Abstract:  For the fault diagnosis problem of the indicator diagram of pumping wells, the feature information in the image is extracted by convolution neural network. In order to ensure the diagnostic performance of the network model, the structure complexity of the network model is reduced. Based on the lightweight convolution neural network, the attention mechanism is introduced to improve the diagnostic performance of the lightweight network model. First, in the network infrastructure the MobileNet-V2 network is used, and the attention ECA (Efficient Channel Attention Module) module is embedded in the inverse residual module of MobileNet-V2. Compared with the ordinary residual network, the features retained after convolution are more complete, so the fault diagnosis capability of the model is improved. Then, the ECA uses 1D convolution to achieve local cross- channel information interaction between adjacent channels and obtain the dependencies between local channels. The resulting channel attention re-calibration weights are multiplied by the corresponding channels of the input feature map of the module, and the attention-weighted feature map is obtained. The MobileNet-V2 accuracy rate is 90. 6% , and the improved MobileNet-V2 diagnostic accuracy rate is 97. 60% . 

Key words: pumping well, fault recognition, indicator diagram, convolutional neural network, attention mechanism