Currently, wind energy is consolidating itself as a cornerstone of global decarbonization. However, this growth poses challenges for the operation and maintenance (O&M) of wind turbines. This work presents a novel method for anomaly detection in wind turbine gearboxes, combining actor-critic reinforcement learning with Bayesian hyperparameter optimization. The model processes vibration signals from the Low-Speed Shaft radial sensor, normalizes them, and segments them into 7-second windows from which 19 statistical and spectral features are extracted. The actor agent determines whether each window corresponds to a normal or anomalous state, while the critic evaluates these decisions based on the temporal difference error. In addition, both the learning rate and the number of update steps are automatically tuned via Bayesian optimization. After training and 10-fold cross-validation, the model achieved an accuracy of 95%, a recall of 96.2%, an F1-score of 95.6%, and an AUC-ROC of 95.6%, with a confusion matrix showing low levels of false positives and false negatives.