Scholarlink

Electric submersible pumps (ESP) often have a wide range of abnormal data when operating under complex conditions, which brings disaster to oilfield production analysis and data mining. In this paper, we propose a vector normalization method based on physical significance constraint from the difficulty of high variability of data distribution of each parameter of ESP The redistributed features fix the lower limit of parameters such as leakage current at −1, which can reflect the equipment significance of the ESP. In order to improve the multimodal temporal feature extraction capability, an improved model of ESP CNN AutoEncoder is proposed. The model uses one dimensional convolution for forward encoding and inverse decoding, and combines reconstruction errors at different scales to dynamically adjust the tolerance of anomaly detection to capture a higher adaptive threshold window. The system is validated on real-time databases in two oil fields and achieved successful application results. Error reconstruction and sensitivity analysis are performed for detection performance to determine tolerance metrics that can dynamically meet different oilfield requirements. In comparison with Isolation Forest model, significant advantages are achieved in the ESP start-up anomalies. Finally, an ESP anomaly detection system with self-learning capability is realized.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Data Anomaly Detection of Electric Submersible Pump Based on Self-Learning Neural Network of Sequential Coding

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • He Ma,
  • Guoqing Han,
  • Jun Wang

摘要

Electric submersible pumps (ESP) often have a wide range of abnormal data when operating under complex conditions, which brings disaster to oilfield production analysis and data mining. In this paper, we propose a vector normalization method based on physical significance constraint from the difficulty of high variability of data distribution of each parameter of ESP The redistributed features fix the lower limit of parameters such as leakage current at −1, which can reflect the equipment significance of the ESP. In order to improve the multimodal temporal feature extraction capability, an improved model of ESP CNN AutoEncoder is proposed. The model uses one dimensional convolution for forward encoding and inverse decoding, and combines reconstruction errors at different scales to dynamically adjust the tolerance of anomaly detection to capture a higher adaptive threshold window. The system is validated on real-time databases in two oil fields and achieved successful application results. Error reconstruction and sensitivity analysis are performed for detection performance to determine tolerance metrics that can dynamically meet different oilfield requirements. In comparison with Isolation Forest model, significant advantages are achieved in the ESP start-up anomalies. Finally, an ESP anomaly detection system with self-learning capability is realized.