Anomaly detection with domain specific shapelet learning for sucker rod pump system.

Area of Science:

• Petroleum Engineering
• Machine Learning
• Artificial Intelligence

Background:

• Reliable operation of Sucker Rod Pump Systems (SRPS) is critical in the petroleum industry.
• Early detection of slow-developing secondary faults in SRPS is essential for preventing failures.
• Conventional methods rely on labeled datasets and mechanistic models, limiting their applicability.

Purpose of the Study:

• To propose an unsupervised, end-to-end learning algorithm for anomaly detection in SRPS.
• To develop a method that does not require labeled datasets or extensive mechanistic information.
• To enable early warning of potential issues for decision-makers.

Main Methods:

• An unsupervised end-to-end learning algorithm, Anomaly Detection with Domain-specific Shapelet Learning (AD-DSL), is proposed.
• AD-DSL utilizes a mechanistic information matrix and a sparsity-promoting objective function.
• Shapelet-based features are learned from motor power time-series data, coupled with a dynamic threshold and anomaly scores for trend monitoring.

Main Results:

• The AD-DSL algorithm demonstrates effective anomaly detection in SRPS motor power data.
• The method provides interpretable, Shapelet-based features for fault identification.
• Quantitative comparisons show AD-DSL outperforms baseline methods in accuracy and delivers competitive F1 scores.

Conclusions:

• The proposed AD-DSL algorithm offers a robust and effective solution for unsupervised anomaly detection in SRPS.
• This approach enables early warnings, enhancing the reliability and operational efficiency of petroleum extraction.
• AD-DSL overcomes limitations of conventional methods by reducing reliance on labeled data and mechanistic models.