Developing light weight models that run on edge devices for detecting anomalies and events in oil and gas extraction using unsupervised learning and similarity methods

Abstract

Oil extraction is an energy intensive process, if any fault happens it may lead to heavy losses. Electrical Submersible Pump (ESP) is the main equipment used in oil extraction process. The objective of this paper is to describe various methods and procedures for detecting the events or anomalous conditions that occur in the ESP equipment during the oil extraction. The proposed models must work on edge devices to reduce the latency time in detecting the events. The ESP equipment generates tera bytes of data every day and manual surveillance of the data is very difficult or almost impossible. ESP equipment has various sensors which measure different data like discharge pressure (DP), intake pressure (IP), temperature (T), current supplied(I) and vibration(V) etc. Each individual sensor measurement is termed as a signal. There are few methods available for event detection in ESP which uses encoders or pattern recognition models[1]. But these models are not compatible to run on edge devices as they require high computation power. The proposed methodology uses unsupervised learning and similarity methods to detect the anomalies. Simple mathematical or statistical techniques are used in building light weight edge device compatible models. Reliability and completeness of the data is important, and the quality engine identifies the data portions with bad quality, so that these portions can be removed prior to the anomaly detection. Any unusual pattern in more than one signal is considered as anomaly. Event is a known anomaly pattern, for example if discharge pressure decreases and vibration increases these are the symptoms for a solid production event happening. The similarity algorithm used in classifying anomalies into events provide the confidence score for the event swith respect to every type of event and the event type with highest score is assigned as the class label. The proposed framework has multiple models, and the size of the models are in KB’s so that the overall application can run on devices that has RAM available in MB’s and processor with a single core.