Method and system for detecting an anomaly of an electrical submersible pump

Abstract

The present invention relates to a method and system for detecting an anomaly of an electrical submersible pump, ESP. Specifically, the proposed invention provides: physics-based anomaly assessment, automatic data imputation using virtual sensors network (203), automatic continual learning based on automatic model drift detection, automatic failure prevention based on reinforcement learning, remaining useful life (RUL) estimation, and cohort analysis.

Description

METHOD AND SYSTEM FOR DETECTING AN ANOMALY OF AN ELECTRICAL SUBMERSIBLE PUMPFIELD OF THE INVENTIONThe present invention relates to a method and system for detecting an anomaly of an electrical submersible pump.BACKGROUND OF THE INVENTIONElectrical Submersible Pumps (ESPs) are one of the most widely employed artificial lift pumping techniques, accounting for over 6o% of the total global oil production. More than 90% of the offshore oil wells require the ESPs for pumping oil and other fluids to the surface.However, the ESP system often reaches the point of production disruptions that result in an economic loss of oil barrels. Given the magnitude of the ESP utilization, the pump failures have a considerable business impact, both in oil production deferment and replacement costs. Specifically, failures of ESP systems can cause huge losses in terms of financial costs, loss of oil production, and unplanned downtime.In energy production applications, achieving high reliability and availability is a crucial task which requires adoption of efficient maintenance policies. The fact that ESPs are complex systems, in addition to limited labeled data availability, have restricted in the past the capability to predict functional anomalies and component degradation phenomena that can lead to failures or may cause unnecessary maintenance interventions with consequent system unavailability. The lack of modelling capabilities combined with the need to mitigate risk of failure led to the adoption of preventive maintenance strategies where ESPs are run to fail with no active monitoring of the degradation phenomena of their components.Several challenges have been faced in identifying abnormal performance to predict and avoid ESP failures timely. Due to the complex nature of the ESP failures, it can be challenging to identify anomalies by simply trending data, as it is affected by several factors, including reservoir characteristics, pump operating conditions, and the installation procedure. Given the high cost of the ESP disruptions, more real-timesurveillance systems are applied to monitor ESP performance. The standard architecture of an ESP well includes proper instrumentation for downhole & surface equipment to monitor and measure performance of the entire system. Most operating oil and gas (O&G) companies currently use their internal knowledge and experience gained in ESP operations over time to monitor the performance. Still, there is a longstanding need for the O&G industry to have an anomaly detection modeling mechanism or predictive failure system that will allow for proactive ESP health condition monitoring.SUMMARY OF THE INVENTIONThe present invention describes a method and system for detecting an anomaly of an electrical submersible pump.A first aspect of the invention is a computer-implemented method for detecting an anomaly of an Electrical Submersible Pump, ESP, comprising the steps of: receiving real-time operational sensor data associated with an operating condition of the ESP; obtaining forecasted operational sensor data and / or reconstructed operational sensor data by using an unsupervised machine learning, ML, model, wherein the ML model has been trained to forecast and / or reconstruct the operational sensor data; and determining an anomaly of the ESP based on the real-time operational sensor data and based on the forecasted operational sensor data and / or the reconstructed operational sensor data.The first aspect of this invention facilitates to measure and transmit data in real-time, providing information about the performance of the well, including flow rate, pressure, temperature, and motor performance. By analyzing this data, well operators can identify potential issues early on and take proactive steps to prevent equipment failures and reduce downtime.In addition, the existing data-driven models are often built using data from a specific set of ESPs or operating conditions, which may limit their applicability to other equipment or situations. Furthermore, changes in operating conditions or equipment design may require the model to be updated or retrained, which can be time-consuming and costly. On the contrary, the first aspect of this invention uses a fully unsupervised method that may be tuned to each specific well parameter and automatically adapt tothe changes in the operating conditions or the design of the pump without the need of labeled training data.According to a second aspect of this invention, according to the first aspect, the anomaly of the ESP is determined if a deviation is detected between the real-time operational sensor data and the combination of the forecasted operational sensor data and the reconstructed operational sensor data.According to a third aspect of this invention, according to the first or second aspects, the unsupervised ML model comprises an LSTM-based forecasting deep neural network configured to forecast the operational sensor data.According to a fourth aspect of this invention, according to any one of the preceding aspects, the unsupervised ML model comprises a GAN-based ML model, preferably a GAN -based variational autoencoder model, configured to generate the reconstructed operational sensor data.According to a fifth aspect of this invention, according to any one of the preceding aspects, the reconstructed sensor data is a new set of operational sensor data generated from historical and / or regular ESP sensor data, which preferably maintains the inherent attributes and patterns of the historical and / or regular ESP sensor data.A limitation of the traditional data-driven approaches is that they are often based solely on historical data and may not consider the specific context of the current situation. For example, a model may predict that an ESP will fail based on historical data, but it may not consider factors such as changes in operating conditions or equipment modifications that could affect the failure likelihood. On the contrary, the third to fifth aspects of this invention use a fully unsupervised abnormal event detection method that automatically adapts to the current data and remains effective in detecting abnormalities due to mechanical or electrical issues.Another key limitation of the existing failure prediction models is that they require high-quality data to be effective. If the data used to train the model is incomplete, inaccurate, or inconsistent, it can lead to inaccurate predictions. The quantity of data available can also limit the accuracy of the predictions. If there are not enough examples of ESP failures to train the model, it may not be able to identify all potential failure modes. On the contrary, the third to fifth aspects of this invention use a fully unsupervised abnormal event detection method that can learn from a small amount ofhistorical information of ESP sensor data and remains effective in detecting abnormalities even in the absence of large amount of labeled training data.According to a sixth aspect of this invention, according to the preceding aspect, the method further comprises the step of generating a failure indicator score to indicate an anomaly based on the anomaly determination step.According to a seventh aspect of this invention, according to any one of the preceding aspects, the method further comprises the step of rendering a user interface for displaying an information about an anomaly and providing a recommendation for a maintenance action.According to an eighth aspect of this invention, according to any one of the preceding aspects, in the user interface rendering step, a heatmap showcasing the real-time health of the ESP is generated, wherein normal ESP conditions, low severity events, and high severity events are rendered with different formats.According to a ninth aspect of this invention, according to any one of the preceding aspects, in the anomaly determination step, physics-driven rules and / or root cause analysis are applied to assess the anomaly.According to a tenth aspect of this invention, according to any one of the preceding aspects, the unsupervised machine learning model has been trained with historical operational sensor data of the ESP.According to an eleventh aspect of this invention, according to any one of the preceding aspects, the method further comprises the steps of: comparing the real-time operational sensor data with the forecasted operational sensor data so as to detect a drift in the operational sensor data due to an external change; and re-training the machine learning model using a drift model to adjust to the drift in the data due to the external change if a difference between the operational sensor data and the forecasted operational sensor data exceeds a certain threshold.Commonly, changes in well parameters, such as pressure and flow rate, can indicate problems in the reservoir, such as sanding, scaling, or water breakthrough. Using the ninth aspect of this invention to monitor these changes, the operators can take corrective actions to prevent further damage to the reservoir and optimize production.According to a twelfth aspect of this invention, according to the preceding aspect, the external change comprises a drift in well choke position, ESP frequency and / or well running status.A thirteenth aspect of the invention is a computer-implemented method for optimizing maintenance of an ESP and failure prevention of an ESP by using a reinforcement learning, RL, model, the method comprising the steps of: defining and observing a state space representing an operating condition of the ESP for the RL model, receiving information on an anomaly of the ESP, initializing a Q-table with arbitrary values, selecting an action for maintaining or preventing failure from the ESP based on the information on the anomaly of the ESP, the operating condition and the Q-table, performing the selected action associated with the ESP operating condition and observing an effect of the selected action to the ESP, updating the Q-value of the Q-table if a positive effect to the ESP is observed; wherein these steps are repeated until a predetermined termination condition is met.According to a fourteenth aspect of this invention, according to the preceding aspect, the actions comprise adjusting an ESP frequency and / or speed.According to a fifteenth aspect of this invention, according to the thirteenth or fourteenth aspect, the termination condition is that the corresponding actions are identified for all different types of ESP failure.According to a sixteenth aspect of this invention, according to the thirteenth to fifteenth aspects, the information on an anomaly of the ESP is generated based on the anomaly detected by the method according to the method of any one of the first to twelfth aspects, preferably the method according to the nineth aspect.A seventeenth aspect of the invention is a computer-implemented method for maintaining an ESP and preventing failure of an ESP, comprising the steps of: inputting information on an anomaly of the ESP detected by the method according to any one of the first to twelfth aspects into an ML model trained by the method according to any one of eleventh to thirteenth aspects; andperforming an action for maintaining the ESP or preventing failure of the ESP, which is output from the ML model.An eighteenth aspect of the invention is a computer-implemented method for restoring real-time operational sensor data of an Electrical Submersible Pump, ESP, comprising the steps of: inputting real-time operational sensor data having missing data of an ESP into a machine learning, ML, forecasting model that has been trained to produce patterns of real-time ESP signals generated by one or more sensors comprised by the ESP; obtaining predicted virtual operational sensor data from the ML forecasting model; and restoring the real-time operational sensor data by imputing the missing data of the real-time operational sensor data with the predicted virtual operational sensor data.The eighteenth aspect of this invention avoids the possibility of mistakenly identifying the unusable data as the data that are actually transmitted from the ESP.According to a nineteenth aspect of the invention, according to the preceding aspect, the ML forecasting model is a virtual sensor network.By training generative ML models to reconstruct ESP sensory signals, the proposed virtual sensor network of the nineteenth aspect of this invention can impute missing data in real-time ESP sensor signals. Furthermore, the proposed virtual sensor network can reconstruct sensors that fail in the downhole gauge equipment due to power losses, electrical problems, and power trips. Real-time data imputation helps the operators to detect problematic well regions or sensors which suffer frequent failures or extreme abnormal deviations from the normal patterns.According to a twentieth aspect of the invention, according to the eighteenth or nineteenth aspects, the operational sensor data having missing data of an ESP is obtained by the steps of: collecting real-time operational sensor data associated with an operating condition of the ESP; detecting non-usable data in the real-time operational sensor data based on an operational status of the ESP; andremoving the non-usable data from the real-time operational sensor data.According to a twenty-first aspect of the invention, according to the preceding aspect, the operating condition of the ESP comprises a well running status, an external choke position, a change of the external choke position and / or a variation in ESP frequency; and the real-time operational sensor data associated with an operating condition of the ESP comprises ESP current, ESP voltage, ESP vibration, well head pressure, the well head temperature, ESP intake pressure, ESP intake temperature, ESP motor temperature, ESP frequency and / or a difference between the well head temperature and the ambient temperature.According to a twenty-second aspect of the invention, according to the twentieth or twenty-first aspects, the non-usable data is detected if a value of the real-time operational sensor data, preferably a maximum value and / or a minimum value of the real-time operational sensor data exceeds or falls below a predetermined threshold.The twenty-second aspect of this invention makes it possible to detect unusable data by using domain knowledge.According to a twenty-third aspect of the invention, according to any one of the twentieth to twenty-second aspects, the method further comprises the step of: generating information about the wear on the ESP motor based on the ESP motor intake temperature and the ESP motor temperature.A twenty-fourth aspect of the invention is a computer-implemented method for detecting an anomaly of an Electrical Submersible Pump, ESP, comprising the steps of the method of any one of first to twelfth aspects, wherein the real-time operational sensor data associated with an operating condition of the ESP is the real-time operational sensor data of an ESP restored according to the method of any one of eighteenth to twenty-third aspects.A twenty-fifth aspect of the invention is a method for obtaining a remaining useful life estimation of ESP, comprising the steps of: obtaining forecasted operational sensor data for a certain period of time by using a machine learning, ML, model, wherein the ML model has been trained to forecast the operational sensor data,obtaining an average useful life estimation based on historical information of ESPs with a same characteristic, obtaining a remaining useful life estimation based on at least the forecasted operational sensor data and the average useful life estimation.According to a twenty-sixth aspect of the invention, according to the preceding aspect, the method, further comprises the steps of: obtaining a total alarm time, a number of pump restarts, and / or a mean value of the forecasted operational sensor data, wherein the remaining useful life estimation is obtained based on the average useful life estimation, a total alarm time, a number of pump restarts, and the mean value of the forecasted operational sensor data.According to a twenty-seventh aspect of the invention, according to the twenty-fifth aspect or twenty-sixth aspect, the average useful life estimation is obtained by performing a cohort-based analysis method and / or a statistical analysis method of historical ESP failure data to estimate the average useful life, wherein the cohort-based analysis method for a system comprising ESPs, comprises the steps of: defining criteria for grouping ESPs comprised by the system; collecting data from the ESPs; grouping the ESPs by using the collected data according to the defined criteria; and examining the grouped ESPs so as to obtain a pattern of the ESPs of each group.According to a twenty-eighth aspect of the invention, according to any one of the twenty-fifth aspect to twenty-seventh aspect, the method further comprises the step of: displaying the remaining useful life estimation and a result of performing the cohort-based analysis method.The twenty-fifth to twenty-eighth aspects enable the detection of performance anomalies and failure patterns to support data-driven decision-making and proactive maintenance, and ultimately enhances ESP pump efficiency and reliability.A twenty-nineth aspect of the invention is an information processing system configured to perform the method according to any one of the first to twenty-fifth aspects.A thirties aspect of the invention is a computer program comprising instructions which, when the program is executed by a computer, cause the computer to carry out the steps of the method of any one of the first to twenty-fifth aspects.BRIEF DESCRIPTION OF THE FIGURESFIG. 1 shows a flow chart of a method and system for detecting an anomaly of an ESP according to one of the embodiments of this invention;FIG. 2 shows a flow chart of an unsupervised abnormal event detection pipeline method for the ESP according to one of the embodiments of this invention;FIG. 3 shows a flow chart of an Al-based virtual sensor network for a real-time ESP data imputation method according to one of the embodiments of this invention;FIG. 4 shows a flow chart of an ESP non-useable data detecting method according to one of the embodiments of this invention;FIG. 5 shows an example of reconstructing the ESP real-time data with the virtual sensor network according to one of the embodiments of this invention;FIG. 6 shows a flow chart of an Al-based ESP surveillance method according to one of the embodiments of this invention;FIG. 7 shows a flow chart of an unsupervised abnormal event detection method according to one of the embodiments of this invention;FIG. 8a and 8b show an example of training and testing an LSTM model for detecting abnormal events of the ESP according to one of the embodiments of this invention;FIG. 9 shows an example of using the unsupervised abnormal event detection method on real-time ESP data according to one of the embodiments of this invention;FIG. io shows a flow chart of a rules-based abnormality assessment method according to one of the embodiments of this invention; andFIG. n shows a flow chart of a Remaining Useful Life (RUL) estimation method according to one of the embodiments of this invention.DETAILED DESCRIPTION OF THE INVENTIONThe invention will now be described based on the figures. It will be understood that the embodiments and aspects of the invention described herein are only examples and do not limit the protective scope of the claims in any way. The invention is defined by the claims and their equivalents. It will be understood that features of one aspect or embodiment of the invention can be combined with a feature of a different aspect or aspects and / or embodiments of the invention.The electrical submersible pumping, ESP, system used for this invention is a type of pumping system that is used to produce fluid from a wellbore in a geological formation. The ESP system comprises several components, such as a submersible pump, a pump intake, a submersible electric motor, a motor controller, a motor protector, a connector, and a sensor unit. The ESP system is deployed in the wellbore by a deployment unit, such as tubing or coiled tubing, and can be powered by a ground surface power system via a power cable.The fluid flows from the formation through perforations in the wellbore casing and into the wellbore near the pump intake. The submersible pump then pumps the fluid up the tubing to the wellhead, which may be located at the seabed or the surface of the earth. The submersible electric motor drives the submersible pump and is protected from the wellbore fluid by the motor protector. The connector connects the ESP system to the deployment unit.The sensor unit collects data from various sensors that are located along the ESP system, the wellbore casing, or other regions of the wellbore. The sensor unit comprises, for example, a motor controller sensor and a downhole gauge equipment sensor. The sensors measure different parameters related to the wellbore environment, such as pressure, temperature, flow rate, vibration, etc. The data from the sensors of the sensor unit is transmitted to the system for detecting an anomaly of an ESP according to this invention (hereinafter, system of this invention), which may be a computer-implemented system at the ground surface or elsewhere. The computer- implemented system comprises Supervisory Control and Data Acquisition (SCAD A) systems for controlling, monitoring, and analyzing the ESP and its processes. Specifically, the sensor unit can collect ESP real-time sensor data, such as the ESPcurrent, the ESP voltage, the ESP vibration, the well head pressure, the well head temperature, the ESP intake pressure, the ESP intake temperature, and / or the ESP motor temperature. The ESP real-time sensor data represents operating conditions of the ESP.The system and method of this invention analyzes the data and uses several subsystems, including Al models, to predict the anomaly of the ESP system or its components. The processing system may also control the surface power system and adjust the operation of the ESP system accordingly.In order to predict the anomaly of the ESP system or its components, the system and method of this invention mainly utilize an artificial intelligence (Al) model, more specifically, a novel physics-information machine learning (ML) framework for predictive maintenance applied to ESP surveillance and condition monitoring. The system and method of this invention utilize a fusion of data-driven analytics and physics-based modelling, where data-driven modeling effectively addresses the challenges of limited available labeled data, real-time data drift, and fast-paced abnormal event detection on a multitude of ESP sensor signals. In addition, the information (e.g., domain knowledge) on the physics-based fluid flow characteristics, the heat transfer characteristics, the electro-hydraulic characteristics and the mechanical stress information of the ESP and the surrounding reservoir is used to assess the detected abnormal events for ESP failure prediction and type classification.The system and method of this invention optimize the maintenance scenario through the combination of physics-driven and data-driven methods that assess the online health status and probability of failure, by detecting abnormal events in the real-time ESP sensor data from the sensor unit. ESP specific fluid flow characteristics, health transfer characteristics, and mechanical stress related information is used to create physics-based models to assess anomalous events and their types. The system and method of this invention can predict ESP failures over the future to optimize asset operation and maintenance, minimizing downtime and residual risk. The integrated physics-driven system empowers long term prediction accuracy, whereas the data- driven analytics ensure fast event prognostics.In order to develop predictive models for ESP failures, two approaches have been made: Physics-based Modelling (Digital Twin) and ML Models.The availability of big data from online monitoring and advanced inspection technologies makes digital twins possible, intended as asset virtual replica that can be trained to detect degradation.In addition, this invention focuses on creating a robust framework that is not currently available in the market, to help O&G companies to predict imminent ESP events ahead of time with an appropriate confidence level and prescribe recommendations to prevent ESP failures.Specifically, the proposed invention provides: physics-based anomaly assessment, automatic data imputation using a virtual sensor network, automatic continual learning based on automatic model drift detection, automatic failure prevention based on reinforcement learning, remaining useful life (RUL) estimation and cohort analysis.[Unsupervised Abnormal Event Detection Pipeline Workflow]Figure 1 shows a workflow of the method and system according to preferred embodiments of this invention.In block 101, the operational sensor data, i.e., the real-time ESP sensor data, is collected, which are associated with the operating conditions of the ESP. Together with other ESP and / or well relevant data, the real-time ESP sensor data is pre-processed as input data for facilitating the subsequent steps.Specifically, the collected data is processed by removing noise, missing values, and incomplete values, and, preferably, by normalizing data, and / or detecting records which are non-useable due to frozen sensors, communication losses, or because the well is not running.More preferably, the input data are recovered by using a virtual sensor network.The failure of motor controller sensors or downhole gauge sensors on ESP installations and inside the well can not only affect the accuracy and reliability of the data collection but also on accurate detection of failure indicators, thereby shortening the lifetime of ESPs. The sensor unit maybe subject to damage from abrasion, corrosion, or other environmental factors, which can affect their performance and accuracy. By using ML models, virtual sensors can be installed in the data pre-processing pipeline which can not only impute missing data in ESP real-time sensor signals but also simulate lost / failed sensors in correlation with other sensory signals and factors such as wellgeometry, flow characteristics, and equipment configuration. Al can be used to create a virtual sensor network for the ESP based on data collected from existing sensors. By using this model, areas of the well that are subject to more frequent sensor failures can be identified.In block 102, the imputed ESP sensor data are fed into the unsupervised abnormal event detection ML models which produce abnormal events along with their failure severity and feature importance score. By using the unsupervised ML model, forecasted operational sensor data are obtained, and accordingly an anomaly of the ESP can be determined based on the operational sensor data and the forecasted operational sensor data, by comparing these two sets of data.The unsupervised ML pipeline (i.e., the unsupervised ML model) can detect abnormal events in the ESP real-time sensor data that may indicate equipment problems due to mechanical or electrical causes and produces a failure indicator to represent the severity of abnormal events. By monitoring the motor's electrical signature, the method and system of this invention can diagnose issues such as pump cavitation, motor overload, or gas locking. The operators can use this information to adjust the pumping parameters of the well and prevent further damage to the equipment. Specifically, the system for detecting abnormal events may comprise a fully unsupervised ML pipeline that forecasts future values and reconstructs normal ESP signals based on historical data, compares the forecasted values and reconstructed signals with the actual values to detect deviations in the complex signals, and generates a failure indicator score to indicate abnormal events that may result in an impending failure.The unsupervised ML pipeline for abnormal event detection comprises an ensemble of a Long Short-Term Memory (LSTM)-based forecasting model and a Generative Adversarial Network (GAN)-based reconstruction model. The two models predict errors by comparing real-time sensor data to their normal behaviors. The models are trained in an end-to-end manner by combining a loss function defined on the forecasting errors and reconstruction errors.Figure 2 shows a specific workflow of the unsupervised abnormal event detection model.In block 201, the above-mentioned sensor unit collects ESP real-time signals (i.e., the ESP real-time data) at the well site. Specifically, the sensor unit collects ESP real-time data from, for example, the motor controller sensor and downhole gauge equipmentsensors and / or the SCADA systems. The input data may further comprise the well geometry, the flow characteristics, the mechanical characteristics, and / or the ESP design configuration. In other words, the system and method of this invention use a large volume of data, including information about the pump's motor, drive, and control system, and identify patterns, trends, and anomalies in the data.In block 202, together with other ESP and / or well relevant data, the real-time ESP sensor data is pre-processed as input data. For example, the collected data is processed to remove noise, missing values, incomplete values, normalize the data, and detects records which are non-useable due to frozen sensors, communication losses, or because the well is not running.In block 203, a virtual sensor network can be implied to recover the ESP real-time data. A suite of ML models that act as virtual sensors reproduces or simulate ESP sensor signals for imputing missing values in the real-time ESP sensor data and generating real-time data for lost / failed sensors.In block 204, an unsupervised abnormal event detection pipeline (i.e., the unsupervised ML models) is deployed to detect abnormal events in the real-time ESP sensor signals using a combination of discriminative and generative ML models. The predicted abnormal events are filtered and assessed using a rules-based pre-failure prediction criterion.Current methods for assessing failures from the ESP may perform badly once operating conditions related to the well or the ESP are changed. The present invention thus provides a model drift detection sub-system for detecting and adapting the drifts, i.e., changes, of the ESP system.Accordingly, in block 205, a drift detection sub-system continuously monitors deviations of the model predictions with the actual data in reference to external inputs / changes in operating conditions. Upon detecting a significant change in operating conditions, the drift detection sub-system automatically triggers a re-training mechanism to re-train the ML models based on the latest historical ESP sensor data enabling the Al pipeline, specifically the unsupervised ML models, to adjust to the changes in the data due to the external inputs.Specifically, the subsystem can automatically detect drift in the real-time data, train the drift model based on an ML based method, re-train the unsupervised abnormalityprediction pipeline using the drift model to adjust to the changes, and apply the trained ML model to real-time data from the ESP to adapt to the latest changes in the data maintaining accuracy of abnormal event detection for failure prediction.The ESP real-time sensor data is highly prone to data drift due to changes in the reservoir conditions, external inputs, and changes in the operating conditions. This results in model drift, making model predictions inaccurate after some periods. The drift detection sub-system and method of this invention monitor external choke position changes, ESP frequency changes, and well running status changes over time, comparing model output with the actual results and detect changes in the model’s performance. Upon detecting significant deviations in the model performance over a period, the proposed drift detection method triggers model retraining causing the model to retrain over the latest historical data to adjust it to the new operating conditions.[Al-based Virtual Sensor Network for Real-time ESP Data Imputation Workflow]Herein, the above-mentioned Al-based virtual sensor network for real-time ESP data imputation system, i.e., the virtual sensor network, is described in more detail.The failure of motor controller sensors or downhole gauge sensors on ESP installations and inside the well can not only affect the accuracy and reliability of the data collection but also the accurate detection of failure indicators, thereby shortening the lifetime of ESPs. Sensors maybe subject to damage from abrasion, corrosion, or other environmental factors, which can affect their performance and accuracy. By using machine learning algorithms, i.e., the virtual sensor network, the virtual sensors can also be a part of the data pre-processing. Not only can this impute missing data in ESP real-time sensor signals but also simulate lost / failed sensors in correlation with other sensory signals and factors such as well geometry, flow characteristics, and equipment configuration. Al models can be used to create a virtual sensor network for the ESP based on data collected from existing sensors. With this model, areas of the well that are subject to more frequent sensor failures can be identified.The virtual sensor network can automatically detect non-useable records in real-time ESP sensor signals based on changes in well running status, changes in external choke positions, and variations in ESP frequency. The virtual sensor network computes rolling statistics of historical data (such as ESP Current, ESP Intake pressure, and ESPdischarge pressure signals, ESP frequency, well head temperature compared to ambient temperature) and applies statistical rules to classify frozen signals and failed sensors. In addition, the virtual sensor network can automatically detect frozen values in ESP realtime signals based on rolling statistics of minimum and maximum values within a specified window. The virtual sensor network can also automatically generate the ESP motor wear features based on the ESP motor intake temperature and ESP motor temperature.Figure 3 shows a workflow diagram illustrating the virtual sensor network. It consists of a suite of ML models that are trained to reconstruct ESP real-time signals. The virtual sensor network is applied on the live ESP data to impute missing sensor values and construct lost / failed sensors in the downhole gauge equipment.In block 301, the above-mentioned sensor unit collects ESP real-time signals (i. e. , the ESP real-time data) from motor controller and downhole gauge equipment sensors and SCADA systems.In block 302, together with the ESP real-time signals, the well geometry, the flow characteristics and the ESP design configurations are fed into the Generative Al model, i.e., the ML models, to simulate the ESP sensor signals. The suite of ML models is trained based on generative Al and learn to reproduce ESP sensory signals in a multivariate manner.In block 303, with the imputed ESP signals that are generated by the virtual sensor network, the real-time ESP data, especially the missing data therein that are caused by failed or lost sensors, are reconstructed.The system and method comprise a non-useable data detecting means to detect unusable data. The non-useable data detecting means detects the unusable data based on the real-time ESP sensor data. Specifically, if the ESP is running, the non-useable data detecting means checks if the received real-time ESP data meet with a predetermined condition. As shown in Figure 4, according to the real-time ESP sensor data, the ESP non-useable data detecting means first determines if the data is null for a period of time; if so, it is very likely that the corresponding sensor is frozen, or the communication is lost. If the parameter is not null, the non-useable data detecting means determines if the parameter is above or below a pre-determined threshold. Specifically, the method and system detect if a maximum value and / or a minimum value of a parameter or parameters of the real-time operational sensor data is above orbelow a predetermined threshold for a period of time. For example, if the ESP current is null for some minutes, the non-useable data detecting means determines that the sensor which relates to the ESP current has a problem or the connection with the corresponding sensors are lost, hence the data is unusable. Similarly, if the difference between the ESP intake pressure and the intake discharge pressure is above 20 psi, the non-useable data detecting means determines that the data is unusable. Similarly, if the well head temperature is below an ambient temperature, the non-useable data detecting means determines that the data is unusable. Similarly, if the ESP frequency is below o, then the non-useable data detecting means determines that the data is unusable. In addition, the non-useable data detecting means checks the running state of the well, so as to determine if the null data is caused by the running state of the well or the sensors / communication therebetween. Other checks may also be deployed as steps in this method, such as changes in well running status, changes in external choke positions, and variations in ESP frequency. Accordingly, the unusable data is removed from the real-time ESP sensor data (which is then the real-time ESP sensor data having missing data) and is replaced with a corresponding segment of the virtual ESP data that is generated by the ML models.In addition, the virtual sensor network can also generate ESP motor wear features based on ESP motor intake temperature and ESP motor temperature.As shown in Figure 5, if unusable data is detected, the system and method of this invention removes the unusable data and replaces the unusable data of that time period with the imputed ESR signal of the same time period.[Al-based ESP Surveillance Framework]Herein, the above mentioned unsupervised abnormal event detection pipeline (model) is described in more detail.Figure 6 shows the workflow diagram illustrating an embodiment of the method and system for detecting an anomaly of an ESP according to this invention. The method and system of this invention utilize the implementation a physics-informed approach that combines machine learning techniques with physics-based domain knowledge that supports abnormal event assessment for failure prediction.In block 601, the above-mentioned sensor unit collects ESP real-time signals (i.e., the ESP real-time data) at the well site and feeds them into a pre-processing pipeline forremoving outliers, imputing missing values, and detecting non-useable records. Preferably, after extracting relevant features from the real-time data, the pre-processed data is fed into a virtual sensor network which detects missing sensor signals and constructs the missing sensors using a set of trained ML models. Then, the input data is fed into the main Al-based ESP surveillance models.In block 602, the physics-based models are generated based on at least some of the input data. The physics-based models are used to create characteristics of fluid flow, heat transfer, and mechanical stress in the ESP and the surrounding reservoir.The physics-based models refer to mathematical or computational representations of the physical behavior and characteristics of the ESP or its sub-systems. These physicsbased models are used to simulate and predict how the ESP or its components will behave under different conditions, such as varying loads, pressures, temperatures, and other operational parameters. Physical models can include equations and algorithms that describe the physics and mechanics of the ESP, including the fluid flow, heat transfer, mechanical stress and other factors. These models help in understanding the structural integrity and performance of the ESP and its components.Specifically, the physics-based models may comprise a flow model, a heat transfer model, and / or a mechanical stress model.The model of the fluid flow represents the fluid flow in the ESP and the surrounding reservoir. By analyzing this model and combining it with sensor data, machine learning algorithms identify changes in flow characteristics that may indicate equipment problems, such as pump wear or sand accumulation.The heat transfer model represents a heat transfer in the ESP and the surrounding reservoir. By analyzing this model and combining it with sensor data, machine learning algorithms can identify changes in temperature that may indicate equipment problems, such as motor overheating or pump cavitation.The mechanical stress model represents mechanical stress in the ESP and its components. By analyzing this model and combining it with sensor data, machine learning algorithms can identify changes in stress that may indicate equipment problems, such as motor vibration or pump wear.In block 603, a fully unsupervised Al pipeline based on discriminative and generative ML models is implemented to detect abnormal events in the real-time ESP sensor data.By analyzing the above mentioned physical-based models and combining them with the real-time sensor data, the ML models identify changes in flow characteristics that may indicate equipment problems, such as pump wear or sand accumulation, identify changes in temperature that may indicate equipment problems, such as motor overheating or pump cavitation, and identify changes in stress that may indicate equipment problems, such as motor vibration or pump wear, respectively. In other words, by combining physics-based models with ML algorithms, the operators can predict when equipment maintenance is required and take proactive steps to prevent equipment failures.Specifically, the Al-based suite of ML models detects abnormal events in real-time sensor data and classify the detected events into different failure modes. By learning from real-world data that represents normal operating conditions, a fusion of generative models and LSTM autoencoders can detect when real-time sensor data deviates from the expected patterns. These abnormal events are later processed by supervised classification models that classify the events in terms of mechanical and electrical failure types.Specifically, as shown in Figure 7, the unsupervised ML pipeline for abnormal event detection comprises an ensemble of an LSTM-based forecasting model and a GAN based reconstruction model. The two models predict errors in real-time sensor data compared to their normal behaviors. The models are trained in an end-to-end manner by combining a loss function defined on the forecasting errors and reconstruction errors. During training, the Al models are fed data from normal ESP sensor data or a normal trend so that the Al models can learn the normal patterns and the correlations between normal trends in ESP data. During prediction, the models generate ESP data as forecasts and reconstructions that are compared with actual ESP sensory readings and produce their difference as errors. The underlying solution used in the Al models is based on the LSTM and GAN based variational autoencoder which works in an ensemble to generate one combined output data.The LSTM based forecasting model is deployed to forecast ESP real-time sensor signals in a multivariate manner. As shown in Figure 8a, the LSTM-based forecasting deep neural network is a specialized deep neural network model used for time-series forecasting. It is based on the Long Short-Term Memory (LSTM) architecture, which is designed to capture long-term dependencies and patterns in sequential data. TheLSTM-based forecasting deep neural network comprises three main parts: the encoder, the latent representation, and the decoder.The encoder part processes the input ESP real-time sensor data (sequence) and extracts the most important features from the real-time ESP sensor data. It compresses the input sequence into a fixed-length vector called the latent representation or code so as to capture the relevant information from the input sequence and create a compact representation that captures the essential characteristics of the ESP real-time sensor data.The latent representation is a compressed representation of the input data. It contains a condensed version of the input data that captures the most important features and patterns. This latent representation is typically a lower-dimensional vector that can be used for further analysis or generation of new data.The decoder part takes the latent representation as input and generates an output sequence. It uses the information encoded in the latent representation to reconstruct or generate new sequences that are similar to the original input sequence. The decoder’s role is to transform the latent representation back into a meaningful output sequence, i.e., output data that represents a normal sensor behavior, so that abnormalities in the real-time ESP data can be detected.The LSTM-based forecasting deep neural network architecture, with its encoder, latent representation, and decoder parts, allows for effective modeling and forecasting of the ESP sensor data by capturing long-term dependencies and patterns in its sequential data.As shown in Figure 8b, the GAN based generative model is deployed to reconstruct the original ESP sensor signals and predicts deviations between the original measurements and the predictions. A Generative Adversarial Network (GAN) based variational autoencoder model is a deep generative model that integrates both the variational autoencoder and a GAN and is a robust deep learning architecture.The variational autoencoder component of the model consists of an encoder and a decoder. The encoder maps the input ESP sensor data to a high-dimensional latent space that captures the underlying structure of the data. The decoder then maps the latent space back to the original input space. The VAE is used to generate content bysampling from the learned latent space. In other words, the VAE component of this ML model is configured to extract the attributes and patterns of normal ESP sensor data.The GAN component of the model has the decoder of the VAE as its generator. Accordingly, the decoder generates new sets of data that are similar to the training data, namely normal or regular ESP sensor data (specifically the data from the VAE that represents the inherent attributes and patterns of normal or regular ESP sensor data), so as to reconstruct the original ESP sensor signals. During the training phase of the GAN -based model, the GAN-based model is reenforced by the discriminator network attempting to distinguish between real (real-time ESP sensor data) and generated (reconstructed ESP sensor data that is generated by the GAN model and has the patterns and attributes of normal or regular ESP sensor data) data. The reconstructed data that is classified as having the patterns and attributes of the original data is then used to reward to the GAN-based model, which allows the GAN-based model to learn from its actions and improve its performance over time.The prediction errors from the two models are combined into a cumulative error which is transformed into a failure indicator value.As shown in Figures 7 and 9, with the user interface arranged in the system for displaying the predicted failure, the failure indicator values are transformed into a heatmap which shows the health of a running ESP, where, e.g., green regions in the heatmap show the normal operating conditions, yellow regions (in lighter shade in the figure) show events with low severity, and red regions (in darker shade) show events with high severity.As shown in Figure 9, first, the real-time ESP sensor signal is compared with the ESP signal that is generated by the forecasting model (LSTM) and generative model (GAN) so as to detect abnormal events (901). Then, the abnormalities from all sensors are combined so as to generate an overall failure indicator score (902). Finally, low-severity (in lighter shade in the figure) and high-severity alarms (in darker shade) are displayed to the operator or the company through a user interface based on the failure indicator score and a pre-defined threshold (903).In block 604, by using the ML model, the real-time ESP data are evaluated, and the detected abnormal events and issues related to the ESP and well are identified, which includes, for example, the presence of sand in the well bore, scaling, gas trapping, water breakage, casing / tubing failure, and plugging.Traditional data-driven approaches are effective for failure assessment when enough historical data is available. However, in the absence of adequate historical operational and maintenance records, pure data-driven models trained in a supervised manner are prone to prediction errors and are not usually interpretable for long time forecasts. Furthermore, the complex interaction of different failure modes that contribute to the overall degradation of the ESP cannot be effectively modelled in a pure supervised data-driven approach. The fact that there is no standard protocol for ESP condition monitoring also makes it difficult to compare data between different ESP systems or operators. This lack of standardization makes it challenging to train a universal ML model for failure prediction.To circumvent these challenges, the method and system of this invention provide a novel unsupervised pipeline to predict failures enabling Al-based predictive analytics generalizable for different pumps. The method and system of this invention uses a set of domain-specific rules applied to the ESP real-time sensor data specific to each abnormal event detected by the ML pipeline. The rules are derived using percentage changes in the ESP sensor signals and are tuned on the real-time ESP data. The proposed abnormality assessment results in an automatic root cause analysis with its associated features importance scores that are used to generate alarms from low to high severity.In preferred embodiments, as shown in block 605, a Root Cause Analysis (RCA) method and system may be incorporated to determine the underlying cause of an abnormal event or a failure. In other words, the system and method of this invention use a combination of statistical and ML-based methods to estimate root causes for the detected abnormal and pre-failure events. Since the system and method of this invention operate on all types of abnormalities detected in the real-time ESP sensor data, the root cause analysis is not limited to only ESP-related issues but also extended to well-related issues. Accordingly, each detected abnormal event is analyzed using a set of domain-specific rules to identify condition indicators in real-time signals (ESP + well) that represent issues such as the presence of sand in the well bore, scaling, gas trapping, water breakage, well tubing / casing issue, or annulus valve failure.Specifically, some of the ESP-related and well-related root causes are described below:Sand production can cause significant damage to ESPs, reducing their lifespan and increasing downtime. By using root cause analysis to detect sand production, operators can take proactive steps to prevent damage to the equipment.Gas locking can cause ESPs to shut down, reducing production and increasing downtime. By using root cause analysis to detect gas in the wellbore, operators can take proactive steps to prevent gas locking.Changes in well performance can indicate problems in the reservoir, such as water breakthrough, scaling, sanding, well casing / tubing failure, well plugging, or annulus valve failure. By using root cause analysis to monitor well-performance, operators can identify potential reservoir problems early on and take corrective actions.In block 6o6, alerts are generated when the predicted probability of ESP failure exceeds a predefined threshold. Accordingly, the system of this invention may notify relevant personnel via email, SMS, or other communication channels. Then, the generated alerts are validated by the relevant personnel or the operators, who can accordingly trigger failure prevention and maintenance scheduling workflow.In block 607, a detailed reinforcement learning based automatic failure prevention and maintenance scheduling, i.e., the workflow of a failure prevention and maintenance scheduling system and method, is illustrated. The schedule maintenance activities to be performed are selected based on information or an output from the ML model for detecting anomalies (i.e., the failure modes), specifically, a result from the ML model by using the root cause analysis of the detected abnormal events, to prevent unplanned downtime and reduce maintenance costs.A reinforcement learning (RL) based ML model is used to make recommendations for adjusting the operating conditions to achieve target decisions such as to minimize system failure risk, maximize well production, minimize maintenance cost, or maximize ESP efficiency. Especially, a user interface arranged in the system is configured to display such recommendations. The input to the system comprises a predicted failure indicator score, related root cause, maintenance costs, crew / component availability, and the target optimization function. The RL based ML model uses a library of historical actions to find the optimal action for the target decision function. The recommended action is monitored by the operators and upon validation, added and updated to the library of historical actions along with the operator feedback.The RL based ML model makes recommendations for adjusting the operating conditions to achieve maximum efficiency. For instance, the method adjusts the ESPspeed to avoid gas locking conditions, improve ESP efficiency, avoid overloading the motor, and for maximum production output. Some example actions are as follows:Automatic ESP speed adjustment: By analyzing sensor data such as pressure, temperature, and flow rates, the machine learning algorithm adjusts the motor speed to achieve optimal efficiency and avoid overloading the motor.Automatic flow rate adjustment: By automatically adjusting the ESP speed, the machine learning algorithm can ensure that the flow rate is optimized for maximum production.Maintenance recommendation in terms of component replacement / cleaning are produced to prevent unplanned downtime and reduce maintenance costs.[Automatic Root Cause Analysis / Rules-based Abnormality Assessment]As mentioned above, the RCA system is preferably arranged with the unsupervised abnormal event detection pipeline model. Each detected abnormal event is examined using a set of conditions (e.g., domain-specific rules) to identify condition indicators in real-time ESP and well signals that represent impending failure or issues such as the presence of sand in the well bore, scaling, gas trapping, water breakage, well tubing / casing issue, or annulus valve failure.Figure io illustrates an abnormal event assessment workflow, i.e., a workflow of an RCA system according to one embodiment of this invention. For each abnormal event detected by the ML pipeline, the real-time ESP signals are run through tree-based rules specifying percentage changes for different failure modes. The percentage change parameters shown in this figure are for illustrative purposes only. The percentage change parameters are tuned with respect to the real-world data. For example, if it is detected that the ESP WHP has a change of less than 3%, the ESP current has a change of more than 8%, and the change of the differential pressure is less than 3%, then the RCA system can determine that there is a fluid recirculation fault failure. This method is called Fault Tree Analysis (FTA). FTA is a method used to identify and evaluate the potential causes of a specific event or failure. It involves constructing a diagram that depicts the logical relationships between events and the underlying causes of the failure. The proposed invention builds an FTA to identify the root cause of ESP failures and the most critical contributing factors.Although not shown in the figures, the RCA system of this invention can use other analysis methods, such as:Fishbone diagram: A fishbone diagram, also known as an Ishikawa diagram, is a visual tool used to identify the potential causes of a problem. The diagram uses a spinelike structure with branches that represent different categories of potential causes, such as equipment, personnel, process, and environment. The proposed method creates fishbone diagrams to identify the potential causes of ESP failures and to narrow down the root cause.Failure Mode and Effects Analysis (FMEA): FMEA is a structured approach used to identify and evaluate potential failures and their effects. It involves systematically analyzing each component of the ESP system and identifying potential failure modes, their effects, and the likelihood of occurrence. The proposed invention builds on FMEA to prioritize the most critical failure modes and identify the root cause.Pareto analysis: Pareto analysis is a technique used to identify the most significant causes of a problem or failure. It involves analyzing data to identify the most common causes of failure and ranking them by frequency. The proposed invention performs Pareto analysis to identify the most critical causes of ESP failures and prioritize corrective actions.Statistical Process Control (SPC): SPC is a method for monitoring and controlling a process to ensure that it operates within a specified range of performance. It involves collecting data on the process and using statistical methods to analyze the data. The proposed invention uses SPC to identify the root cause of a failure by identifying any process deviations that may have contributed to the failure.[Failure Prevention and Maintenance Scheduling]The outcome of the prescriptive and recommendation method, i.e., the method and system of the failure prevention and maintenance scheduling, is the provision of dispositions for an optimized maintenance planning, by identifying and prioritizing actions to be performed by field operators. The proposed method examines the ranking of the anomalies and failures mode identified by the abnormal event detection model sub-system. For each failure mode, the method identifies the most likely successful dispositions extracted from a library of historical prescriptions and collected feedback. The library is continuously updated with new cases, and it reports, for each failuremode, maintenance actions performed and feedback that site operators and diagnostic analytics provide on the effectiveness of executed actions. The RL model is used for failure prevention by training an RL agent to take actions that minimize the risk of system failure. Below is an example workflow of the RL-based failure prevention method and system: a. State Representation step: The first step in using RL for the ESP failure prediction (i.e., the abnormal event detection ML model) is to define the state space that represents the current state of the system. This includes preferably variables such as temperature, pressure, vibration, and flow rate. b. Action Selection step: The action space is defined by the maintenance actions that can be taken by the RL agent. These actions preferably include, for example, replacing a component, cleaning the system, and / or adjusting the operational parameters. The RL agent then selects an action to take based on the current state of the system or a result based on the output of the abnormal event detection model. This action can be to adjust the pump speed, change the frequency of maintenance, or take other actions to prevent system failure. c. Reward Function step: The RL agent receives a reward based on the action taken and the resulting system state. The reward function is designed to encourage the agent to take actions that minimize the risk of system failure. In the context of ESP maintenance, the reward function may be defined as the reduction in the probability of mechanical failure or the increase in the system's efficiency. d. Initialize the Q-table step: The Q-table is initialized. The Q-table is a lookup table that stores the expected cumulative reward for each possible state-action pair. e. Learning step: The RL agent uses the reward signal to learn which actions are most effective in preventing system failure. The agent updates its policy over time to improve its ability to predict and prevent failures. The RL agent interacts with the environment by taking actions based on the current state and the Q-table. The Q-table is updated after each action based on the reward received and the expected future rewards. f. Deployment step: Once the RL agent has been trained, it can be deployed in a production environment to continuously monitor the system and take actions to prevent failures.The Q-learning algorithm is a type of RL algorithm that can be used in this system. A Q- learning algorithm learns the optimal action to take in each state by maximizing the expected cumulative reward. In the context of ESP maintenance, the Q-learning algorithm can learn the optimal maintenance strategy based on sensor data from the ESP system and feedback on the effectiveness of previous maintenance decisions. Below are three example strategies that can be used for automatic failure prevention of this system: a. Pump speed control: The RL-based model can adjust the speed of the ESP motor based on the flow rate and other variables to optimize the pump's performance. By examining sensor data such as pressure, temperature, and flow rates, the machine learning algorithm can adjust the motor speed to achieve optimal efficiency and avoid overloading the motor. b. Pump efficiency prediction: The ML-based model can predict the efficiency of the ESP based on real-time data, such as temperature, pressure, and power consumption. The machine learning algorithm can then make recommendations for adjusting the operating conditions to achieve maximum efficiency. c. Flow rate control: Al can optimize the flow rate of the ESP by analyzing real-time data such as reservoir pressure, production rates, and temperature. By adjusting the pump speed and other operating parameters, the machine learning algorithm can ensure that the flow rate is optimized for maximum production and efficiency.[Remaining Useful Life (RUL) estimation]In order to further facilitate the O&G companies to have a better understanding of the ESPs, a method for predicting the Remaining Useful Life (RUL) is provided for the ESP. RUL estimation is essential to prevent system failures by assessing the remaining lifespan of the pump and performing necessary maintenance or replacements.This method consists of a forecasting model that forecasts the ESP parameters for N number of days in the future. Preferably, the forecasting model is the above-mentioned unsupervised ML model for forecasting the ESP sensor signals. The forecasted signals are combined to generate a cumulative forecast. In addition, the cohort analysis and / or the statistical analysis of ESP historical data, e.g., the historical ESP failures, is used toestimate an average useful life of the pump. Finally, the RUL of the pump is estimated in real-time using a RUL formula.Specifically, as shown in Figure 11, a forecasting model, e.g., the above mentioned unsupervised ML model, is trained to forecast values of ESP parameters (in dotted lines) for a given forecasting period N (e.g., 3 months). The figure on the left shows the forecasting period and the forecasted ESP signals 1101. Accordingly, the forecasted signals for that period of time are combined to generate a combined forecasted signal 1102. Cohort analysis and statistical analysis of historical failures are then used to estimate an Average Useful Life (AUL) of the pump. Finally, the RUL is computed based on the AUL, the total alarm time (e.g., the total time of the alarms generated by the above anomaly detection method and system), the number of pump restarts, and the mean value of the combined forecast. It is given by the following formula:RUL = _ { A U L tota i_a I a r m t i me) _1 +( n u m b e r of resta rts * mea n ( co m b i n ed_f o recast) )In order to help the operators to analyze the RUL, a web application is preferably used to visualize the RUL of the ESP in real-time along with the cohort analysis results based on historical failure data.[Cohort analysis]Hereinafter, a more detailed explanation of the cohort analysis method and system is described.The cohort analysis method is a powerful analytical technique used to gain insights into the behavior and performance of specific groups or cohorts of customers, users, or entities over time. This method is very suitable for the maintenance and management of well ESP. Other well pumps can also use this method for management.For ESPs, the cohort analysis method involves categorizing pumps into groups based on certain characteristics or criteria, and then tracking the performance and behavior of these groups over time. The cohort analysis method starts by defining criteria for grouping the ESPs, followed by collecting data from any one of the other sub-system of this invention, grouping the ESP by using the collected data according to the predetermined criteria; examining the cohorts (i.e., the grouped ESPs) so as to obtain an pattern of the ESPs of each cohorts.The following is a more detailed step-by-step description of how the cohort analysis method can be applied to the ESP:Cohorts Defining step: Identify the criteria or characteristics that will be used to categorize the ESPs into cohorts. These criteria could include pump type, pump manufacturer, well location, and well reservoir.Cohorts Creating step: Group the ESPs into cohorts based on the chosen criteria. For example, cohorts can be created based on the pump manufacturer, such as pumps manufactured by a certain manufacturer vs. pumps manufactured by another manufacturer.Data Collection step: Gather data on the performance, failure and maintenance history of each well pump in the cohorts.Time Period Defining step: Define the time periods over which the cohorts require analysis. Cohort analysis typically looks at data over a series of time intervals, such as months, quarters, or years.Performance Metrics Step: Determine the key performance metrics which need to be analyzed for each cohort. For well pumps, this could be Survival Rate.Analysis step: Analyze the data for each cohort over the specified time periods. Compare the performance metrics between cohorts to identify trends, differences, and patterns. For example, the ESPs from a certain manufacturer may have a higher Survival Rate compared to pumps from another manufacturer according to an analysis result of the method.For example, as shown in the following table, in order to calculate survival rates, once the cohort data is prepared, the function calculates survival percentages for each year to failure. For each cohort (installation year), the function computes the remaining equipment count for each year to failure. This is done by subtracting the cumulative number of failures from the total cohort count. The survival Rate for each year is calculated by dividing the remaining equipment count by the total cohort count and then multiplying by too. This provides the percentage of equipment that survived up to a specific year.Insights Identification step: Examine the analysis result and obtain a pattern based on a result of the examination. The cohort analysis method helps to identify insights that may not be apparent when looking at the data as a whole. For example, that ESPs installed in a particular year tend to have similar performance patterns and that ESPs in certain locations require more frequent maintenance.Decision-Making: Use the insights gained from cohort analysis to make data- driven decisions. For ESPs, this could involve optimizing maintenance schedules, selecting more reliable pump models, or planning for replacements based on the expected lifespan of different cohorts.Iteration and Improvement step: Cohort analysis is an iterative process. Continuously monitor and update the cohorts as new data becomes available and refine the analysis to make more informed decisions and improve the performance of the well pump systems.Therefore, the cohort analysis method is a valuable tool in failure detection for the ESPs by helping identify patterns, anomalies, and early warning signs of potential issues. Specifically, the system and method for cohort-based analysis of ESP systems enables the detection of performance anomalies and failure patterns to support data-driven decision-making and proactive maintenance, ultimately enhancing ESP pump efficiency and reliability.The described approach can be implemented through computer-based methods, computer-readable media with instructions, and computer systems with memory and processors.Implementations of the subject matter and the functional operations described in this specification can be implemented in digital electronic circuitry, in tangibly embodied computer software or firmware, in computer hardware, including the structures disclosed in this specification and their structural equivalents, or in combinations of one or more of them. Software implementations of the described subject matter can be implemented as one or more computer programs, that is, one or more modules of computer program instructions encoded on a tangible, non-transitory, computer- readable computer-storage medium for execution by, or to control the operation of, data processing apparatus. Alternatively, or additionally, the program instructions can be encoded in / on an artificially generated propagated signal, for example, a machinegenerated electrical, optical, or electromagnetic signal that is generated to encode information for transmission to suitable receiver apparatus for execution by a data processing apparatus. The computer-storage medium can be a machine-readable storage device, a machine-readable storage substrate, a random or serial access memory device, or a combination of computer-storage mediums. Configuring one or more computers means that the one or more computers have installed hardware, firmware, or software (or combinations of hardware, firmware, and software) so that when the software is executed by the one or more computers, particular computing operations are performed.The term "real-time," "real time," "realtime," "real (fast) time (RFT)," "near(ly) realtime (NRT)," "quasi real-time," or similar terms (as understood by one of ordinary skill in the art), means that an action and a response are temporally proximate such that an individual perceives the action and the response occurring substantially simultaneously. For example, the time difference for a response to display (or for an initiation of a display) of data following the individual's action to access the data may be less than i ms, less than i sec, or less than 5 sees. While the requested data need not be displayed (or initiated for display) instantaneously, it is displayed (or initiated for display) without any intentional delay, taking into account processing limitations of a described computing system and time required to, for example, gather, accurately measure, analyze, process, store, or transmit the data.The terms "data processing apparatus," "computer," or "electronic computer device" (or equivalent as understood by one of ordinary skill in the art) refer to data processing hardware and encompass all kinds of apparatus, devices, and machines for processing data, including by way of example, a programmable processor, a computer, or multiple processors or computers. The apparatus can also be, or further include special purpose logic circuitry, for example, a central processing unit (CPU), an FPGA (field programmable gate array), or an ASIC (application-specific integrated circuit). In some implementations, the data processing apparatus or special purpose logic circuitry (or a combination of the data processing apparatus or special purpose logic circuitry) may be hardware- or software-based (or a combination of both hardware- and software-based). The apparatus can optionally include code that creates an execution environment for computer programs, for example, code that constitutes processor firmware, a protocol stack, a database management system, an operating system, or a combination of execution environments. The present disclosure contemplates the use of data processing apparatuses with or without conventional operating systems, for example LINUX, UNIX, WINDOWS, MAC OS, ANDROID, IOS, or any other suitable conventional operating system.A computer program, which may also be referred to or described as a program, software, a software application, a module, a software module, a script, or code can be written in any form of programming language, including compiled or interpreted languages, or declarative or procedural languages, and it can be deployed in any form, including as a stand-alone program or as a module, component, subroutine, or other unit suitable for use in a computing environment. A computer program may, but need not, correspond to a file in a file system. A program can be stored in a portion of a file that holds other programs or data, for example, one or more scripts stored in a markup language document, in a single file dedicated to the program in question, or in multiple coordinated files, for example, files that store one or more modules, sub-programs, or portions of code. A computer program can be deployed to be executed on one computer or on multiple computers that are located at one site or distributed across multiple sites and interconnected by a communication network.

[0106] While portions of the programs illustrated in the various figures are shown as individual modules that implement the various features and functionality through various objects, methods, or other processes, the programs may instead include a number of sub-modules, third- party services, components, libraries, and such, as appropriate. Conversely, the features and functionality of various components can be combined into single components, asappropriate. Thresholds used to make computational determinations can be statically, dynamically, or both statically and dynamically determined.The methods, processes, or logic flows described in this specification can be performed by one or more programmable computers executing one or more computer programs to perform functions by operating on input data and generating output. The methods, processes, or logic flows can also be performed by, and apparatus can also be implemented as, special purpose logic circuitry, for example, a CPU, an FPGA, or an ASIC.Computers suitable for the execution of a computer program can be based on general or special purpose microprocessors, both, or any other kind of CPU. Generally, a CPU will receive instructions and data from and write to a memory. The essential elements of a computer are a CPU, for performing or executing instructions, and one or more memory devices for storing instructions and data. Generally, a computer will also include, or be operatively coupled to, receive data from or transfer data to, or both, one or more mass storage devices for storing data, for example, magnetic, magneto-optical disks, or optical disks. However, a computer need not have such devices. Moreover, a computer can be embedded in another device, for example, a mobile telephone, a personal digital assistant (PDA), a mobile audio or video player, a game console, a global positioning system (GPS) receiver, or a portable storage device, for example, a universal serial bus (USB) flash drive, to name just a few.Computer-readable media (transitory or non-transitory, as appropriate) suitable for storing computer program instructions and data includes all forms of permanent / non- permanent or volatile / non-volatile memory, media and memory devices, including by way of example semiconductor memory devices, for example, random access memory (RAM), read-only memory (ROM), phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), erasable programmable read-only memory (EPROM), electrically erasable programmable readonly memory (EEPROM), and flash memory devices; magnetic devices, for example, tape, cartridges, cassettes, internal / removable disks; magneto-optical disks; and optical memory devices, for example, digital video disc (DVD), CD-ROM, DVD+ / -R, DVD- RAM, DVD-ROM, HD-DVD, and BLURAY, and other optical memory technologies. The memory may store various objects or data, including caches, classes, frameworks, applications, modules, backup data, jobs, web pages, web page templates, data structures, database tables, repositories storing dynamic information, and any otherappropriate information including any parameters, variables, algorithms, instructions, rules, constraints, or references thereto. Additionally, the memory may include any other appropriate data, such as logs, policies, security or access data, reporting files, as well as others. The processor and the memory can be supplemented by, or incorporated in, special purpose logic circuitry.To provide for interaction with a user, implementations of the subject matter described in this specification can be implemented on a computer having a display device, for example, a CRT (cathode ray tube), LCD (liquid ciystal display), LED (Light Emitting Diode), or plasma monitor, for displaying information to the user and a keyboard and a pointing device, for example, a mouse, trackball, or trackpad by which the user can provide input to the computer. Input may also be provided to the computer using a touchscreen, such as a tablet computer surface with pressure sensitivity, a multi-touch screen using capacitive or electric sensing, or other type of touchscreen. Other kinds of devices can be used to provide for interaction with a user as well; for example, feedback provided to the user can be any form of sensory feedback, for example, visual feedback, auditory feedback, or tactile feedback; and input from the user can be received in any form, including acoustic, speech, or tactile input. In addition, a computer can interact with a user by sending documents to and receiving documents from a device that is used by the user; for example, by sending web pages to a web browser on a user's client device in response to requests received from the web browser.The term "graphical user interface," or "GUI," may be used in the singular or the plural to describe one or more graphical user interfaces and each of the displays of a particular graphical user interface. Therefore, a GUI may represent any graphical user interface, including but not limited to, a web browser, a touch screen, or a command line interface (CLI) that processes information and efficiently presents the information results to the user. In general, a GUI may include a plurality of user interface (UI) elements, some or all associated with a web browser, such as interactive fields, pulldown lists, and buttons. These and other UI elements maybe related to or represent the functions of the web browser.Implementations of the subject matter described in this specification can be implemented in a computing system that includes a back-end component, for example, as a data server, or that includes a middleware component, for example, an application server, or that includes a front-end component, for example, a client computer having a graphical user interface or a Web browser through which a user can interact with animplementation of the subject matter described in this specification, or any combination of one or more such back-end, middleware, or front-end components. The components of the system can be interconnected by any form or medium of wireline or wireless digital data communication (or a combination of data communication), for example, a communication network. Examples of communication networks include a local area network (LAN), a radio access network (RAN), a metropolitan area network (MAN), a wide area network (WAN), Worldwide Interoperability for Microwave Access (WIMAX), a wireless local area network (WLAN) using, for example, 802.11 a / b / g / n or 802.20 (or a combination of 802.1 lx and 802.20 or other protocols consistent with this disclosure), all or a portion of the Internet, or any other communication system or systems at one or more locations (or a combination of communication networks). The network may communicate with, for example, Internet Protocol (IP) packets, Frame Relay frames, Asynchronous Transfer Mode (ATM) cells, voice, video, data, or other suitable information (or a combination of communication types) between network addresses.The computing system can include clients and servers. A client and server are generally remote from each other and typically interact through a communication network. The relationship of client and server arises by virtue of computer programs running on the respective computers and having a client-server relationship to each other.While this specification contains many specific implementation details, these should not be construed as limitations on the scope of any invention or on the scope of what may be claimed, but rather as descriptions of features that may be specific to particular implementations of particular inventions. Certain features that are described in this specification in the context of separate implementations can also be implemented, in combination, in a single implementation. Conversely, various features that are described in the context of a single implementation can also be implemented in multiple implementations, separately, or in any suitable sub-combination. Moreover, although previously described features maybe described as acting in certain combinations and even initially claimed as such, one or more features from a claimed combination can, in some cases, be excised from the combination, and the claimed combination may be directed to a sub-combination or variation of a sub-combination.Particular implementations of the subject matter have been described. Other implementations, alterations, and permutations of the described implementations are within the scope of the following claims as will be apparent to those skilled in the art.While operations are depicted in the drawings or claims in a particular order, this should not be understood as requiring that such operations be performed in the particular order shown or in sequential order, or that all illustrated operations be performed (some operations maybe considered optional), to achieve desirable results. In certain circumstances, multitasking or parallel processing (or a combination of multitasking and parallel processing) maybe advantageous and performed as deemed appropriate.Moreover, the separation or integration of various system modules and components in the previously described implementations should not be understood as requiring such separation or integration in all implementations, and it should be understood that the described program components and systems can generally be integrated together in a single software product or packaged into multiple software products.Accordingly, the previously described example implementations do not define or constrain this disclosure. Other changes, substitutions, and alterations are also possible without departing from the spirit and scope of this disclosure.Furthermore, any claimed implementation is considered to be applicable to at least a computer-implemented method; a non-transitory, computer-readable medium storing computer-readable instructions to perform the computer-implemented method; and a computer system comprising a computer memory interoperably coupled with a hardware processor configured to perform the computer-implemented method or the instructions stored on the non-transitory, computer-readable medium.

Claims

CLAIMS1. A computer-implemented method for detecting an anomaly of an Electrical Submersible Pump, ESP, comprising the steps of: receiving real-time operational sensor data associated with an operating condition of the ESP; obtaining forecasted operational sensor data and / or reconstructed operational sensor data by using an unsupervised machine learning, ML, model, wherein the ML model has been trained to forecast and / or reconstruct the operational sensor data; and determining an anomaly of the ESP based on the real-time operational sensor data and based on the forecasted operational sensor data and / or the reconstructed operational sensor data.

2. The method according to claim 1, wherein the anomaly of the ESP is determined if a deviation is detected between the real-time operational sensor data and the combination of the forecasted operational sensor data and the reconstructed operational sensor data.

3. The method according to any one of the preceding claims, wherein the unsupervised ML model comprises an LSTM-based forecasting deep neural network configured to forecast the operational sensor data.

4. The method according to any one of the preceding claims, wherein the unsupervised ML model comprises a GAN -based ML model, preferably a GAN -based variational autoencoder model, configured to generate the reconstructed operational sensor data.

5. The method according to any one of the preceding claims, wherein the reconstructed sensor data is a new set of operational sensor data generated from historical and / or regular ESP sensor data, which preferably maintains the inherent attributes and patterns of the historical and / or regular ESP sensor data.

6. The method according to any one of the preceding claims, further comprising the step of generating a failure indicator score to indicate an anomaly based on the anomaly determination step.

7. The method according to any one of the preceding claims, further comprising the step of rendering a user interface for displaying an information about an anomaly and providing a recommendation for a maintenance action.

8. The method according to the preceding claim, wherein in the user interface rendering step, a heatmap showcasing the real-time health of the ESP is generated, wherein normal ESP conditions, low severity events, and high severity events are rendered with different formats.

9. The method according to any one of the preceding claims, wherein, in the anomaly determination step, physics-driven rules and / or root cause analysis are applied to assess the anomaly.

10. The method according to any one of the preceding claims, wherein the unsupervised machine learning model has been trained with historical operational sensor data of the ESP.

11. The method according to any one of the preceding claims, further comprising the steps of: comparing the real-time operational sensor data with the forecasted operational sensor data so as to detect a drift in the operational sensor data due to an external change; and re-training the machine learning model using a drift model to adjust to the drift in the data due to the external change if a difference between the operational sensor data and the forecasted operational sensor data exceeds a certain threshold.

12. The method according to the preceding claim, wherein the external change comprises a drift in well choke position, ESP frequency and / or well running status.13- A computer-implemented method for optimizing maintenance of an Electrical Submersible Pump, ESP and failure prevention of an ESP by using a reinforcement learning, RL, model, the method comprising the steps of: defining and observing a state space representing an operating condition of the ESP for the RL model, receiving information on an anomaly of the ESP, initializing a Q-table with arbitrary values, selecting an action for maintaining or preventing failure from the ESP based on the information on the anomaly of the ESP, the operating condition and the Q-table, performing the selected action associated with the ESP operating condition and observing an effect of the selected action to the ESP, updating the Q-value of the Q-table if a positive effect to the ESP is observed; wherein these steps are repeated until a predetermined termination condition is met.

14. The method according to the preceding claim, wherein the actions comprise adjusting an ESP frequency and / or speed.

15. The method according to claims 13 or 14, wherein the termination condition is that the corresponding actions are identified for all different types of ESP failure.

16. The method according to any one of claims 13 to 15, wherein the information on an anomaly of the ESP is generated based on the anomaly detected by the method according to the method of any one of claims 1 to 12, preferably the method according to claim 9.

17. A computer-implemented method for maintaining an Electrical Submersible Pump, ESP and preventing failure of an ESP, comprising the steps of: inputting information on an anomaly of the ESP detected by the method according to any one of claims 1 to 12 into a machine learning, ML, model trained by the method according to any one of claims 11 to 13; and performing an action for maintaining the ESP or preventing failure of the ESP, which is output from the ML model.

18. A computer-implemented method for restoring real-time operational sensor data of an Electrical Submersible Pump, ESP, comprising the steps of: inputting real-time operational sensor data having missing data of an ESP into a machine learning, ML, forecasting model that has been trained to produce patterns of real-time ESP signals generated by one or more sensors comprised by the ESP; obtaining predicted virtual operational sensor data from the ML forecasting model; and restoring the real-time operational sensor data by imputing the missing data of the real-time operational sensor data with the predicted virtual operational sensor data.

19. The method according to the preceding claim, wherein the ML forecasting model is a virtual sensor network.

20. The method according to claims 18 or 19, wherein the operational sensor data having missing data of an ESP is obtained by the steps of: collecting real-time operational sensor data associated with an operating condition of the ESP; detecting non-usable data in the real-time operational sensor data based on an operational status of the ESP; and removing the non-usable data from the real-time operational sensor data.

21. The method according to the preceding claim, wherein the operating condition of the ESP comprises a well running status, an external choke position, a change of the external choke position and / or a variation in ESP frequency; and the real-time operational sensor data associated with an operating condition of the ESP comprises ESP current, ESP voltage, ESP vibration, well head pressure, the well head temperature, ESP intake pressure, ESP intake temperature, ESP motor temperature, ESP frequency and / or a difference between the well head temperature and the ambient temperature.

22. The method according to claims 20 or 21, wherein the non-usable data is detected if a value of the real-time operational sensor data, preferably a maximumvalue and / or a minimum value of the real-time operational sensor data exceeds or falls below a predetermined threshold.

23. The method according to any one of the claims 20 to 22, further comprising the step of: generating information about the wear on the ESP motor based on the ESP motor intake temperature and the ESP motor temperature.

24. A computer-implemented method for detecting an anomaly of an Electrical Submersible Pump, ESP, comprising the steps of the method of any one of claims 1 to 12, wherein the real-time operational sensor data associated with an operating condition of the ESP is the real-time operational sensor data of an ESP restored according to the method of any one of claims 18 to 23.

25. A method for obtaining a remaining useful life estimation of an Electrical Submersible Pump, ESP, comprising the steps of: obtaining forecasted operational sensor data for a certain period of time by using a machine learning, ML, model, wherein the ML model has been trained to forecast the operational sensor data, obtaining an average useful life estimation based on historical information of ESPs with a same characteristic, obtaining a remaining useful life estimation based on at least the forecasted operational sensor data and the average useful life estimation.

26. The method according to the preceding claim, further comprising the steps of: obtaining a total alarm time, a number of pump restarts, and / or a mean value of the forecasted operational sensor data, wherein the remaining useful life estimation is obtained based on the average useful life estimation, a total alarm time, a number of pump restarts, and the mean value of the forecasted operational sensor data.

27. The method according to claims 25 or 26, wherein the average useful life estimation is obtained by performing a cohort-based analysis method and / or astatistical analysis method of historical ESP failure data to estimate the average useful life, wherein the cohort-based analysis method for a system comprising ESPs comprises the steps of: defining criteria for grouping ESPs comprised by the system; collecting data, preferably performance, failure and / or maintenance history data, from the ESPs; grouping the ESPs by using the collected data according to the defined criteria; and examining the grouped ESPs so as to obtain a pattern of the ESPs of each group.

28. The method according to any one of claims 25 to 27, further comprising the step of: displaying the remaining useful life estimation and a result of performing the cohort-based analysis method.

29. An information processing system configured to perform the method according to any one of claims 1 to 28.

30. A computer program comprising instructions which, when the program is executed by a computer, cause the computer to carry out the steps of the method of any one of claims 1 to 28.

Images