Abstract: ‍ ‍A method for identifying individual OFDM radiation sources based on a deep residual shrinkage network （DRSN） and integrated shrinkage network was investigated with the goal of improving the low identification rate for communication radiation sources under low signal noise. First， starting from the principle of OFDM-transmitter signal generation， the sources that can produce fingerprint differences in OFDM transmitters were analyzed， and the coherent accumulation of OFDM signals in adjacent frames was performed to effectively improve the signal-to-noise ratio of the OFDM signals. The impact of differences in the transmission content was reduced by intercepting the leading OFDM signals. The power accumulation and bispectral diagonal slicing signal processing of the leading OFDM signals were performed to construct a multi-domain data set of the leading OFDM signals. Then， this multi-domain data were sent to a DRSN network with an automatic soft threshold denoising and cross-layer connection structure to prevent gradient disappearance for training， which effectively reduced the interference of noise on the transmitter fingerprint signal and prevented the problem of a poor recognition effect caused by a poor threshold setting. In addition， a 5-fold cross-validation strategy was adopted to train the DRSN network. Overfitting was prevented during the network training， and the integrated learning concept was used to integrate the primary prediction results of the three DRSN networks. Finally， the fusion results were sent to a logistic regression sub-linear classifier as secondary data， the ECOC strategy was used to convert the multi-classification task into a binary classification task， and the sample category was encoded. After the test sample passed through the binary classifier， a set of predictive category codes was obtained. By calculating the Euclidean distance between the sample category code and predicted category code， the final classification result was determined based on the minimum Euclidean distance. Experimental results on public data sets showed that compared with other deep learning methods， when the SNR was 5 dB and 0 dB， the accuracies of the DRSN and integrated fusion OFDM radiation source identification were 97% and 95.88%， respectively， with low complexity， which verified the effectiveness of this method at a low signal-to-noise ratio.