Abstract
Mechanical vibration is one of the most used signals for traditional condition monitoring in rotating machines due to its sensitivity to mechanical vibrations; however, its deployment in industrial facilities is often hampered by significant cost and practical limitations. In addition, the harsh and variable environmental conditions typical of industrial settings can adversely affect accelerometer performance, leading to issues such as signal degradation or sensor failure. Therefore, using new signals is a challenge in condition-based maintenance. This work demonstrates that the electric torque given by the three-phase induction motor is a sensitive signal with respect to faults of its mechanical load, particularly a broken tooth in a gearbox. Exhaustive frequency analysis shows the sensitivity of the electric torque for different levels of fault severity. Then, a systematic methodology using machine learning approaches such as Random Forest and KNN classifiers is conducted to perform supervised fault diagnosis with statistical and Poincaré plot (PP) features. The classification accuracy and precision show that the electric torque is as good as the vibration signal to diagnose different levels of tooth breakage in a gearbox. By using statistical features, classifier models achieve 100% accuracy and precision, and with PP features the maximum performance is 98.15% accuracy and 98.40% precision.