The planetary gearbox is a critical part of a wind turbine. It is of great significance for safety and reliability. Therefore, a fault diagnosis method for a planetary gearbox based on a deep feature fusion network is proposed to operate under variable speed conditions, with limited numbers of samples and in strong noise scenarios. First, the original signal is expanded into multiple characteristic domains. Then a multi-dimensional stack sparse autoencoder is used to extract the features of each domain. Finally, to solve the problem that the traditional Softmax classifier is not capable of classifying the fusion information, an echo state network optimized by competitive swarm optimizer is proposed to perform feature fusion and output the diagnostic results. The experimental results of fault diagnosis in multiple scenarios show that the proposed method has a good classification performance in variable speed conditions. In addition, it is also robust to a reduction in the number of training samples and external noise. This work is supported by the National Natural Science Foundation of China (No. 51977128).