FAULT CLASSIFICATION IN HELICOPTER VIBRATION SIGNALS

In Autoregressive (AR) modeling of machinery signals large model orders are usually required to adequately capture the characteristics of the data. For fault classification only a small subset of the coefficients are relevant. A new approach is presented for machinery fault classification that uses AR coefficient vectors but performs dimension reduction before classification, utilizing techniques from the statistics and pattern recognition communities. The hybrid approach is illustrated by analyzing the Westland helicopter transmission vibration signal data set. The data consist of 8 accelerometer signals for each of 68 no-fault and seeded-fault runs on the aft transmission of a CH-46E helicopter. The effect of different dimension-reduction and classification techniques on fault classification is studied. For the comparison the Kolmogorov-Smirnov separability test and classification performance are used. This approach drastically reduces the training data requirements and complexity of classifiers. These results should advance the performance of HUMS systems and bring us closer to adequately addressing prognostics for complex machinery.