Abstract: With the large-scale development of offshore wind power, the condition prediction of offshore wind turbines has attracted widespread attention. Accurate and timely condition prediction will help reduce the economic losses that may be caused by the deterioration of the turbines. In order to improve the accuracy of early fault warning, this paper visualized the SCADA (supervisory control and data acquisition) data, and input it into the neural network as a whole to fully reflect the correlation between the faults of different components of the offshore wind turbine and the multi-state coupling of the SCADA data. For the problem of failure identification caused by the sparse sample data of some fault types, the CycleGAN (cycle generative adversarial networks) with double-layer generator and double discriminator was used to enrich the fault label samples. In order to improve the timeliness of unit fault warning and make fault warning as early as possible, this paper adopted correlation analysis to reduce the dimension of high-dimensional SCADA data to simplify the structure of RBF (radial basis function) neural network, accelerate the convergence of neural network, and improve the training speed. The results of a calculation example for a practical offshore wind farm in China show that the method proposed in this paper can effectively predict the occurrence of faults in advance and at the same time can effectively identify the types of faults, which is beneficial for wind farms to deal with faults in advance, and arrange operation, maintenance and repair plans reasonably, so as to avoid heavy losses.