As a state self-sensing element, smart gear has high industrial application value. However, the traditional intelligent gear only realized the status data collection. When a large amount of raw data is uploaded to the industrial internet of things, it is not only limited by bandwidth, which reduces the timeliness of fault diagnosis, but also occupies the computing and storage resources of the cloud server. To solve those problems, this study proposes an intelligent lightweight model for smart gear edge diagnosis, which named LWDCN. Firstly, the deep convolutional neural network with wide first-layer kernels (WDCNN) is improved with an operation of depthwise separable convolution process, which can realize an efficient operation for lightweight sensing. The channel attention is introduced to improve the expressive ability of network in feature representation. Once the proposed LWDCN was trained on the cloud end, this designed model would be further deployed on an edge-side hardware prototype with STM32F429 chip, which named smart gear edge diagnosis unit (SGEDU). This edge hardware prototype can be deployed directly on the edge of the gear to implement data acquisition, data processing, model reasoning, and condition diagnosis. This means that, a cloud-edge collaborative diagnosis mechanism can be built for smart gear. The experimental results shown that the proposed identification accuracy, and SGEDU can accurately diagnosis the gear state on-line. It is foreseeable that this study can consume a large amount of state data at the edge end and improve the intelligent level of smart bearing and smart gear.