Aiming at the problem that the poor feature extraction of the oil well indicator diagram leads to the low accuracy of condition diagnosis, an oil well condition diagnosis method based on multi-branch fusion embedded attention feature extraction is proposed. Firstly, in order to make the extracted implicit feature information of the indicator diagram more comprehensive, based on the convolutional autoencoder, a multi-branch and multi-scale encoder structure is designed to extract and fuse the feature information of the displacement-load data of the indicator diagram. Then, in order to strengthen the local features after multi-branch fusion, an embedded channel attention mechanism is designed to add global maximum pooling on the basis of global average pooling, so that it can pay attention to the global and local features of the indicator diagram at the same time. Meanwhile, in order to further enhance the ability of implicit feature extraction for key information of the indicator diagram, the channel attention mechanism module is embedded after the channel squeeze and before the excitation, so that the squeezable channel is pre-adjusted with the weights once and the weights are adjusted twice after the excitation. Finally, the extracted features are put into the long short-term memory network model for well condition diagnosis. The results show that the well condition diagnosis method based on multi-branch fusion embedded attention feature extraction improves the effective feature extraction capability of the indicator diagram to a certain extent, increases the well condition diagnosis rate, and meets the actual needs of the oilfield.