Research was carried out on the problem of the decrease of diagnostic accuracy due to the imbalance of the collected data during the fault diagnosis of rolling bearings.Facing the characteristics of the multi-scale complexity of the original one-dimensional vibration signal,an imbalance fault diagnosis method based on multi-scale cost-sensitive convolutional neural network was proposed.A multi-scale one-dimensional convolution feature extraction layer with a series-parallel structure was constructed,and multi- feature extraction was achieved by designing the connection mode of different convolution layers and selecting different convolution kernel sizes; the attention mechanism was used to adaptively set cost matrix of the Adacost to realize the adaptive allocation of weights.Experiments on rolling bearing data sets of Western Reserve University with various imbalance ratios show that this method can effectively improve the classification performance of the model in different imbalance data sets,and has stronger generalization ability.