Deep learning has been widely used in fault diagnosis, especially the convolutional neural network (CNN). However, traditional CNN models usually use single-scale kernels to extract features, which ignores the multiscale features of input data. This article develops a novel residual neural network named multiscale feature fusion deep residual networks for fault diagnosis of aerostat. The designed multiscale feature fusion block (MFF Block) realizes automatic extraction, fusion and compression of multiscale features. The series connection of multiple MFF Blocks makes the proposed model able to extract deeper and wider features from raw signal segments. Then a multiscale pooling layer is developed to extract the most effective features for accurate fault diagnosis. The proposed model is evaluated on an exclusive aerostat strain signal dataset. The comparison results illustrate that the proposed model achieves superior diagnostic performance than several popular models.