Research and Application of an Intelligent Cable-Controlled Injection–Production Integration and Control System

Abstract

During offshore oilfield development, traditional injection–production processes commonly suffer from delayed regulation, low operational efficiency, and heavy reliance on manual intervention. Achieving real-time diagnosis of injection–production anomalies and dynamic optimization under complex geological conditions and harsh marine environments represents a core scientific challenge. This study presents the development and field deployment of an intelligent cable-controlled injection–production integrated management system. The work is positioned as an application- and system-oriented study, focusing on addressing practical challenges in offshore oilfield operations through the integration of established machine learning techniques into a cohesive operational platform. The system employs a cloud-native microservice architecture and integrates nine functional modules, enabling closed-loop management from data acquisition to intelligent decision making. Key methodological contributions include: (1) a weighted ensemble model combining Random Forest and SVM for blockage diagnosis, balancing global feature learning with boundary sample discrimination to achieve 92% diagnostic accuracy; (2) a Bayesian fusion framework that integrates static geological priors with dynamic sensitivity analysis for probabilistic quantification of injector–producer connectivity, achieving 85% identification accuracy with rigorous uncertainty propagation; and (3) a three-stage human–machine collaborative mechanism that substantially reduces anomaly response latency while ensuring field safety. Field application in Bohai oilfields demonstrates that the system shortens the injection–production response cycle by approximately 42%, reduces anomaly response time from over 72 h to less than 2 h (a 97% reduction), decreases water consumption per ton of oil by 27.6%, and increases injection–production uptime by 11.3 percentage points. This study provides an interpretable, extensible, and closed-loop technical solution for intelligent offshore oilfield development, with future directions including digital twin predictive simulation and reinforcement learning for real-time optimization.

1. Introduction

With the deepening global energy transition and the rapid advancement of digital technologies, oil and gas field development is undergoing a significant shift from traditional to intelligent models [1]. The “1534” overall development strategy proposed by China National Offshore Oil Corporation (CNOOC) explicitly identifies digitalization and intelligence as critical directions for future oilfield development [2]. Within this context, traditional injection–production technologies face prominent issues, such as delayed regulation, low efficiency, and strong dependence on manual intervention, struggling to meet the demands for efficient offshore oilfield development. Particularly under complex geological conditions and harsh marine environments, achieving precise regulation and dynamic optimization of injection–production parameters becomes a key technical bottleneck constraining development effectiveness.

Intelligent injection–production completion systems can significantly enhance the overall recovery factor of oil and gas reservoirs through real-time monitoring and feedback control mechanisms [3]. According to the latest CNOOC statistics, by the end of 2023, cable-controlled intelligent injection–production technologies had been deployed in 421 wells across oilfields such as Liaodong, Bonan, and Qinhuangdao, with 181 wells in the Liaodong operating area and 110 in Bonan, forming a large-scale application demonstration [4]. The deployment of these intelligent tools provides a massive data foundation for building new intelligent oilfields characterized by unmanned operations, visualized wellbores, collaborative operations, and scientific decision making, also creating favorable conditions for this research study.

Scholars worldwide have conducted extensive frontier research in intelligent injection–production. Smith et al. [5] proposed a data-driven injection–production optimization method, using machine learning algorithms to analyze historical production data and establish parameter optimization models; Li et al. [6] developed a well blockage diagnosis model based on pressure transient analysis, enabling quantitative evaluation of blockage severity; Wang et al. [7] investigated connectivity analysis methods between injection and production wells based on dynamic production data, providing theoretical support for injection–production regulation. More recently, Faizah et al. [1] demonstrated the transformative potential of real-time edge intelligence for intelligent completion injectivity, showing that processes previously taking months can now be completed in minutes through automated systems. However, existing research often focuses on isolated technical problems, lacking systematic integrated solutions, particularly in terms of real-time capability, accuracy, and practicality.

This paper presents an application- and system-oriented intelligent cable-controlled injection–production integrated management system. Rather than proposing novel algorithmic methods, this work focuses on the integration of well-established machine learning techniques into a cohesive, field-deployable platform that addresses real-world operational challenges in offshore oilfields. The system innovatively combines machine learning algorithms with reservoir engineering theory, achieving closed-loop management encompassing data acquisition, anomaly diagnosis, and intelligent regulation. Compared with existing technologies, this system exhibits three significant advantages: (1) employing a microservice architecture to enable flexible expansion of functional modules [8]; (2) establishing multi-algorithm fusion intelligent diagnosis models to enhance analysis precision [9]; (3) developing standardized data interfaces for seamless integration with field equipment [10,11].

2. System Architecture Design

2.1. Overall Architecture Design

The system adopts the industry-leading cloud-native three-tier architecture (IAAS, PAAS, and SAAS) and microservice system to construct a highly modular and scalable technical framework (Figure 1). The overall architecture is divided into three main parts: data layer, service layer, and application layer. Each part is interconnected through microservice interfaces and supplemented with auxiliary service functions such as user permissions, security authentication, and microservice management. The architecture aligns with recent industry trends in cloud–edge collaboration for intelligent oilfield monitoring [12].

• Data Layer: The data engine provides storage capabilities for system-related data and offers a complete data processing workflow. It supports various data types, particularly the loading and parsing of data types relevant to the oil and gas industry, and delivers data-related services such as data collection, data querying, data preprocessing, and data loading. As the foundational support of the system, the data layer integrates multi-source heterogeneous data, including real-time monitoring data, static geological data, production dynamic data, and operational history data. This layer employs a distributed storage architecture, with structured data stored in the Kingbase database and unstructured data managed through the Minio object storage system. Specifically designed for the oil and gas industry, the data layer features a professional data processing engine that supports the parsing and processing of specialized data types such as well logging curves and well testing data, providing a complete Extract, Transform, Load (ETL) process.
• Technical Service Layer: This layer utilizes three public services from the oilfield cloud platform: user service, permission service, and log service. Serving as the intelligent core of the system, the technical service layer comprises two key components: first, the AI computing engine, which integrates various machine learning algorithms such as clustering, classification, regression, SKLearn, and TensorFlow within the algorithm model service; second, the professional graphics engine, which implements basic geometric visualization and specialized visual scenario business rules through the graphic interaction operation service.
• Application Layer: The presentation method is web-based, utilizing technologies such as JavaScript, Vue, and HTML5. The client requires Internet Explorer 10 or above, or a browser based on the Chrome kernel. This layer provides a user-friendly interface for end-users, ensuring a seamless experience across different terminal devices. The application layer implements nine functional modules: real-time injection–production monitoring, dynamic pre-alarming, digital well history, blockage diagnosis, injection–production dynamic analysis, connectivity analysis, intelligent regulation, auxiliary analysis tools, and system management. Each functional module adopts a component-based design, allowing for flexible configuration based on user requirements.

To address key offshore operational constraints—such as limited on-site personnel, high communication latency, and intermittent network connectivity—the architecture incorporates several design features. The cloud-native microservice framework enables modular deployment and remote maintenance, reducing the need for on-site interventions. Local data caching and edge-compatible preprocessing modules ensure that core functions, including real-time monitoring and threshold-based alarming, remain operational during network interruptions. These design considerations ensure the system’s reliability and usability under typical offshore working conditions.

Communication between the system layers is facilitated through well-defined interface protocols, with data exchange conducted in JSON format to ensure loose coupling and scalability. Additionally, the system architecture design fully considers the unique environment of offshore oilfields, ensuring the continued operation of core functionalities even under abnormal conditions such as network interruptions.

2.2. Technical Architecture Implementation

The system is implemented using a Spring Boot + Vue.js frontend/backend separation technology stack (Figure 2). Key technology selections are as follows:

• Development Language: Java JDK 1.8 (Oracle Corporation, Austin, TX, USA).
• Platform Frontend: The frontend uses the Vue development framework, specifically Vue 2, which provides reactive data binding and a component system, making frontend development more efficient and modular.
• Frontend UI Framework: Element UI (Version 2.1, Eleme Inc., Shanghai, China), a desktop-oriented component library based on Vue 2.
• Backend: The backend uses the Spring Boot framework (Version 2.7.1, VMware, Inc., Palo Alto, CA, USA), an open-source Java-based framework designed to create standalone, production-level Spring-based applications. It simplifies Spring application development by following the principle of “convention over configuration,” allowing developers to quickly start and run Spring applications.
• Database: Renmin Jincang (Kingbase) (Version V8R6, Beijing Kingbase Data Technology Co., Ltd., Beijing, China) is a database management system that supports SQL standards, suitable for transactional processing, analytical processing, data warehousing, and other application scenarios.

Recent advancements in edge computing have demonstrated the value of integrating real-time data processing capabilities with cloud-based analytics. The system’s architecture is designed to accommodate future edge computing modules that could enable downhole real-time decision making, following the paradigm established in recent intelligent completion deployments [1]. The development tools and version numbers used in this study are shown in

Table 1. Tools and versions used.

No.	Tool Name	Purpose	Version
1	IntelliJ IDEA	Java development tool	Ver. 2018
2	Python	Provides the foundational environment for running machine learning algorithms	3.6.3
3	Renmin Jincang	Provides data storage services	9.5.21
4	Nginx	Load balancing, responsible for server forwarding	1.20.1
5	JDK	Provides the foundational runtime environment for backend applications	1.8
6	Minio	Used for storing large-capacity unstructured data	2020-12-29T23-29-29Z
7	Redis	Caching database for data caching operations	6.0.12

.

3. Key Technology Research

3.1. Intelligent Blockage Diagnosis Technology

3.1.1. Algorithm Model Research

Well blockage diagnosis is a core function of the intelligent injection–production system. Traditional diagnostic methods rely heavily on manual experience, suffering from subjectivity and inefficiency. This paper innovatively proposes a hybrid diagnostic algorithm integrating Random Forest and Support Vector Machine (SVM), significantly improving diagnostic accuracy.

The Random Forest algorithm avoids overfitting inherent in single decision trees through ensemble learning by constructing multiple trees. It uses information gain as the feature selection criterion:

$$g\left(D,A\right)=H\left(D\right)-H\left(D\right|A)$$

(1)

where $H\left(D\right)$ is the empirical entropy representing the dataset’s disorder and $H\left(D\right|A)$ is the conditional entropy representing the disorder given feature A [13]. By recursively selecting features with the highest information gain for node splitting, a highly generalizable classification model is ultimately constructed.

The SVM algorithm maps nonlinear problems in low-dimensional space to linearly separable problems in high-dimensional space by using the kernel trick. This paper employs the Radial Basis Function (RBF) kernel:

$$K\left(x_i,x_j\right)=\mathrm{e}\mathrm{x}\mathrm{p}(-\gamma {\left|\left|x_i-x_j\right|\right|}^2)$$

(2)

Balancing model complexity and generalization performance is achieved by adjusting the penalty coefficient C and kernel parameter $\gamma$ [14]. SVM is particularly suitable for small-sample, high-dimensional classification problems, complementing Random Forest’s strengths.

The model ensemble strategy uses weighted voting, fusing the predictions of Random Forest and SVM. Weight coefficients were determined via grid search and cross-validation, set to 0.6 for Random Forest and 0.4 for SVM. This specific weighting was determined via grid search and cross-validation, and it was found to be optimal, as it leverages Random Forest’s strength in capturing global feature interactions while still incorporating SVM’s discriminative power on boundary samples to correct edge-case errors. This approach preserves Random Forest’s ability to capture global features [9] while leveraging SVM’s discriminative power on boundary samples.

3.1.2. Feature Engineering Processing

High-quality feature engineering is fundamental to model performance. Three categories of features were extracted from raw production data:

• B Basic Features: They include directly monitored indicators such as tubing pressure, casing pressure, daily water injection volume, and daily liquid production rate. These reflect the well’s fundamental production status and are primary inputs for blockage diagnosis.
• Derived Features: Obtained through specialized calculations, including apparent water absorption index, productivity index, pressure derivative, etc. the Apparent Water Absorption Index (AWAI), is defined as:

$$AWAI={\displaystyle \frac{Q_w}{P_t-P_c+ϵ}}$$

(3)

where: $Q_w$ is daily water injection Volume, $P_t$ is tubing pressure and $P_c$ is casing pressure. $ϵ$ is a small positive constant to avoid division by zero. These features incorporate reservoir engineering expertise, providing deeper insights into changes in formation flow characteristics.

• Temporal Features: Generated via sliding window statistics, including 3-day moving averages, week-over-week change rates, etc. These capture dynamic trends in production parameters, crucial to early blockage warning.

All features underwent rigorous preprocessing before model input: missing values were filled using linear interpolation; outliers were identified and corrected using the 3σ principle; numerical features were normalized using Min–Max scaling as

$$x^{,}={\displaystyle \frac{x-\mathrm{m}\mathrm{i}\mathrm{n}\left(x\right)}{\max \left(x\right)-\mathrm{m}\mathrm{i}\mathrm{n}\left(x\right)}}$$

(4)

and categorical features were converted into numerical form via one-hot encoding.

3.1.3. Model Training and Optimization

Model training employed a 10-fold cross-validation strategy: the dataset was randomly split into 10 subsets; models were trained on 9 subsets and validated on the remaining 1; this process was repeated 10 times; average performance across all folds was used as the evaluation metric. This strategy maximizes the use of limited data while ensuring reliable assessment.

Hyperparameter optimization used Bayesian Optimization, building a probabilistic model of the parameters to guide the search direction, significantly improving efficiency. Key hyperparameters included: Random Forest (n_estimators, max_depth); SVM (C, γ).

Model performance was evaluated using multiple metrics: accuracy reflects overall classification effectiveness; precision measures the reliability of positive predictions; recall assesses the ability to identify positive samples; F1-score, the harmonic mean of precision and recall, is suitable for imbalanced classification problems. Final model performance on the test set is shown in

Table 2. Blockage diagnosis model performance evaluation results.

Algorithm	Accuracy	Precision	Recall	F1-Score
Random Forest	0.89	0.91	0.87	0.89
SVM	0.90	0.88	0.92	0.90
Ensemble Model	0.92	0.93	0.91	0.92

. As shown in Table 2, the ensemble model achieved the highest F1-score of 0.92, demonstrating its superior balance between precision and recall compared with the individual algorithms.

3.1.4. Field Application Validation

The data were labeled according to the method shown in the wellbore blockage reference (Figure 3): Taking the oil pressure index as an example, mark the declining phase and the low-level fluctuation phase as label 0 (normal production); mark the rising phase as label 1 (mild blockage); and mark the peak phase and the stable phase after the peak as label 2 (severe blockage).

Taking Well B23 as an example (Figure 4), comparing model results with field operations, the model successfully predicted the blockage progression. Field operations performed acid stimulation on 10 November 2022. The blockage severity prediction chart (where the blue line represents the actual operating parameters, namely oil pressure, daily water injection volume, and visual water absorption index, and the red line represents the predicted values of these parameters, consistent with the method in Figure 3. The blockage severity can be marked as normal production, mile blockage, and severe blockage based on the shape of the curve) shows that, excluding misjudgments caused by shut-ins due to other operations, key instances of slight and severe blockage could be identified by the model, which consequently generated an alarm, verifying its accuracy.

The annotation method based on tubing pressure trends served as the initial coarse label generation stage. This approach originates from the empirical knowledge of field engineers: when wellbore blockage occurs, the injection pressure typically exhibits a “rise–peak–stabilization” pattern. However, this was merely the starting point of label generation, not the final stage. To transform these preliminary labels into “ground truth labels” suitable for supervised learning, we conducted rigorous cross-validation based primarily on the following two categories of hard data: direct evidence from downhole operation records and skin factor inversion. Through the aforementioned comparison, the labels generated solely by the rules shown in Figure 3 achieved approximately 80% consistency with operational records. For samples where the rules conflicted with the records, we first eliminated operational interferences and then incorporated expert experience combined with other dynamic data such as injection volume and water cut to conduct final arbitration for the disputed periods.

It is important to acknowledge the potential impact of labeling uncertainty on model performance. The initial labels generated based on tubing pressure trends are subject to uncertainty arising from operational interferences (e.g., well shut-ins) and ambiguous blockage conditions. To mitigate this uncertainty, we cross-validated the labels against two independent data sources: downhole operation records and skin factor inversion. Approximately 80% of the rule-based labels were consistent with operation records. For conflicting samples, expert arbitration was conducted using additional dynamic data (e.g., injection volume, water cut) to resolve discrepancies. This multi-source validation process enhances the reliability of the training labels and, consequently, the robustness of the trained model.

The model also provides quantitative evaluation of blockage severity and treatment recommendations. For wells with slight blockage (damage degree < 30%), pressure increase/enhanced injection is recommended; for severe blockage (damage degree ≥ 30%), acidizing is suggested. This functionality provides a scientific basis for field decisions, significantly improving treatment effectiveness.

3.2. Pollution Radius Calculation Model

3.2.1. Theoretical Basis

Pollution radius is a key parameter for evaluating formation damage, directly influencing the design and effectiveness prediction of stimulation treatments. This paper establishes a pollution radius calculation model based on the Dupuit formula and composite reservoir theory.

For an ideal well, production is given by the Dupuit formula:

$$Q={\displaystyle \frac{2\pi kh(p_e-p_w)}{\mu \mathrm{l}\mathrm{n}\left(\frac{r_e}{r_w}\right)}}$$

(5)

When formation damage occurs, the additional flow resistance is characterized by the skin factor S, and the production rate becomes:

$$Q^{\prime }={\displaystyle \frac{2\pi kh(p_e-p_w)}{\mu \ln \left(\frac{r_e}{r_w}\right)+S}}$$

(6)

Treating the damaged zone as the inner region of a composite reservoir, the skin factor can be expressed as:

$$S=({\displaystyle \frac{k}{k_s}}-1)\mathrm{l}\mathrm{n}({\displaystyle \frac{r_s}{r_w}})$$

(7)

By solving these equations simultaneously, the pollution radius $r_s$ and damaged zone permeability $k_s$ can be determined [15].

3.2.2. Numerical Simulation Validation

Numerical simulations were conducted for reservoirs with different permeabilities (

Table 3. Simulated production decline under different damage degrees.

Permeability (mD)	Damage Degree	Production Decline (%)
2000	0.3	12.5
2000	0.5	20.8
2000	0.7	29.2
1000	0.3	15.6
1000	0.5	26.3
1000	0.7	37.5
500	0.3	18.9
500	0.5	31.4
500	0.7	44.8

) to validate model reliability. Simulation results show the following: The lower the reservoir permeability, the more significant the impact of damage on production. For the same permeability, the production decline exhibits an approximately linear relationship with the damage degree.

Based on the simulation results, a pollution radius correction formula was established:

$$r_s=r_{sc}\mathrm{e}\mathrm{x}\mathrm{p}({\displaystyle \frac{{RQ}_a}{{RQ}_r}}-1)$$

(8)

where $r_{sc}$ is the reference pollution radius (20 m), ${RQ}_a$ is the actual production decline, and ${RQ}_r$ is the theoretical production decline. This correction accounts for reservoir heterogeneity and fluid properties, making the results more consistent with real conditions.

3.2.3. Application Case Analysis

Taking Well A in a Bohai oilfield as an example, its production declined from 85 m³/d to 62 m³/d while the bottomhole flowing pressure was kept constant. Model calculations yielded: skin factor S = 3.2, pollution radius rs = 15.8 m, and damage degree Ds = 0.42.

Based on the calculation results, a small-scale acidizing treatment was recommended and implemented for the well. After the operation, the production of the well was restored to 80 m³/d, close to the initial production capacity, verifying the model’s reliability in assessing pollution degree and guiding treatment design.

3.3. Connectivity Analysis Technology

3.3.1. Static Connectivity Analysis

Static connectivity analysis is based on geological and reservoir static data, including the following:

(1)

Sandbody Distribution Analysis: Determining sandbody connectivity between injectors and producers through depositional facies studies and sandbody correlation.

(2)

Fault Seal Evaluation: Analyzing fault impact on fluid flow using methods like Shale Gouge Ratio (SGR).

(3)

Well Trajectory Analysis: Considering the spatial positional relationships of horizontal well sections within the reservoir.

Based on static and dynamic data (well locations, deviations, layers, sandbodies, production dynamics, and well histories), built-in algorithms automatically determine static injector–producer relationships using predefined rules. Expert knowledge is then used to calibrate these results. Subsequently, big data techniques deeply mine patterns within injection–production dynamic data to build dynamic correlation models [16], ultimately achieving quantitative characterization of injector–producer relationships driven by the fusion of geological and injection–production data (Figure 5).

3.3.2. Dynamic Connectivity Analysis

Dynamic connectivity analysis utilizes production dynamic data and employs the Sobol Global Sensitivity Analysis method [17]. This method is a time-series model that requires using lagged injection volumes as inputs (for example, employing the NARX model—Nonlinear Autoregressive with Exogenous Inputs model). This method decomposes the output variance of a model into contributions from individual input variables and their interactions:

$$f\left(x\right)=f_0+{\sum }_i{f}_i\left(x_i\right)+{\sum }_{i<j}{f}_{ij}\left(x_i{,x}_j\right)+\dots +f_{1,2,\dots ,n}\left(x_1{,\dots ,x}_n\right)$$

(9)

The connectivity coefficient between injector i and producer j is defined as:

$$S_{ij}=D_{ij}/D$$

(10)

where D denotes the total variance and $D_{ij}$ represents the contributed variance from injection well i to production well j.

Implementation steps:

(1)

Construct a Multi-Layer Perceptron (MLP) neural network model with injector rates as inputs and producer liquid rates as output;

(2)

Generate input samples via Monte Carlo sampling;

(3)

Calculate the main effect index and total effect index for each injector;

(4)

Obtain standardized connectivity coefficients through normalization.

3.3.3. Integrated Static–Dynamic Analysis Method

An innovatively proposed comprehensive evaluation method integrates static and dynamic analyses:

(1)

Static analysis provides geological constraints, excluding obviously non-connected well pairs;

(2)

Dynamic analysis quantifies connection strength, reflecting actual flow characteristics;

(3)

Results are fused using a Bayesian framework to obtain posterior connectivity probability:

$$P(C_{ij}|D)={\displaystyle \frac{P\left(D|C_{ij}\right)P\left(C_{ij}\right)}{P\left(D\right)}}$$

(11)

where $P\left(C_{ij}\right)$ is the prior connectivity probability derived from static geological analysis, and $P\left(D|C_{ij}\right)$ is the likelihood function inferred from dynamic production data [17]. This integrated approach aligns with emerging hybrid modeling techniques that combine physics-based understanding with data-driven insights [16].

Bayesian updating and propagation process:

(1)

On the Numerically Stable Quantification of $P\left(C_{ij}\right)$

We agree that converting geological interpretations (e.g., fault seal capacity) into a stable numerical probability is challenging. To minimize subjectivity, we employed a multi-attribute scoring system combined with a frequentist statistical approach.

(2)

On the Propagation of Geological Uncertainty through the Bayesian Update

The Bayesian framework is inherently designed to propagate uncertainty. This is handled in our model as follows:

• Prior as a Distribution: The static prior $P\left(C_{ij}\right)$ is not implemented as a single scalar but as a probability distribution. High uncertainty in the geological interpretation (e.g., a fault with sparse calibration data) is represented by a prior distribution with a larger variance (e.g., a Beta distribution with appropriate parameters).
• Dominance of the Likelihood: The dynamic likelihood $P\left(D|C_{ij}\right)$ acts as an “uncertainty filter:”

  ○

  If the dynamic data are strong (e.g., clear interference test results), the likelihood function is sharp. This dominant evidence will override the prior uncertainty, causing the posterior distribution to converge towards the dynamic indication. Strong flow data correct geological ambiguity.

  ○

  If the dynamic data are also weak, the posterior distribution retains the wide uncertainty from the prior, effectively warning the user that the conclusion is based on weak evidence.

• Sensitivity Analysis: We conducted sensitivity tests on the most subjective prior parameters (e.g., fault seal probability). As shown in

  Table 4. Results of integrated static–dynamic connectivity analysis.

  Well Pair	Static Probability	Dynamic Coefficient	Fused Probability	Actual Result	Conclusion
  B01	0.85	0.92	0.91	Connected	Correct
  B02	0.92	0.15	0.33	Not Connected	Correct
  B03	0.78	0.87	0.85	Connected	Correct
  B04	0.95	0.05	0.22	Not Connected	Correct
  …	…	…	…	…	…
  B32	0.95	0.88	0.93	Connected	Correct

  , the final identification accuracy (85%) remained robust across a reasonable range of prior values. This suggests that in our specific case study, the dynamic likelihood dominated the updating process, further mitigating concerns about prior subjectivity.

Our workflow ensures that the static prior is quantifiable and traceable. The Bayesian framework then rigorously propagates geological uncertainty into the final posterior, providing a more honest assessment of connectivity risk.

Application examples (Table 4) show that this method correctly identified connectivity for 27 out of 32 total well pairs in Oilfield B, achieving an identification accuracy of 85%, significantly higher than single-method analyses.

4. System Implementation and Applications

4.1. Implementation of Core Functions

4.1.1. Real-Time Monitoring Module

The Real-Time Monitoring Module integrates monitoring data from four main categories: oil wells, water injection wells, electrical submersible pumps (ESPs), and zonal injection–production tools (Figure 6). Key features include the following:

(1)

High-Frequency Data Acquisition: Key parameters are sampled at 1 Hz to ensure capture of dynamic features.

(2)

Multi-Dimensional Visualization: It supports curves, dashboards, 3D wellbore views, and other formats.

(3)

Anomaly Detection: It implements real-time anomaly identification based on Statistical Process Control (SPC) methods.

Data Downsampling and Aggregation: High-frequency raw data (e.g., once per second) are aggregated over fixed time windows and converted into low-frequency feature data for use by upper-level intelligent modules.

(1)

Mean/Extreme Value Aggregation: Calculate the average, maximum, or minimum values per minute, hour, or day. For example, convert 3600 s level pressure points into a single hourly average.

(2)

Sampling Storage: Retain data points only at fixed intervals (e.g., save one point every 5 min), and discard the intermediate points.

(3)

Interval Statistics: In addition to the mean, store statistical features such as variance and rate of change within the time period to preserve the fluctuation characteristics of the data.

The module is connected to over 200 wells, processing more than 20 million data points daily, providing the data foundation for intelligent analysis [18].

4.1.2. Dynamic Early Warning Module

The Dynamic Early Warning Module employs a multi-level warning mechanism (Figure 7):

(1)

Threshold Warning: It sets upper/lower limits for single parameters.

(2)

Trend Warning: It identifies abnormal trends using time-series analysis.

(3)

Composite Warning: It combines multiple parameters using logical rules.

Warning rules support graphical configuration, allowing users to flexibly define various warning scenarios. The system also provides a warning analysis dashboard displaying key metrics like warning statistics and handling progress.

4.1.3. Digital Well History Module

The Digital Well History Module (Figure 8) manages full lifecycle data:

(1)

Dynamic Information: production curves, operational history, etc.

(2)

Static Information: well configuration, log interpretations, etc.

(3)

Monitoring Data: production profiles, injection profiles, etc.

(4)

Laboratory Data: fluid properties, core analysis, etc.

Adopting a “Well–Layer–Parameter” three-tier architecture, the module supports multi-dimensional data correlation analysis, significantly improving historical data utilization.

4.2. Three-Stage Human–Machine Collaboration

We have reduced the response time from >72 h to <2 h precisely through the following standardized workflow.

Stage 1—Automated Alerting and Diagnosis (Duration: <1 min). System Action: Real-time data streams feed into the model. When the blockage probability exceeds a threshold (e.g., 90%), the system automatically triggers a red alert on the interface and generates the diagnostic recommendation panel mentioned above.

Human Status: The engineer receives a notification on their mobile device or workstation but does not take immediate action, waiting instead for the system to complete its preliminary diagnosis.

Stage 2—Manual Review and Confirmation (Duration: <30 min). Human Action: Upon seeing the red alert, the duty engineer first reviews the diagnostic basis at Level 3. If SHAP indicates that the primary contributing factor is “pressure derivative upwarping” and the engineer confirms this by retrieving the pressure buildup curve, they click “Confirm Diagnosis.”

Collaboration Mechanism: If the basis provided by the system conflicts with the engineer’s experience (for example, if the system mistakenly identifies pressure gauge drift as blockage), the engineer can click “Reject.” The sample is then flagged and returned to the data lake for subsequent model retraining. This step is critical to ensuring field safety, as humans retain ultimate decision-making authority.

Stage 3—Instruction Issuance and Closure (Duration: <1.5 h). System Action: Once the diagnosis is confirmed, the system automatically displays contingency plans based on the severity of the blockage: for mild blockage, it recommends “pressurized water injection + monitoring”; for severe blockage, it recommends “acidizing operation” and pushes an operation design template.

Human Action: The shift manager reviews the plan and issues instructions with an electronic signature. In cases of severe blockage, the operation team begins preparing equipment; for mild blockage, the control room remotely adjusts the injection pressure.

Result Feedback: After the operation is completed, personnel manually enter the operation record into the system. This record serves as new labeling data, forming a closed loop of “alert–diagnosis–operation–feedback.”

Before the system was implemented, engineers had to manually retrieve data, plot curves, and convene meetings for discussion, a process that typically took 1–3 days. The system handles “monitoring” and “preliminary diagnosis,” compressing 72 h of data analysis time into just 1 min. Humans retain “decision making” and “review,” streamlining the previously meeting-intensive discussion process into a simple “confirm” or “reject” action on a unified interface, typically completed within 2 h.

4.3. Field Application Results

The system was deployed in multiple oilfields in the Bohai Sea, achieving significant results:

(1)

Injection–Production Efficiency Improvement: Water injection utilization increased from 68% to 78.5% through intelligent regulation.

(2)

Anomaly Diagnosis Timeliness: Diagnosis time reduced from an average of 3 days to 2 h.

(3)

Treatment Success Rate: Success rate of treatments recommended by the model increased to 92%.

(4)

Labor Cost Savings: Reduced field inspection frequency by over 30%.

(5)

Typical Application Cases:

• Case 1—L Oilfield (12 Well Patterns): After the full application of intelligent regulation in 2023, the response time for water injection well groups was shortened by approximately 42% (

  Table 5. Injection response time comparison in L Oilfield.

  Stage	Traditional (Pre-Application)	Intelligent (Post-Application)
  2021	85 days	-
  2022	78 days	48 days (System Pilot)
  2023	72 days	42 days (Full Deployment)

  ). Connectivity analysis was used to optimize injection–production correspondence, effectively reducing ineffective circulation. Precision de-plugging based on blockage diagnosis models significantly improved single-well oil production [10]. These improvements align with recent industry reports on digital transformation benefits in offshore operations [12].
• Case 2—Q Oilfield: The application of the Dynamic Early Warning Module enables proactive prediction of equipment failures (e.g., electric submersible pumps), while the pollution radius model optimizes treatment design. The integrated use of these technologies led to a marked improvement in overall operational efficiency: anomaly response time was reduced from an average of over 72 h to less than 2 h, a decrease of over 97%; single-well regulation frequency decreased by 72%; the injection–production time rate increased by 11.3%; and water consumption per ton of oil dropped by 27.6% (

  Table 6. Key application performance indicators in Q Oilfield.

  Evaluation Metric	Pre-Application	Post-Application	Improvement	Calculation Method
  Injection–Production Uptime	83.2%	94.5%	+11.3%	(Production Time/Calendar Time) × 100%
  Anomaly Response Time	>72 h	<2 h	−97%	Mean Time from Detection to Action
  Regulation Frequency per Well	4.3 times/year	1.2 times/year	−72%	Total Annual Regulation Commands/Well Count
  Water Consumption per Ton of Oil	5.8 m3/t	4.2 m3/t	−27.6%	Annual Water Injection (m3)/Annual Oil Production (t)

  ) [15]. The 97% reduction in response time parallels the 90% improvement reported in recent edge intelligence deployments for similar offshore applications.

5. Conclusions

This study presents an application- and system-oriented intelligent cable-controlled injection–production integrated management system developed for offshore oilfield operations. Rather than proposing novel algorithmic methods, this work focuses on the integration of well-established machine learning techniques—including ensemble learning, Bayesian inference, and SHAP-based interpretability—into a cohesive, field-deployable platform that addresses real-world operational challenges. Based on a “data-driven + knowledge-driven” integrated approach, an intelligent cable-controlled injection–production integrated management system was constructed, yielding three new insights:

(1)

Multi-source feature fusion combined with ensemble learning significantly enhances blockage diagnosis accuracy. To overcome the limitations of single algorithms in global feature extraction and boundary sample discrimination, a weighted ensemble model integrating Random Forest and Support Vector Machine was constructed. Through grid search and cross-validation, the weight coefficients were optimized (RF:SVM = 0.6:0.4), enabling effective capture of global feature interactions while maintaining precise discrimination of boundary samples. Field validation demonstrates that the ensemble model achieves a diagnostic accuracy of 92% and an F1-score of 0.92, significantly outperforming single algorithms. The model also enables quantitative evaluation of blockage severity and intelligent recommendation of treatment strategies, providing a scientific basis for precise de-plugging operations.

(2)

The static–dynamic integrated Bayesian framework enables probabilistic quantitative characterization of injector–producer connectivity. To address the dual challenges of subjectivity in static geological interpretation and uncertainty in dynamic production responses, a connectivity quantitative evaluation method integrating Sobol global sensitivity analysis with Bayesian inference was established. Static geological information—such as fault seal capacity and sandbody connectivity—is quantified as probabilistic prior distributions, while dynamic production responses serve as the likelihood function. Through Bayesian updating, rigorous propagation and fusion of uncertainty are achieved. Sensitivity analysis shows that when dynamic evidence is strong, the posterior probability converges toward the dynamic indication, effectively overcoming prior subjectivity; when dynamic evidence is weak, the posterior retains a wide distribution, providing users with a quantified risk warning. This method achieved an identification accuracy of 85% across 32 well pairs in Oilfield B, significantly outperforming single-method analyses.

(3)

The human–machine collaborative closed-loop mechanism substantially reduces anomaly response latency. To address the trust barrier and responsibility allocation challenges faced by intelligent models in industrial field applications, a three-stage human–machine collaborative decision-making framework was designed and validated. Stage 1 enables automated alerting and preliminary diagnosis, compressing data response time to within 1 min. Stage 2 introduces SHAP explainability analysis as the basis for manual review, allowing engineers to verify the model’s diagnostic logic based on physical mechanisms, thereby establishing a human–machine trust bridge. Stage 3 enables instruction issuance and operational feedback, forming a closed loop of “alert–diagnosis–operation–feedback.” This framework reduces anomaly response time from over 72 h to less than 2 h—a reduction of 97%—while ensuring field safety, significantly improving the timeliness and accuracy of anomaly management.

Through the above technical integration and mechanism innovation, systematic field application validation was conducted in Bohai oilfields. Application results demonstrate that the injection–production response cycle was shortened by approximately 42%, the injection–production uptime increased by 11.3 percentage points, water consumption per ton of oil decreased by 27.6%, and the regulation frequency per well decreased by 72%. This system provides an interpretable, extensible, and closed-loop technical solution for the intelligent development of offshore oilfields, effectively addressing the critical challenge of transitioning from “algorithm accuracy” to “field practicality.”

Future Directions: Although this study has achieved significant application results in Bohai oilfields, the generalization capability of the AI models under different geological and fluid conditions requires further validation. Future research will focus on the following directions: (1) integrating digital twin technology to construct virtual injection–production systems for real-time simulation of pressure and saturation distributions, enabling predictive “what-if” analysis for regulation strategies [19]; (2) exploring reinforcement learning for real-time optimization of injection–production parameters to achieve adaptive dynamic control; (3) developing edge computing modules to support downhole real-time decision making, reducing dependence on network transmission [20]; and (4) constructing cross-regional intelligent injection–production cloud platforms to facilitate large-scale replication and promotion of the technical achievements [21,22,23].