A Review of Data-Driven Intelligent Monitoring for Geological Drilling Processes

Abstract

The exploration and development of resources and energy are fundamental to human survival and development, and geological drilling is a key method for deep resource and energy exploration. Intelligent monitoring technology can achieve anomaly detection, fault diagnosis, and fault prediction in the drilling process, which is crucial for ensuring production safety and improving drilling efficiency. The drilling process is characterized by complex geological conditions, variable working conditions, and low information value density, which pose a series of difficulties and challenges for intelligent monitoring. This paper reviews the research progress of the data-driven intelligent monitoring of geological drilling processes, focusing on the above difficulties and challenges. It mainly includes multivariate statistics, machine learning, and multi-model fusion. Multivariate statistical methods can effectively handle and analyze complex geological drilling data, while machine learning methods can efficiently extract key patterns and trends from a large amount of geological drilling data. Multi-model fusion methods, by combining the advantages of the first two methods, enhance the ability to handle complex multivariable and nonlinear problems. This review shows that existing research still faces problems such as limited data processing capabilities and insufficient model generalization capabilities. Improving the efficiency of data processing and the generalization capability of models may be the main research directions in the future.

1. Introduction

Geological resources, including petroleum, natural gas, minerals, and water, are indispensable natural resources for human social development. They are not only the raw material basis for industrial production but are also directly related to national energy security and economic independence. With the continuous growth of the global economy, the demand for geological resources is increasing, making the exploration and development of geological resources particularly important.

Since the implementation of the “14th Five-Year Plan”, China’s coalbed methane exploration and development have entered a higher stage. Breakthrough achievements have been made in exploring new fields and strata such as deep coal seams and thin coal seams, while significant results have been achieved in the enhancement and transformation of old gas fields [1]. Among these, the development of deep coalbed methane is particularly important. China’s deep coalbed methane resources are abundant, with deep coalbed methane resources of 29 major basins (groups) estimated to reach 40.71 trillion cubic meters, significantly exceeding the shallow coalbed methane resources within 2000 m [2]. Especially in the depth range of 1500 to 2000 m, the proportion is 31.5%.

China’s deep coalbed methane resources are mainly concentrated in the Junggar, Ordos, and Turpan–Hami–Santanghu basins, with these three regions accounting for 37%, 32%, and 27% of the total, respectively [3]. According to the China Petroleum Exploration and Development Research Institute, coalbed methane resources at depths of 2000 to 3000 m in China total about 18.47 trillion cubic meters [4]. Therefore, it is necessary to strengthen the exploration of deep geological resources. Moreover, the decision made at the 2023 National Natural Resources Work Conference also emphasized the importance of a new round of strategic mineral prospecting actions [5]. This policy aims to achieve breakthroughs in mineral prospecting by promoting technological projects and strengthening technical support in the field of resource exploration, thereby ensuring the economic and social development needs of the country.

In this context, the geological drilling process plays a crucial role in the development of deep coalbed methane resources. Precise drilling not only enables the acquisition of critical data on underground coalbed methane reserves but also provides core samples for analyzing their physical and chemical properties, thereby assessing the storage capacity and extractability of coalbed methane. However, current drilling technologies face numerous challenges; complex geological conditions significantly increase risks and costs, particularly in deep environments where the risk of tool wear and breakage escalates [6]. These issues can impact the efficiency of the geological drilling process and even compromise its safety. Therefore, the comprehensive monitoring of the geological drilling process must be implemented to ensure that operations are conducted safely and efficiently.

For an extended period, the detection of deep geological environments has encountered significant challenges due to limitations in technology and equipment. Current sensing technologies for deep geological detection often struggle to accurately capture the complexity and variability of subsurface structures, resulting in low information density during the drilling process [7]. This limitation hampers the comprehensive monitoring capabilities essential for effective geological drilling. To mitigate these challenges, the exploration and application of intelligent monitoring technologies have become crucial for achieving the precise monitoring of inefficiencies and abnormal statuses in drilling. Such intelligent monitoring can analyze drilling parameters in real time, predict and avoid potential risks, and offer considerable scientific and economic benefits in enhancing drilling efficiency, reducing energy consumption and ensuring operational safety [8].

Intelligent monitoring in geological drilling integrates anomaly detection, fault diagnosis, and fault prediction. Anomaly detection serves as the foundation of intelligent monitoring, continuously tracking key indicators such as drilling speed, pressure, and torque to identify deviations from normal operating patterns, thereby facilitating early fault detection. Fault diagnosis entails a thorough analysis of these detected anomalies to pinpoint specific fault types and their locations, which is essential for resolving issues and implementing targeted corrective actions. Finally, fault prediction leverages both historical and real-time data, alongside diagnosed fault types, employing advanced data analysis and machine learning models to forecast the likelihood and timing of future faults. This comprehensive approach enhances safety, efficiency, and reliability in the geological drilling process through real-time monitoring, prompt fault diagnosis, and precise predictions of potential future faults [9].

In summary, scholars have conducted extensive research in the field of data-driven intelligent monitoring, which plays a crucial role in the exploration and development of geological resources. These studies not only reduce the cost and risk of geological drilling but also improve the efficiency of data processing and decision making. This paper will review the academic contributions in the field of intelligent monitoring methods in the geological drilling process, including anomaly detection, fault prediction, and fault diagnosis, while introducing representative methods in these three aspects. By summarizing valuable experiences in the research process, it points out the existing problems and prospects for the future development of intelligent monitoring in the geological drilling process.

2. Descriptions and Analysis of Drilling Process

Geological drilling is a complex engineering process that faces numerous challenges due to uncertain geological conditions and the demanding nature of equipment operation. In response, key technologies like Measurement While Drilling (MWD), Logging While Drilling (LWD), Managed Pressure Drilling (MPD), and Rotary Steerable Systems (RSSs), along with advancements in intelligent monitoring, have played a crucial role in improving both the efficiency and safety of drilling operations [10,11]. These technologies enable the real-time monitoring of critical drilling parameters and fault diagnosis, allowing for more precise control, the prediction of drilling performance, and overall success in the drilling process.

2.1. Characteristic Analysis of Drilling Process

Geological drilling involves a highly complex process, requiring the coordination of multiple components and technologies to drill into the Earth’s crust. As shown in Figure 1, the system consists of a traveling block, draw works, a rotary table, and other components essential for controlling the drill pipe’s movement. The drill bit, positioned at the bottom of the drill string, penetrates underground formations. The bottom-hole assembly includes various tools for guiding and monitoring the drilling process. Mud is pumped down the drill pipe via the slurry pump and returns through the mud pit, carrying debris back to the surface [12]. This process involves managing various parameters such as drilling speed, weight on the bit, and torque to ensure efficiency and safety. Intelligent monitoring plays a critical role in tracking these factors and optimizing drilling operations in real time.

(1)

Multi-condition Characteristics

Multi-condition characteristics refer to the different behaviors and performance features exhibited by systems or equipment under different operating conditions, environmental settings, or working states. The geological drilling process is a highly complex and variable engineering environment, involving various working states and conditions. As drilling progresses, systems or equipment often need to operate under multiple conditions, including different loads, speeds, temperatures, pressures, or environmental conditions. Each condition can uniquely impact the system’s performance and behavior, posing significant challenges to the data-driven intelligent monitoring of the drilling process.

Parameter Coupling: One significant challenge faced by the monitoring system in the geological drilling process is the diversity of the drilling environment. Due to the varying geological formations, such as transitioning from sandstone to shale or encountering fractured zones and increasing pressure and temperature with drilling depth, multiple drilling parameters become interdependent. For instance, an increase in the drilling fluid density to control formation pressure can affect the rate of penetration and equivalent circulating density, which in turn influences the risk of wellbore instability. These parameters are not independent but are influenced by other parameters and environmental conditions, interacting closely. This parameter coupling requires an intelligent monitoring system to analyze the interrelationships and linkage effects among multiple parameters, such as how changes in rotary speed and torque affect bit wear and drilling efficiency, rather than simply tracking changes in individual parameters [13].

High-dimensional Data: The diverse conditions involved in geological drilling result in numerous parameter variables that the monitoring system needs to handle. These variables include rock hardness, type, drilling depth, drilling technology, drilling speed, and more, each directly or indirectly affecting drilling efficiency and safety [14]. High-dimensional data not only increase information volume but also introduce challenges in data analysis as data dimensions increase.

In summary, the multi-condition characteristics of the geological drilling process reflect its complexity and dynamic variability. Different conditions may result in completely different data distributions, and condition changes are frequent and complex. This necessitates high adaptability in data-driven intelligent monitoring to accommodate multiple condition switches. Real-time adaptation to changing conditions is essential, capturing different condition changes accurately to ensure drilling process efficiency and safety.

(2)

Non-stationary Characteristics

Non-stationary characteristics, a key concept in time series analysis, refer to data not maintaining stability over time, exhibiting trends, seasonality, volatility, and autocorrelation. This means that the data’s mean, variance, and correlation change over time, complicating analysis and prediction [15]. In the drilling process, the non-stationary characteristics impact data-driven intelligent monitoring in the following ways:

Temporal Dynamics: As the drill bit penetrates deeper, it encounters diverse geological challenges, leading to changes in statistical properties of monitored parameters. For instance, the drilling pressure may increase due to harder rock formations, while the drilling speed might decrease as the bit encounters more resistance. This variability necessitates that monitoring systems adapt to these fluctuations, ensuring real-time adjustments in analysis models. Continuous changes in physical parameters—such as increased mud flow to maintain borehole stability and variations in temperature and the density of the drilling fluid—highlight the need for a responsive monitoring system to maintain accuracy throughout the drilling cycle [16].

Spatial Dynamics: The spatial variability in geological conditions, such as differences in rock types, fault distributions, and stratigraphic features, directly impacts the selection of drilling parameters and overall drilling performance. For instance, drilling through varying rock types may require adjustments in the weight on the bit and rate of penetration. Intelligent monitoring systems need to account for these spatial variations to accurately detect anomalies. The proper interpretation of spatial data, including real-time measurements of rock hardness and fracture density, is essential for anticipating risks and enhancing safety. This knowledge enables optimized parameter settings and drilling strategies, improving efficiency while reducing hazards [17].

Due to the non-stationary characteristics of the geological drilling process, parameters and conditions continuously change in a nonlinear manner. This means the drilling process exhibits temporal and spatial dynamics, with a high degree of spatiotemporal information coupling. This complexity and unpredictability demand high sensitivity in intelligent monitoring to adapt to these dynamic characteristics.

(3)

Low Information Value Density

In the geological drilling process, despite collecting vast amounts of data, the proportion of truly valuable information is relatively low, indicating low information value density. This is mainly due to two reasons:

Transmission delay: The significant distance between surface and downhole monitoring sensors may result in data transmission delays. While these delays are generally short, they can still affect real-time monitoring and decision making in fast-changing drilling environments. Even slight delays in data transmission can reduce the timeliness of information and may occasionally limit the effectiveness of prompt adjustments, potentially impacting decision accuracy [18].

Noise interference: Drilling equipment generates significant mechanical and acoustic noise, which can interfere with various sensors used in Measurement While Drilling and Logging While Drilling systems, such as acoustic, pressure, and vibration sensors. Mechanical noise is produced by the movement and operation of the drilling equipment, while acoustic sensors in Measurement While Drilling systems may struggle to differentiate between background noise and meaningful signals. Pressure sensors in Logging While Drilling systems can be affected by mud flow variations or sudden pressure changes [10,11]. As the drilling depth increases, the higher temperature and pressure present additional challenges, leading to potential signal drift or distortion. Furthermore, equipment wear, such as drill bit damage or irregular drill rod movement, introduces additional vibration noise, complicating the analysis of sensor data [19].

In summary, the low information value density in the geological drilling process is primarily due to transmission delays and noise interference. Delays reduce information timeliness, and noise can lead to inaccurate or distorted data, affecting decision accuracy. Addressing this challenge requires intelligent monitoring to handle large data volumes and quickly extract valuable information to support effective drilling decisions and operations.

2.2. Functions of Intelligent Monitoring in Geological Drilling

Intelligent monitoring plays a crucial role in the geological drilling process, with key functions including anomaly detection, fault diagnosis, and fault prediction. Anomaly detection monitors real-time drilling parameters such as the torque, weight on bit, drilling speed, and mud flow rate to detect deviations from normal patterns, signaling potential issues like bit wear or formation changes [20]. Fault diagnosis then identifies the root causes of these anomalies and recommends corrective actions to maintain operational continuity and efficiency [21]. Fault prediction leverages historical and real-time data to anticipate future faults, enabling proactive adjustments to prevent incidents like drill string failure or wellbore instability [22]. Together, these functions form the foundation of intelligent monitoring, making it indispensable for modern geological drilling operations.

(1)

Anomaly Detection

In geological drilling, data-driven anomaly detection techniques play a crucial role. In analyzing drilling process data in real time, these techniques identify behaviors that may cause deviations from normal operations, enhancing safety and efficiency. Real-time data analysis forms the core of data-driven anomaly detection.

To achieve real-time anomaly detection, Reeber et al. proposed a drilling tool wear-monitoring method based on Extreme Gradient Boosting and autoencoders. They utilized Extreme Gradient Boosting to model complex nonlinear relationships within drilling data and employed autoencoders to detect anomalies by identifying deviations between original and reconstructed data. This combination allows for the efficient processing and accurate detection of tool wear in real time [23]. These methods process and analyze drilling data in real time for efficient and accurate anomaly detection. Similarly, Alsaihati et al. proposed an intelligent system using real-time data analysis and machine learning models to predict surface torque during drilling, using the Mahalanobis distance for the anomaly detection of downhole issues like stuck pipes [24].

And Zhong et al. employed a Convolutional Long Short-Term Memory Neural Network model for real-time anomaly detection in drilling data. This model integrates convolutional layers to capture spatial features and utilizes long short-term memory units to model temporal dependencies in sequential data, enhancing the detection of anomalies by learning both spatial and temporal patterns in drilling data [25]. Li et al. proposed an anomaly detection method based on the relationship between input and output signals in the drilling process, developing mathematical models to establish normal operational behavior and detecting deviations from this behavior to identify anomalies, emphasizing the critical role of real-time data analysis in data-driven anomaly detection [26].

These studies demonstrate that data-driven anomaly detection in geological drilling can process vast amounts of data to identify behaviors deviating from normal drilling operations. As the initial part of intelligent monitoring, the goal is to quickly extract potential anomaly information from real-time drilling data, providing immediate optimization guidance for drilling teams.

(2)

Fault Diagnosis

Data-driven fault diagnosis involves analyzing specific causes, nature, and solutions for faults after anomaly detection. In geological drilling, fault diagnosis is a key function of intelligent monitoring. It helps drilling teams promptly identify and resolve issues, enhancing overall safety and efficiency. This includes determining fault types, locating sources, and proposing corresponding repair or adjustment measures.

Fault diagnosis in geological drilling has evolved significantly, starting from theoretical model establishment to incorporating data-driven methods, particularly machine learning and deep learning techniques. Reiss provided the theoretical foundation for fault diagnosis in geological drilling [27]. As technology advanced, data-driven methods were introduced. Shen et al. developed a condition monitoring and fault diagnosis system by integrating serial communication protocol bus technology [28]. Zhang et al. advanced the field with an automatic fault diagnosis system based on drilling parameters [29], using Principal Component Analysis and Self-Organizing Maps for accurate fault diagnosis. Additionally, deep learning techniques enable fault classification from sound signals, reducing dependence on expert experience and improving diagnosis accuracy [30]. Vununu et al. combined Principal Component Analysis and Artificial Neural Networks to develop an automatic machine fault diagnosis system based on sound, demonstrating the application value of machine learning in fault diagnosis [31].

These studies show the evolution from theoretical model-based methods to modern automated diagnosis based on data. This progression has enhanced fault diagnosis accuracy and efficiency, providing reliable decision support for drilling operations. Fault diagnosis offers in-depth problem analysis and solutions, minimizing downtime and reducing potential safety risks, making it an indispensable core function in intelligent monitoring.

(3)

Fault Prediction

Data-driven fault prediction is crucial in geological drilling, involving various technical applications to predict faults based on anomaly detection and fault diagnosis results. It significantly reduces unplanned downtime and improves operational safety and efficiency, making it a research hotspot in geological drilling. Advanced data analysis, machine learning algorithms, and artificial intelligence extract valuable information from drilling data to identify potential risks and fault signs.

To predict potential faults during drilling, scholars proposed a method for predicting open-hole cable logging faults [32], aiming to reduce costs and time increases caused by faults and improve drilling efficiency. They employed three machine learning techniques—support vector machine, naive Bayes, and decision tree—to predict open-hole cable logging results based on drilling process data. The support vector machine showed an optimal prediction accuracy. This predictive capability is closely linked to the drilling and logging processes. Logging provides essential data about the geological formations encountered, which can influence drilling parameters and decision making. By integrating fault prediction with logging, operators can better anticipate potential issues, ensuring a smoother drilling operation. This connection underscores the importance of fault prediction not only in improving efficiency but also in enhancing safety during drilling activities.

Similarly, Noshi et al. used supervised and unsupervised learning data mining algorithms [33], including logistic regression, hierarchical clustering, and decision tree, to analyze comprehensive data from eighty land wells for predicting casing failure. Zhai et al. developed an intelligent prediction model for drilling complexity based on case-based reasoning, integrating adjacent well data, computer technology, artificial intelligence, and data mining [34] to diagnose and predict potential faults before drilling operations.

3. Intelligent Monitoring in Drilling Process

Currently, data-driven intelligent monitoring research for geological drilling processes can be categorized into multivariate statistical methods, machine learning techniques, and multi-model fusion methods. Multivariate statistical methods can effectively reveal key patterns and trends in data [35], aiding in the detection of abnormal drilling conditions. However, while multivariate statistical methods are effective in data simplification and interpretation, they may struggle with nonlinear complex patterns. In contrast, machine learning methods are highly powerful in pattern recognition [36] and prediction [37], but their opacity [38] presents challenges for result interpretation and validation. Further, combining multivariate statistical methods with machine learning techniques, known as multi-model fusion methods, can leverage their respective strengths, enhancing the accuracy and reliability of intelligent monitoring. This section will analyze current application research status of intelligent monitoring in geological drilling processes from these three directions, exploring their advantages and existing issues.

3.1. Intelligent Monitoring Based on Multivariate Statistics

Multivariate statistical analysis involves using mathematical and statistical methods to analyze and interpret relationships and patterns in multivariable datasets. In geological drilling, this includes analyzing multiple related variables such as the drilling pressure, speed, torque, and mud flow rate [13]. Multivariate statistical analysis is essential in the monitoring of the geological drilling process as it can reveal complex, multidimensional relationships within the data, helping to identify geological features, optimize drilling efficiency, and detect potential risks in real time [39]. This analysis supports intelligent monitoring systems by providing a scientific basis for data interpretation, thus enhancing the accuracy and reliability of prediction models, as shown in

Table 1. Overview of methods for geological drilling monitoring.

Considering Issue	Method	Characteristics	Application Scenarios
Multi-condition Characteristics	DB Clustering [39]	Handling data abruptly and slow changes	Local similarity analysis of multi-condition data in drilling process
	MB-SFA, MB-ICA [40]	Considering static, dynamic, and large-scale characteristics	Complex condition monitoring in modern industrial processes
	SFA, BN [41]	Extracting static and dynamic features and clustering analysis	Anomaly detection in multi-mode switching during drilling process
	DIPCA [42]	Extracting dynamic, linear, and nonlinear features	Real-time monitoring of nonlinear dynamic processes
	Multi-step DSFA [43]	Precisely partitioning dynamic conditions, and changing control limits	Full-condition monitoring of dynamic systems
Non-stationary Characteristics	MCC [44]	Capturing complex dynamic characteristics	Multivariate anomaly detection in non-stationary processes
	IPCA [45]	Block processing and handling dynamic characteristics	Process monitoring under non-stationary characteristics
	Deterministic Alg. [46]	Extracting the slowest varying features	Monitoring dynamic changes in time series data
	KPCA, KL Div. [47]	Handling minor shifts	Detecting subtle changes in non-stationary processes
	CA [48]	Extracting non-stationary features from historical data	Predicting fouling in steam generator pipes
	CA [49]	Constructing a stationary feature data set	Dynamic monitoring of data non-stationary characteristics
	ACA [50]	Distinguishing true faults from normal variations	Fault identification in dynamically changing environments

.

(1)

Considering Multi-condition Characteristics

Applying multivariate statistical methods is a key step in analyzing geological drilling data, especially considering the multi-condition characteristics of the drilling process. Geological drilling is influenced by various complex factors, including the physical and chemical properties of geological formations, technical parameters of drilling tools, and operational conditions, collectively forming the multi-condition characteristics. Multivariate statistical analyses, such as principal component analysis (PCA), cluster analysis, and factor analysis, can effectively handle and analyze these complex multivariable data.

Given the correlations among various conditions in the geological drilling process, Zhang et al. used time series feature extraction and density-based (DB) clustering methods to analyze the extracted features, addressing the data fluctuations and slow variations due to multi-condition characteristics, particularly in detecting and diagnosing wellbore instability issues such as lost circulation and kick [39]. Additionally, Huang et al. proposed a dual-layer distributed monitoring structure [40] based on multiblock slow feature analysis (MB-SFA) and multiblock independent component analysis (MB-ICA), to handle complex static, dynamic, and large-scale characteristics in modern industrial processes. Xu et al. proposed multiple subspace slow feature analysis (SFA) [41], suitable for addressing issues arising from multi-condition characteristics. This method uses domain subtraction clustering to divide different modes, further dividing each mode into Gaussian and non-Gaussian subspaces, extracting static and dynamic features, and performing Bayesian network (BN) fusion monitoring for anomaly detection. Guo et al. proposed a multi-feature extraction technique based on principal component analysis [42] for nonlinear dynamic process monitoring, combining dynamic internal principal component analysis (DIPCA), PCA, and kernel principal component analysis (KPCA) in a serial structure to extract dynamic, linear, and nonlinear features. Ma et al. proposed a new multi-step dynamic slow feature analysis (DSFA) algorithm [43], carrying out full-condition monitoring for dynamic systems, precisely dividing dynamic conditions, and adjusting control limits based on condition changes.

(2)

Considering Non-stationary Characteristics

Data in the drilling process typically exhibit significant non-stationarity, including trends, periodicity, seasonality, and random noise in time series. These characteristics make traditional statistical methods and models inadequate, necessitating the use of multivariate statistical methods to better understand and analyze these non-stationary characteristics. Messaoud et al. achieved anomaly detection in the drilling process through time series analysis and multivariate control charts (MCCs) [44], capturing complex dynamic characteristics of non-stationary processes. Fan et al. addressed the challenges of non-stationarity by proposing a distributed monitoring method based on integrated probabilistic component analysis (IPCA) and the minimal redundancy–maximum relevance algorithm [45], effectively handling the complex dynamic characteristics due to non-stationarity.

To further address non-stationary characteristics, Zafeiriou extended slow feature analysis (SFA) [46], including a novel deterministic algorithm (Alg.) and an expectation maximization (EM) algorithm to extract the slowest varying features from multiple time-varying data sequences. Cai et al. combined kernel principal component analysis (KPCA) with Kullback–Leibler (KL) divergence [47] to handle changes due to small shifts in non-stationary processes, validated on typical experimental datasets. Kwak et al. used cointegration analysis (CA) to extract non-stationary features accumulated in historical data [48], successfully applied in the fouling prediction of DC steam generator pipes. Wen et al. combined extracted non-stationary features with stationary features to form a new stationary feature dataset, updating monitoring indicators [49].

Given the frequent change patterns in geological drilling monitoring, Zhang et al. proposed an adaptive cointegration analysis (ACA) method [50] to distinguish real faults from normal changes, updating the model with normal samples and adapting to gradual changes in cointegration relationships. Alternatively, Rao et al. proposed a non-stationary process monitoring method based on alternating conditional expectations (ACE) and CA [51], maximizing the linear correlation of transformed variables to handle nonlinear relationships between variables. Zhao et al. proposed a sparse CA-based total variable decomposition and distributed modeling algorithm [52] for non-stationary processes, fully decomposing different cointegration relationships between non-stationary variables and exploring the close linear correlations through local cointegration vectors in each block.

In summary, detection methods based on multivariate statistics are increasingly attracting attention in drilling process monitoring. These techniques allow for the in-depth analysis of complex multivariable data, revealing intrinsic relationships between variables for more precise state monitoring of the geological drilling process. However, implementing these methods often requires determining the types of data to be monitored first and constructing corresponding monitoring models based on data characteristics, leading to multiple modeling processes to address multi-condition characteristics. Future research directions will need to deepen our understanding and application of multivariate statistical methods and innovate algorithms and technologies to meet high-standard monitoring requirements for geological drilling, achieving efficient and accurate monitoring and analysis.

3.2. Intelligent Monitoring Based on Machine Learning

With the rapid development of artificial intelligence technology, machine learning, as a core branch, is increasingly applied in various fields. In geological drilling monitoring, the introduction of machine learning techniques provides new perspectives and methods for traditional geological drilling operations. Machine learning enables computer systems to learn from data and make decisions without explicit programming. In the context of geological drilling process monitoring, machine learning techniques analyze historical geological drilling data, geological information, and real-time monitoring data to learn complex relationships between geological features and drilling process parameters [53]. This enables the accurate detection of anomalies, real-time fault diagnosis, and predictive maintenance, enhancing the overall safety and efficiency of drilling operations, as shown in

Table 2. Machine learning methods applicable to monitoring the drilling process.

Task	Method	Characteristics	Application Scenarios
Anomaly-Detection	NN [53]	Multi-param. fusion, real-time monitoring	Identifying different states in the drilling process
	LOF, NN [54]	Detecting local anomalous data	Anomaly detection in NN monitoring
	RF [55]	Dimensionality reduction, improving efficiency	Extracting effective features for anomaly detection
	DBN, GAN [56]	Reconstructing missing data, feature selection	Anomaly detection in high-dimensional data
	Cascade monitoring, CNN [57]	Analyzing spatial–temporal info, combining sub-models	Comprehensive anomaly detection in industrial processes
	GMM, stacked denoising AE [58]	Initial mode identification, extracting deep nonlinear features	Robust monitoring under steady-state modes
Fault-Diagnosis	AC-GAN, Bayesian algo. [59]	Mitigating data scarcity issues	Automatic diagnosis of downhole drilling accidents
	CNN [60]	Incremental learning, including new samples	Dynamically updating fault diagnosis
	Multi-task learning, CNN [61]	Simultaneous anomaly localization and fault classification	Fault diagnosis in complex processes
	CNN [62]	High-precision classification	High-precision fault diagnosis
	DT [63]	Clear rules, easy to interpret	Fault diagnosis and alarm design in industrial processes
	Optimal ELM, Bernoulli transform coyote opt. [64]	Improving classifier performance	Enhancing fault diagnosis accuracy
Fault-Prediction	BN, LSTM [65]	Combining time series prediction	Early fault warning for steam turbines
	MLP, ANN, BN [66]	Combining multiple models, enhancing prediction accuracy	Fault prediction in production processes
	BN [67]	Handling uncertainty	Early warning of wellbore loss and influx accidents
	BN [68]	Flexible modeling, handling complex relationships	Monitoring operational parameters in oil wells

.

(1)

Anomaly Detection

With the development of machine learning technology, its application in drilling anomaly detection is becoming more mature. Machine learning models can analyze historical geological drilling data and real-time monitoring data, learning the distinctions between normal operations and anomalies, enabling the automatic detection of potential anomalies in the drilling process. Compared to traditional rule-based and experience-based detection methods, machine learning offers higher flexibility and accuracy, effectively reducing the risk of human error. Liao proposed a neural network (NN)-based [53] model, emphasizing the ability to distinguish between normal and abnormal drilling states, optimizing network performance through algorithm improvements. Yang et al. developed a local outlier factor (LOF) anomaly detection algorithm to detect various anomalies [54], validated in NN monitoring models.

Addressing low information value density, Li et al. proposed a feature simplification random forest (RF) algorithm [55] to extract effective features, reducing dimensionality and improving anomaly detection efficiency. Tian et al. proposed a feature-based deep belief network (DBN) method [56], using generative adversarial networks (GANs) to reconstruct random and non-random missing data, selecting feature variables using Spearman’s rank correlation coefficients from high-dimensional data, and successfully employing a DBN for deep abstraction, learning, and tuning in anomaly detection.

Additionally, Yu et al. proposed a cascade monitoring network to simultaneously analyze spatiotemporal information for detecting industrial process anomalies [57]. This method uses convolutional neural networks (CNNs) to extract spatiotemporal information from each variable, combining multiple sub-models into a final monitoring model. Gao et al. used Gaussian mixture models (GMMs) for preliminary mode identification [58], employing stacked denoising autoencoders (AEs) to extract deep nonlinear features embedded in process variables, establishing robust monitoring models for each steady-state mode.

(2)

Fault Diagnosis

Machine learning techniques learn complex relationships between normal and abnormal states in geological drilling by analyzing historical and real-time data, enabling automatic fault identification and classification. Wang et al. proposed a downhole drilling accident diagnosis method [59] using an auxiliary classifier generative adversarial network (AC-GAN) to expand the dataset and a Bayesian algorithm for the diagnosis model, addressing low information value density. Yu et al. proposed an incremental-learning general CNN [60], updating itself with newly collected abnormal samples and fault categories for fault diagnosis. Zhao et al. proposed a multi-task learning CNN model [61] for simultaneous abnormal variable localization and fault classification, applicable in geological drilling fault diagnosis.

Further, Glaeser et al. successfully achieved high-precision fault diagnosis using advanced CNN classifiers [62]. Dorgo et al. proposed a decision tree (DT) classifier-based alarm information design method [63] for industrial process fault diagnosis, applicable in geological drilling fault diagnosis. Hu et al. proposed a new fault diagnosis method based on optimal extreme learning machine (ELM) [64], using a Bernoulli transform–coyote optimization algorithm to optimize the kernel ELM classifier, improving fault diagnosis accuracy. These methods, employing typical machine learning techniques, vary in strategies but are suitable for geological drilling fault diagnosis, enhancing fault diagnosis precision.

(3)

Fault Prediction

Timely and accurate fault prediction is crucial in geological drilling for ensuring operational safety, reducing costs, and improving efficiency. Machine learning techniques analyze historical and real-time monitoring data, learning complex relationships between normal and abnormal states, enabling the early identification and prediction of potential faults. Bayesian networks (BNs), which are a typical machine learning algorithm, are often used for fault prediction.

For instance, Zhang et al. proposed a method for predicting wellbore loss and influx accidents in drilling processes by constructing a BN-based prediction model [67]. They selected critical drilling parameters, such as mud weight, formation pressure, and drilling rate, that represent accident characteristics and considered the uncertainty of parameter changes during accidents. Their model effectively predicted potential wellbore instability incidents, providing valuable guidance for drilling operations.

Similarly, Mamudu et al. utilized Bayesian networks to develop fault prediction models for monitoring operational parameters in oil wells [68]. By incorporating various operational data like pressure, temperature, and flow rates into the BN model, they could predict potential faults in real time, enhancing the reliability and safety of oil well operations.

Liu et al. proposed an early fault warning method based on a combination of Bayesian networks and long short-term memory (LSTM) neural networks [65]. They developed an LSTM prediction model that addresses data uncertainty and considers complex equipment operations. Tested with real steam turbine data, their method provided accurate early warnings during fault creep stages. Although applied to steam turbines, this approach is applicable to geological drilling, where equipment complexity and data uncertainty are significant challenges.

In another study, Mamudu et al. also proposed a method combining multilayer perceptrons (MLPs) and artificial neural networks (ANNs) with BN techniques for effective production fault warning [66]. By integrating MLP and ANN models with a BN, they improved the fault prediction accuracy in production systems. This method is applicable to geological drilling fault prediction due to similar operational complexities and the need for accurate fault forecasting.

These methods employ machine learning, especially Bayesian network techniques, differing in technical approaches and focuses but are applicable in geological drilling fault prediction. They demonstrate the effectiveness of advanced machine learning models in handling complex relationships between drilling parameters and predicting potential faults.

In summary, with the rapid advancement of artificial intelligence, especially machine learning, its application in geological drilling monitoring is becoming increasingly widespread. Through deeply analyzing historical geological drilling data, geological information, and real-time monitoring data, machine learning techniques can identify complex relationships between drilling parameters, achieving effective state monitoring. Despite significant potential in enhancing safety, efficiency, and accuracy, challenges remain in data quality and availability, model generalization and adaptability, real-time computational efficiency, and model interpretability. Future research must explore data processing, model optimization, and technological innovation to address these challenges, further advancing geological drilling monitoring technology.

3.3. Intelligent Monitoring Based on Multi-Model Fusion

As shown in

Table 3. Multi-model fusion for the drilling process.

Task	Method	Characteristics	Application Scenarios
Anomaly Detection	Hybrid PCA, multivariate CNN-LSTM [69]	Enhancing anomaly detection performance	Anomaly detection and optimization
	Ensemble learning, KCVA, Bayesian inference [70]	Improving monitoring performance	Monitoring complex industrial processes
Fault Diagnosis	Enhanced RF, SFA [71]	Analyzing static and dynamic nodes	Dynamic fault classification
	Ensemble learning, ICA [72]	Enhancing model generalization	Monitoring non-Gaussian processes
	EWC, PCA [73]	Continuous learning, preventing forgetting	Monitoring complex and variable industrial processes
Fault Prediction	Machine learning models, multivariate statistics [74]	Integrating multiple techniques	Predicting potential faults in the drilling process

, multi-model fusion involves organically combining multivariate statistical analysis and machine learning methods, providing comprehensive and precise monitoring for geological drilling. This approach integrates various data analysis techniques, retaining the interpretability of multivariate statistical methods while leveraging machine learning’s strength in handling nonlinear features, better managing complex data, and offering robust fault prediction and precise fault diagnosis. Consequently, the effectiveness of multi-model fusion methods in intelligent monitoring applications in geological drilling has been widely researched.

To validate the effectiveness of multi-model fusion methods, Islamov used machine learning models combined with multivariate statistical methods to predict potential faults in geological drilling [74]. The study compared various machine learning algorithms, including logistic regression, naive Bayes classifier, K-nearest neighbors, decision trees, support vector machines, RF, gradient boosting, and NNs, to identify and classify abnormal states. Multivariate statistical methods were used to evaluate different machine learning algorithms’ performance, including accuracy, recall, and F-score, to determine the most suitable fault prediction model for geological drilling. Barbosa emphasized machine learning’s potential in predicting and optimizing drilling rates, highlighting multivariate statistical methods’ importance in model performance evaluation and feature selection [75].

While machine learning techniques may outperform traditional models in fault prediction accuracy, multivariate statistical analysis remains crucial in feature selection and model evaluation, demonstrating the effective fusion of both methods. Chai proposed an enhanced RF [71], analyzing static and dynamic nodes simultaneously, classifying faults, and using a modified SFA method to design new slow indices for supervised fault classification, reflecting the fusion of machine learning and multivariate statistics in logic and standard design, applicable in drilling fault classification.

Further, more in-depth multi-model fusion methods combine machine learning and multivariate statistics. Tariq proposed using hybrid probabilistic PCA combined with multivariate convolutional long short-term memory (CNN-LSTM) models [69], integrating neural networks and probabilistic clustering to enhance anomaly detection performance. Li and Yan proposed an independent component analysis (ICA) method based on ensemble learning [72] for non-Gaussian process monitoring, improving model generalization by integrating ensemble learning logic. Wang and Wu proposed a similar method, introducing ensemble learning and kernel canonical variable analysis (KCVA) to develop a novel ensemble kernel canonical variable analysis method [70], combining multiple KCVA models using Bayesian inference to improve process monitoring performance.

The fusion of machine learning and multivariate statistical modeling addresses their respective shortcomings. For instance, Zhang used elastic weight consolidation (EWC) to solve catastrophic forgetting in PCA models [73], achieving continuous learning, applicable in complex and variable industrial process monitoring, including intelligent geological drilling monitoring.

In summary, multivariate staistical analysis and machine learning each have their own strengths in geological drilling monitoring. Multivariate statistical analysis excels at identifying relationships between multiple variables, providing clear insights into trends and correlations in drilling data. Machine learning, on the other hand, is powerful in processing large datasets and detecting complex patterns, allowing for real-time anomaly detection and predictive fault diagnosis. The fusion of these two methods combines their strengths, resulting in a more robust approach to monitoring, where data interpretation and fault prediction are both enhanced. Together, these three approaches play a critical role in improving the accuracy, reliability, and efficiency of geological drilling monitoring, as shown in Figure 2.

4. Challenges and Prospects

In recent years, significant progress has been made in the application of multivariate statistical methods and machine learning techniques in the field of geological drilling data processing and monitoring. These technologies leverage their unique advantages in handling complex data, providing new opportunities to enhance the intelligent monitoring of geological drilling processes. Moreover, their integration has the potential to facilitate real-time decision making and improve operational efficiency. Despite achieving a range of results, some challenges remain in practical applications. Future research directions will focus on exploring more effective solutions to address these challenges and promote further advancements in this field.

4.1. Challenges

Multivariate statistical methods play a crucial role in the processing of geological drilling data, effectively revealing multidimensional relationships within the data and providing a scientific explanation for intelligent monitoring. These methods demonstrate unique advantages in addressing the multifactorial and non-stationary characteristics of the drilling process. Concurrently, the introduction of machine learning techniques has significantly enhanced the level of intelligence in monitoring, allowing for more accurate predictions of drilling performance and potential risks through the analysis of historical and real-time data, thereby facilitating effective anomaly detection, fault diagnosis, and prediction.

The integration of multivariate statistical analysis and machine learning techniques through a multi-model fusion approach offers a more comprehensive and precise solution for geological drilling monitoring. This innovative method not only improves data processing capabilities but also enhances the accuracy of fault warnings and the adaptability of diagnostic models. Nevertheless, current intelligent monitoring technologies still face several limitations, which constrain their effectiveness and further development, primarily manifesting as the following:

(1)

Lack of Comprehensive Consideration of Global and Local Features

Current drilling process-monitoring technologies face two main issues when addressing multi-condition and non-stationary characteristics. On one hand, focusing on the global features of the drilling process often overlooks the importance of local features, which may have a decisive impact on drilling efficiency and safety under specific conditions. On the other hand, when concentrating on local features, the relationships between these features and their collective impact on the overall drilling process might be missed. In geological drilling monitoring, maintaining a dynamic balance between global and local features is crucial. Global features provide an overall trend of the drilling process, while local features reveal subtle, short-term changes, which are often key to predicting anomalies and avoiding potential risks.

(2)

Scarcity and Low Information Value Density of Drilling Data

Existing data-driven intelligent monitoring methods can somewhat mitigate the challenges of data scarcity and low information value density in geological drilling. Nevertheless, these methods still face challenges in processing complex geological drilling process data. The uncertainty and variability of the drilling environment require intelligent monitoring systems to handle large volumes of data and possess high generalizability. Drilling data may become scarce due to equipment limitations, costs, and external environmental factors, and the valuable information density within the data might be low. This necessitates more effective identification and utilization of potential value in sparse data during intelligent geological drilling monitoring.

(3)

Lack of Spatiotemporal Information Coordination

Although recent data-driven intelligent monitoring methods for geological drilling processes have made significant technological advancements, particularly in addressing the temporal characteristics and interrelations of local variables, they still have limitations. These methods often focus on either temporal analysis or spatial feature analysis, failing to effectively combine these two critical dimensions. However, geological drilling is a highly complex and dynamic process involving temporal evolution and multiple spatial variables, such as drilling equipment and formation characteristics. These spatiotemporal interactions collectively determine the efficiency and safety of the drilling process. Thus, relying solely on single-dimensional analysis makes it challenging to fully capture and understand the complex phenomena in the drilling process, limiting the accuracy and applicability of monitoring methods.

4.2. Future Directions and Solutions

To address these challenges, future research should focus on several key areas. First, enhancing the integration of global and local features is essential for effectively capturing their interactions. Second, the development of intelligent monitoring technologies capable of managing large datasets with high generalizability will enhance the analysis of the drilling performance. Lastly, incorporating robust spatiotemporal analyses will facilitate a deeper understanding of the complexities inherent in the drilling process. These directions will significantly contribute to advancing intelligent monitoring capabilities in geological drilling and may outline future development paths. The following aspects will be the focus of ongoing research:

(1)

Intelligent Monitoring Based on Multi-scale Information Granulation

To address the lack of comprehensive consideration of global and local features in geological drilling process monitoring, multi-scale information granulation methods can be adopted. Through analyzing data at different granularity levels, this approach can effectively capture both the overall trends and local detail changes in the drilling process, providing a comprehensive understanding of drilling data. This method can identify subtle anomalies that might be overlooked in conventional data analysis and explore their relationships with the overall drilling process, deepening the understanding of anomaly causes and their potential impacts on drilling efficiency and safety.

(2)

Intelligent Monitoring Based on Sample Augmentation and Transfer Learning

To tackle the issues of data scarcity and low information value density in drilling processes, the combination of sample augmentation and transfer learning offers an effective method to overcome traditional data limitations. Since drilling data are often scarce and have low information value density, transfer learning methods can significantly enhance the monitoring model performance. This approach leverages pre-trained deep learning models from other domains or related tasks to achieve knowledge transfer, reducing the dependency on large labeled datasets and enhancing model generalization under limited data conditions. Sample augmentation techniques further supplement this by artificially expanding the training set, improving model generalization.

(3)

Intelligent Monitoring Based on Spatiotemporal Correlation Analysis

To address the lack of spatiotemporal information coordination in geological drilling process monitoring, intelligent monitoring methods based on spatiotemporal correlation analysis can be employed. Through analyzing the relationships between temporal sequence data and spatial distribution data, this approach can deeply capture the dynamic changes and interactions in the drilling process across time and space. Detailed spatiotemporal analysis not only helps to understand the changing characteristics of the drilling process more precisely but also identifies potential spatiotemporal anomalies that might be overlooked. This provides richer and more accurate data support for the drilling process, significantly improving efficiency and safety.