This paper proposes an improved multilayer self-organizing map (MLSOM) network-based method for wind turbine condition monitoring, addressing the lack of consideration for the interrelationship and information transmission between the unit and its key components in existing methods. Initially, the Pearson correlation coefficient is employed to select features from the supervisory control and data acquisition (SCADA) system, with the feature information serving as the input for the wind turbine′s tree structure′s bottom node. Recognizing the nonlinear and timeseries nature of wind turbine data, long short-term memory (LSTM) models are developed and trained using historical data to predict SCADA feature values. The prediction residuals replace the feature information as input to the bottom node of the self-organizing map (SOM) network within the MLSOM model, creating a normal behavior model for each component. The minimum quantization error serves as an indicator for assessing component health using the trained SOM model. Monitoring models are established for key components such as the generator, gearbox, and converter. These component health indicators are then integrated as top-level node information and used as input for the top-level SOM model in the MLSOM to form a normal behavior model for the entire wind turbine, yielding a comprehensive health indicator for condition monitoring. Case study results of two wind turbines demonstrate that the proposed method effectively transmits and aggregates component information step by step, enabling the condition monitoring of the entire wind turbine.