Abstract: A new DNN model，called multi-channels weighted convolutional neural network (MCW-CNN) is proposed in order to solve the feature extraction problem of CNNs that use single-channel signal as input，which can not express the fault characteristics hidden in the vibration signals effectively. Firstly，empirical mode decomposition (EMD) is used to obtain multi-channel one-dimensional signals to highlight the fault characteristics，which is converted into multi-channel image as input of MCW-CNN. So that the image feature extraction of CNN can be fully utilized，and the gearbox diagnosis is further converted into an image recognition issue that CNN is good at dealing with. Secondly，due to the difference of frequency and bandwidth of each channel，a dynamic receptive field is proposed in MCW-CNN to extract the detailed features from these images. In addition，the intensity of the impact feature of each channel image is different，so the fusion weight is determined based on the kurtosis，which can highlight the channels that have strong impact characteristics. Finally，the effectiveness of the proposed method is verified on a gearbox fault diagnosis case. Experimental results show that the MCW-CNN can effectively extract the fault characteristics of the vibration signal. It has a significantly higher recognition accuracy rate than other typical deep learning methods and traditional classifiers.

Key words: gearbox fault diagnosis, multi-channels signal, feature learning, convolutional neural network, information fusion