Search

• Classifier performance drastically increases by including pairwise comparisons. • Zero-impurity Decision-Tree classifier based on number of positives is designed. • Paper shows a formal and understandable formulation of neural network algorithms. • Decision tree and multilayer neural networks can give rise to similar results. • Issues appearing during the real inspection system implementation are described. This paper presents a modification to be added to multiclass classifiers, that improves their performance when classifying, in this case, defects appearing in polyethylene films. It aims to classify a new defect by confronting every defect type against each of the other types. In a simplified way, the type that results winner in more matches is the type that the defect belongs to. Different ways of implementing neural networks have been tested, using Gradient Descent and techniques for backpropagation. These techniques have been formally and understandably explained. In addition, a method based on decision trees has been included for comparison. Different issues related to the practical implementation of the detection and identification system within an installed production chain are addressed. The resulting system has been incorporated as a real inspection automatism in a polyethylene manufacturing line, and trained with defects previously obtained from the same line. [ABSTRACT FROM AUTHOR]