Table of Contents
ESP failure prediction uses data that an electric submersible pump already sends to surface, such as motor current, intake pressure, motor temperature and vibration, to flag wells whose pump is likely to fail soon. In a published field study the warning was up to 7 days: enough to plan a workover and protect the pump, not to replace surveillance engineers.
Key takeaways
- Across 13 industry papers covering more than 15,000 ESPs, average reported run life ranged from 10 to 66 months, with a mean of 27 months (Oil & Gas Journal, 2010).
- Motor current is measured at surface by the drive or switchboard. A downhole gauge adds intake pressure, motor winding temperature, vibration and current leakage, sampled every 5 to 25 seconds on one vendor's gauge (Baker Hughes).
- Labels are the hard part. Pull reasons and teardown findings need consistent codes, such as the ESP-RIFTS nomenclature built on ISO 14224 and API RP 11S1 (Alhanati et al., 2001).
- In one field study, a PCA and XGBoost model flagged electrical failures up to 7 days ahead with a precision of 0.80 and a recall of 0.60 (Abdalla et al., 2022).
- Measure run life with survival curves and score models per failure event, not per data point.
Why do electric submersible pumps fail?
An ESP is a multistage centrifugal pump driven by an electric motor at the bottom of the well, with a seal section between them, a power cable to surface and a drive or switchboard that powers it. ESPs are considered the second-most widely used artificial lift method (Abdalla et al., 2022). A failure means pulling the string with a workover rig, while the well produces little or nothing.
A Baker Hughes engineer listed ten factors that drive ESP run life (Vandevier, 2010). The main ones are summarised below, with the signals where they may show up.
| Cause | What it does | Where it may show in the data |
|---|---|---|
| Operation outside the pump's flow range | Serious downthrust or upthrust | Intake pressure and frequency far from the design point |
| Poor power quality | Outages, poor voltage regulation and harmonics break down motor insulation over time | Voltage imbalance, frequent trips and restarts |
| High temperature | Each 8 °C (18 °F) above the motor's thermal rating halves the dielectric life expectancy | Rising motor winding temperature |
| Free gas | Gas interference and gas lock | Unstable, fluctuating current; shutdowns |
| Solids and deposits | Sand, scale, asphaltenes and paraffin wear or plug stages and reduce motor cooling | Falling flow, rising motor temperature |
| Corrosion and cable damage | Holes in tubulars and metal parts; insulation failure | Rising current leakage, ground faults |
Starts and stops matter too. A study of wells in Mossoró, Brazil, notes that the lifespan of an ESP is affected by repeated shutdowns, and that frequent shutdowns and restarts eventually damage the motor (Brasil et al., 2023). In a polymer flood, Abdalla et al. link electrical failures to cable damage and to stator overheating, with polymer deposits on the motor reducing heat dissipation.
How long should an ESP last, and how do you measure it?
The 27-month average quoted above hides a measurement problem. The industry uses several definitions, and they give different answers (Oil & Gas Journal, 2010):
- Average run life of failed units ignores pumps still running, many of which may have run for a long time.
- Mean time between failures (MTBF) divides the running time of all units by the number of failures.
- Mean time before pulling (MTBP) counts every pull, for any reason, not only equipment failures.
Survival curves based on Kaplan-Meier analysis avoid these traps. A Baker Hughes paper argues that simple run life statistics are often used as the central key performance indicator but may not give decision makers enough information, and applies survivability curves to compare the factors that affect run life (Bryan and Irausquin, SWPSC).
Agree on the metric before the project starts, so that every baseline is calculated the same way.
What data does ESP failure prediction need?
| Source | Typical signals | Notes |
|---|---|---|
| Drive or switchboard | Motor current, frequency, voltage, trips | Measured at surface; current alone has been used to classify operating conditions |
| Downhole gauge | Intake pressure and temperature, motor winding temperature, vibration, current leakage; discharge pressure on some models | On one model, intake pressure every 5 s and most other channels every 25 s (Baker Hughes) |
| Wellhead | Tubing and casing head pressure, wellhead temperature | Used alongside gauge data in the Abdalla study |
| Well tests | Oil, water and gas rates, water cut | Sparse; a virtual flow meter can fill the gaps |
| Workover and teardown records | Pull date, reason for pull, failed component, teardown findings | The source of your labels |
Mind the gaps. ESP gauges typically send their data through the pump's power cable. Baker Hughes notes that when a ground fault occurs on the cable, the gauge signal is lost even if the motor keeps running (Baker Hughes, Zenith GFI). A gauge outage can therefore be a symptom of cable trouble, so do not simply fill it in with interpolated values.
Our guide to extracting process historian data for machine learning covers resampling and timestamps, and the oil and gas data readiness checklist helps check the records first.
How do you label ESP failures?
A failure label needs the well, the failed component, the failure mode and the time it started, not just the workover date. The ESP-RIFTS joint industry project, run by C-FER Technologies, defined a general data set and a standard nomenclature for coding ESP failures. It takes failure definitions and attributes from ISO 14224, and component and teardown terms from API RP 11S1 (Alhanati et al., 2001). Mapping your own codes to it makes data comparable across fields. The project also studies specific failure mechanisms, such as power quality, wells with high gas volume fractions, shaft seals, electrical penetrators and pothead connectors (C-FER Technologies).
Three points matter in practice:
- Separate pulls from failures. A pump pulled for a recompletion did not fail.
- Use teardown findings. The reason recorded at the pull is a first guess; the teardown says which component failed.
- Expect few examples. In the Brasil study, gas locking was not found in the field data at all, so the authors simulated those cases.
In Algerian fields, workover reports may mix French, Arabic and abbreviations; see documents that mix Arabic, French and English. Our guide on labelling failures in time-series data covers onset marking, label confidence and weak supervision in detail.
Which methods are used, and what do published studies show?
Approaches range from simple to complex:
- Ammeter chart review and rules. A classic Southwestern Petroleum Short Course paper compares the ammeter chart to an electrocardiogram of the motor and says timely analysis can catch minor problems before they become costly (Cunningham and Riling). Rules written with surveillance engineers remain the baseline any model must beat.
- Operating condition classification. Brasil et al. classified 24-hour current records from 20 wells into normal operation, normal operation with gas, gas interference and gas locking; a support vector machine reached 93% accuracy on 54 test records (Brasil et al., 2023).
- Anomaly detection on normal operation. PCA with T² and SPE, or isolation forest, trained on healthy periods of each well. No failure labels are needed to build it; see our anomaly detection methods guide.
- Supervised failure-window classifiers. Abdalla et al. resampled gauge and drive data to hourly steps, compressed it with PCA and trained XGBoost to recognise the 7 days before a workover for electrical downhole failures (Abdalla et al., 2022).
- Vibration spectrum diagnosis. The public ESPset dataset holds 6,032 vibration spectra from ESPs used in offshore oil exploration, under a CC BY 4.0 licence (ESPset, Mendeley Data).
| Study | Data | Target | Reported result | Caution |
|---|---|---|---|---|
| Abdalla et al., 2022 | Intake and discharge pressure, motor temperature, frequency, drive current, wellhead and casing pressures; polymer-flooded field | 7 days before an electrical downhole failure | Precision 0.80, recall 0.60, F1 0.71 on the pre-failure class | 339,089 normal points against 1,728 pre-workover points; the paper does not say the test set came from a later period |
| Brasil et al., 2023 | 24-hour motor current records, 20 wells in Mossoró, Brazil | Four operating conditions | 93% accuracy (SVM) | 179 records in total; gas locking cases simulated |
Neither result transfers directly to your field.
How should you evaluate an ESP failure model?
Point-by-point accuracy misleads when, as in the Abdalla data, more than 99% of the points are normal. Score the model on failure events instead:
- Events caught: of the failures in the test period, how many raised an alert before the trip?
- Warning time: was the alert early enough to plan a rig, order parts and adjust the drive?
- False alerts per well per month: surveillance engineers will stop reading a list that cries wolf.
- Time split: train on older years and test on the latest one. Hourly points from the same well are strongly correlated, so a random split can overstate results.
Compare the result with what engineers already catch with trip settings and ammeter charts.
How do you run an ESP failure prediction pilot?
- Pick a fleet and one or two failure modes, such as electrical downhole failures in wells with gauges, with a named owner in production engineering.
- Agree the run life metric and compute a baseline survival curve for the fleet.
- Collect several years of drive and gauge data from SCADA or the historian, covering as many failures as possible, with workover and teardown records.
- Rebuild the labels: separate pulls from failures and mark the onset where engineers can see it.
- Start with rules and anomaly detection; add a supervised model only where enough reviewed failures exist.
- Test on the most recent period, per event, against the current surveillance practice.
- Run in parallel for several months with alerts reviewed weekly by engineers.
- Plan retraining. Water cut, gas and solids change over a well's life; see model drift monitoring.
Where should the model run?
ESP fields are often far from the head office, with limited links. Run the model close to the data, on servers inside your network, with read-only access to SCADA or the historian and no write access to the drives. Our guide to AI at remote sites with poor connectivity covers the architecture. The same approach applies to other rotating equipment; see compressor and turbine health monitoring.
What we have seen at KABAS AI
Our oil and gas page lists predictive maintenance among its use cases, and DALILE connects production data, maintenance records and equipment histories in one system that runs on your infrastructure. Our data science service builds predictive models that anticipate failures and deploys them with monitoring on-premise. We have not published an ESP project, so what follows is guidance, not a reported result.
- Start with the workover archive, not the algorithm. Clean pull dates, reasons and teardown findings decide what a model can learn.
- Put engineers' time into the test set. A small set of well-documented failures, each reviewed by two engineers, is a more reliable test than a large set of uncertain labels.
- Search the text you already have. Teardown reports and workover notes hold the failure history; our article on root cause analysis with an LLM over maintenance history shows how to use it.
For the wider picture, read AI predictive maintenance in oil and gas.
Sources
- Abdalla et al., Machine Learning Approach for Predictive Maintenance of the Electrical Submersible Pumps (ESPs), ACS Omega 7(21), 2022
- Brasil et al., Diagnosis of Operating Conditions of the Electrical Submersible Pump via Machine Learning, Sensors 23(1), 2023
- Vandevier, ESP—1: Run-time analysis assesses pump performance, Oil & Gas Journal, October 2010
- Vandevier, ESP—Conclusion: Multiple factors affect electrical submersible pump run life, Oil & Gas Journal, November 2010
- Alhanati, Solanki and Zahacy, ESP Failures: Can We Talk the Same Language?, SPE ESP Workshop, 2001 (C-FER Technologies)
- C-FER Technologies, ESP-RIFTS research
- Bryan and Irausquin, Impact from Analyzing the Run Life Statistics Using Survivability Curves Methodology on ESP Key Performance Indicators, Southwestern Petroleum Short Course
- Cunningham and Riling, Submersible Electrical Pump Ammeter Chart Interpretation, Southwestern Petroleum Short Course
- Baker Hughes, Zenith E-Series ESP gauge, technical data sheet
- Baker Hughes, Zenith GFI ESP gauge, specification sheet
- ESPset: vibration dataset for ESP fault diagnosis, Mendeley Data, 2024
Frequently asked questions
How far in advance can an ESP failure be predicted?
It depends on the failure mode and the data. In one published field study, a model flagged electrical downhole failures up to 7 days before the workover, and caught a majority but not all of the pre-failure days. Not every failure leaves an early signature in the data, so a prediction model supports surveillance engineers rather than replacing them.
Do we need downhole gauges to predict ESP failures?
Not to start. Motor current is measured at surface by the drive or switchboard, and a published study classified gas interference and gas locking from current alone. Downhole gauges add intake pressure, motor temperature, vibration and current leakage, which help to separate well problems from equipment problems.
How many past failures do we need to train a model?
There is no fixed number. With only a few documented failures per mode, train an anomaly detector on normal operation and keep the failures to test it. A supervised model that predicts a specific failure mode needs enough reviewed examples of that mode, which usually means pooling similar wells.
What is the difference between ESP run life and MTBF?
Run life is often reported as the average running time of failed units, which ignores pumps that are still running. Mean time between failures divides the total running time of all units by the number of failures. If every pull is counted, not only failures, the result is a mean time before pulling. Survival curves give a fuller picture than any single average.
Should the prediction model be able to stop or adjust the pump?
No. Keep it advisory, reading data from SCADA or the historian and sending alerts to production and maintenance engineers. Trips and protections stay in the drive and the control system, and people decide on speed changes and workovers.



