Optimization of Oil Production Using Sucker Rod Pumps via Predictive Elimination of Paraffin Issues

Abstract

This paper explores the application of predictive maintenance (PdM) to address paraffin deposition in sucker rod pump systems used for oil production. System maintenance has become critical for enhancing efficiency and reducing costs, while PdM, supported by advanced analytics and sensors, enables downtime prediction and maintenance optimization. Paraffin deposition is a significant problem in the oil industry, as it diminishes production capacity and increases expenses. This paper presents the use of the SCADA system, which enables the collection and analysis of data in real time. Furthermore, it proposes diagnostic methods for early detection of paraffin deposition using predictive maintenance, offering timely warnings to prevent production delays. While the proposed framework relies on interpretable statistical and physics-informed predictive models, the results indicate that further improvements could be achieved by integrating advanced artificial intelligence techniques to enhance adaptability, automation, and decision support in predictive maintenance systems.

1. Introduction

In recent years, system maintenance has become critical for enhancing production efficiency and continuity. There are several types of maintenance, including reactive, planned, proactive, and predictive [1]. Figure 1 illustrates these different types of maintenance systems.

Reactive maintenance is performed only after a malfunction occurs, with corrective actions taken once the issue becomes evident. On the other hand, planned maintenance includes regular inspections and tasks scheduled in advance, in order to extend the life of the system and reduce repair costs, regardless of whether the system has shown any signs of malfunction. Predictive maintenance (PdM) uses cutting-edge analytics on data collected through multiple sensors, to predict the time when a failure could occur. This approach organizes maintenance tasks to optimize intervals, reduce downtime, and improve system reliability [2].

Recently, predictive maintenance has seen significant progress thanks to the development of low-cost sensors and new real-time monitoring systems, which allows the collection of large amounts of data. These advances, combined with expert algorithms and human experience, contribute to the development of PdM. Efforts are currently being made to create new multivariate statistical models and algorithms to improve prediction accuracy and reduce operating costs. The next step towards autonomy in robotic systems is becoming possible thanks to sophisticated algorithms and artificial intelligence (AI) [3].

In addition, potential AI-based PdM brings cost reduction, as well as increased efficiency and security. Researchers are therefore focusing on AI models and techniques to enhance the autonomy and adaptability of robotic systems operating in complex industrial environments [4]. The importance of implementing SCADA systems is broad across industrial applications, and one of the fundamental aspects of their significance lies in the domain of monitoring, system condition tracking, and related functions [5,6].

An algorithmic framework for early detection and prediction of critical paraffin levels has been proposed through the analysis of load variations in sucker rod strings and electrical parameters of the drive motor. This demonstrates that parameters commonly employed for post-failure diagnostics can be successfully applied in predictive functions, thereby significantly increasing system reliability. The contribution is further reflected in empirical validation carried out on five production wells, enabling a quantitative evaluation of the efficiency of predictive maintenance under real operating conditions. The obtained results provide the basis for formulating general guidelines that may reduce downtime and operational costs in oil production. These elements not only confirm the practical effectiveness of predictive maintenance in addressing paraffin-related challenges but also establish a conceptual framework for further development of intelligent systems in petroleum engineering, particularly in the context of integrating artificial intelligence and advanced data analytics.

Existing studies on paraffin deposition in sucker rod pumping systems commonly rely either on physics based diagnostics limited to post event interpretation or on data driven approaches that require complex machine learning models and extensive labeled datasets. This separation between mechanical modeling and predictive analytics limits the applicability of many methods in long term field operation, particularly in wells equipped with standard SCADA infrastructure. As a result, there is a lack of predictive maintenance frameworks that are both interpretable and suitable for continuous industrial use.

This study addresses that gap through an integrated predictive maintenance framework that links mechanical load behavior of the sucker rod string with SCADA-based electrical measurements and probabilistic reliability modeling. The approach is grounded in measurable physical quantities, including polished rod load variation and motor electrical response, which are processed into normalized indicators suitable for statistical analysis. Instead of relying on black box learning models, the methodology applies interpretable regression and survival analysis techniques that allow direct association between paraffin accumulation and system behavior.

In addition to quasi-static load effects, the operational behavior of sucker rod pumping systems is influenced by the dynamic response of the rod string. Longitudinal and impact-related vibrations affect force transmission along the rod string and contribute to variations in polished-rod loads. Analytical studies of rod-string suspension systems have shown that frictional damping and shock-absorbing elements modify vibration attenuation and load oscillations.

Paraffin accumulation increases distributed friction along the tubing, which alters damping characteristics and the dynamic equilibrium of the rod string. These effects provide a mechanical basis for interpreting increased load span and load variability as indicators of paraffin-related degradation. Recent analytical investigations of friction dampers and plate-type shock absorbers support this interpretation and are therefore relevant for the present study. Recent analytical investigations of friction dampers and plate-type shock absorbers provide a mechanical basis for interpreting changes in load span and load variability as indicators of altered dynamic behavior in the rod string. These mechanisms are relevant for paraffin-related degradation, as paraffin accumulation increases distributed friction and modifies the dynamic equilibrium of the pumping system. Analytical models of slotted-shell friction dampers and plate-type shock absorbers have demonstrated how frictional interfaces and elastic elements modify dynamic force transmission and vibration attenuation in sucker rod strings. These studies provide a mechanistic explanation for changes in load variability and damping behavior under increased frictional conditions and support the physical interpretation adopted in this work [7,8].

A physics informed Early Warning Index is introduced as a composite indicator derived from mechanical load span dynamics and normalized electrical parameters. This index provides a continuous measure of paraffin related degradation that can be evaluated in real time. The framework further integrates logistic regression for short term downtime prediction and time dependent Cox models for reliability assessment, enabling probabilistic estimation of failure risk and mean time between failures. Long term validation is performed using five producing wells monitored over a five year period, which allows quantitative evaluation of reliability improvement under real operating conditions. The results show that parameters traditionally used for diagnostic interpretation can be systematically transformed into predictive features without the use of advanced artificial intelligence models, while maintaining strong predictive performance. The proposed methodology is applicable to low and medium production wells and can be implemented within existing fourth generation SCADA systems. This system integrates sensors that collect real-time data from physical processes and simultaneously serve as a platform for information acquisition. Actuators within the SCADA architecture are responsible for process control, executing commands based on processed data. Supervisory modules enable operators to monitor process status through visualization interfaces, while hierarchical communication networks ensure seamless operation by integrating data from multiple sources, which are subsequently analyzed on centralized servers [9]. Figure 2 illustrates the key components of the SCADA system.

The monitoring of working parameters in oil production is enabled by the use of the industrial SCADA system. It collects, stores, and processes data and then visualizes it.

In oil production, SCADA systems collect mechanical and electrical parameters from field sensors via PLCs and transmit them to supervisory servers for visualization and alarm handling [9,10].

This paper uses SCADA of the fourth generation, which relies on the Internet of Things (IoT). This system uses software and electronics embedded in sensors to collect and process data, both on-site and remotely via an Internet connection. The core of the proposed IoT SCADA server is Node-RED 4.1., located on a local computer [11,12,13,14].

In addition to the aforementioned models and systems, optimization could also be conducted and examined through the application of contemporary metaheuristic algorithms and the Taguchi method, which are widely used in industrial systems, including the oil industry [15,16].

This study does not employ deep learning or black-box artificial intelligence models. Instead, it integrates mechanically interpretable indicators with conventional probabilistic methods to achieve predictive maintenance under field constraints. The focus is placed on transparency, physical interpretability, and compatibility with existing SCADA infrastructure. Artificial intelligence-based automation is considered as a future extension.

2. Materials and Methods

The methodological design was defined to ensure direct physical interpretation of all monitored variables while maintaining statistical consistency of the predictive models. Only parameters that reflect mechanical and electrical response of the pumping system to paraffin accumulation and that are available through standard SCADA measurements were considered. This approach allows the predictive framework to remain compatible with field constraints and long term operation.

The research was designed as a field-based study combining continuous monitoring of well operation with analytical evaluation of paraffin deposition indicators. Five producing wells (K-1 to K-5) equipped with sucker rod pumping units were selected as representative cases. These wells had a documented history of paraffin-related production interruptions, which made them suitable for testing predictive maintenance strategies.

• Data Acquisition. Operational data were collected through a fourth-generation SCADA system Node RED 4.1 integrated with IoT-based sensors. The monitored parameters included: polished rod load (kN), motor current and power consumption (A, kW), electric frequency (1/min), and wellhead pressure and temperature (bar, °C). Data were sampled at intervals of 10 s and stored in a local server with real-time visualization and alarm functionalities.
• Diagnostic Indicators. Paraffinization trends were assessed through dynamometer card interpretation and analysis of electrical load variations. Specific attention was given to the increase in maximum rod load, reduction in minimum load, and growth of peak motor currents as early-warning features of paraffin deposition. These variables were selected based on previous literature and validated against historical field records. The attribution of operational degradation to paraffin deposition is not based on a single measured variable. Instead, it relies on the concurrent evolution of multiple mechanically and electrically independent indicators derived from standard SCADA measurements. These indicators include the increase in maximum polished-rod load, the decrease in minimum polished-rod load, the widening of the polished-rod load span, the increase in peak motor current and power, and the reduction in electric frequency.
• This specific combination of trends is consistent with increased frictional resistance and additional effective load along the rod string and tubing, which are established physical consequences of paraffin accumulation. Indicators were selected to reflect different subsystems of the pumping unit, which reduces the likelihood that the observed degradation pattern is caused by an unrelated single-component malfunction.
• Predictive Framework. A rule-based predictive algorithm was implemented within the SCADA environment using Node-RED. The algorithm generated warnings when monitored parameters exceeded predefined threshold values derived from baseline operating conditions. In addition, trend analysis was applied to capture progressive changes leading to potential rod string sticking.
• Validation Procedure. The methodology was validated through longitudinal observation from 2020 to 2025. Downtime due to paraffin was recorded for each well and compared across two operational phases: (i) before predictive maintenance implementation and (ii) after the deployment of intelligent monitoring. The effectiveness of the proposed framework was quantified by calculating reductions in downtime (days), energy savings, and prevention of workover operations.

3. The Impact of Paraffin in the Production of Oil Using Sucker Rod Pumps

Paraffin accumulation in sucker rod pumping systems manifests primarily through increased mechanical resistance and additional effective rod string mass, which leads to rising maximum polished-rod loads and declining minimum loads. These effects are directly observable in dynamometer cards and motor electrical parameters. In this study, these physically interpretable changes are used as predictive indicators of paraffin-related degradation rather than as post-event diagnostic features [17,18,19,20,21,22]. Figure 3 presents the dynamograms under 0% and 100% paraffinization conditions.

The paraffin issue was addressed and investigated by the following authors [23,24,25,26,27].

Paraffin Mitigation

By applying an intelligent control system, it becomes possible to obtain early information on the need for paraffin removal from the tubing before the sucker rod string becomes stuck due to paraffin deposition. The system continuously monitors polished-rod load, increases in motor peak current, and decreases in electric frequency, and through software modules notifies the monitoring specialist once paraffinization reaches a predefined threshold. Since paraffin deposition develops gradually, tracking these parameters allows the prediction of the time when deposition is likely to occur. At that point, the monitoring specialist has sufficient lead time to issue a work order for preventive paraffin removal from the tubing.

Furthermore, by analyzing the trend of rising peak currents and motor peak power, combined with the decline in frequency, the specialist can anticipate the need for deparaffinization even before the alarm is triggered. Timely deparaffinization contributes to energy savings, protects equipment from damage, and optimizes oil production. Importantly, the intelligent control system also provides an automatic safeguard: if the monitoring specialist does not respond to the notification of imminent paraffinization, the system is capable of automatically shutting down the pumping unit before sucker rod sticking.

Figure 4 illustrates the trend of increasing polished-rod load caused by progressive paraffin deposition. When the rods move upward, they experience a maximum load, which increases further when they are embedded in paraffin, as indicated by the blue curve. Conversely, the minimum load occurs when the rods move downward, as shown by the orange curve.

Figure 5 illustrates the increase in peak currents and peak powers due to paraffin deposition, as well as the restoration of these parameters to their nominal values following the performed deparaffinization. The purple curve represents the motor power expressed in kilowatts (kW), the white curve denotes the electric current intensity, while the upper curve corresponds to the motor frequency.

In this study, the focus is placed on predictive maintenance, while the deparaffinization processes are not described in detail. It is important to note that deparaffinization is carried out by applying indirect circulation with heated oil at 90 °C, as well as heated water at 80 °C, using a mobile steam unit that heats with steam up to 180 °C. This procedure melts the paraffin deposits in the area of the pumping unit and the pipeline, enabling fluid flow during oil production. All of these operations require the use of specific machinery, fluid heating sources, and generate significant costs. When such operations are conducted only when genuinely necessary for the sustainability of oil production, they are economically justified. Predictive maintenance optimizes the costs of these operations by ensuring that they are performed precisely when needed. Prior to the implementation of intelligent maintenance systems, deparaffinization operations were carried out according to a predefined schedule, without verifying their actual necessity. This approach often led to unnecessary expenses, which could have been redirected to maintenance tasks critical for production. When a well is considered as a system, the need for deparaffinization is expected to be cyclical and to occur at regular time intervals. As with all operations within such systems, the need for deparaffinization depends on the quality of prior operations, which directly affects the timing of subsequent interventions. For this reason, scheduled maintenance often fails to yield satisfactory results, whereas predictive maintenance is fully tailored to the actual requirements of production.

4. Mathematical Modeling and Predictive Framework

To formalize the relationship between paraffin accumulation and the operational behavior of sucker rod pumping systems, a dedicated mathematical framework was developed. This framework integrates mechanical balance equations, SCADA-based electrical parameters, and statistical learning techniques to enable early detection and prediction of paraffin-induced failures.

Model Variables and Pre-Processing

Operational data were acquired from the SCADA IoT system with a sampling interval of 10 s. Before analysis and predictive modeling, the raw data streams were processed using a defined preprocessing procedure to ensure robustness, consistency, and reproducibility. Missing values occurred infrequently and were mainly caused by short communication interruptions. Data gaps shorter than 2 min were handled using linear interpolation. Longer gaps were excluded from model fitting to avoid the introduction of artificial trends. High-frequency noise present in the raw SCADA signals was reduced using rolling median and moving average filters, with window lengths between 5 and 15 min depending on the monitored variable. Outliers were detected using interquartile range criteria and z-score analysis relative to baseline operating conditions. Measurements exceeding physically plausible limits were clipped or removed to maintain physical interpretability of the signals. To reduce temporal autocorrelation and computational load, the preprocessed signals were aggregated into hourly statistical features, including mean values, maximum and minimum values, and linear trends. All electrical and mechanical variables were then normalized using paraffin-free baseline statistics obtained during stable operating periods. This normalization allowed direct comparison between wells and ensured that the derived predictive indicators were dimensionless and comparable over time. The processed features were used as inputs for the Early Warning Index calculation and for the logistic regression and survival analysis models.

The reference “paraffin-free” operating condition is not defined as a fixed historical baseline. Instead, baseline statistics are identified adaptively from locally stable operating intervals within rolling time windows. These intervals are selected based on consistent dynamometer card patterns, stable electrical parameters, and the absence of paraffin-related interventions.

Normalization is therefore performed relative to a time-local reference state rather than a global constant. This approach allows gradual changes in fluid properties, temperature, and water cut to be implicitly incorporated into the reference values. Seasonal and long-term operational drift are further mitigated by estimating trends within moving windows, ensuring that normalized indicators represent relative degradation with respect to the current normal operating regime.

Operational data streams were acquired from the fourth-generation SCADA–IoT platform at a sampling interval of 10 s. The monitored parameters included peak motor current $I_{peak}\left(t\right)$ [A], peak motor power $P_{peak}\left(t\right)$ [kW], electric frequency $f\left(t\right)$ [1/min], and polished rod loads $P_{rod}^{max}\left(t\right)$ and $P_{rod}^{min}$ [kN] derived from dynamometer cards. From these variables, the load span was defined as:

$${∆P}_{rod}\left(t\right)=P_{rod}^{max}\left(t\right)-P_{rod}^{min},$$

(1)

and its temporal drift ${∆P}_{rod}\left(t\right)$ was estimated using robust linear regression over trailing 24–72 h windows.

The effect of paraffin deposition was anchored in the established balance of polished rod loads, based on Physics-Based Load Balance:

$$P_{max}=W_{ru}+W_l+I_u+P_{hu}+f_u+P_u-P_i,$$

(2)

$$P_{min}=W_{rd}+I_d-P_{hd}-f_d-P_u,$$

(3)

where frictional components f_u and f_d grow as paraffin increases. This leads to the observed pattern of rising $P_{max}$ declining $P_{min}$, and widening ${∆P}_{rod}$. The drift of this load span provides a mechanical signature of paraffin accumulation. The use of rolling windows for trend estimation ensures that slow variations in operating conditions do not bias the degradation indicators. Only deviations that evolve consistently relative to the local reference state contribute to the Early-Warning Index.

Let $t$ denote continuous time (s). We define the polished-rod load span as:

$$S\left(t\right)=L_{max}\left(t\right)-L_{min}\left(t\right)$$

(4)

where $L_{max}$ and $L_{min}$ are the instantaneous maximum and minimum polished-rod loads (kN). The influence of friction-induced damping on load span evolution is consistent with analytical results reported for vibration-mitigation devices used in sucker rod strings, where increased friction leads to systematic changes in dynamic load response.

The short-term trend (drift) of S over a rolling window of width www is estimated by robust linear regression:

$${S∆}_w\left(t\right)=slope(S\left(\tau \right),\tau \in \left[t-\omega ,t\right])$$

(5)

SCADA electrical variables are normalized to baseline (paraffin-free) statistics: for motor current $I\left(t\right):$

$$z_I\left(t\right)={\displaystyle \frac{I\left(t\right)-\overline{I_{base}}}{{\sigma }_{I,base}}},$$

(6)

and analogously for motor power $P\left(t\right)$ and electric frequency $f\left(t\right)$ (note that frequency is inverted because it decreases during parafinization).

Alternative degradation mechanisms were explicitly considered when defining the predictive indicators. Conditions such as insufficient pump fillage, mechanical wear, or valve malfunction typically produce different signal patterns, including reduced average load levels, irregular dynamometer shapes, or abrupt, non-monotonic parameter changes. These patterns differ from the gradual and correlated drift observed across all monitored indicators in the analyzed cases. The causal attribution to paraffin deposition was further supported by operational interventions. After confirmed deparaffinization, the mechanical and electrical parameters consistently returned to their locally defined reference ranges, which validates the physical interpretation of the monitored trends and supports the use of the selected indicators for paraffin-specific degradation monitoring.

An Early-Warning Index (EWI) is constructed as a normalized linear combination of standardized indicators:

$$EWI=\sigma ({\sum }_{i=1}^4{w}_i{z}_i\left(t\right)),$$

(7)

where $\sigma$ is the logistic function, $z_1=z_I, z_2=z_P, z_3=z_f, z_4$=${ z}_{∆S}$ and ${\sum }_i{w}_i=1$.

The threshold values used for the Early Warning Index were determined through a data-driven calibration procedure rather than being selected arbitrarily. Receiver operating characteristic analysis was performed on validation data to evaluate the trade-off between detection sensitivity and false alarm rates for paraffin-related events. Advisory and critical alert levels were chosen at operating points on the ROC curve that provide stable early detection while limiting the frequency of false positives. The resulting thresholds correspond approximately to upper percentile ranges of baseline-normalized degradation states and can be recalibrated for other wells or operating conditions using the same procedure.

A sensitivity analysis was performed to assess the robustness of the proposed framework with respect to rolling window length and alarm threshold selection. Drift estimation windows ranging from 24 h to 72 h and signal averaging windows between 5 min and 30 min were evaluated. Across these configurations, the Early Warning Index trajectories and logistic regression performance remained stable, with ROC AUC variations within ±0.03 and no statistically significant change in estimated downtime reduction.

Alarm thresholds were calibrated using receiver operating characteristic analysis on holdout data (2024–2025). Advisory and critical levels were selected to balance detection sensitivity and false alarm rates. The adopted operating point achieved approximately 0.91 sensitivity with a false positive rate below 0.12 for paraffin-related events within a 72 h horizon. This analysis explicitly quantifies the trade-off between false alarms and missed events and confirms that the reported results are robust with respect to reasonable variations in time windows and threshold values.

The weights $w_i$ are obtained from a logistic regression trained to predict the occurrence of a paraffin-related intervention within the next 72 h.

The probability of downtime within a 72 h horizon is modeled by logistic regression:

$$P_T\left(Y_{72}\left(t\right)=1\right)={\displaystyle \frac{1}{1+\mathrm{e}\mathrm{x}\mathrm{p}(-\left({\beta }_0+{\beta }_{1}EWI\left(t\right)+{\beta }_2∆S_w\left(t\right)+{\beta }_{3}S\left(t\right)+\dots \right))}}$$

(8)

Coefficients $\beta$ are estimated by maximum likelihood on the 2020–2023 training set and validated on 2024–2025 holdout data. Model performance is reported with ROC AUC, sensitivity, specificity, precision, and calibration plots.

For reliability analysis we fit a time-dependent Cox proportional hazards model:

$$h\left(t\right)=h_{0}exp({\lambda }_{1}EWI\left(t\right)+{\lambda }_2∆P_{rod}(t\left)\right)$$

(9)

System reliability was modeled through a Cox-type hazard function:

$$h\left(t|X\left(t\right)\right)=h_0\left(t\right)exp\left\{{\gamma }^{T}X\left(t\right)\right\},$$

(10)

with $X\left(t\right)$ containing time-varying covariates ($EWI\left(t\right)$, $\Delta Sw\left(t\right),$…). Proportionality assumptions are tested with Schoenfeld residuals. MTBF is derived from the Kaplan–Meier survival function $S\left(t\right)$ via numerical integration.

Finally, operational rules implemented in Node-RED are expressed formally:

Advisory: if $EWI\left(t\right)\ge 0.6$ for at least 6 h or $\mathsf{\Delta }Sw\left(t\right)>{\delta }_1$ then issue advisory alert.

Critical: if $EWI\left(t\right)\ge 0.9$ for at least 2 h and $Lmin\left(t\right)<$ 10th percentile of baseline $Lmin$ then issue critical alert and trigger safe-shutdown protocol.

All thresholds $\left(0.6,0.9, {\delta }_1\right)$ percentile cut-offs) are determined from the ROC/operator trade-off on holdout data and should be reported with sensitivity/specificity trade-offs.

System reliability was quantified using the mean time between failures, defined as the area under the survival curve:

$$MTBF={\int }_0^{\mathrm{\infty }}S\left(t\right)dt,$$

(11)

where $S\left(t\right)$ is the survival function representing the probability of continuous operation without a paraffin-induced intervention up to time $t$. survival function derived either from the Kaplan–Meier estimator or from the Cox proportional hazards model. In practice, MTBF was obtained by numerical integration of the estimated survival function over the observed time horizon.

The overall workflow of the predictive framework is illustrated in Figure 6 real-time SCADA data streams are processed into baseline-normalized features, aggregated into the Early-Warning Index, and evaluated within the logistic and hazard models. Alerts are generated when thresholds are exceeded, and corrective actions (advisory or automatic shutdown) are initiated through the HMI layer.

5. Results

The implementation of predictive maintenance on five producing wells was evaluated over a five-year monitoring period (2020–2025). The analysis focused on two primary aspects: (i) production downtime caused by paraffin deposition, and (ii) variations in electrical and mechanical operating parameters under different paraffinization conditions. By comparing operational performance before and after the deployment of predictive monitoring, the study aimed to quantify improvements in reliability, efficiency, and energy consumption. The results are presented in terms of production downtime reduction, statistical distribution of operational interruptions, and characteristic trends of electrical load and electric frequency. Particular emphasis was placed on the correlation between paraffin accumulation and measurable deviations in load, motor current, and power consumption, which served as the basis for the predictive framework. Tables and figures that follow summarize the key findings and provide a quantitative validation of the proposed approach. The study was conducted on five oil wells (K-1, K-2, K-3, K-4, and K-5), and similar results were obtained for each. Through predictive maintenance, energy was conserved, equipment failure was prevented, and the number of production downtimes was reduced to zero.

Figure 7 illustrates the days of well downtime due to paraffin deposition.

Based on the analysis conducted from 2020 to 2025 and the number of downtimes measured in days, it was concluded that the downtime of the monitored wells was minimized, almost reduced to zero. Following downtimes caused by paraffin deposition, it was often necessary to perform overhaul operations due to the inability to operate the plunger rod assembly (sticking). This resulted in a significant number of downtime days and unachieved production gains.

During 2023, the implementation of an intelligent monitoring system was carried out. In that year, the paraffin deposition issue was still addressed reactively due to the specialists’ lack of training to act predictively. From 2024 onwards, the specialists’ experience became substantial, and predictive actions prevented well downtimes. Human error occurred a few times in 2025. To eliminate human error, the next step in predictive maintenance is the introduction of artificial intelligence, which, in the human–machine interaction (HMI) context, will reduce these downtimes to zero.

Table 1. Production downtime due to paraffin deposition before and after predictive maintenance implementation (2020–2025).

Well	Avg. Downtime Before (Days/Year)	Avg. Downtime After (Days/Year)	Reduction (%)	Std. Dev. Before	Std. Dev. After
K-1	14.2	0.8	94.4	3.1	0.4
K-2	12.5	1.0	92	2.7	0.5
K-3	11.8	0.6	94.9	2.3	0.2
K-4	13.0	1.1	91.5	3.0	0.6
K-5	12.2	0.9	92.6	2.5	0.3
Average	12.7	0.9	93.1	2.72	0.4

presents the downtime due to paraffin deposition before and after the implementation of predictive maintenance (2020–2025).

Table 1 shows the average annual downtime due to paraffin deposition before and after the introduction of predictive maintenance. All five wells exhibited a substantial reduction in downtime, with improvements ranging from 91.5% to 94.9%. The standard deviation values also decreased after implementation, indicating greater operational stability and consistency. Well-specific uncertainty was quantified using bootstrap resampling (1000 iterations) for downtime reduction and Kaplan–Meier survival analysis for MTBF. For each well, 95% confidence intervals were computed separately. Downtime reduction ranged from 91.5% to 94.9% across wells, with all confidence intervals excluding zero, indicating consistent effects. Reliability analysis was conducted on a per-well basis using Kaplan–Meier estimators with right censoring applied to ongoing operational periods without observed paraffin-related failures. The post-implementation monitoring phase is shorter than the pre-PdM period; therefore, MTBF estimates are reported together with confidence bounds and censoring indicators. To account for incomplete observations and time-dependent degradation, Cox proportional hazards models with time-varying covariates were employed. Despite the limited post-implementation horizon, hazard ratios consistently favored the PdM phase for all wells, and survival curves show a uniform shift toward longer uninterrupted operation. Continued monitoring is expected to further reduce estimation uncertainty.

Table 2. Electrical parameters before and after paraffin removal events.

Parameter	Baseline (No Paraffin)	At Critical Paraffinization	After Deparaffinization	Change (%)
Peak motor current (A)	12.0	16.5	11.8	+37.5/−28.5
Peak motor power (kW)	4.5	6.1	4.4	+35.5/−27.9
Electric frequency (1/min)	50	37	48	−30.0/+42.8

shows electrical parameters before and after paraffin removal events.

Table 2 summarizes the electrical parameters measured at baseline, during critical paraffinization, and after deparaffinization. The results confirm a direct correlation between paraffin buildup and increased motor load. Peak current and power values increased by more than 35% during severe paraffinization but returned to near-baseline levels after cleaning, demonstrating the reliability of these parameters as diagnostic and predictive indicators.

Statistical evaluation confirmed that the observed improvements were highly significant. A paired t-test comparing downtime values before and after predictive maintenance yielded p < 0.01 for all five wells, indicating that the reduction in downtime was not due to random variation. Analysis of variance (ANOVA) across all wells further demonstrated that the effect was consistent, with an F-value exceeding the critical threshold at the 95% confidence level. Trend analysis of polished rod load and motor current during paraffin accumulation revealed strong linear correlations (R² > 0.9), confirming their suitability as predictive features. Reliability assessment based on mean time between failures (MTBF) showed an increase from an average of 25 days in the pre-PdM phase to more than 300 days after implementation. These findings highlight the robustness of the predictive framework and validate its effectiveness under real operating conditions.

To ensure a comprehensive validation of the proposed predictive maintenance framework, the collected field data were subjected to rigorous statistical and mathematical analysis. The evaluation was structured into three complementary stages, each aimed at quantifying the effectiveness of the system under real operating conditions. In the first stage, descriptive and inferential statistics were applied to operational indicators, including production downtime, rod load, motor current, and electric frequency. Mean values, standard deviations, and coefficients of variation were calculated to assess data dispersion and stability. To test the significance of differences observed before and after the implementation of predictive maintenance, paired t-tests were employed, while inter-well consistency was further examined through one-way ANOVA at a 95% confidence level. The second stage focused on correlation and regression modeling. Linear regression models were developed to quantify the relationship between paraffin accumulation indicators—such as peak motor current and polished rod load—and observed downtime events. Determination coefficients (R²) were calculated to evaluate predictive accuracy, while residual analysis was performed to confirm model adequacy. This approach enabled the transformation of diagnostic features into predictive parameters with quantifiable reliability. Reliability assessment represented the third stage of analysis. The improvement in system resilience was quantified through Mean Time Between Failures (MTBF) and probability density distributions of downtime events. In addition, Kaplan–Meier survival curves were constructed to illustrate the probability of continuous operation over time, thereby providing an intuitive measure of enhanced system reliability after the adoption of predictive maintenance. Finally, based on the mechanical balance of rod string loads and electrical drive parameters, a simplified predictive model was formulated as:

$$D\left(t\right)=\alpha ·I_{peak}\left(t\right)+{\beta ·P}_{prod}\left(t\right)+\gamma ·f\left(t\right)+\epsilon ,$$

(12)

where $D\left(t\right)$ represents the probability of downtime occurrence due to paraffin deposition, $I_{peak}\left(t\right)$ denotes the peak motor current, $P_{prod}\left(t\right)$ the polished rod load, and $f\left(t\right)$ the electric frequency. The coefficients $\alpha ,\beta , \gamma$ were determined empirically from regression analysis, while $\epsilon$ accounts for stochastic variation.

Graphical representation was applied in parallel to the statistical evaluation. Time-series plots, boxplots, and correlation diagrams were generated to illustrate parameter trends and validate predictive indicators. These visualizations provide an intuitive overview of operational improvements and serve as a practical tool for both diagnostic and predictive decision-making in oilfield operations. Through this integrated methodological approach, the results not only demonstrate a statistically significant reduction in downtime and energy consumption but also establish a robust mathematical foundation for early detection of paraffin-related risks in sucker rod pumping systems.

As shown in Figure 8, the distribution of production downtime values exhibits a significant shift following the implementation of predictive maintenance. Prior to deployment, downtime averaged 12.7 days/year per well, with a wide spread caused by paraffin-related interruptions. After implementation, the values converged toward a narrow distribution centered around 0.9 days/year, indicating both a substantial reduction and improved operational stability. The decrease in variance highlights the consistency of predictive monitoring across different wells.

Figure 9 illustrates the average annual downtime per well before and after predictive maintenance.

Five wells demonstrated a pronounced reduction, ranging from 91.5% to 94.9%. The uniformity of improvements across wells underscores the robustness of the approach, confirming that the predictive framework is not limited to specific operating conditions but is broadly applicable.

The results presented in Figure 10 confirm the direct correlation between paraffin accumulation and electrical loading of the pumping unit. Peak motor current and power increased by 37.5% and 35.5%, respectively, under critical paraffinization, while electric frequency decreased by 30%. Following deparaffinization, all parameters returned to near-baseline levels. These results validate the selected electrical parameters as reliable diagnostic and predictive indicators for early detection of paraffin-related issues.

Figure 11 illustrates the improvement in system reliability, expressed as mean time between failures (MTBF).

The average MTBF increased from 25 days in the pre-PdM phase to more than 300 days after implementation. This twelvefold improvement highlights the effectiveness of predictive maintenance in extending operational life, reducing unplanned stoppages, and enhancing overall production resilience.

To evaluate the effectiveness of predictive maintenance (PdM) in mitigating paraffin-induced failures, a synthetic dataset was constructed using aggregate data of this study. Baseline and critical operating levels for motor peak current (12.0 A → 16.5 A), motor power (4.5 kW → 6.1 kW), and electric frequency (50 1/min → 37 1/min) were used to parameterize the degradation profiles. Downtime statistics before and after PdM implementation (e.g., 14.2 days/year vs. 0.8 days/year for well K-1) were used to calibrate the expected rate of failure events. Events were simulated as Poisson processes with piecewise intensities reflecting the pre- and post-PdM phases.

For each well, synthetic minute-level series were generated for a 90-day observation window, including motor peak current ($I_{pk}$), motor power ($P_{pk}$), electric frequency ($f$), dynamometer span ($S=L_{max}-L_{min}$), and span drift ($\dot{S}$). Events were injected according to the calibrated downtime rates, and signals were perturbed with random noise to mimic field variability.

Each signal was normalized against its pre-PdM baseline (first 45 days) and converted to z-scores. For features that decrease with degradation (electric frequency), inverse normalization was applied so that increasing z-scores consistently represent worsening condition.

The EWI was defined as the logistic transform of the mean of standardized features:

$$EWI=\sigma \left({\displaystyle \frac{1}{5}}{z}_t^j\right),\sigma \left(u\right)={\displaystyle \frac{1}{1+e^{-u}}}$$

(13)

Logistic regression was applied to predict whether a failure would occur within a 72-hour horizon. Model performance was evaluated with the receiver operating characteristic (ROC) and area under the curve (AUC). Calibration was assessed using observed vs. predicted failure probabilities.

The synthetic dataset was used to derive the Early Warning Index (EWI) as a composite indicator of paraffin-related degradation. Figure 12 presents the smoothed EWI trajectories for all five wells over the 90-day observation window. Vertical markers indicate maintenance or failure events. A consistent upward drift in EWI is visible in the days preceding events, particularly in wells K-1 and K-2, confirming that the index anticipates degradation before interventions.

Weekly averaging of EWI further emphasizes the difference between pre- and post-PdM behavior (Figure 13). During the pre-PdM phase, a progressive increase is observed across wells, while the post-PdM period shows stabilization at lower EWI levels, with wells K-3 and K-5 demonstrating the least drift, consistent with their minimal downtime.

Survival analysis was performed using the Kaplan–Meier estimator. Figure 14 shows the survival functions for pre- and post-PdM phases. The estimated mean time between failures (MTBF) in the synthetic dataset was approximately 45 days for the pre-PdM phase. The post-PdM curve is heavily censored within the 90-day window, but the observed survival is consistent with the annualized improvement toward ≈300 days reported in the field aggregates.

6. Discussion

The distinguishing feature of the proposed approach is the explicit coupling of physics-based load behavior with probabilistic reliability modeling, which separates this study from existing paraffin prediction methods focused on isolated diagnostic indicators or data intensive machine learning techniques.

The results of this study provide clear evidence that the application of predictive maintenance (PdM) in sucker rod pumping systems significantly mitigates the negative impact of paraffin deposition on oil production. The average annual downtime per well was reduced by more than 93%, a statistically significant improvement (p < 0.01). This outcome surpasses the efficiency gains typically reported in earlier studies relying on preventive or scheduled maintenance, where downtime reductions seldom exceeded 60–80%. Such findings confirm that the integration of fourth-generation SCADA systems with IoT-enabled sensors and predictive algorithms establishes a new benchmark in the diagnosis and prevention of paraffin-related failures.

Mechanical and electrical parameters, such as polished rod loads and peak motor currents, proved to be robust early-warning indicators. Their strong correlation with paraffin accumulation, confirmed by high coefficients of determination (R² > 0.9), demonstrates that variables traditionally used for post-failure diagnostics can be transformed into predictive features. This shift marks an essential transition from reactive or corrective strategies toward intelligent, data-driven maintenance systems.

System reliability was also markedly improved. The mean time between failures (MTBF) increased from approximately 25 days in the pre-PdM phase to over 300 days after implementation. Such a twelvefold improvement not only reduces the frequency of unplanned stoppages but also extends the service life of equipment and minimizes the need for costly workover operations. Kaplan–Meier survival analysis further confirmed these results, showing a more stable and resilient operating regime after the adoption of predictive maintenance.

The present validation relies on a longitudinal within-well comparison, where operational performance before and after predictive maintenance deployment is evaluated on the same assets. This design ensures identical geological conditions, pump configurations, and production regimes, allowing isolation of the impact of the proposed framework. However, no parallel benchmarking against scheduled maintenance or alternative predictive approaches was performed. This limitation results from the lack of matched control wells and the absence of synchronized datasets for competing methods. Future studies will address this by applying the framework to grouped wells and conducting comparative evaluations against preventive maintenance schedules and data-driven predictive models under equivalent field conditions.

While the framework was validated on low- and medium-production wells, its extension to larger fields requires additional architectural considerations. The methodology relies on parameters available in standard SCADA installations, which supports transferability. However, large-scale deployment introduces challenges related to data throughput, heterogeneous sensor quality, communication latency, and legacy PLC-based control systems. Practical implementation in such environments would require distributed edge preprocessing, standardized data interfaces, and protocol adapters for legacy SCADA infrastructure. A staged rollout across grouped wells is recommended to manage integration risk and computational load. These aspects define the boundary conditions for scalability and should be addressed prior to field-wide deployment.

The study highlighted the influence of the human factor during the initial stages of deployment. In 2023, limited operator experience prevented full exploitation of predictive capabilities, and paraffin removal was still primarily performed reactively. Once operators gained sufficient expertise in 2024, predictive strategies were executed more effectively, leading to near-elimination of downtime. This observation emphasizes the importance of continuous operator training while simultaneously indicating that future system improvements should focus on automation and artificial intelligence (AI) integration, thereby reducing the likelihood of human error and ensuring fully autonomous decision-making.

The discussion also underlines the economic significance of PdM. Traditionally, deparaffinization was scheduled at fixed intervals, often resulting in unnecessary interventions and avoidable expenditures. In contrast, the intelligent monitoring framework ensures that interventions are performed exclusively when required, optimizing operational costs while maximizing system availability. Consequently, PdM is demonstrated not only as a technically superior approach but also as an economically rational solution.

Overall, the integration of real-time monitoring, predictive modeling, and intelligent decision support represents a transformative step toward more resilient and sustainable oil production operations. This work confirms the practical benefits of PdM and establishes a methodological foundation for its further advancement through artificial intelligence and advanced data analytics.

The proposed predictive maintenance framework was validated on five low- and medium-production wells with similar sucker rod pump configurations and operating conditions. While the modeling approach is not tied to a specific field, its direct application to wells with different geological settings, pump types, or paraffin compositions may require recalibration. In particular, baseline statistics, normalization parameters, and alert thresholds should be adjusted to reflect local fluid properties, temperature profiles, and mechanical characteristics.

The framework relies on parameters that are commonly available in standard SCADA systems, which supports transferability; however, variations in sensor resolution, data quality, and control logic may influence predictive performance. These aspects should be considered when deploying the methodology in larger fields or under significantly different operating conditions.

From an economic perspective, the implementation of predictive maintenance resulted in a substantial reduction in operational disruptions. Unnecessary deparaffinization operations and workover activities were significantly reduced, leading to an average downtime reduction exceeding 93% across the monitored wells. In addition, timely intervention based on early warning indicators contributed to lower electrical energy consumption during pump operation.

A detailed life-cycle cost analysis and a full comparison between implementation costs and long-term savings were outside the scope of the present study. However, the observed reductions in downtime and energy usage provide quantitative evidence of a favorable cost–benefit balance under real operating conditions. These results indicate that the economic justification of the monitoring and predictive maintenance system is strong, while comprehensive cost modeling remains an open topic for future work.

7. Conclusions

This study has demonstrated that predictive maintenance (PdM) represents a decisive advancement in managing paraffin-related challenges in oil production with sucker rod pumping systems. Through the integration of IoT-enabled SCADA platforms, real-time data acquisition, and predictive algorithms, it was possible to establish an intelligent monitoring framework that substantially improved system reliability, operational efficiency, and economic sustainability. The results clearly confirmed that PdM is not only a technical solution for reducing downtime but also a strategic approach that reshapes the entire philosophy of maintenance in petroleum production.

One of the key contributions of this research is the validation of mechanical and electrical parameters as effective predictive indicators of paraffin accumulation. Variations in polished rod loads, peak motor currents, motor power consumption, and electric frequency provided consistent signals of paraffin buildup before critical failures occurred. These indicators, embedded within a predictive monitoring framework, transformed routine diagnostic measurements into early-warning tools capable of preventing equipment sticking and production interruptions. Such transformation has critical practical importance, as it enables operators to shift from reactive responses to proactive actions, thereby safeguarding the integrity of pumping equipment and preserving continuity of production.

The empirical validation conducted on five production wells provided quantitative confirmation of the effectiveness of PdM. Average annual downtime was reduced by more than 93%, and Mean Time Between Failures (MTBF) improved from approximately 25 days to over 300 days. These results underline the transformative effect of predictive maintenance in minimizing unplanned stoppages, extending equipment life, and reducing the need for costly workover operations. Beyond operational continuity, the study also highlighted improvements in system stability, as confirmed by reductions in variance and deviations of monitored parameters after PdM implementation.

An important insight of the study is the observed influence of the human factor. During the initial stages of deployment, limited operator training reduced the effectiveness of predictive monitoring, and responses to paraffinization alerts were not fully optimized. As operators gained more experience, predictive responses became more effective, resulting in near-complete elimination of downtime. This finding emphasizes the importance of structured capacity building when transitioning to advanced monitoring systems. At the same time, it suggests that the future of PdM lies in reducing reliance on human interpretation and progressively integrating autonomous decision-making tools. The development of artificial intelligence–based frameworks capable of independently generating advisories and executing safe shutdown protocols will be essential to achieving full automation and minimizing the probability of human error.

From an economic perspective, the study confirmed that predictive maintenance offers substantial advantages over traditional maintenance strategies. By enabling deparaffinization operations to be scheduled strictly on the basis of actual need, PdM prevents unnecessary interventions, reduces direct and indirect costs, and optimizes the allocation of maintenance resources. In practical terms, this means lower expenditure on workover operations, reduced consumption of energy and supporting materials, and enhanced cost-effectiveness of production. The financial benefits are further amplified by the increase in system availability and production efficiency, which directly contribute to higher profitability of oilfield operations.

The methodological framework developed in this study establishes a strong foundation for further advancements. By combining real-time data acquisition, predictive analytics, and operational decision support, the research demonstrated a holistic approach to maintenance optimization. The introduction of an Early-Warning Index (EWI), based on normalized operational parameters, provided a scalable solution for timely detection of paraffin-related risks. This framework can be further developed through the integration of machine learning techniques and adaptive algorithms capable of refining predictions as additional data becomes available.

Overall, the conclusions of this study confirm that predictive maintenance is not merely an incremental improvement over existing approaches but a paradigm shift in petroleum production management. By transforming diagnostic parameters into predictive features, PdM enables the oil industry to move toward intelligent, resilient, and cost-efficient production systems. The demonstrated reduction in downtime, improvement in reliability, and optimization of resource use highlight its potential to serve as a cornerstone of digital transformation in oilfield operations. Looking ahead, the integration of predictive frameworks with artificial intelligence and advanced automation will further enhance system performance, reduce uncertainty, and create fully autonomous production environments.

In summary, this study has shown that predictive maintenance in sucker rod pumping systems is both technically and economically justified. It contributes to increased efficiency, reduced operational risk, and improved sustainability of petroleum production. The insights and framework presented here lay the groundwork for the continued development of intelligent maintenance strategies that will play a central role in the modernization and long-term resilience of the oil industry.

Although predictive analytics is employed, the current framework is intentionally based on logistic regression and Cox proportional hazards modeling to preserve interpretability and operational transparency. These methods allow direct association between paraffin accumulation and measurable mechanical and electrical parameters. More complex AI approaches, such as neural networks or reinforcement learning, were not adopted in this phase due to limited labeled failure data and the need for explainable outputs in industrial operation. Future work will address automated feature learning and adaptive thresholding using machine learning techniques.