FAILURE PREDICTION FOR ARTIFICIAL LIFT SYSTEMS

Predicting a well failure before it occurs can dramatically improve performance and reduce operating cost. It potentially can provide operators enough lead time to adjust operating parameters to avert failures, or to schedule maintenance to minimize downtime and to avoid unplanned repairs. This paper presents a failure prediction framework and corresponding failure prediction algorithms for rod pump artificial lift systems. It adapts state‑of‑the‑art data mining approaches to learn pre‑failure dynamic record patterns by comparing known failure cases against normal cases in the training data. Then, it applies these patterns to testing cases for failure prediction. In this framework, each well is represented by a set of attribute value time series. These attribute values include dynamometer card area, peak/min surface loads, stroke length and so on. Multiple training records are generated from these time series by segmentation using a specific time window (e.g. 14 day) that is statistically determined by cross‑correlation and autocorrelation analysis of the raw data. Then, multiple features are extracted from these segments using robust polynomial regression. Next, in order to cover more feature space and to adjust the training process avoid the bias caused by limited training samples, we use "random peek" mechanism which randomly selects one sample from each testing well and labels them as normal to improve the performance. We use moving average approach to predict failures by sliding a time window from beginning to the end of all testing data by wells and extract their features. The data set from this paper is taken from a real‑world asset using rod pump artificial lift systems. The results show that the failure prediction framework is capable of capturing future rod pump and tubing failures. Furthermore, this framework shows good potential for different event predictions.