Classifying Defects in Pallet Stringers by Ultrasonic Scanning

Detecting and classifying defects are required to grade and sort pallet parts. Use of quality parts can extend the life cycle of pallets and can reduce long-term cost. An investigation has been carried out to detect and classify defects in yellow-poplar (Liriodendron tulipifera, L.) and red oak (Quercus rubra, L.) stringers using ultrasonic scanning. Data were collected for sound and unsound knots, bark pockets, decay, holes, and wane using rolling transducers in a pitch-catch arrangement. Data from eight ultrasonic variables—energy, pulse length, time of flight (TOF)-amplitude, TOF-energy, TOF-centroid, energy value, energy pulse value, and peak frequency—were used to classify defects. Three different types of classifiers were used to categorize defects—a multi-layer perceptron network (MLP), a probabilistic neural network (PNN), and a k-nearest neighbor (KNN) classifier. Mean values for the energy variables demonstrated statistically significant differences between clear wood and defects and among defect types. Mean values for the TOF variables did not differ significantly between clear wood and knots. All three types of classifiers were able to distinguish defected from clear wood in oak with accuracies above 95%; accuracies for yellow-poplar were somewhat lower for the MLP and PNN classifiers. Among the defect classes, decay exhibited the highest recognition rate for both yellow-poplar and oak. Wane and holes in oak were readily confused owing to their common loss of transducer contact. Overall accuracy at the data-point level varied from 69-78%. Simple post-processing operations are expected to improve that substantially. Based on accuracy performance alone, the MLP and KNN appear equally preferable for this task