The operation of the motor is complicated and there will be a lot of noise under actual operating conditions. The noise causes low-accuracy of bearing fault diagnosis. In order to improve this problem， a bearing fault diagnosis method based on the multi-head self-attention mechanism of one-dimensional fully convolutional self-encoding network （MHSACAE） combined with convolutional neural network （CNN） is proposed. Firstly， we use the MHSACAE for noise reduction. And then we use CNN for fault diagnosis. Particularly， the MHSACAE adopts an unsupervised training method. The method fully considers actual working conditions and the inherent connection of the sequence data， while the ability to filter noise is achieved and the original fault infor? mation is retained to the greatest extent. So that CNN can realize high-precision diagnosis of motor bearing faults under noise condi? tions. Finally， the comparison with other bearing fault diagnosis methods under noisy conditions proves that the proposed method has better results.