In order to solve the problem that the diagnosis of rolling bearing fault in the actual complex environment requires comprehensive performance such as accuracy, robustness and generalization, a lightweight adaptive CNN network (1D-LECA-Inception) is proposed, which integrates the attention mechanism. The Inception module is reconstructed by 1-D depth-separable convolution and the scale of the convolution kernel is broadened, and the unimportant information is screened out by the efficient channel attention (ECA) module. The residual structure, batch normalization (BN) and adaptive activation function AdaptH _ Swish are integrated to improve the stability and generalization ability of the overall network model, and the Paderborn and Case Western Reserve bearing data sets are used to compare with other classification models. The results show that the method performs better than other classification models under the same load, variable load and noise interference conditions.