Identifying maximum imbalance in datasets for fault diagnosis of gearboxes

Research into fault diagnosis in rotating machinery with a wide range of variable loads and speeds, such as the gearboxes of wind turbines, is of great industrial interest. Although appropriate sensors have been identified, an intelligent system that classifies machine states remains an open issue, due to a paucity of datasets with sufficient fault cases. Many of the proposed solutions have been tested on balanced datasets, containing roughly equal percentages of wind-turbine failure instances and instances of correct performance. In practice, however, it is not possible to obtain balanced datasets under real operating conditions. Our objective is to identify the most suitable classification technique that will depend least of all on the level of imbalance in the dataset. We start by analysing different metrics for the comparison of classification techniques on imbalanced datasets. Our results pointed to the Unweighted Macro Average of the F-measure, which we consider the most suitable metric for this diagnosis. Then, an extensive set of classification techniques was tested on datasets with varying levels of imbalance. Our conclusion is that a Rotation Forest ensemble of C4.4 decision trees, modifying the training phase of the classifier with a cost-sensitive approach, is the most suitable prediction model for this industrial task. It maintained its good performance even when the minority classes rate was as low as 6.5 %, while the majority of the other classifiers were more sensitive to the level of database imbalance and failed standard performance objectives, when the minority classes rate was lower than 10.5 %.