Considering the fact that most of the tested objects are in the normal state under practical working conditions can cause the scarcity of fault samples and differences within fault data, which further leads to low accuracy of fault classification, a DenseNet-based fault diagnosis model of unbalanced samples for the pressure-reducing valve is proposed, called weighted dense convolutional neural network(W-DenseNet). Firstly, the input data of the model is obtained by reconstructing original data of one-dimensional pressure signal and converting it into a two-dimensional grayscale image. Next, a feature extraction network is built based on the DenseNet. Then, to realize the weighted average of unbalanced sample errors, penalty coefficients are added to different types of samples in the loss function. Finally, the data acquisition system of the pressure-reducing valve is built and the classification performance experiment is carried out to validate the proposed model. The experimental results show that the W-DenseNet model exhibits good classification performance on data sets of pressure-reducing valves with different degrees of balance. When the sample imbalance occurs among each fault class, the recall rate of the model for the three fault types is still up to 95.18%, 95.47%, and 96.89%, respectively.