Artificial neural network approach to process diagnosis.

摘要

A prototype artificial neural network (ANN) system consisting of a visual system (VS) and an ANN classifier for attribute classification and variable control, with parallel information processing capability, was successfully developed. The ANN provides a means for improving the performance of on-line process control in a way that is superior to traditional pattern recognition approaches.; The VS was used to acquire data for the ANN classifier. A digital camera was employed in the VS to capture images of a metal piece. These were later transformed to binary images. Two types of data sets were generated from the VS and then served as input for the ANN classifier. One set of data was used for attribute classification while the other was used for variable classification and control. The use of the VS suggests a feasible alternative for data extraction and thus has the potential to reduce the amount of human intervention in a discrete manufacturing environment.; The ANN classifier, which had input, hidden, and output units, was first trained to classify unacceptable and acceptable patterns (attribute classification). The number of hidden units, ranging from one to forty, was changed during the experimental process for evaluating the effect on classification rate. Different random number seeds were used to generate the weights for connections between units. Number of training iterations, training time, classification time, and RMSE (root mean square errors) were used for assessing the performance of the ANN classifier. The results showed that the ANN was able to achieve 100% convergence within 1,000 iterations of training trials based on a tolerance level of 0.009. However, RMSE did not show a consistent trend in predicting the performance of ANN.; The ANN classifier was then trained to classify unnatural patterns/trends (e.g., trends, cycles, and mixtures) on a variable control chart. Seven unnatural patterns were examined. Training time, number of training iterations, classification time were also used for assessing the performance of the ANN classifier. The results showed that the ANN achieved 100% convergence in a limited number of training trials based on the tolerance value of 0.009.; The stability of classification of the ANN for attribute classification and variable control was also assessed by introducing data with noise. The results showed that the ANN was able to handle certain noise levels without a significant misclassification rate.