Automated classification of electrical appliances on a production line

Abstract

User simulation inspection of every appliance on a production line is time consuming and expensive. A more effective way is to use sensors for fast indirect measurements of selected quality indexes, ie appliance features, which may be used for automatic on-line inspection and classification, by correlation to manual inspection. Feature vectors for classifying electrical appliances tend to form overlapping, irregular amorphous clusters in a multidimensional feature space. Three classifier algorithms were formulated to address this difficult classification problem, which is aggravated by the requirement that almost no bad units should be misclassified as good ones. The discriminatory power of two or all three classifiers is combined by a classifier voting strategy. The different classifiers and voting strategies are compared in terms of four performance indexes, predicting the percentage of bad units sent to the customer, percentage of good units rejected as bad ones, a cost-weighted class contamination index and the expected percentage of correctly classified units. Practical application is implemented on a feature data base of several hundred labelled refrigerators, whereby it is demonstrated that three classifier c voting will practically never misclassify a bad unit as a good one.