Rotary machines are key equipment in many industrial sectors, from mining operations to advanced manufacturing. Among the critical components of these machines are bearings, gearboxes, rotors, among others. These components tend to present failures, which can be catastrophic, with economic, safety and/or environmental consequences. Among the most established methods for classifying bearing faults, the envelope method has been widely used, with relative success, for several years. However, this method and its variations are difficult to automate and require extensive experience on the part of the analyst. We report that, while traditional methods (e.g., envelope) successfully classified bearing failures less than 45% of the time, machine learning methods were successful more than 62% of the time, and in some cases reaching 67%. This work differs from others in the sense that it uses all the available measurements from a well-known database, not just a subset. In addition, the measurements are taken at the motor base, which are more difficult to classify, and avoid using different segments of the same signal in both training and validation, thus reducing the possibility of overfitting. As a consequence, the results obtained are apparently inferior to those reported elsewhere, but probably closer to what one might expect in practical applications.