Cryptanalysis is predicated on the recognition of cipher algorithms, but in practice, researcher often do not know the cipher algorithm used. This paper focuses on block cipher algorithms identification and proposes an interpretable fusion model. The model extracts feature from the ciphertext based on bit segmentation and finds the optimal bit segmentation length. For the high-dimensional nonlinear ciphertext feature data, UMAP is used for dimensionality reduction, and the processed data is converted into image features. The model is based on CNN-Transformer, which separately inputs the image features and the ciphertext data itself, and then fuses them. Eight block ciphers including AES, DES, SM4, etc. were selected as the experimental objects. Under random key, a total of 72,000 encrypted ciphertext files of eight block cipher algorithms were constructed for algorithm identification. The experimental results show that, compared to existing research, under random keys, the binary classification accuracy is about 91%, and the eight-classification accuracy is about 70%, which are 8 and 19% higher than the average accuracy of existing research, respectively. Meanwhile, the extracted ciphertext bit rules are analyzed to provide model’s interpretability. It offers a new perspective for in-depth research on cipher algorithm recognition.