Abstract
Designing systems that can detect irregularities in process monitoring data and which can be implemented efficiently is a challenge. The ability to detect and diagnose such events will help ensure stability of operations, avoiding excursions that cross thresholds reflecting safe operating limits and consequently improving process reliability and availability. Typically, using classical outlier techniques to detect anomalies assumes that a ‘normal' pattern of behavior, in a statistical sense, can be assumed and/or confirmed and used to identify what is an extreme observation. In practice, simple univariate measures of normality seldom result in effective solutions in systems whose dynamic behavior generates wide-ranging time varying data.
This paper presents a multi-criteria, data-driven approach for detecting irregularities in multivariate, time-series, dynamic systems with irregular changing patterns. The capability to automatically detect and flag such data will reduce the risk of crossing operational boundaries and improve levels of operations availability.
We developed a semi-supervised machine learning framework that provides a blue print for configuring application-specific irregularity detection. The framework takes into account the extreme dynamics typically encountered in Upstream operations that limit standard learning models' capabilities to accurately represent and hence draw inferences about the state of the underlying system. A hybrid, adaptive, non-linear, probabilistic, reasoning model learns iteratively over multiple operating windows to identify breakouts in typical behavior and to detect ‘remarkable' divergence of reconstructed signal from original space.
The method can identify single or multiple instances of irregular activity, since the underlying causal factors, resulting in such behavior, are likely to generate a number of irregularities reflecting nonconformity to normal pattern. Unexpected sequential changes in observed system behavior can be leveraged as a first-level filter of the basic anomaly signature. Results show that applying the framework to detecting irregularity in wells operations enabled near real-time flagging of irregular patterns. An operator's adjustable alert can be set based on a threshold value of the calculated score of the streaming signal.
This new approach is a significant departure from conventional methods of establishing and using baselines for normal process behavior, such as signature-based, profiling, fingerprinting, etc., to detect anomalies. Its key advantage is that it avoids the use of single-valued thresholds intrinsic to the competing methods of the types listed above. Such single-valued thresholds have the effect of limiting the practical application of such techniques in Oil and Gas applications.
