Radio frequency fingerprint identification (RFFI) can classify wireless devices by analyzing the signal distortions caused by the intrinsic hardware impairments. State-of-the-art neural networks have been adopted for RFFI. However, many neural networks, e.g., multilayer perceptron (MLP) and convolutional neural network (CNN), require fixed-size input data. In addition, many IoT devices work in low signal-to-noise ratio (SNR) scenarios but the RFFI performance in such scenarios is rarely investigated. In this paper, we analyze the reason why MLP- and CNN-based RFFI systems are constrained by the input size. To overcome this, we propose four neural networks that can process signals of variable lengths, namely flatten-free CNN, long short-term memory (LSTM) network, gated recurrent unit (GRU) network and transformer. We adopt data augmentation during training which can significantly improve the model's robustness to noise. We compare two augmentation schemes, namely offline and online augmentation. The results show the online one performs better. During the inference, a multi-packet inference approach is further leveraged to improve the classification accuracy in low SNR scenarios. We take LoRa as a case study and evaluate the system by classifying 10 commercial-off-the-shelf LoRa devices in various SNR conditions. The online augmentation can boost the low-SNR classification accuracy by up to 50% and the multi-packet inference approach can further increase the accuracy by over 20%.