to determine the probability of misclassification of the multilayer perceptron (MLP) neural model, subject to weight errors. The type of applications considered are classification/recognition tasks involving binary input-output mappings. The analytical models are validated via simulation of a small illustrative example. The theoretical results, in agreement with simulation results, show that, for the example considered, Gaussian weight errors of standard deviation up to 22% of the weight value can be tolerated. The theoretical method developed here adds predictability to the fault tolerance capability of neural nets and shows that this capability is heavily dependent on the problem data.
RESUMEN
This paper introduces an approach to cosmetic surface flaw identification that is essentially invariant to changes in workpiece orientation and position while being efficient in the use of computer memory. Visual binary images of workpieces are characterized according to the number of pixels in progressive subskeleton iterations. Those subskeletons are constructed using a modified Zhou skeleton transform with disk shaped structuring elements. Two coding schemes are proposed to record the pixel counts of succeeding subskeletons with and without lowpass filtering. The coded pixel counts are on-line fed to a supervised neural network that is previously trained by the backpropagation method using flawed and unflawed simulation patterns. The test workpiece is then identified as flawed or unflawed by comparing its coded pixel counts to associated training patterns. Such off-line trainings using simulated patterns avoid the problems of collecting flawed samples. Since both coding schemes tremendously reduce the representative skeleton image data, significant run time in each epoch is saved in the application of neural networks. Experimental results are reported using six different shapes of workpieces to corroborate the proposed approach.