Improved prototypical network for fault diagnosis of pumping unit

The indicator diagram is an important method to judge the working condition of rod pumping unit. Due to the large amount of data, the diagnosis accuracy of common fault conditions can reach 99%, while the diagnosis accuracy of a small part of the fault data is only 90%-92%. To improve the fault diagnosis accuracy of pumping unit when the amount of data is small, an improved prototypical network based on the Squeeze and Excitation module is proposed for fault diagnosis of pumping unit. The residual connection and the Squeeze-and-excitation module are introduced to extract the characteristics of the indicator diagram. Then the indicator diagram images of different categories are mapped to the feature space. After measuring the distance, input the Softmax classifier to realize the fault diagnosis of the pumping unit under the condition of small sample. The experimental results show that the improved prototypical network model can improve the classification accuracy on different small sample data sets. Especially when the sample data is less than 50, the diagnostic accuracy of the six fault types of normal, insufficient liquid supply, pump collision, gas influence, floating Vanal leakage and sand resistance is increased by 3%-20% compared with AlexNet and ResNet34 models.