In order to solve the problem that the diagnostic model cannot assign different weights to the features in the one-dimensional vibration signal according to their importance, which leads to the failure to extract representative features and thus affects the accuracy and robustness of the diagnostic model. A feature highlighting method based on reverse attention amplification mechanism ( RA) is proposed to reduce the proportion of non-important features by reversing attention and pruning the features, so as to highlight the important features. The long short term memory (LSTM) network is used to learn the temporal information between the features and to classify the fault types through the fully connected layer. The optimal data interception length, pruning hyperparameters and the stability of the model after adding noise to the signal are selected experimentally. The optimal data interception length, pruning hyperparameters are experimentally selected and verified the stability of the model after adding noise to the signal. The experimental results show that the proposed RA-LSTM bearing fault diagnosis method has excellent fault diagnosis performance, with fault diagnosis accuracy reaching 100% and excellent robustness even after the addition of noise.