In practice, the scarcity of failure data for bearings makes it challenging to meet the extensive training requirements of deep learning models. This paper leverages the fine-grained feature extraction capabilities of Convolutional Neural Networks and the classification ability of fuzzy clustering without the need for training, proposing a small-sample bearing fault diagnosis method based on fuzzy clustering and an improved Densenet network. Initially, the pretrained Densenet network is modified by removing the classification layer and retaining only the feature extraction layers, and designing an Adaptive Global Average Pooling (GAP) layer to replace the Fully Connected (FC) layer. Subsequently, fuzzy clustering is utilized instead of the Densenet networks softmax classification layer, eliminating the need for training to achieve classification. Experimental results demonstrate that by training the GAP layer parameters with small-sample data, the model significantly reduces the requirement for training samples. During diagnosis, bearing time-domain images are input into the network, outputting 1 920 feature data at the GAP layer. Feature vectors matrices are constructed from the feature data of different fault states. Fuzzy similarity matrices and fuzzy equivalence matrices are obtained using fuzzy clustering methods. As the confidence factor changes from high to low, dynamic clustering diagrams are derived from the corresponding Boolean matrices, thereby achieving bearing fault classification.