Prognostic Health Management of Pumps Using Artificial Intelligence in the Oil and Gas Sector: A Review

Abstract

A system’s operational life cycle now includes an integrated health management and diagnostic strategy due to improvements in the current technology. It is evident that the life cycle may be used to identify abnormalities, analyze failures, and forecast future conditions based on current data. Data models can be trained using machine learning and statistical ideas, employing condition data and on-site feedback. Once data models are trained, the data-processing logic can be integrated into onboard controllers, allowing for real-time health evaluation and analysis. Interestingly, the oil and gas industries may encounter numerous obstacles and hurdles as a result of the integration, highlighting the need for creative solutions to the perplexing problem. The potential benefits in terms of challenges involving feature extraction and data classification, machine learning has received significant research attention recently. The application and utility in pump system health management should be investigated to explore the extend it can be used to increase overall system resilience or identify potential financial advantages for maintenance, repair, and overhaul activities. This is seen as an evolving research area, with a variety of application domains. This article present a critical analysis of machine learning’s most current advances in the field of artificial intelligence-based system health management, specifically in terms of pump applications in the oil and gas industries. To further understand its potential, various algorithms and related theories are examined. Based on the examined studies, machine learning shows potential for prognostics and defect diagnosis. There are, few drawbacks that is seen to be preventing its widespread adoption which prompt for further improvement. The article discussed possible solutions to the identified drawbacks and future opportunities presented. This study further elaborates on the commonly available commercial machine learning (ML) tools used for pump fault prognostics and diagnostics with an emphasis on the type of data utilized. Findings from the literature review shows that the neural network (NN) is the most prevalent algorithm employed in studies, followed by the Bayesian network (BN), support vector machine (SVM), and hybrid models. While the need for selecting appropriate training algorithms is seen to be significant. Interestingly, no specific method or algorithm exists for a given problem instead the solution relies on the type of data and the algorithm’s or method’s aptitude for resolving the provided errors. Among the various research studies on pump fault diagnosis and prognosis, the most frequently discussed problem is a bearing fault, with a percentage of 46%, followed by cavitation. The studies rank seal damage as the third most prevalent flaw. Leakage and obstruction are the least studied defects in research. The main data types used in machine learning techniques for diagnosing pump faults are vibration and flow, which might not be sufficient to identify the condition of pumps and their characteristics. The various datasets have been derived from expert opinion, real-world observations, laboratory tests, and computer simulations. Field data have frequently been used to create experimental datasets and simulated data. In comparison to the algorithmic approach, the data approach has not received significant research attention.

1. Introduction

Pumps are a common item of hydraulic equipment, used in both industrial and residential settings, and their primary function in high-volume fluid handling is to control oil-flow operations. Their areas of application include power plants, oil refineries, building services industry, water engineering, chemical and process industries, and biomedical applications, to mention a few [1]. However, the persistent use of high-speed pumps without maintenance causes serious damage to the entire assembly by destroying their mechanical components. Meanwhile, using classical models to predict reliability is no longer viable. A typical reliability evaluation methodology relies primarily on the pumps’ lifespan data, and failure data may be challenging to obtain.

Furthermore, prior lifespan statistics may not fully depict the pump’s behavior in a specific application or environment [2]. As a result, studies reveal that maintenance expenditures account for over a third of overall operating costs in most cases. According to Alraghbi [3], maintenance costs account for about fifteen to seventy percent (15% to 70%) of the production. In another study by Wang, it was revealed that maintenance costs is about forty percent (40%) of the total cost [4]. The performance characteristics of most industrial pumps may be connected to their reliability estimations as they degrade over timeDue to this finding, degradation assessment methods have recently gained popularity due to their capability to monitor the operational status of a system over time. The traditional maintenance solutions usually applied are divided into two categories, namely, corrective and preventative maintenance. Corrective maintenance is used to fix complex systems and equipment only after they have failed, leading to an increase in the processes direct expenses.

In contrast, preventive maintenance is carried out regularly to avoid equipment breakdowns. Thus, preventative repairs to machinery or components are performed when their remaining useful life is unknown, resulting in unplanned downtime and increased operating costs [5]. Consequently, a scheduled maintenance or preventive approach should be provided to enhance overall equipment condition while lowering the rate of equipment failure, reducing the cost of maintenance, and increasing the useful lifetime of equipment [6]. During the fourth Industrial Revolution (Industry 4.0), industries continuously seek ways to improve production processes while minimizing costs [7].

Besides these examples, various terminologies and categories of maintenance management strategies may be found in other research [8,9,10,11,12,13,14,15,16,17,18,19,20]. The study considered the categories provided by the research works of Susto et al. [21,22]. In their study, they categorized the maintenance procedures into three groups, compared with the existing researchers. These maintenance procedures can be described as follows.

Run-to-Failure (R2F) maintenance, also known as corrective maintenance: this maintenance type occurs only when a piece of equipment fails to function correctly. It is the most basic maintenance method, as it necessitates a halt in production and the repair of the components to be replaced, both of which add a direct cost to the operation. On the other hand, preventive maintenance (PvM), often referred to as time-based maintenance (TBM) or scheduled maintenance (SMS) is a preventative maintenance approach that is conducted regularly, according to a predetermined schedule in time or process iterations, to anticipate equipment problems. It is a typically successful method of avoiding failures. Whilst, the predictive maintenance (PdM) method employs predictive technologies to identify when maintenance operations are required. It is based on the constant assessment of the reliability of a machine or process, allowing maintenance to be conducted only when necessary. In addition, it enables early failure identification using prediction tools, based on historical data, for instance, machine learning techniques, integrity factors such as visual features, wear, statistical inference methods, and the engineering approach [23]. Figure 1 depicts the types of maintenance strategies and procedures explained above.

The advantages of PdM include extending the period when equipment is utilized and operated. Examples include delaying/reducing maintenance operations and cutting material and labor costs [6]. According to Jardine et al. [24], there are three sorts of PdM approaches suitable for monitoring equipment conditions for diagnostic and prognostic purposes. These are model-based techniques, statistical approaches, and artificial intelligence approaches [25]. Figure 2 depicts the extension of the life of equipment using PdM.

The authors demonstrated in Figure 2 how the use of PdM in the prognosis of pumps might extend the life of the equipment by detecting faults or the onset of issues at an early stage, prior to the equipment’s total failure. Early detection permits early maintenance to be performed to prevent failure. As a result of the early employment of maintenance, the pumps’ lifespan is extended. Additionally, the oil and gas industry views oil field digitization as a brand-new possibility for greater production efficiency. The crucial issue is how to put these tools into practice in a way that all known risks are managed, value is truly delivered, the actual outcomes can significantly affect the operation’s profitability, and, of course, they are applicable to specific, set production optimization targets. Certain aspects of oil production applications, such as reservoir management, production optimization, artificial lift, flow assurance, and predictive maintenance were among the important areas where prior research in this field had the potential to further modernize technology. This study primarily focuses on the application of machine learning to comprehend equipment, specifically pumps status, in order to assist in predictive maintenance and save operational downtime.

This study presents an overview of artificial intelligence (AI) approaches and focuses on the use of machine learning algorithms for system health monitoring. In recent years, machine learning has been used for fault diagnosis and prediction. This expansion has, so far, included mechanical equipment monitoring, electrical systems, power installations, and aerospace disciplines. This comprises techniques for diagnosing electromechanical equipment faults, classifying degradation and pattern identification, and predicting component Remaining Useful Life (RUL). The current state-of-the-art is a compilation of articles that were examined as part of an ongoing study on the use of machine learning. The authors provided a thorough assessment on ML by taking a practical approach and focusing their work on pump-system health management. Interestingly, there are other publications offering far more in-depth research, wherein researchers conducted thorough analyses on machine learning within the oil and gas industry, e.g., see [27,28,29,30,31,32,33]. However, design engineers and researchers who work in the fields of testability, diagnostic algorithm design, and health monitoring technologies within the context of artificial intelligence research should find this paper to generally be of interest. It can be used to explain the idea of machine learning and identify applications for it. When current implementation technology is taken into account, the study also identifies several problems. Although the solutions being developed for pump applications are the primary focus, other disciplines should be able to find the information in this article to applicable. The objectives of this study can be summed up as follows:

• ✓ Provide an overview of the concepts of conventional maintenance strategies applied in previous studies.
• ✓ Analyze current trends in the state of the art regarding the application of machine learning works related to fault diagnosis and prognosis using various fault detection techniques in pumps.
• ✓ Develop a coherent understanding of machine learning in supervised, unsupervised, and semi-supervised algorithms.
• ✓ Develop an appreciation of the core merits of machine learning through a sector-wise view of the technology.
• ✓ Provide the current progress improvements, including their basics, pros and cons, and the standard commercially available Machine Learning (ML) tools.
• ✓ Provide practical insights for academia and industries on which algorithms are best-suited for certain problems and how to approach the problems, while enabling quicker and more efficient maintenance decision0making.

Other related works studying ML from 2010 to 2022 and the type of data used in the studies were reviewed. The findings were elaborated to help operators meet life-extension requirements while optimizing their cost structure and conclusions; future recommendations are also made.

The study offers an organized and comprehensive overview of multi-application domain machine learning research on system health management. The existing literature on the subject either focuses on one or more research areas or on a specific application domain. The writers first outline the fundamentals of each technique, before highlighting its many variants. Accordingly, developments in the literature are explored. This template offers a clearer and more concise overview of the various strategies while noting their advantages and disadvantages. While some of the current research focuses on applications for pumps in the oil and gas industries, interest from the point of view of machine learning is growing. The various applications of these strategies are listed in this publication. However, a significant feature of real-time applications is the computational complexity of these techniques, which is not discussed in many of the evaluated studies. As it appear to be no clear way to choose, build, or implement a machine learning architecture because the majority of ML implementations in this field are application- or equipment-specific, according to the authors. As a result, one of the main objectives for future ML implementations is to spread defect diagnosis and prognosis toward lower design levels so that detections or forecasts can happen more closely to the actual occurrence and, consequently, localization is made possible.

Additionally, there are no any developed end-to-end solutions or adequate benchmarking provisions in the outcomes. Which algorithms will perform better for certain applications are not always obvious because these strategies frequently call for extensive parameter and framework customization. It is significant to highlight the paper’s main contributions.

This paper offers a critical analysis of the growing body of research on the use of ML for pump health management. ML-based fault detection is gaining popularity, although the majority of methods are application- or equipment-specific, making it unclear how to choose, construct, or use ML algorithms. Application complexity, end-to-end learning solutions, suitable benchmarking, and estimating the costs of adopting the algorithms are all areas lagging behind in research and deep understanding. The remainder of the paper is organized as follows: Section 2 focuses on the diagnostics and prediction of faults in pumps and the relevant studies for the PHM of pumps in the oil and gas sectors. The literature review process is presented in Section 3, using the standard PRISMA methodology. Section 4 contains the ML algorithms adopted in related studies, backgrounds, and applications, including their particular advantages and disadvantages. The section further explains the primary outcomes for the ML algorithm queries and data sources. Section 5 brings this study to its conclusion and Section 6 offers our recommendations and future prospects for research.

2. Diagnostics and Prediction of Faults in Pumps

Prognostics and health management (PHM) is a significant field of study since it forms the basis for advanced prediction technologies [34]. It aids not only by determining how well the apparatus functions but also by predicting when a failure will occur and minimizing the effects of these unforeseen failures [34]. Forecasting is a significant focus of PHM applications, which also aids in determining the system’s remaining useful life (RUL). Additionally, it supports creating a current maintenance strategy [34]. The conceptual difference between a diagnosis and a prognosis is significant in terms of time. Diagnosis is identifying the nature and cause of a specific phenomenon. This method is used to identify the underlying cause of a system failure after it has occurred [34]. Prognosis, on the other hand, is derived from a Greek word that denotes foreknowledge and fore-sensing [34]. It aids in the prediction of faults before they can occur. The primary goal of the prognosis is to forecast the event before it may happen; hence, time is a crucial aspect in this scenario, rather than diagnosis. Meanwhile, Figure 3 represents an overview of condition-based maintenance (CBM), relating to diagnostic and prognostic maintenance, as in the work of Tchakoua et al. [35]. CBM is achieved in three primary processes, as illustrated in Figure 4, including data collecting with sensors, signal processing employing various data approaches, and feature extraction via the collection of characteristics that will assist in determining the present condition of the observed equipment.

As demonstrated in Figure 3, the system’s information from current and past status, obtained from the data, can be used to detect or predict faults in pumps. Corrective maintenance is performed once a defect is diagnosed, to address the failures. If, on the other hand, a failure is predicted, preventative maintenance is performed before the fault occurs. Additionally, faults in pumps may be operationally created, system faults, mechanically caused, or any combination of these. Mechanical pump problems are brought on by malfunctioning components, such as bent rotors, misalignments, and bearing issues. System flaws, on the other hand, involve improper installation and leaking. However, operationally generated faults—primarily obstruction, cavitation, and flow-related issues—occur while the pump is functioning [36].

According to Tiwari [1], due to the pressure below the vapor pressure suction, cavitation is reported to be the typical fault that most frequently develops in pumps, as described in several studies [1,36,37,38,39,40,41,42,43]. Apart from cavitation, another recurring problem in pumps is damage to the casing and impeller, which can be caused due to the pulsation of pressure that is associated with internal recirculation. A further problem is the blockage of the suction and discharge pipe caused by solid elements and impurities in the liquid being pumped. Regardless of the pump’s condition, these faults can progress at any phase of its operation. A study by Grudfos [44] shows a breakdown of the causes of failure in pumps, with a description of the repair cost for each failure, as demonstrated in Figure 4.

Traditionally, fault diagnoses have been achieved by physically evaluating the health of equipment. This adds to the labor intensity while also affecting the accuracy of the diagnosis. At the most fundamental level, fault diagnosis methods may be categorized as either model-based or data-driven. Model-based fault diagnosis methods utilize a mathematical model of the observed system. Using such a model, estimations of system/process outputs can be developed, which are then contrasted against actual process outputs to provide a residual signal or innovation. Based on the values and attributes of the produced residual signal, probable fault situations are identified, based on a comparison between the model outputs and the actual system outputs.

Figure 5 shows the perceptions of a typical model-based approach to fault diagnostics. However, advanced signal processing algorithms can assist in determining what sorts of failures occurred and where they occurred in the equipment. Additionally, the diagnosis outcomes from signal processing techniques are too technical for machine users to comprehend. Therefore, during the fault detection procedure, parameters indicating the state and functionality of the pumps are monitored. These parameters can be identified as process parameters, including fluid flow, temperature, pressure, electric current, power, and efficiency.

Another set of parameters used in fault detection is that of fault character parameters, including the peak value, virtual value, amplitude spectrum, phase spectrum, and power spectrum [45]. In addition, oil analysis parameters are also vital to study. This is carried out by analyzing the material, concentration, and the wear particle’s dimension distribution, using oil analysis and the characteristic statistical parameters, such as the mean time before failure (MTBF), reliability, availability, and maintainability. These parameters can be considered fault symptom signals, indicating the characteristics of potential faults [35]. As a result, current industrial applications select defect diagnostic technologies that can automatically categorize machine health statuses [45].

According to Olsen et al. [46], PHM is now the safest method for managing the safety state of equipment. These are accomplished by the methodical implementation of recent test results in AI and IT technologies. The authors stated that PdM could save maintenance expenses and extend the RUL. PdM primarily focuses on utilizing predictive data to plan future maintenance activities correctly. In addition to the process data and parameters, PdM strives to capture data on the physical condition of a machine or component., including data on pressure, vibration, temperature, viscosity, acoustics, viscosity, and flow rate [47]. At the same time, this information is now often used for equipment health assessments, defect identification, early problem detection, and forecasting future equipment conditions. In addition, intelligent fault diagnosis (IFD) is anticipated to achieve the goals mentioned earlier using machine learning theories [48]. The traditional theories of machine learning, such as support vector machines (SVM) [36,37,39,49,50,51,52,53,54,55] and artificial neural networks (ANN) [27,45,49,56,57,58,59,60,61,62,63,64], have been employed to analyze machine faults in the oil and gas industries. The use of ML algorithms in industrial applications for fault detection has been the subject of numerous comprehensive literature reviews and research, but there are few concerning the oil and gas sector.

The research conducted by Carvalho et al. [6] include a systematic review of the literature (SLR) that concentrated on machine learning methodology, utilized alongside predictive maintenance methods. Lei et al. [65] also established a field roadmap and an SLR on machine learning applications for machine malfunction diagnosis. Del Ser et al. [66] analyzed the most recent progress in data analysis techniques and machine learning algorithms in the Industry 4.0 paradigm. The authors noted the sharp rise in publications since 2010 on machine fault diagnostics using machine learning. The authors reported locating more than twenty articles on predictive prognostic models, but only one on the oil and gas sector. Interestingly, only few studies, have specifically addressed the use of ML in the oil and gas industry. This includes research study by Costello et al. [55], who presented a data-driven model for gas circulator unit health tracking using real vibration data as part of this body of work, while Dong et al. [48] also revised the research on the intelligent diagnosis of oil transfer pump failures. Lee et al. [67] examined PHM design for rotary machinery systems, discussing the various approaches for categorizing important components and tools for choosing the most suitable algorithm for a particular application. Similarly, review by Srivyas et al. [68] on various maintenance approaches, types of failure, and failure detection techniques used in hydraulic pumps. A significant number of the previous studies focused on faults that occur one at a time, this is evident based on the above findings. However, concurrent faults are an essential consideration in practice [67]. Beckerle et al. [69] classified the methods of detecting faults, based on the signals evaluated and models used, as depicted in Figure 6. Furthermore, Nagendra [70] classified the fault detection techniques into model-based, as in [71], data-driven, as in [72], and hybrid fault detection, as in [73].

In the signal-based approaches, only the output signals are examined. Meanwhile, signal-model-based methods extract features using a specific signal model that matches the predicted output signal characteristics. The appropriate models include Fourier transform models for harmonic signals, statistical representations for random signals, correlation analysis, and wavelet analysis.

Table 1. Signal-based approaches.

Authors	Parameter	Method	Fault
[74] Wang et al., 2007	Statistical parameters	Neuro-Fuzzy classifier	Cavitation, impeller damage, and imbalance
[75] Sakthivel et al., 2010	Statistical features extracted from vibration signals	Decision tree and rough sets Fuzzy classifiers	Bearing, seal, and impeller faults
[76] Soylemezoglu et al., 2011	Sensors for lateral acceleration, vertical acceleration, and inlet and outlet pressure, as well as outlet flow	Mahalanobis–Taguchi System	Clogging of filter and seal faults or impeller
[77] Xie Gang et al., 2012	Inlet upper flow rate of the tank, outlet flow of the pump, flow detection unit, inlet valve of the pump, tank level, the inlet pressure of the pump, liquid level detection unit, and outlet valve of the pump	Signed directed graphs	Motor circuit damage and power outage of the impeller and the drive shaft
[78,79] Kallesoe 2006	Current spectrum signatures	Signal analysis	Different fault detection methods for blockages, cavitation, and damaged impellers

presents the signal-based methods applied in studies.

The signal-based approach was adopted in several studies for the detection of various faults, such as cavitation, blockages, impeller damage, inadequate bearing seals, and power outage, as shown in Table 1. Using methods such as signal analysis, neuro-fuzzy classifier signed directed graphs, and the Mahalanobis–Taguchi system, statistical parameters were taken from vibration analysis or current spectrum analysis for the given problems. Conversely, process model-based approaches compared the process’s input and output signals with the system model. The process model-based approaches include parameter estimation, neural networks, observers, state estimation, and parity equations.

Table 2. Signal model-based approaches.

Authors	Parameter	Method	Fault
[80] Brandenburgischen 2015	Pressure fluctuation and mechanical vibration signals	Neuro-fuzzy classifier	Impeller misalignment
[81] Kafka T 1999	Pressures, temperatures, power, and flow	Fast Fourier transformation (FFT)	Imbalance, blockade, and wear
[82] Kollmar 2002	Pressures, temperatures, power, and flow	Neural networks, discriminants, and Bayes classifiers	Imbalance, blockade, and wear
[83] Kiggen 2006	Flow, temperature, angular velocity, pressure, and the magnetic field	Decision tree classifier	Gas fault states in fluid, cavitation, and blockade
[84] Wang et al., 2007	Flow, temperature, angular velocity, pressure, and the magnetic field	Discrete wavelet transforms, a neuro-fuzzy classifier based on rough set theory	Cavitation, misalignment, and impeller damage
[85] Mendel et al., 2008	Horizontal, axial, and vertical accelerations	K-nearest-neighbor approach, as well as a multi-layer perceptron neural network	Bearing faults
[86] Stopa et al., 2014	Currents and rotor positions	Load torque signature analysis	Cavitation
[87] Yunlong et al., 2012	Frequency spectra	Bayesian belief network, as well as a support vector machine	Incorrect alignment, imbalance, and foundation looseness

shows the studies that applied this method for detecting different faults.

As can be seen in Table 2, model-based strategies were employed with mathematical models in studies to discover faults. Methods such as fast Fourier transformation, Bayes classifiers, neural networks, Bayesian belief networks with support vector machines, load torque signature analysis, discrete wavelet transforms, neuro-fuzzy classifiers, and decision tree classifiers were applied to resolve the given problems. The majority of the parameters are pressure fluctuation, temperature, flow, angular velocity, and frequency spectra. Based on the table, most of the research focused on model-based solutions that leverage signal or process models [6]. However, Harihara and Parlos [88] argued that a data-driven strategy is more capable of detecting faults than a model-based approach. This is because the data-driven method is based on examining and processing raw signals collected from the system. As a result, Bachschmid [89] recommended investigating numerous rotor system failures.

Table 3. Process model-based approaches.

Authors	Parameter	Method	Fault
[90] Geiger 1985	Input, output, and differences in pressure	Parameter estimation scheme	Bearing without lubrication, polluted bearing, increased ring backlash, impeller defect, volute defect, cavitation, insufficient ventilation, increased fluid temperature, and changed valve position
[91] Nold 1991	Pressure, temperature, and flow signals	Parameter estimation	Wear of suction-sided seal gap, deposits, wear at the impeller outlet, impeller blade fracture, cavitation erosion at the impeller inlet
[92] Liu et al., 1994	Motor current, voltage, angular velocity, and fluid speed	Observer-based fault diagnosis	Deviating coefficients in process and actuator, sensor faults
[93] Patton et al., 1997	Motor current, suction, and discharge pressures	Observer-based fault diagnosis	Detection of multiple faults occurring at once

presents the studies based on the process-model techniques applied in detection, along with what faults were addressed.

It has been observed that parameter estimation and observer-based fault diagnosis are the most commonly used methods in studies, as shown in Table 3. It has been challenging to source earlier studies on the raw data from operational machinery, despite a thorough literature analysis. Hajizadeh et al. [94] recently described the recent application of AI for defect identification in the oil and gas industry. The importance of evaluating fault detection, and its advantages it is widely used for strategic management and technological enabling. However, as a result of the inadequate readiness of industries to share ownership of data, notably in the area of petrochemical applications, various studies rely on information derived through simulations or experimental tests performed in a laboratory. The following section provides an overview of the machine learning methods used in predictive maintenance for pumps.

3. Methodology

Understanding the current research regarding ML algorithms and pump health management is one of the objectives of this study. The study investigates previously published materials that provided knowledge of the potential applications and demonstrated academic curiosity about the main trends, important works, and future directions of research. As a result, the authors have made an effort to establish an organized reference point for the expanding body of literature on this new topic. The scope of this project spans the years 2010 to 2022 because there is a growing practical need for research in this area. This research is based on reviewing a range of journal articles and conference papers that are all closely relevant to pump health management principles and ML applications, in order to achieve the study’s objectives. The articles were discovered to be dispersed throughout a variety of sources due to the breadth and diversity of these methodologies.

This paper discusses the findings of the investigation in regard to the following questions: (1) What machine learning approaches and variables are available for predicting the reliability and integrity of oil and gas pumps? (2) Which datasets are significant for predicting the dependability and structural integrity of oil and gas pumps? The trajectory of the publications included in this study is depicted in Figure 7 illustrating how rapidly the use of machine learning methods for assessing the dependability of pumps has increased. Modeling based on machine learning is a developing and novel method. Prior to 2009, few publications on the topic existed. After that point, however, the number of publications began to increase, a trend that is expected to continue into the foreseeable future.

However, as emphasized in this study, many review studies focused on general issues instead of specific context. Hence, the trend of ML applications for pumps can be found in Figure 7. A total of three hundred and fifty-three (353) articles were found from 2010 to 2022. It is evident that there has been a rapid increase in publication numbers over the years, with 2022 having the highest publication rate in ML, recording eighty-eight (88) publications.

3.1. Search Methodology

The literature for this review was gathered using a systematic search technique. This method of searching was designed with databases such as the Web of Science, Science Direct, and Scopus in mind. The most commonly used keywords were determined initially by reading the abstracts of the majority of research publications on the issue, as shown in

Table 4. Search strategy keywords.

Principal Terms	Oil and Gas Pumps	Prognostic and Diagnostic	Reliability	Failure	Machine Learning
Derived terms	Crude oil pumps	Detection	Remaining useful life	Degradation	Artificial intelligence
	Hydrocarbon pumps	Prevention	Integrity	Critical failure
	Gas pumps		Probability of failure

. The following keywords were used in the search queries: ((“Machine learning*”) AND (“pumps*”) AND (“oil” OR “gas”) AND (“prognosis” OR “diagnosis”)). The search included the years 2010 to 2022 and included journal articles and conference proceedings that were published in English. Figure 7 entails the process flow of the methodology of the search.

3.2. Selection Criteria

The PRISMA declaration was utilized as the foundation for the criterion that was selected. In the course of the investigation, the primary objective was to map the existing body of engineering, maintenance, and reliability literature on the topic of “machine learning-based prognostics and health management of oil and gas pumps.” Articles on artificial neural networks, support vector machines, fuzzy logic, Bayesian networks, deep learning, and other methods used in research are included in this collection. The years 2010 through 2022 were included in the scope of the search. We did not include any of the publications that were published prior to 2010.

3.3. Criteria for Inclusion and Exclusion

This section establishes the inclusion and exclusion criteria shown in

Table 5. Article filtering criteria based on machine learning.

Inclusion and exclusion criteria	The papers are based on machine learning models The papers that predict the reliability of oil and gas pumps The papers predict the parameters that contribute to failed oil and gas pump reliability
Exclusion criteria	The papers are not based on machine learning The papers do not predict the reliability of pumps for oil and gas The papers are not peer-reviewed

. These inclusion and exclusion criteria were constructed in response to the research questions raised in Section 3 of this study.

3.4. Quality Evaluation

The research was conducted, utilizing primary materials, journal articles, and conference proceedings. All duplicates were extensively examined to ensure the quality of the review. Throughout the review process, the abstracts of the publications were thoroughly evaluated and then refined to verify the quality and relevance of the academic content. Each item was thereafter examined in depth. The quality of publications included in this study was determined using the filtering criteria in Table 5. The authors used models based on machine learning to perform a diagnosis or prognosis of oil and gas pump failures.

3.5. Extraction of Data

After scooping and deleting the duplicates, specific papers were chosen that were in accordance with the subjects of the research. Four hudred and ninty-seven (497) articles were initially returned. The chosen works must comprise original journal articles, conference papers, or conference proceedings. Engineering, reliability, and integrity-related topics must be covered in English-language articles. To reach this study’s goals, a systematic literature review methodology was used. This focused on peer-reviewed publications of the most recent studies on the dependability of oil and gas pumps. Due to its systematic approach, which includes a thorough explanation of the methods used to choose, scan, and evaluate the literature, with the purpose of eliminating bias and promoting transparency, a structured literature review is more narrative in nature. SLR was first used in the medical and healthcare fields by Cook et al. [95]. SLR is typically seen as superior to more conventional and less systematic review procedures since it makes it simple for other researchers to corroborate the study’s conclusions. SLR gives authors the ability to conduct a thorough and rigorous analysis of the literature. It has been employed from 2002 until the end of 2020.

4. Machine Learning (ML) Algorithms

Machine learning (ML), as described by Amruthnath et al. [96], is a subclass of AI that can learn with little or no extra help from human beings. Consequently, ML aids in the resolution of numerous problems, including vision, large data, robotics, and iris recognition, as discussed in [97,98,99]. ML algorithms are categorized into four groups: Figure 7 illustrates these categories as supervised, unsupervised, semi-supervised, and reinforcement learning (RL) groups. Furthermore, ML is centered on condition-monitoring data, which is utilized to meet a variety of objectives. For instance, discussions on failure diagnostics can be found in [100,101,102,103]. Based on these studies, diagnostics are a situation in which the existence of a failure may be identified, according to its parameters without halting and dismantling the asset being monitored. At the same time, a more complicated case is the prediction of failures before they occur [103,104]. After a defect has been identified, the equipment’s remaining useful life (RUL) is evaluated [105,106]. This section discusses the relevant literature on successful predictive maintenance applications, highlighting the concept of algorithms. An overview of the ML algorithms can be seen in Figure 8.

According to Wuest et al. [107] (see Figure 8) defined unsupervised machine learning as an algorithm that finds clusters, depending on the data that is available, while unsupervised ML is essentially a specific ML approach that seeks to learn the system architecture without a recognized output in supervised ML or feedback, as in RL. The three fundamental types of unsupervised learning are clustering, self-organizing maps, and association rules. In comparison, Çınar et al. [108] categorized ML into classification, regression, and clustering.

Furthermore, several commercial tools have been employed in studies. To identify and characterize the failures, the researchers examined the frequency spectrum visualizations of vibration data. Amihai et al. [109] suggested several measures that aid this analysis, such as noise sensitivity and the estimation of spinning speeds.

Table 6. Common ML tools.

Tools	Annotations
TensorFlow IBM	Collection of open-source software for performing computational calculations.
IBM Watson Studio RapidMiner	ML framework developed specifically for an AI-driven business.
RapidMiner	Embraces the entirety of the data science lifecycle, including but not limited to data planning, machine learning, and the deployment of prediction algorithms.
Google Cloud AI Platform	Machine learning can learn from any data, in any amount.
Box skills	Organize your data, then, at scale, draw insights from it.
Google Cloud AutoML	Develop superior models that are tailored to their company’s demands.
SAS Enterprise Miner	Simplifies the data mining procedure so that models may be quickly created and the most important patterns for processes can be found.
MATLAB	A software package made by MathWorks that can be used for programming, modeling, and simulation.
IBM Watson Machine Learning	Use preexisting data for model development, testing, and production use in machine learning and deep learning.
Anaconda Enterprise	Leverage ML/AI/data science techniques.
Amazon SageMaker	Models may be constructed, trained, and deployed rapidly, regardless of scale.
IBM Decision Optimization	Uses both mathematical and AI methods to improve planning and scheduling decisions.
IBM Cloud Pak for Data	Transforms its data gathering and data analysis practices to integrate AI throughout its organization.
BigML	Learning through programmable machines.
H2O	Early disease identification, medication discovery, and individualized medical care.
Oracle Data Science Cloud Service	Model training, deployment, and management in the Oracle Cloud.
Domino	Accelerate the process of creating and deploying predictive models.
Deep Cognition	Build ML models without code.
KNIME Analytics	A platform for data analytics, reporting, and integration
Qubole	A platform for ML, streaming, and ad hoc analytics using data lakes.

describes the available standard ML tools.

4.1. Progress of Machine Learning in Pumps Application

As science and technology advance, artificial intelligence technology is increasingly being applied to mechanical manufacturing and automation. Artificial intelligence technology creates a production model through a computer simulation system and conducts extensive data analysis to make the relevant precautionary measures in the event of an emergency, which ensures an orderly production system, reduces the potential capital loss of manufacturing enterprises, and greatly improves manufacturing efficiency and accuracy [110]. Industrial pumps are crucial components of all industries and require adequate maintenance, often known as condition monitoring. Additionally, the oil and gas industry views oil field digitization as a brand-new possibility for greater production efficiency. The crucial issue is how to put these tools into practice in such a way that all known risks are managed, value is truly delivered, the actual outcomes actually affect the operation’s profitability, and, of course, they are applicable to specific, set production optimization targets.

Aspects of oil production applications, such as reservoir management, production optimization, artificial lift, flow assurance, and predictive maintenance were among the important areas where prior research in this field had promised to further modernize technology. This primarily is focused on the application of machine learning to comprehend equipment status, in order to assist in predictive maintenance and save operational downtime. Electrical submersible pumps are one of the most extensively utilized artificial lift methods (ESP).

Machine learning (ML) techniques have gained popularity in numerous academic domains, including CBM, in recent years [33]. Techniques such as support vector machines (SVM) and random forest (RF) were popular at first. Then came artificial neural network-based techniques (ANN), which performed well in both diagnostics and prognostics [33]. The benefits of ML include the ability to evaluate enormous amounts of data and to identify particular trends and patterns that people would not notice. For example, for an e-commerce site such as Amazon, knowing its users’ browsing patterns and past purchases enables the site to offer them the most appropriate goods, discounts, and reminders. It takes advantage of its findings to show them relevant adverts. The user no longer has to supervise a project at every stage, due to the advent and robustness of ML. Giving machines the ability to learn enables them to make predictions and independently improve their algorithms. As ML algorithms mature, their accuracy and productivity continue to increase. They can consequently make quick and intelligent selections. Borrowed from the example of weather forecast model, the algorithms become faster while making more accurate predictions as the dataset expands. In dynamic or uncertain contexts, machine learning algorithms are adept at managing data that is multidimensional and multivariate. In the end, it is used in numerous applications. However, ML also has a few drawbacks, including the fact that AI tools require a sufficient quantity of high-quality training data before they can function well in operational mode. While smarter algorithms may aid in obtaining better results from smaller datasets, bad data cannot be improved by manipulation. Thus, access to large and high-quality data is a major facilitator and obstacle for the effective development of AI systems. Oil and gas fields produce infinite quantities of raw data. In the oil and gas business, there are identified challenges regarding the quality and accuracy of field data, as well as a general absence of huge volumes of labeled data. Training datasets must be meticulously gathered using a well-planned workflow and circumstance-specific multi-year procedure. To increase the value of the data that oil and gas firms own or have access to, they must restructure and modify their organizational structures and procedures [111]. Another major challenge is the ability to accurately interpret the results generated by the algorithms. The researcher must also carefully choose the algorithms for their purposes. Machine learning is autonomous but is also highly susceptible to errors.

According to an average simulation, by 2030, over 70% of organizations may accept the use of at least one form of AI technology, but fewer than half may have fully absorbed the five categories [112]. The number of people will increase in the future. Monitoring the status of a machine is a procedure that cannot avoid failure but can forecast the potential of failure or a fault condition by measuring certain parameters. If an algorithm for machine learning can be developed, the system will be more efficient and it will be possible to detect faults at the ground level, thereby extending the pumping system’s lifespan. Among the numerous machine learning techniques, classification and regression analysis are utilized in the majority of cases to find anomalies. When discrete systems are present, classification is typically employed, whereas regression is utilized when continuous functions are present. There are further machine learning algorithms that, with the aid of a predictive control model, can assess system anomalies. The predictive control hybrid model is seen as a new area of research, whereby researchers can predict anomalies to reduce the loss of energy, resources, and time, and make the system faultless.

Based on the Onepetro platform’s search results, the number of AI-related articles has increased significantly since 2000, with the artificial neural network (ANN), fuzzy logic, support vector machine (SVM), hybrid intelligent system (HIS), genetic algorithm (GA), particle swarm optimization (PSO), decision tree (DT), random forest (RF), k-nearest neighbors (KNN), recurrent neural network (RNN), CNN, and deep belief network (DBN), etc., being the most prominent algorithms that are widely used in the different applications related to oil and gas. This indicates that researchers are becoming increasingly interested in the application of artificial intelligence in the oil business, and among all algorithms, the ANN method is the most researched.

4.2. Artificial Neural Network (ANN)

Haykin [113] defines a neural network as a massively parallel distributed processor, composed of simple processing units, which has a natural predisposition for accumulating and making available experiential information. The network acquires knowledge from its environment through a learning process, while interneuron connection strengths, also known as synaptic weights, are utilized to retain the gained information. An ANN is composed of interconnected neurons that are arranged in layers. Each layer consists of a number of simple neuron-processing units, known as nodes or neurons, which communicate with one another via weighted numerical connections. It is composed of n layers of neurons, two of which are input and output layers, respectively. The former is the first and only layer that receives and transmits external signals, while the latter is the final layer and the one responsible for transmitting the results of computations. Hidden layers are the n-2 innermost layers that, in relays, extract pertinent characteristics or patterns from incoming signals. The output layer is then directed toward those properties that are deemed essential. Complex neural networks may contain a number of hidden layers, feedback loops, and time-delay components that are designed to make the network as successful as possible at identifying significant information or patterns. Figure 9 depicts the simplified construction of a typical ANN with a single input layer, hidden layer, and output layer [114].

According to Bello et al. [113], ANNs are a useful tool for analyzing and diagnosing the nonlinear behavior of complex systems and can be used by operators and decision-makers as a beneficial performance assessment tool. Even if the underlying links are difficult to articulate or are unknown, ANNs are programmed to learn from past examples and recognize nuanced functional relationships among the presented data. Complex physical processes with nonlinear, high-order, and time-varying dynamics, as well as those for which analytic models may not yet exist, can be modeled without difficulty using these methodologies. Neural networks are superior to the classic linear approaches for modeling and forecasting nonlinear time series. A lack of systematic techniques when creating neural network models is likely the leading reason for variations in the reported results [113].

4.3. Support Vector Machine (SVM)

Vapnik introduced SVMs in the late 1960s, based on the notion of statistical learning. However, beginning in the middle of the 1990s, methods for SVMs began to emerge as processing power became more accessible, paving the way for several practical applications. The fundamental SVM addresses two-class situations in which the data are separated by a hyperplane and a set of support vectors. For completeness, a brief introduction to SVM is offered below. The SVM can be regarded as a way to build a classification line or hyperplane between two datasets. In a two-dimensional setting, the SVM’s operation may be explained simply and without the loss of generality. In Figure 10, a succession of circles (class A) and squares (class B) represent two distinct data classes (class B). The SVM attempts to position a linear boundary (solid line) between the two classes and orient it so that the margin (shown by dashed lines) is maximized. The SVM attempts to position the boundary so that the distance between it and the nearest data point in each class is maximized. The border is then positioned in the center of this distance between the two spots. As support vectors, the nearest data points are used to establish the margins (SV, represented by the gray circle and square). Once the support vectors have been chosen, the remainder of the feature set can be discarded, as the SVs provide all the information required for classification.

4.4. Genetic Algorithms

These algorithms are used to seek the solution space by simulating the “survival of the fittest” evolutionary process. They are used to solve linear and nonlinear problems by examining all parts of state space and utilizing the potential regions by applying mutation, crossover, and selection procedures on individuals in a population. The use of a genetic algorithm necessitates the consideration of six fundamental issues: chromosome (genome) representation, selection function, genetic operators such as mutation and crossover for reproduction function, initial population creation, the termination criterion, and evaluation function. The interest in real-coded or floating-point genomes for multidimensional parameter optimization problems is increasing because of the proximity of the second type of representation to the problem space, greater average performance, and more efficient numerical implementation. The representation of the genome depends on the specific topic being considered. Real-coded genomes and the accompanying genetic operators were utilized to identify the features and classifier parameters in this study. In the GAs, a population of ten individuals with randomly produced genomes was employed. This population size was designed to ensure a relatively substantial exchange between genomes within the population and to limit the chance of population convergence [115].

4.5. Particle Swarm Optimization (PSO)

Particle swarm optimization (PSO) is a population-based algorithm, inspired by the behavior of swarms during foraging. When using a PSO, each data point will remember the optimal location at which it was initially placed. With some variation in outcomes, the PSO system has been put to use in a wide variety of empirical applications.In particle swarm optimization, each solution to a given optimization problem is considered a particle. Throughout the iterations of PSO, the positions of the particles advance toward the global optimum of the fitness (objective) function, as shown in Figure 11 [116].

4.6. K-Means

The k-means method is a common clustering approach in data mining that is generally used to cluster huge datasets. The k-means approach was first presented by MacQueen in 1967; it was one of the simplest unsupervised learning algorithms and was used to tackle the problem of a well-known cluster. It is a partitioning clustering technique designed to categorize data, which iteratively divides k-date objects into distinct clusters, convergent on a local minimum. Therefore, the clusters created thereby are compact and independent. The algorithm comprises two distinct steps. The first phase randomly selects k-centers, where k is a predetermined value. The subsequent step is to transport each data object to the closest center. In general, the Euclidean distance is used to calculate the distance between each data object and the cluster centers. When all data objects are included in certain clusters, the initial stage is complete, and early grouping is performed, recalculating the average of the first clusters to form. This process is repeated until the criterion function reaches its minimum value [60]. The k-means method is one of the most straightforward unsupervised learning algorithms that solve the well-known clustering problem. A simple and straightforward method is used to classify a given dataset into a predetermined number of clusters (k-clusters are assumed) as depicted in Figure 12. The k-means technique is utilized in the absence of labeled data. The general method transforms heuristics into very precise prediction algorithms. Given a weak learning method that can regularly find classifiers (rules of thumb) at least marginally better than random, accuracy of, for instance, 55%, a boosting algorithm can provably generate a single classifier with extremely high accuracy of, for instance, 99% [117].

4.7. Decision Trees

Decision trees (DT) are ‘trees’ that classify instances by ordering them in accordance with their feature values. Each node in a decision tree represents a feature of a classifiable instance, and each branch represents a possible value for that node. Beginning with the root node, the instances are categorized and arranged, depending on their feature values. In data mining and machine learning, decision-tree learning employs a decision tree as a prediction model that maps observations about an item to make judgments about its target value. Such tree models are often known as classification trees or regression trees. Classifiers based on decision trees typically apply post-pruning procedures that assess the performance of decision trees after they have been pruned using a validation set. Any node may be removed and allocated to the class with the highest frequency among the examples of training that are sorted to it [117]. Figure 13 shows an example of a decision tree.

4.8. KNN

K-nearest neighbors (KNN) is a method proposed by Fukunaga et al. that bases its classification on distance, wherein the training phase is the storing of labeled samples and the classification of a sample as belonging to a class, based on samples that have been labeled as belonging to that class [118]. An effective strategy for both classification and regression is to add weights to the contributions of the neighbors so that closer neighbors contribute more to the average than more distant neighbors. A popular weighting technique, for instance, assigns each neighbor a weight of 1/d, where d is the distance between them. The neighbors are selected from a set of objects where the class (for k-NN classification) or property value (for k-NN regression) is known. This can be considered as the training set for the algorithm, but training is not explicitly necessary. An illustration of KNN is depicted in Figure 14.

4.9. Random Forest (RF)

The random forest (RF) classifier consists of a group of decision trees that employ a bagging method in the training phase, in which the input-labeled data is utilized to execute predictions through the various sets of rules produced in each decision tree by selecting the best forecast [119].

Understanding the advantages and disadvantages of each algorithm is more significant than simply learning the history of artificial intelligence, in order to select the most appropriate algorithm and use it correctly. To enable a clear perspective of each method’s advantages and limitations, we provide a summary of their pros and cons in

Table 7. The pros and cons of commonly used algorithms in the oil and gas field Adapted from Li et al. [120].

Algorithms	Pros	Cons
Artificial Neural Network (ANN), multi-layer perceptron (MLE), feed-forward (FF), radial basis function (RBF) convolutional (CN) functional, (FN), and probabilistic (PN)	Powerful parallel distributed processing ability, strong distributed storage and learning ability, strong robustness and fault tolerance to noisy nerves, full approximation of complex nonlinear relations, associative memory function, etc.	Numerous parameters are required, including network topology, initial weight and threshold settings; output that is difficult to read; lengthy training period, etc.
Particle swarm optimization (PSO)	Free from the problem information, solve problems with real numbers, strong universality, few parameter adjustments, straightforward theory, simple implementation, collaborative search, and rapid convergence.	Low precision, prone to divergence, and dependent on parameters; the theory is imperfect.
Fuzzy logic	Robustness is strong and straightforward to accomplish; mathematical precision is unnecessary.	Low precision and absence of systematic design.
Support vector machine (SVM)	Appropriate for small-sample machine learning issues; can improve generalization performance, tackle high dimensional problems, address nonlinear problems, and avoid neural network structure selection and the local minimum point problem.	Nonlinear issues are sensitive to missing data, and no universal solution exists.
Genetic algorithm (GA)	Possibility of parallelism and resilience, simplicity, use of a probability mechanism to perform iterations with a given degree of unpredictability, extensible, and simple to connect with other algorithms.	After determining the ideal solution, complex programming, issue decoding, and more training time are necessary to arrive at an exact solution.
Decision tree (DT)	Nonlinearity among parameters has no effect on DT performance; explainable and interpretable.	Complex; duplication may occur for identical subtrees of different routes.
K nearest neighbors (KNN)	Classes do not require linear divisibility; modest and strong; understandable and easy to implement technique; can be trained quickly; robust in case of associated noise; it is mainly well suited for multimodal classification.	Tends to disregard the attributes’ importance; slow and expensive, and sensitive to local data structure; memory restriction.
Random forest	High performance gives rise to variable measures; quick in implementation; less complex; robust with noisy data and can learn in increments.	Computationally expensive over fit issues; does not depend on variables; disregards the original geometry of data; low performance with attribute-related training data.
k-Means	Each data point may exist in several clusters; normalization of gene behavior depiction; a modifiable model that can accommodate varying dataset distributions; if the training data rise, the number of parameters does not alter.	Need to define cluster count c; membership cutoff value must be established; the initial assignment of centroids influences cluster formation; the convergence is slow in some cases.
Recurrent neural network (RNN)	Can record the information as timed activations; can manipulate consecutive data where lengths are arbitrary.	It is affected by the vanishing gradient type; it is incompatible with extra-deep modeling stacking.
Convolutional neural network (CNN)	Capable of detecting only the relevant features from a given dataset; similar parameters can be applied to distinct challenges; rapid training.	The tuning of parameters is challenging, and the network needs a substantial amount of data
Deep belief network (DBN)	A layer-by-layer learning technique enables it to learn the features; deals with unlabeled data and is immune to overfitting and underfitting problems; not affected by the fragmentation of training data, thus, it reduces the over-smoothing problem.	Some pre-training strategies degrade when the input data is constrained; runtime is extensive; the quality of output is low.

for the regularly used oil-field algorithms. ANN is the most popular and simplest method; however, it has stringent input parameter requirements. PSO is simple to implement and does not require problem-specific data, but its accuracy is quite poor. The application of fuzzy logic does not require a mathematical model, although its precision is low. SVM is suitable for learning from small samples and is sensitive to real-world data. The GA technique has excellent parallelism, can search quickly, and can be easily integrated with other algorithms, but it has a more complex programming procedure and a longer training period. Although the aforementioned algorithms have their limits, the key to problem-solving is algorithm selection and application.

4.10. Analysis

This section begins with reviewing articles about machine learning applications in pumps. In terms of scholarly and pertinent information, numerous review papers on ML in pumps have been published recently. The methodology of a systematic literature review was used in this study. The systematic review of the literature (SLR), a tried-and-true method, is widely employed to identify, evaluate, and interpret key research findings for a certain issue, area, or phenomenon of interest. SLR is secondary research that makes an effort to review studies with comparable objectives, critically evaluate their methodologies, and synthesize them using statistical analysis and, if practical, meta-analysis. In order to implement SLR, we used the methods outlined in a previous paper [121]. Figure 15 shows the statistics of the results.

The significance of this analysis is to demonstrate the development of published articles in ML from 2010 to 2022. Although many studies have been presented in the literature, the results demonstrate the critical significance that the data classification method has in achieving accurate detection. Therefore, research is still ongoing to identify a suitable algorithm for addressing any given problem in a pump. The available data type depends on the data’s amount and the chosen algorithm’s flexibility. In this regard, to address the issues mentioned above, Samanta et al. [115] examined the performance of the genetic algorithms NN and SVM on bearing failure data, to maximize the performance of these algorithms. The results demonstrated that the selection of characteristics substantially impacts the classifier’s performance.

On the other hand, Zouari et al. [57] utilized the NN and neuro-fuzzy approaches to identify faults in centrifugal pumps, using vibration data. For this purpose, an accelerometer was used to collect data concerning air injection, cavitation, and partial flow defects. Rajakarunakaran et al. [122] utilized the ANN to identify the fault in a centrifugal pump. Vibration signals were employed by Nasiri et al. [123], along with a feed-forward network, back-propagation algorithm, and binary adaptive resonance network model, to gauge the severity of cavitation. The authors hypothesized that the vibration sensor’s radial placement produces the best outcomes. By fusing SVM and ANN, Azadeh et al. [124] created an algorithm for categorizing two distinct centrifugal pump failures. Noisy data may be handled using the suggested method, which is also one of the flaws.

To obtain more valuable information from the raw data, Jia et al. [125] designed a complex architecture for an NN, to accelerate the raw data. Datasets for planetary gearboxes and rolling element bearings were utilized to identify the faults. The outcome of the method used is more accurate than the earlier approaches. Meanwhile, Zhao et al. [126] employed unsupervised learning to diagnose faults using vibration signals, while Azizi et al. [127] suggested an extended regression neural network-based approach for estimating the cavitation severity. The study considered no cavitation, minimal cavitation, and developed cavitation. Bordoloi and Tiwari [1] utilized the SVM technique to assess the suction obstruction and cavitation severity, using vibration data from the pump casing and bearing block. Hence, to obtain the most successful performance from the SVM classifier, kernel parameters were carefully selected in the work by Panda et al. [36] to forecast cavitation and flow obstruction in the pump, while Rapur and Tiwari [128] worked on the cover plate problem, discharge blockage, suction blockage, and impeller fault. Their algorithm identified the different rates of mechanical defects, flow disturbance, and the various combinations of these problems. Figure 16 summarizes the failure issues in studies and the common measures used, along with the typical features selected for analysis using the standard algorithms.

According to Figure 14, the typical failure issues addressed by several research works include leakage, defective bearings, cavitation, misalignment, and seal breakage. The algorithms employed in most studies to address these failures are ANN. SVM, principal component analysis (PCA), the neuro-fuzzy hybrid model, fuzzy logic, rough sets, and Fourier transform. Furthermore, the study’s standard parameters include vibration pressure analysis and acoustic emissions. Lastly, the common features considered in most studies are the vibration characteristics and the frequency time-domain characteristics. Understanding the properties of the data obtained and the potential cause-and-effect relationships between the qualities of the data is required before using certain algorithms in a study. In addition, it is important to comprehend the pump’s nature in terms of its working circumstances, servicing frequency, system dynamics, and any other pertinent characteristics [67]. Depending on the application and setup, researchers typically have varying algorithm choices. Concerning this, Lee et al. [67] elaborated on these standard algorithms, along with their applications, advantages, and disadvantages.

The most commonly used algorithms in PHM studies have been summarized, along with each algorithm and its applications, benefits, and limitations. The selection and application of the algorithms depend on their suitability and adaptability to specific characteristics of the data in a study. For further details, the reader may refer to Lee et al. [67]. However, the data usually utilized in various studies are categorized into two types. These are actual data, which are generally taken from real-world machinery. The simulated or generated data, on the other hand, is the second form of data, and it is intended to fulfill particular requirements, such as the validation of models in machine learning. Therefore, this study divides the datasets into two groups, based on the current research work.

Table 8. Data used in the various studies.

Author/Year	Algorithm Used	Findings	Parameters	Platform Used	Data Source
[38] Chen et al., 2021	k-nearest neighbor algorithm (KNN), based on the Mahalanobis distance, using the ReliefF weight analysis algorithm	Cavitation, impeller damage, and machine seal damage	Root mean square, peak factor, skewness coefficient, and kurtosis	Python	Synthetic data
[40] Giro et al., 2021		Gaussian mixture model (GMM)	Pressure		Real data
[49] Pier et al., 2020	Support vector machine (SVM) and multilayer perceptron (MLP)	Early fault prediction of a centrifugal pump	Temperature, pressure, and vibration probes	KNIME	Real and synthetic data
[119] Hu et al., 2020	Gaussian naive Bayes (GNB), support vector machines (SVM), random forest (RF), multilayer perceptron (MLP), and k-nearest neighbors (KNN)	Cavitation	Vibration	LINDA	Synthetic data
[129] Moleda 2020	Polynomial regression	Fault detection and diagnostics	Temperature, Pressure, Mass Flow, Current		Real data
[130] Zhou et al., 2020	Monte Carlo modified NN	Monitoring and failure prediction of lube oil pumps	Time stamp, pressure transmitter, oil pressure, oil quality, oil temperature, alternating current, and direct current		Real data
[131] Wang et al., 2019	Gaussian mixture model (GMM) clustering	Operation modes	Vibration		Synthetic data
[132] Yang et al., 2019	Feature-based transfer neural network (FTNN), convolutional neural network (CNN)	Identify the health states of bearings used in real-case machines BRMs	Raw vibration data		Real data
[131] Wang et al., 2019	Adaptive order particle filter (AOPF)	Estimate RUL	Vibration		Synthetic data
[45] Li et al., 2019	Sparsity and neighbor- hood preserving deep extreme learning machine (SNP-DELM), extreme learning machine-autoencoder (ELM-AE)	Bearing fault diagnosis	Number of balls, Pitch diameter, Ball diameter, Outer-race fault order, Inner-race fault order	MATLAB	Real data
[133] Luis. P et al., 2019	Temporal convolutional network (TCN), KNN, LSTM,	Internal pump leakage	Pressure, temp. volume flow, vibration, cooling efficiency, cooling power, the efficiency factor
[134] Janssens et al., 2018	Deep neural networks, convolutional neural networks, deep learning. (DNN-CNN, DL)	Machine fault detection and oil level prediction	Accelerometer, thermocouple, and thermal camera measurements		Real data
[135] Wu 2018 et al.	Self-organized mapping (SOM) neural network (NN); a long short-term memory (LSTM) network is utilized	Fault detection	Vibration		Synthetic data
[28] He et al., 2018	Ladder network, multi-layer perceptron, and denoising autoencoder (LN, MLP, DAE)	Estimate RUL	Flow, pressure, stress
[37] Dutta et al., 2018	SVM	Cavitation detection	Speed, pressure	Lab view	Real data measurement
[132] Yang et al., 2018	Feature-based transfer neural network (FTNN), convolutional neural network (CNN)	Health status of bearings	Normal (N), inner race fault (IF), roller fault (RF), and outer race fault (OF).		Real data
[61] Zhang et al., 2018	Convolutional neural networks (CNNs)	Fault diagnosis in rolling bearing	Healthy condition (H), outer race fault (OF), inner race fault (IF), and ball fault (BF).		Synthetic data
[136] Samir et al., 2017	Intrinsic time-scale decomposition (ITD)	Motor current signals from different pump fault cases.	RMS values of the first proper rotation component (PRC) with the raw signal RMS values
[137] Jason and David, 2017	Deep belief network feed-forward neural network (DBN- FNN)	Remaining useful life (RUL) of the rotating components			Real data
[61] Zhao et al., 2016	Stacked denoising autoencoder (SDA) and softmax regression	Identify possible failure modes			Real data
[138] Han et al., 2016	Decision-tree C4.5 algorithm	Internal thermo-electric potential faults	Low-temperature superheat, high-temperature superheat, low energy discharge, high energy discharge, arc discharge with overheating	KNIME, MATLAB	Synthetic data
[139] Liu 2015	Adaptive hidden semi-Markov model (AHMM)	Estimate RUL	Oil flow		Real data

depicts the data used in the various studies.

In Table 8, numerous studies on pumps are listed, and the table also describes the various parameters used in the studies. The findings show the multiple platforms used for the analyses in each work. The table further elaborated on what type of data was acquired by the authors (synthetic or real data) for the study, and some types were not mentioned in the studies. The table also enumerates the parameters and algorithms employed for each identified problem chosen to be solved by the authors.

In this study, an observation has been made on the frequency of usage of the algorithms used in studies. Figure 17 shows that ANN is the most commonly used algorithm in studies, followed by BN and SVM. The studied literature emphasizes the importance of selecting proper training algorithms to speed up the training process and achieve excellent classification accuracy. It has been observed that there is no specific method or algorithm chosen for a particular problem. The choice depends on the type of data and the suitability of the algorithm or methodology to solve the given faults.

A comprehensive literature analysis of fault prognosis and diagnosis on pumps is provided herein, and Figure 18 shows that previous studies concentrate more on fault diagnostics than prognostics. Currently, fault diagnostic systems are predominantly constructed as a collection of discrete components, such as data collecting, feature extraction, dimensionality reduction, and fault recognition, with little regard for the diagnostic system as a whole.

In this study, another significant observation is that among the numerous research studies on pump fault diagnosis and prognosis, the most frequent fault discovered or addressed is the bearing fault, as shown in Figure 19, with a percentage of 46%, followed by cavitation. Seal damage is the third-highest fault in studies. Leakage and blockage are the least common faults considered in studies. Another important finding in this review is that most of the accomplished projects depend on auditory and vibration data. Additionally, existing issues, such as the unpredictability of the failure form, and the lack of sample data in the data system, must all be addressed.

5. Conclusions

In this study, the research explored the most recent advancements in the field of AI and machine learning and their applications in the oil and gas industries. This study indicated that the frequency of application of artificial intelligence (AI) for the accurate diagnosis and forecasting of pump defects has risen dramatically over the past decade, particularly in the oil and gas industries. Artificial intelligence has gained popularity and is being widely used in mechanical manufacturing and automation, as a result of the quick pace of life and the arrival of Industry 4.0. The manufacturing industry benefits greatly from its powerful data processing, computational, and storage capabilities. It is beneficial for encouraging work efficiency and quality control, frees up more worthwhile tasks for humans to complete, creates much safer workplaces, offers problem diagnosis and predictive maintenance, and creates more intelligent supply chains. The advancement of mechanical manufacturing assists artificial intelligence and vice versa. The manufacturing sector is undergoing the fourth Industrial Revolution, and the introduction of 5G and the use of artificial intelligence will hasten this change. This comprehensive study will be able to pave the way for future research works, by highlighting the common faults identified, the algorithms employed in previous studies, and the parameters that are taken into consideration. Therefore, this simplifies identifying other vital areas for further investigation in the respective studies. Hence, multiple challenges that need to be addressed in future research were identified, as follows:

• It has been noticed that no specific method or algorithm exists for a given problem. The process relies on the type of data and the algorithm’s or method’s aptitude for resolving the provided errors. Among the various research studies on pump fault diagnosis and prognosis, the most frequently treated problem is that of a bearing fault, with a percentage of 46%, followed by cavitation. The published studies rank seal damage as the third most prevalent flaw. Leakage and obstruction are the least studied defects in research.
• Artificial neural networks, support vector machines, and hybrid models are the most frequently used machine learning models for evaluating the health of damaged oil and gas pumps.
• Due to their ability to incorporate the strengths of multiple machine learning models rather than just one, hybrid models were found to be the most successful of all the models tested. Depending on their location and medium, a wide range of factors can affect the health of oil and gas pumps; however, the most frequently used factors include pressure, temperature, flow, angular velocity, noise, vibration signals, and so on. However, neither reliability, nor precision, nor processing time was taken into account.
• The research question found that four different types of datasets were primarily used to create machine learning models, particularly datasets based on expert judgment, field data, experimental data, and simulation-based data. Simulated data and experimental datasets have frequently been produced using field data. The data side is currently under-researched, compared to the algorithmic side.
• The data types employed in the machine learning methods for pump fault diagnosis are mainly limited to vibration and flow, which may not be adequate to characterize the condition of pumps and their attributes. This can lead to false alarms and a lack of confidence in the predictions under the dynamic operational environments of pumps.

6. Future Directions

This paper reports a literature review of machine learning, addressing current research gaps in the PdM of pumps. A review was carried out to evaluate the potential of the ML algorithms in enhancing the current PdM of pumps. The ultimate objective is to provide a comprehensive system, with a number of interconnected components as the research interest grows. Also, it is necessary to speculatively ask if the advantages of employing machine learning for health monitoring outweigh the efforts necessary to realize it in existing applications. If so, is there a methodical way to develop and apply the solution? This review has shown that machine learning methods have the potential to provide appropriate outcomes, which makes them more alluring. More effective health monitoring algorithms must be implemented as the oil and gas sectors grow more automated and technologically advanced, in order to maintain them. However, in order to forecast system behavior and implement remedial measures, these advancements rely on high-quality data from numerous sources, frequently spread across different geographic areas. The objective is to understand the constraints that machine learning brings to the application and then make an effort to make the issue simpler.

The ML technique possesses high potential in terms of big data analysis. Its efficacy relies on abundant sensing data, which may not be accessible for aging pumps. Attention should be given to the availability and quality of condition monitoring data when employing ML techniques. Future research should include factors such as reliability, precision, and processing time to ensure that the condition monitoring system conforms to industry requirements. Future diagnostic and prognosis processes should leverage and tightly integrate not only data-driven AI technologies but also the evaluation of failure mechanisms and past knowledge to increase the accuracy and reliability of the outcomes.