Options
2012
Conference Paper
Titel
Processing and evaluation of gear data using statistical classifiers
Abstract
Quality assessment of gear wheels is an important task for the automobile industry and suppliers. Various kinds of faults have to be detected: cracks, structural damage, tiny inclusions or any contamination of the raw material. We propose a pattern recognition approach of ultrasonic signals induced in the test objects which is capable of distinguishing between good and bad parts. A signature analysis device was built at Fraunhofer IZFP Dresden. It contains an ultrasonic actuator for excitation and two sensors for data acquisition. From every sensor signal a short-term power spectrum is calculated. In a training stage models are created for two classes (good and bad) from a classified training set of gear wheels. These models are evaluated against a disjoint test set. A correlation coefficient based classifier was used as baseline system. Its class models ("reference patterns") consist of a number of partial short-term power spectra which have to be selected manually. We compared this approach to a statistical classifier based on hidden MARKOV models which can be trained without any user input. We used a good class and measured the divergence of all test parts from the good class. The final decision on a test part is made by an threshold. Below this threshold a part is classified as good, above it as bad. Only the HMM classifier was able to classify all parts in our experiments correctly.