Abnormal-heart-rhythm classifying method based on deep transfer learning

The invention discloses an abnormal-heart-rhythm classifying method based on deep transfer learning. The method includes the steps that a processed ECG signal is passed through a 1D-CNN-and-LSTM cascade network structure, and spatial features and time features of the signal are combined to extract features of the ECG signal; through a multi-scale-fusion connection mode, the small-scale features are up-sampled and interpolated into an adjacent feature graph in the same scale to conduct feature fusion; feature difference adjustment is conducted on features of data of a source domain and a targetdomain through a self-adaptation layer, classification losses of the features output from the self-adaptation layer through the data of the source domain are calculated through a fully-connected layer and a Softmax classifier, and MMD losses between the source-domain features and the target-domain features output through the self-adaptation layer are calculated; finally, network parameters are jointly adjusted by combini