Intelligent diagnosis method of torque-angle dynamometer cards for beam pumping units based on transfer learning

Abstract

Remote monitoring and intelligent diagnosis of operating conditions are two critical aspects for achieving cloud intelligent diagnosis of the beam pumping unit. However, the current manual detection method restricts real-time diagnostic efficiency. Due to poor sensor real-time performance and high maintenance costs, the technical conditions of the dynamometer card for remote monitoring are inferior to those of motor power. This paper establishes a torque-angle dynamometer card fault diagnosis method based on transfer learning to diagnose working conditions effectively based on motor power. Firstly, based on the existing production data and mechanism model, the torque-angle dynamometer card dataset was built by converting motor parameters and dynamometer cards. Secondly, five convolutional neural networks (AlexNet, VGG16, ResNet_V2, MobileNet_V3_large, and DenseNet) were constructed, and three pre-training models (ResNet_V2, MobileNet_V3_large, and DenseNet) based on ImageNet dataset were developed based on the transfer learning. Finally, utilizing the dataset above, the recognition models were trained. The overall recognition performances of the models and the recognition performance under each working condition were evaluated and compared. The results indicated that, excluding MobileNet_V3_large, the recognition accuracies of these base models achieved 90%, while the recognition performance of each working condition generally fell below 80%. This highlighted the significant impact of the dataset quality on the recognition performance of convolutional neural networks. In contrast, the pre-trained models based on transfer learning achieved a recognition performance of 95% overall and 90% for each working condition. The recognition accuracy, speed, and generalization performance of the pre-trained models were significantly better than the base models.