Wind energy has become a key source for diversifying Brazil’s energy matrix, significantly contributing to the generation of clean and sustainable energy. Due to its importance and the large investments being made in this sector, there is an increasing need to anticipate failures in wind turbines. The exponential increase in the number of installed turbines, along with the aging of the generation fleet, has intensified the competition to reduce operation and maintenance costs, which means minimizing unplanned downtime and reducing large, costly corrective maintenance. The objective of this study is to use vibration data available from Condition Monitoring and Management Systems (CMMS) to identify turbines with significant condition deviations that present a high risk of failure. The data processing approach, using algorithms such as Condensed Nearest Neighbor (CNN) and Principal Component Analysis (PCA) in the pre-processing stage, along with Support Vector Machines (SVM) for health state classification, has demonstrated excellent accuracy, above 90 percent, both in single-turbine tests and in multiple-turbine tests, making it suitable for managing wind farms with a large number of turbines. The experiments conducted with a combination of five different turbines allowed the identification of the best performance scenarios, maintaining results with over 90 percent accuracy in the proposed model according to the goal of early fault detection in a fleet, even when using reduced training data for the applied model. It is also important to highlight scenarios where performance was not adequate, impacting both accuracy and the rate of evaluated false positives.