A multi-feature fusion model based on the mixture of experts (MoE) model is proposed, which further improves the recognition accuracy of RF fingerprint identification algorithms in low SNR scenarios. Different signal processing and data representation methods were used for training. Experiments are conducted to demonstrate the performance advantages of our method, and ablation studies on the data representations are carried out. Experimental results show that the identification accuracy for 150 transmitting devices still exceeds 90% at a SNR of 4 dB.