To address the issues of unknown parameters and difficult modeling in current model-based proportional servo valve fault diagnosis methods, we propose the Attention Convolution Capsule Network (ACCN) algorithm for the first time and design a new proportional servo valve fault diagnosis method based on this network. This algorithm first uses convolutional neural networks and an efficient channel attention mechanism to extract and fuse features from the original multi-channel one-dimensional time-domain signals. The fused results are then input into the subsequent capsule network for further feature extraction and fault classification, yielding the corresponding diagnosis results. The effectiveness of the proposed algorithm and diagnostic method was validated on a proportional servo valve fault simulation test bench. The proposed method achieved 100% accuracy on the test set, and the inclusion of the channel attention mechanism and capsule network enhanced the diagnostic performance of the model to varying degrees.