Power generators are one of the critical assets of power grids. The early detection of faults in power generators is essential to prevent cutoffs of the electrical supply in the power grid. This work presents a comparative analysis of machine learning (ML) models for the generator fault diagnosis. The objective is to show the ability of simple and ensemble ML models to diagnose faults using as attributes partial discharges and dissipation factor data. For this purpose, a generator fault database was built, gathering information from operational data curated by power generator experts. The hyper-parameters of the ML models were selected using a grid search (GS) and cross-validation (CV) optimization. ML models were evaluated with class imbalance and multi-classification metrics, a correspondence analysis, and model performance by class (fault type). Furthermore, the selected ML model was validated by experts through a diagnosis system prototype. The results show that the gradient boosting model presented the best performance according to the performance metrics among single and ensemble ML models. Likewise, the model showed a good capacity to detect type 3 and 4 faults, which are the most catastrophic failures for the generator and must be detected in a timely manner for prompt correction. This work gives an insight into the need and effort required to implement an online diagnostic system that provides information about the power generator health index to help engineers reduce the time taken to find and repair incipient faults and avoid loss of power generation and catastrophic failures of power generators.