Aiming at the problem that it is difficult to realize effective fault diagnosis for weak signals of rolling bearings in the complex environment of strong noise and variable working conditions, a bearing fault diagnosis method based on improved convolutional Lenet-5 neural network is proposed. Firstly, the collected one-dimensional time-domain bearing vibration signals are preprocessed and converted into two-dimensional grayscale images which are convenient for convolution operation. Secondly, the continuous one-way traditional convolutional layers in the most basic Lenet-5 model are improved into Block1 module, Block2 module and Block3 module to extract more concrete and accurate feature information. Finally, L2 regularization and Dropout optimization are used to avoid overfitting. In order to verify the robustness and generalization performance of the proposed method in complex working conditions, experimental validation was carried out using the rolling bearing dataset and the gearbox experimental dataset. The experimental results of the bearing dataset show that the average accuracy of the proposed method in the variable noise experiments is 99. 3%. In the variable load experiments, the average accuracy of fault diagnosis is higher than 90. 26%. In the variable operating conditions experiments, the average accuracy of fault diagnosis is higher than 89. 01%. In the gearbox dataset experiments, the fault diagnosis accuracy of anti-noise is up to 96. 3%. The improved Lenet-5 method has the better ability of fault diagnosis for 12 fault types of rolling bearings, and has better anti-interference and generalization performance under variable working conditions.