Fault diagnosis method for sucker rod well with few shots based on meta-transfer learning

Abstract

In the actual production process of the oil field, the functionality of the oil well pumps will be negatively affected by many factors such as manufacturing quality, installation quality, sand, wax, water, gas, heavy oil, and corrosion, which will cause great loss to the production. Therefore, it is very important to analyze the working conditions of the rod pumping systems. In actual oilfield production, the working conditions of deep well pumps are analyzed based on the measured surface indicator diagrams. However, traditional computer diagnosis of pumping wells relies on necessary mathematical methods, or deep networks with many parameters. These methods require a lot of data, with complex analysis processes, long testing time and low efficiency. This article studies the application of meta-transfer learning in the diagnosis of rod pump wells in few-shot scenarios. Meta-transfer learning combines the advantages of both meta-learning and transfer learning. It can not only provide good initial parameters for learners based on deeper networks through the pre-training stage of transfer learning, but also achieve automatic adjustment of hyperparameters with the help of meta-learning. This enables fast gradient iteration and reduces the probability of overfitting, thereby improving model performance. We also conduct comparative experiments to compare the experimental performance of this method with classical meta-learning methods and deep convolutional networks on the classification problem of indicator diagrams. According to the experimental results, the accuracy rate of meta-transfer learning in the diagnosis of few-shot working conditions in practical problems is close to 80%, which is better than the 70% accuracy rate of the comparative experiments. In the actual oil field, there are not many indicator diagrams for pumping unit diagnosis, so this method can well meet the needs of fault detection.