Deep Learning approaches for novelty and concept drift detection in oil industry

Anomaly detection. Novelty detection. Deep Learning. Auto Encoders.

This work presents strategies for anomaly detection and novelty detection in oil production. Oil wells are designed with pressure and temperature sensors located in strategic points to supervise the operations. Detecting unexpected events using these sensors is a field of interest in the oil and gas companies to improve operational safety and reduce costs associated with non-productive time and failure repair. Anomalies are a combination of sensor data that escape from the normal range of values. There are classes of known anomalies that are possible to detect and anomalies with an unknown pattern, called novelties. The proposed work takes advantage of the properties of the Auto Encoder architectures, allowing us to find meaningful representations of the data. We achieve a reduction in data dimensionality and computational complexity of the classification algorithms to determine anomalies and novelties. Because the normal states change over time in an oil production well, known as concept drift, the strategy intends to overcome this issue without new training. In this context, the two significant contributions of the proposed work are I) study and overcome the concept drift in anomaly detection; and II) detect and save novelties to perform anomaly classification in oil production. The solutions are evaluated in oil wells located in Brazil with actual, simulated, and extrapolated data. We have pressure and temperature sensor data from oil wells in this data set. The work will use the existing 3W dataset, which is available for download on the Internet, and It also uses Petrobra’s uncharacterized production data. The strategy intends to be robust and scalable enough to process real-time data from several oil wells simultaneously.