How Artificial Intelligence Response Prioritization facilitates action-focused surveillance at scale

Part I: Introduction and Case Studies of Events Detected 

Avalon Surveillance AiRP (AI Response Prioritization), is a state-of-the-art real-time surveillance system for ESPs (Electrical Submersible Pumps). This 3-part article series will present several case studies showcasing the most important features and aspects that will benefit your operation.

In Part I of the series, we’ll discuss how the AiRP engine proactively detects time series patterns that can lead to low/no-flow conditions that inflict damage to the pump or lead to production loss. Part II will cover case studies showcasing how the engine helps surveillance teams respond to events with timely actions and to optimize operations. Finally, in Part III, we discuss how the AiRP engine achieves high availability and robustness to various realistic data issues.

Important features of the AiRP engine:

• AiRP is designed to raise attention in real-time by proactively detecting developing conditions that may escalate to imminent, critical low flow events. This informs timely actions to prevent further escalation or mitigate impact. These conditions are a function of multiple measurement channels. 
• Unlike traditional approaches applying thresholds on individual channels, AiRP is easily scalable by being efficient (low ratio of events to threshold alarms) and offers consistent interpretable behavior and excellent event detection rate. A recent audit of hundreds of events detected by the AiRP engine, showed that almost 95% of detected events were valid, with the remaining false events attributable to data pipeline or problems unrelated to the events detection engine.
• AiRP is fully automated. A plug-and-play solution with no pre-requisites for design or installation information. The underlying machine-learning (ML) engine is trained and tested against a significant field dataset covering various locations, well-types, ESP designs and duration. This enables the solution to be deployed flexibly without requiring retraining or tuning. The engine automatically learns “normal” operating conditions for a given well and adapts to changing conditions. On a case-by-case basis, the overall solution can be configured at the application level (e.g., sensitivity) and in a slower loop, the core ML can be fine-tuned against more targeted datasets and/or specific event types. (You can read more about the underlying ML system in this paper).
• AiRP learns and reports reference values for each signal. These are what constitute the most recent “normal” conditions. Reference signals are useful for diagnostics and interpreting detections and help users in optimizing operational strategies. They can also set more advanced alarms on gradual operational changes, such as slow developing periods of decreasing production.

Case 1.1: Insufficient Lift (IL) event resulting in shutdown   Description: The AiRP engine raised the first IL event (A) indicating a developing low-flow condition, within minutes of a discernable and sustained change. The engine relied on multiple signatures to confirm the event: i) an increase in motor temperature, ii) a significant increase in pump intake pressure (>300 psi), and iii) a decrease in motor current. Tubing head temperature and pressure also further confirm the low-flow condition on the surface. The event continued to escalate for several hours without any action resulting in a pump shutdown. Even after the pump restart, the drawdown was not at the expected normal level and, consequently, the engine raised another IL event (B) until sufficient flow was established.   Impact: The prolonged periods of low-flow, shutdown, and time to establish stable flow from subsequent restart (total >12 hours) resulted in a significant loss of production. Additionally, stress was experienced by the pump with wasted energy and an unnecessary restart, which would contribute to cumulative damage.

Case 1.2: Gas Interference (GI) events resulting in shutdown   Description: The AiRP engine raised a series of GI events (A, B, and C) indicating systematic low-flow conditions; GI events are characterized by cyclic/repetitive behavior synchronized across multiple channels. In addition to this behavior, there are other signatures indicating escalating low-flow conditions - increasing motor temperature and pump intake pressure (>200% by the end of escalation) accompanied by more than 15% drop in motor current. No remedial action was taken after the first event for ~6 hours, eventually leading to a pump shutdown.   Impact: The prolonged periods of low-flow, shutdown, and time to establish stable flow from subsequent restart (total >12 hours) resulted in a significant loss of production. Additionally, stress was experienced by the pump with wasted energy and an unnecessary restart, which would contribute to cumulative damage.

Case 1.3: Insufficient Lift (IL) event with very late action resulting in shutdown   Description: The AiRP engine first raised an IL event (A) at point (1) primarily due to i) a sudden rise in motor temperature (>32 degF over its expected reference value), ii) a sudden rise in pump intake pressure (>35% over its expected reference value), and iii) a sudden drop in discharge pressure (>60% of its expected reference value). At point (2), we can observe the event severity declining as motor temperature and intake pressure start coming down with discharge pressure going up. However, the situation then takes a turn in the wrong direction with intake pressure levels remaining above reference and even increasing, sustaining the event for many more hours. No action was taken for close to 9 hours. Around Point (3), the pump speed was manipulated in an attempt to improve the situation. However, as the situation had escalated to that point, this was unsuccessful and ultimately led to a sudden spike in motor temperature that tripped the system.   Impact: There was sufficient time to take early action to avoid this shutdown. The prolonged periods of low-flow, shutdown, and time to establish stable flow from subsequent restart (total >15 hours) resulted in a significant loss of production. Additionally, stress was experienced by the pump with wasted energy and an unnecessary restart, which would contribute to cumulative damage.

Case 1.4: Insufficient Lift (IL) event triggered by operations in adjacent well   Description: Near Point (1), there was a sudden increase in the pump intake pressure in this well by more than 80% of its reference value accompanied by 40ºF overheating of motor temperature compared to its expected reference value. These triggered a timely detection of an IL event (A). During the event, the intake pressure suddenly dropped sharply before spiking back up. Pump discharge pressure also increased noticeably.  The surveillance team noted that these unintended disturbances were caused by operations on an adjacent well located in the same pad. Eventually, there was a pump shutdown due to system trips based on high motor temperature and low intake pressure.   Impact: This case shows how pump operations can be influenced by operations on adjacent or connected wells. There were ~30 minutes for actions to be taken to avoid a shutdown. The periods of low-flow, shutdown, and time to establish stable flow from subsequent restart (total ~5 hours) resulted in loss of production. Additionally, stress was experienced by the pump with wasted energy and an unnecessary restart, which would contribute to cumulative damage.

Case 1.5: Insufficient Lift (IL) event after startup resulting in prolonged period of lower production     Description: The pump was restarted at the same speed as before the shutdown. However, intake pressure did not decline normally as it started increasing at Point (1), indicating that the pump struggled to lift fluid at previous levels. Appropriately, the AiRP engine raised an IL event (A). The event lasted for several hours, and once intake pressure stopped increasing, the engine started learning the new normal level and removed the event at Point (2).   Impact: The well was under low flow for about 16 hours, resulting in notable production loss. Additionally, if the event had continued to escalate, there would have been additional risks of cumulative damage to the pump and likely a pump shutdown.

Case 1.6: Dead Head (DH) event caused by surface operation   Description: The AiRP engine triggered a DH event (A) despite no operational changes in the pump speed. More specifically, this is a DH-2 (Type 2) event, which indicates the flow restriction is at the surface. The engine used the following signatures to confirm this event – a significant increase in tubing head pressure (>200% over reference value), accompanied by an increase in intake pressure, an increase in discharge pressure, and a noticeable drop in motor current.   Impact: The surveillance team confirmed in retrospect that the cause for the event was a surface operation (choke manipulation) that created an unintended flow restriction for this well. In this case, the restriction was fortunately removed, before the event escalated towards a shutdown. However, the event lasted ~1.5 hours leading to additional stress on the pump; capturing such events will be important for any run life analysis. Going forward, the ability to be alerted to such conditions will enable close monitoring and attention during similar operations in the future.

Case 1.7: Broken Shaft ESP failure event   Description: The AiRP engine detected a Broken Shaft failure event (C) based on two main signatures – discharge pressure did not build up at pump restart (1), and motor current remained significantly below its reference level. The signatures also correspond to no/low-flow conditions, and hence, the engine also concurrently raised an IL event (B). In a previous pump restart (2), the pump had also encountered a brief IL event (A).   Impact: The surveillance team confirmed the Broken Shaft failure from the subsequent workover. Such events can help in failure diagnostics and provide additional (independent) confidence for operations to decide the next steps – whether to plan for a workover immediately or in parallel to initiate well interventions. When such events are captured automatically over a significant population, further analytics can be used to derive insights and optimize operations. For example, this well had several restart attempts with low flow events in a relatively short period of time before this failure (not shown in the figure); these conditions would have certainly accelerated the accumulative stress on the pump.