Exploratory data analysis and interpretable machine learning for anomaly detection in Epb tunnel boring machine operations

Abstract

Introduction:

Earth-Pressure-Balance Tunnel Boring Machines (EPB TBMs) generate high-dimensional sensor streams during excavation, yet operational anomalies are difficult to identify early because the signals are noisy, coupled, and change with geology and operating mode.

Methods:

This study develops an exploratory data analysis framework for EPB TBM operations using ring-based data from a 7 km tunneling project comprising more than 4,400 rings. After preprocessing and z-score normalization, Pearson and Spearman correlation analyses confirmed strong torque–thrust coupling (r ≈ 0.85). Time-series visualization with z-score outlier screening identified rings showing high thrust but low penetration, indicating potential tool inefficiency, wear, or challenging ground conditions. An XGBoost regression model was used to predict thrust force, and SHAP analysis quantified feature contributions after excluding thrust-derived and internally redundant indices, identifying cutterhead torque, alignment position, and penetration per revolution as the principal independent predictors.

Results:

The model achieved R² = 0.872, RMSE = 381.8 kN, and an a20-index of 0.993 on the held-out test set.

Discussion:

The framework captures both expected physical couplings and orthogonal behaviors, enabling interpretable anomaly identification and providing actionable insights for monitoring and decision support in EPB TBM operations.

1 Introduction

Tunnel Boring Machines (TBMs) have become indispensable to modern underground construction, enabling the development of subways, sewage networks, utility corridors, and mining passages. These highly mechanized systems, often exceeding 100 m in length and weighing several hundred tons, operate in complex geological, hydraulic, and spatial environments where even minor mechanical faults can lead to costly delays or safety hazards. Continuous monitoring and analysis of TBM performance have therefore become fundamental to maintaining construction efficiency and operational reliability.

Among TBM technologies, Earth-Pressure-Balance (EPB) machines have been particularly prominent in urban soft-ground tunneling, where minimizing ground loss and surface settlement is paramount. According to Apoji (2023), EPB TBMs maintain face stability by regulating pressure within the cutterhead chamber through the conditioning of excavated soil. Additives such as foam or bentonite are introduced to transform the spoil into a plastic, flowable medium, allowing chamber pressure to be balanced against the surrounding ground and groundwater conditions. This pressure-control mechanism is central to safe and efficient EPB tunneling.

Modern EPB TBMs generate vast quantities of operational data through onboard sensors that continuously record thrust force, cutterhead torque, penetration rate, cutter rotation speed, chamber pressure, and grout volume. These measurements are systematically logged on a per-ring basis which is a structuring approach recommended by (Fu et al., 2022) as optimal for capturing the cyclic nature of excavation and segment-installation sequences. The ring-based data structure used in this study encompassed approximately 4,400 rings over a 7 km tunnel alignment, providing an ideal temporal resolution for systematic exploratory analysis and anomaly detection.

Operational anomalies of the TBMs, such as excessive thrust, inefficient cutting, or grout overuse, often signal underlying mechanical degradation, cutter wear, or adverse ground conditions. However, the multivariate and nonlinear nature of TBM sensor interactions makes anomaly identification complicated and unreliable using basic threshold logic. To address this limitation, the present study developed a structured, machine learning based Exploratory Data Analysis (EDA) framework to uncover latent relationships, temporal deviations, and interpretable behavioral patterns from TBM operations.

The framework developed in this research consisted of four integrated analytical modules. First, the data preprocessing was conducted which handled missing or erroneous entries and harmonized scales through imputation and z-score standardization, ensuring statistical comparability across parameters (Guo et al., 2025). Second, correlation analysis was conducted using Pearson and Spearman coefficients that quantified mechanical interdependencies, revealing a strong positive relationship between torque and thrust (r ≈ 0.85), thereby validating the physical coherence of the dataset. Third, ring-wise time-series visualizations were generated followed by anomaly risk scoring. This emphasis on temporal structure is consistent with recent work showing that multi-ring time-series frameworks can capture dynamic excavation effects more effectively than static formulations while remaining interpretable through SHAP-based analysis (Zou et al., 2026). This module revealed statistically significant deviations in the operational trends. Rings characterized by high thrust, yet low penetration were indicative of inefficient cutting or unfavorable strata, a pattern consistent with density-based anomaly detection principles demonstrated in EPB systems Apoji (2023), where chamber-pressure irregularities often precede mechanical intervention.

The SHAP analysis, conducted after excluding all features mathematically derived from thrust or from other predictors, identified Cutterhead Torque as the dominant predictor consistent with the torque–thrust coupling independently established in the correlation analysis (r ≈ 0.85). Tunnelmeter (alignment position) and Cutterhead Penetration per Revolution followed as the next most influential contributors, reflecting geological heterogeneity along the 7 km alignment and cutting-efficiency conditions at the face.

By integrating correlation mapping, temporal anomaly visualization, and SHAP-based interpretability into a coherent analytical framework, this study demonstrated a transparent and data-centric approach to identifying abnormal TBM behavior. This structured EDA methodology bridged empirical tunneling expertise with modern machine-learning interpretability. This interpretability focus is increasingly emphasized across infrastructure-management research, where explainable AI is regarded as essential for transparency, trust, and engineering adoption (Hu et al., 2026) offering a reproducible foundation for performance diagnostics and anomaly detection in mechanized tunneling operations.

The main contributions of this study are summarized as follows: 1) Development of a structured Exploratory Data Analysis (EDA) framework tailored for ring-based EPB TBM operational datasets. 2) Integration of statistical correlation mapping, time-series anomaly detection, and interpretable machine learning to reveal mechanical and geotechnical dependencies in tunneling operations. 3) Application of SHAP-based interpretability to quantify the influence of operational parameters on TBM thrust behavior, bridging data-driven analytics with tunneling engineering knowledge.

Despite the growing volume of ML-based TBM performance prediction studies, three limitations persist in the current literature. First, prior work has largely bypassed the exploratory stage, deploying predictive models on raw sensor data without systematic preprocessing, feature-dependency analysis, or temporal characterization of the data structure. Second, existing anomaly detection approaches including density-based and deep-learning methods identify anomalous conditions but do not attribute them to specific operational drivers in terms that practicing engineers can interpret and act upon. Third, feature selection in published TBM ML models has rarely been audited for circular dependencies between predictor variables and the target, a methodological gap that inflates apparent model performance without providing genuine mechanistic insight. The present study addresses all three limitations through the structured EDA framework described above: systematic preprocessing ensures data integrity before any modelling is attempted; ring-wise time-series screening with z-score thresholds exposes temporal anomaly patterns within the operational record; and SHAP-based interpretability conducted after rigorous removal of circularly derived features identifies the genuine independent drivers of thrust force in physically interpretable terms. Together, these contributions advance the state of the art by demonstrating that exploratory interpretability is a prerequisite for reliable TBM diagnostics rather than an optional analytical supplement.

The remainder of this paper is organized as follows: Section 2 reviews relevant literature and positions this study within the existing research landscape; Section 3 describes the dataset; Section 4 presents the methodology; Section 5 reports the results and observations; Section 6 discusses the findings; Section 7 outlines directions for future research; and Section 8 presents the conclusions.

2 Literature review

Data-driven methodologies for Tunnel Boring Machine (TBM) performance monitoring and predictive maintenance have evolved into a rapidly expanding research frontier, driven by the substantial financial and safety repercussions associated with unscheduled tunneling downtime. Over 2 decades of published work, however, a consistent pattern emerges: the field has invested heavily in predictive accuracy while systematically underinvesting in the exploratory, diagnostic, and interpretability infrastructure that makes predictions actionable for practicing engineers.

The first generation of TBM ML research focused primarily on deploying models directly against operational and sensor data to forecast key performance indicators such as penetration rate, advance rate, and machine loads (Sun et al., 2018; Flor et al., 2023; Shan et al., 2022). These studies demonstrated that ML algorithms could achieve competitive accuracy on historical TBM datasets. Some models tend to lose robustness and transferability when geological conditions, project-specific operating regimes, and error distributions shift between projects (Li et al., 2024; Zhang et al., 2024). The root cause, as Hastie et al. (2009) and Kuhn and Johnson (2019) identified in the broader ML literature, is that without systematic exploratory analysis, preprocessing, and explicit examination of variable interrelationships, even strong predictive models can overfit localized patterns and fail to generalize. Harandizadeh et al. (2021)’s hybrid ANFIS-PNN model optimized by an imperialism competitive algorithm achieved correlation coefficients above 0.96 on a TBM penetration-rate dataset from Malaysia, representing a high-water mark of predictive performance. Similarly, Li et al. (2021) applied a hybrid gene expression programming and whale optimization framework to TBM penetration prediction in granitic rock and demonstrated that soft-computing hybridization could match or exceed empirical formulations.

A second generation of studies recognized this limitation and began developing multi-stage workflows that combine statistical data understanding with predictive modelling. Devarajan et al. (2022) produced one of the earliest demonstrations of this paradigm shift: beginning with association-rule mining to uncover co-occurrence patterns among cutterhead torque, thrust force, and slurry pressure in metro TBM data, their workflow then constructed interpretable decision trees capable of flagging operational anomalies in real time, before finally training a compact neural network for penetration-rate forecasting. By calibrating association-rule thresholds to a minimum support of 0.2% and a confidence level of 75%, the framework successfully isolated infrequent but operationally significant anomalies including transient torque surges during soft-soil transitions and contributed to a reported 30% reduction in unplanned maintenance stoppages. Apoji (2023) extended this staged approach specifically to EPB TBMs, applying PCA to reduce dimensionality before deploying a Local Outlier Factor algorithm, achieving an 85% true-positive rate at a 5% false-alarm rate across 6 months of continuous tunneling data. More recently, Karahan and Alkaya (2026) demonstrated the value of systematic EDA prior to ML deployment in a hard-rock TBM context: their ensemble framework incorporating boxplot analysis, correlation assessment, VIF-based multicollinearity inspection, and SHAP feature importance substantially improved both predictive stability and interpretability when applied to a small heterogeneous dataset. Further extending interpretable ML to sequential TBM sensor data, Zhou et al. (2026) developed a causally interpretable AI framework for TBM load time-series prediction, and Lei et al. (2026) applied Transformer-based architectures to continuous TBM sensor sequences for surrounding rock classification during active tunneling.

Beyond TBM-specific contributions, the broader predictive maintenance and anomaly detection literature has established several foundational principles that the tunneling field has been slow to absorb. Rousseeuw and Leroy (2005) and the comprehensive review by Zimek et al. (2012) established the conceptual basis for robust outlier detection across industrial domains. Subsequent advances in deep learning including autoencoders, variational networks, and Gaussian mixture models achieved considerable success in manufacturing and process monitoring settings (Benhanifia et al., 2025), and Zhang et al. (2021) applied hybrid deep-learning frameworks to tunneling datasets. However, interpretability remained a persistent challenge. The response increasingly prominent in the recent TBM literature has been to integrate SHAP-based interpretability with predictive models. Hu et al. (2025) demonstrated this approach for TBM tunnel crown convergence prediction: their Bayesian-optimized LightGBM model with SHAP analysis explicitly ranked the contribution of each geological and design feature to predicted deformation. Zhang et al. (2026) extended interpretability into operational control by coupling residual-based prediction with risk-graded early-warning logic for TBM advance-rate monitoring. At the intersection of geological perception and operational optimization, Dong et al. (2025) demonstrated that self-supervised learning applied to TBM muck images could reliably identify lithology during active excavation, and Zhou et al. (2025) showed that ML-based decision-making frameworks for TBM control parameters could translate data-driven insights into concrete operational recommendations.

What remains absent from the published literature, despite this progress across two generations of research, is a structured exploratory data analysis framework that systematically integrates all four analytical functions data-quality assessment, correlation diagnostics, temporal anomaly detection, and interpretable machine learning within a single coherent pipeline designed specifically for EPB TBM ring-based operational data at the scale of a full tunneling project. Prior studies have advanced individual components in isolation: Apoji (2023) addressed unsupervised anomaly detection; Devarajan et al. (2022) demonstrated multi-stage staging; Karahan and Alkaya (2026) demonstrated EDA-informed ML in a hard-rock, geological-input context; Hu et al. (2025) demonstrated SHAP-based interpretability for a convergence prediction problem; and Zhou et al. (2026) and Lei et al. (2026) advanced interpretable time-series modelling and geological perception respectively. There is a need for a study to combine all four elements preprocessing, correlation mapping, z-score-based temporal screening, and SHAP-attributed model interpretation into a single reproducible EDA architecture applied to a large-scale EPB TBM project. This gap constitutes the primary motivation and contribution of the present research.

3 Dataset

The available EPB TBM logs typically include continuous readings of thrust, torque, RPM, and pressures for each ring which is similar to the standard data monitored for different TBMs in different projects. Importantly, all data here is aggregated per ring. Rodríguez et al. (2024) explicitly recommend using excavation rings as time units when modeling TBM data, to better reflect the actual tunneling process. Likewise, Bai et al. (2025) describe a ring-based summation-averaging strategy to handle TBM time-series inputs. Thus, in our dataset each record is one ring of excavation, consistent with these practices.

Each numeric column has at most very few missing values (typically <0.5%). For example, nearly all rings have valid Thrust and Torque (4445/4447 records), while filtered fields like “Ring install time (<8h)” have ∼4323 valid entries (the rest exceeded 8 h and were blank). A couple of trivial columns (e.g., “Unnamed: 10”) are mostly empty and can be ignored.

A statistical summary of key fields is shown below in Table 1.

In addition to the above, the statistical distributions of six key EPB TBM operational parameters across all 4,453 excavation rings were visualized through violin plot as shown in Figure 1. Each panel displays the full empirical distribution (blue violin), the interquartile range and median (white box with red line), whiskers extending to 1.5×IQR, and individual outlier rings (red dots). Descriptive statistics are annotated within each panel.

FIGURE 1

For context, mean thrust ∼9.21 × 10³ kN (std ∼1.07 × 10³) and mean cutter torque ∼1.454 × 10³ kNm (std ∼249) in this data. The average ring advance was ∼39 min. The field penetration index spans ∼100–400 kN/mm, consistent with hard-to-soft rock variability (Note: the unusually high maximum grout volume (30.6 m³) likely reflects an outlier event).

As expected, position-related fields are strongly interdependent: Station, Tunnelmeter and Ring all correlate nearly perfectly (linear along the tunnel alignment). Operational metrics also correlate via physical laws (e.g., higher thrust often yields higher penetration, up to a point). Exploratory plots of scatter (not shown) can reveal these trends and any anomalies (for example, rings where thrust is high but advance is low may signal a stuck ring). We will apply outlier/anomaly detection to these multivariate patterns.

Similar ring-based TBM datasets and analyses have been reported. Bai et al. (2025) explicitly use “excavation rings as units” for time-series modeling of TBM wear. SAALG Geomechanics, a Barcelona-based geotechnical engineering and data analytics company, notes that modern TBM data platforms update soil parameters and predictive models ring-by-ring. In practice, TBM logs (torque, thrust, RPM, pressure, etc.) are routinely collected per ring and fed into predictive models (e.g., for penetration rate). These systems use retrospective analytics to flag anomalous performance zones and recalibrate assumptions. Consequently, our online research focused on comparable EPB tunneling projects and machine-learning studies that utilized ring-by-ring TBM logs (e.g., TBM performance or cutter wear prediction). Prior work on TBM anomaly detection often dealt with cutter wear or geological events using the same kinds of ring-based features, however, without the use of ML techniques.

4 Methodology

A structured Exploratory Data Analysis (EDA) framework was developed to extract actionable insights from TBM operational data and uncover hidden anomaly patterns. The methodology consisted of the following sequential stages as shown in the flow chart (Figure 2).

FIGURE 2

4.1 Data preprocessing

The raw EPB TBM dataset contained numerous data-quality issues common in tunneling operations. As noted in prior reviews, TBM “big tunneling data are full of invalid, discrete, missing, noisy, inconsistent, and redundant information”. In our case, sporadic sensor dropouts and spurious readings (e.g., negative thrust values or unphysical spikes) were present. These problems reduced effective sample size and biased statistics. Indeed, missing values and outliers “compromise the statistical power” of the dataset and significantly biased estimates of mean and variance. Therefore, systematic cleaning was performed: all missing or anomalous values were identified and either imputed or removed before analysis.

Missing data (often caused by instrument faults, communication failures, or human input errors) were handled by interpolation or simple imputation. For example, isolated gaps in a sensor time series were filled by linear interpolation, using the average of neighboring readings:

This technique preserved continuity across the gaps without introducing bias. In cases of longer missing intervals or at the boundaries, nearest-neighbor or mean imputation was used to avoid discarding entire segments. In total, less than 1% of the data were missing; however, even these gaps were treated carefully because, as Kwak and Kim (2017) emphasize, omitted values can “reduce the data available to be analyzed” and degrade the reliability of downstream results.

After cleaning the time series, several derived features were engineered and rescaled the data for analysis. First differences and rolling statistics were computed to highlight dynamic changes (e.g., the cycle-to-cycle change in thrust or torque). For example, the first-difference feature d_t = x_t - x_(t-1) captures abrupt shifts. “Moreover, each numeric feature was normalized to remove scale effects, using z-score standardization:

This ensured all features have zero mean and unit variance, which is important for many anomaly detection algorithms. In summary, the preprocessing pipeline which consisted of imputation, outlier handling, smoothing, and normalization produced a clean, consistent dataset suitable for exploratory analysis.

Z-score standardization (zero mean, unit variance) was selected as the normalization method after systematic consideration of the available alternatives. Min-max scaling was considered but was found less impactful because the dataset contains extreme outlier events. For instance, grout volume has values up to 30.6 m³ against a mean of 5.3 m³ which would compress the data from majority of rings into a narrow band and distort relative distances between normal operational values. Similarly, Log transformation was not utilized because several sensor channels record legitimate zero values during non-operational phases, for which log transformation is undefined. Robust (median/IQR) scaling was also evaluated but not used because the ±3 standard deviation threshold applied in the subsequent time-series outlier screening step relies on the standard deviation unit having a direct probabilistic interpretation a value beyond ±3σ occurs with probability less than 0.3% under a Gaussian distribution which the IQR-based scale does not provide. Z-score normalization satisfies all three requirements of this analysis: it is robust to the extreme values documented in this dataset (Guo et al., 2025; Kwak and Kim, 2017), internally consistent with the ±3σ screening threshold, and produces SHAP values in standard deviation units that enable direct cross-feature comparability across variables measured in physically different units (kN, bar, mm/rev, mm, etc.).

Table 2 illustrates an example of the preprocessing workflow. Raw sensor readings with missing or anomalous entries are shown alongside their cleaned values as an example of derived feature. For instance, a missing penetration rate value was interpolated linearly, and a short spike in thrust was smoothed by a rolling average. Such preprocessing transformed the raw TBM logs into well-behaved features that supported the subsequent anomaly detection analysis.

4.2 Exploratory data analysis

To reveal the important characteristics and hidden structures in the TBM sensor data, this EDA framework is organized and carefully developed.

4.2.1 Correlation analysis

The correlation analysis was performed directly on the ring-aggregated sensor dataset without prior dimensionality reduction. Each observation in the dataset corresponds to a single excavation ring, with variables representing average or representative sensor readings recorded during that ring’s excavation. Pearson and Spearman correlation coefficients were computed pairwise across all 4,453 rings using the cleaned, z-score normalized feature matrix. No statistical feature extraction, principal component transformation, or other dimensionality-reduction step was applied before computing the correlations. The Pearson coefficient captured linear associations between variables, while the Spearman coefficient captured monotonic non-linear relationships. Computing both from the same original ring-level data allows direct comparison of their respective patterns and ensures that any differences between the two heatmaps reflect genuine non-linearities in the data rather than artefacts of pre-processing. This approach is consistent with the ring-based data structuring strategy recommended for EPB TBM operational datasets (Rodríguez et al., 2024; Bai et al., 2025), which treats each excavation ring as an independent statistical unit.

The Pearson correlation heatmap, as shown in Figure 3, revealed several strong linear relationships, particularly among mechanical and alignment parameters. Ring Number, Station, and Tunnelmeter showed near-perfect correlation, confirming that they can be used interchangeably as temporal or positional indicators of excavation progress. Thrust Force and Cutterhead Torque were also tightly coupled, reflecting the mechanical interdependence between forward thrust demand and rotational resistance during cutting. Moderate positive correlations were observed between Penetration per Revolution and Advance Rate, indicating that greater penetration generally translates into faster advancement. Cutterhead Rotation Speed and Torque exhibited a moderate association as well, consistent with resistance-based feedback control during operation. Notable negative relationships were also evident: the Penetration Index (Ip) was inversely related to penetration and positively related to thrust, which is expected given its proportional relationship to thrust over penetration, while Specific Energy decreased as penetration rate increased, suggesting that inefficient cutting consumes more energy. Overall, these correlations highlight deterministic linkages among core excavation variables and support the use of linear dependency analysis for maintenance-oriented monitoring and energy optimization.

FIGURE 3

The correlation heatmaps revealed a highly significant positive linear relationship between key mechanical parameters: Thrust force (TBM#1 Thrust force [kN]) and Cutterhead torque (TBM#1 CH Torque [kNm]).

Both Pearson and Spearman coefficients exceeded +0.89, confirming that increases in thrust are strongly associated with corresponding increases in torque. This mechanical coupling is expected during high-resistance excavation phases and validates the physical integrity of the sensor data.

FIGURE 4

These relationships are critical for detecting complex subsurface behaviors where linear assumptions break down, e.g., variable soil types, fluctuating moisture levels, or suboptimal foam injection. Spearman correlation enables the model to partially identify latent non-linear trends, especially between penetration and specific energy (ρ = −0.305), but reveals weak monotonicity between penetration and pressure (ρ = −0.054) and energy and pressure (ρ = +0.105). This highlights the need for multivariate models that go beyond simple rank-based correlations to detect complex interactions in TBM performance data.

Several variables exhibited near-zero association under both Pearson and Spearman metrics, indicating statistical and functional independence. For example, Keystone position [A°] showed negligible correlation with Grout Pressure L01 [bar], and Vertical deviation offset ref [mm] showed negligible correlation with TBM#1 CH Torque. Such orthogonal channels are valuable for anomaly detection because they reflect independent subsystems and broaden the detection space to include fault modes not directly tied to cutting performance.

The correlation structure also revealed highly correlated feature clusters, including thrust and torque, grout pressures (L01–L05), and articulation offsets (articulation points 1 and 2). These clusters indicate redundancy; retaining all variables can introduce multicollinearity and reduce model stability. Redundant variables can therefore be reduced by keeping a representative feature from each cluster or aggregated using dimensionality-reduction methods such as PCA to form composite indices. For instance, thrust and torque may be replaced by a single mechanical-load component, reducing feature dimensionality without sacrificing predictive power.

In contrast, low-correlation variables provide complementary information and are particularly useful for discriminating anomalies. Features such as Vertical deviation offset ref [mm], Tailshield clearance top right [mm], and Keystone position [A°] capture behaviors unrelated to cutterhead mechanics, including machine misalignment, overbreak or shield instability, and uneven backfilling or segment misplacement. Finally, the observed correlation patterns provide practical guidance for PCA: groups with shared variance can be compressed efficiently, while unrelated domains should not be mixed, as combining mechanically independent features (e.g., torque with keystone angle) would reduce interpretability. By compressing high-correlation groups and retaining orthogonal channels, the resulting model remains compact while preserving the most informative signals for monitoring and anomaly detection.

4.2.2 Time-series analysis

Five critical parameters of the EPB TBMs were examined through time-series plots to clarify excavation behaviors and to locate operational anomalies with greater spatial clarity. In this updated presentation, the excavation ring number is kept in its original form on the horizontal axis, while the parameter values are shown as Z-scores on the vertical axis. This approach preserves the true excavation sequence and makes it easier to identify the exact ring intervals at which abnormal behaviors occurred.

The selected parameters are Thrust Force, Cutterhead Torque, Ring Advance Time, Penetration per Revolution, and Total Grout Volume. Together, these variables describe machine resistance, excavation efficiency, cycle performance, and backfill response. Because only the vertical axis is standardized, each curve still expresses relative deviation from its own mean, while the raw ring number retains direct engineering meaning along the tunnel alignment. The dashed reference lines in the figures mark an absolute Z-score of 3, which highlights intervals of unusually high or unusually low response.

Figure 5 presents the thrust force response across the full excavation record. The initial rings show a pronounced negative departure, which is consistent with start-up instability and early system adjustment. After this opening stage, the thrust signal remains mostly near the central band, but several sharp low events appear around rings 1208 to 1209 and again around rings 2407 to 2409. In the later part of the alignment, especially from about ring 3300 onward, the signal becomes more energetic and reaches its highest positive levels near rings 4041 to 4050. This pattern suggests that the machine encountered intervals of higher resistance in the final portion of the drive, interrupted by a few short-duration loss events.

FIGURE 5

The next parameter employed in this analysis is TBM’s cutterhead torque. Figure 6 shows cutterhead torque against the same raw ring sequence. The strongest positive excursion occurs near ring 935, where torque rises far above the normal operating band. Additional positive clustering is visible near rings 3428 to 3430, and another high-activity zone appears around the 4000-ring range. A notable low event occurs near ring 3114, with earlier negative values also visible during the opening stage. When Figure 6 is read together with Figure 5, several intervals display concurrent increases in thrust and torque, which strengthens the interpretation of increased cutting resistance rather than random signal noise.

FIGURE 6

Figure 7 illustrates ring advance time, which reflects the total time required to complete one ring cycle. The first few rings contain extremely high values, indicating that the opening portion of the tunnel included substantial delay and adjustment. Outside this start-up zone, most rings lie close to the mean level, which shows that the cycle time was generally stable. However, strong delay spikes reappear near ring 3114 and again in the region around rings 4020 to 4060. These isolated peaks imply that prolonged cycle times were linked to localized operational interruptions, difficult ground, or short periods of reduced efficiency rather than to a persistent decline throughout the drive.

FIGURE 7

Low penetration with consistent or rising thrust and torque strongly indicates mechanical inefficiency or adverse geological zones. This pattern aligns with previous outlier and anomaly detection findings. Monitoring this parameter is essential for detecting cutter dullness or poor soil conditioning.

Figure 8 presents penetration per revolution. This parameter is especially important because it reflects how much forward movement is achieved for each cutterhead revolution. The record shows repeated downward excursions, with particularly severe reductions near ring 3114, in the interval around rings 4026 to 4050, and again near the final section around ring 4443. A strong positive peak appears near ring 3729. The repeated low-penetration episodes suggest short zones of poor cutting efficiency. Where penetration decreases while thrust or torque remains elevated, the most likely interpretation is that the machine required more effort to achieve less advance, which may indicate harder ground, poorer conditioning, or increasing cutter wear.

FIGURE 8

Finally, the fifth parameter utilized in this analysis is the Total Grout Volume [m³]. Grout volume measures the quantity of cementitious material injected behind the segmental lining. It serves to fill the annular gap and prevent ground subsidence.

Grout volumes were mostly within 5–6 m³, with several excursions well above 10 m³. One extreme spike exceeded 30 m³, suggesting an acute grouting event.

Spikes in grout volume may signify over-excavation, voids, or loose ground requiring excessive fill. Such events are correlated with collapse risk zones and may also point to improper ring installation or tail voids. Grout metrics are crucial for assessing the long-term structural and geotechnical stability of the tunnel.

Figure 9 shows the total grout volume response. The most striking feature is the extreme early spike near ring 11, which stands well above the rest of the record and indicates an exceptional grouting event during the opening stage. Apart from this initial response, the grout signal remains much more compact, although smaller positive excursions are visible at several points and a marked low event occurs around rings 2407 to 2409. The later portion of the alignment displays a slightly higher baseline than the early and middle sections, which may reflect a gradual change in annular fill demand or ground response. Overall, the revised plot shows that grout anomalies are highly localized and can now be tied to exact excavation intervals more clearly because the horizontal axis preserves the original ring numbering.

FIGURE 9

Figure 10 combines thrust, torque, and penetration in a single multivariate view and marks all outliers beyond the reference band. The opening rings from two to eight form the clearest common anomaly cluster, which is consistent with start-up effects. Further clustered behavior appears around rings 1208 to 1213, where thrust and torque both fall sharply, around rings 2407 to 2409, where another shared low event is visible, and near ring 3114, where torque and penetration both show strong negative departures. In the interval from rings 4041 to 4050, thrust reaches its highest positive values while penetration drops strongly, indicating a combination of high machine effort and weak cutting efficiency. The coincidence of these responses across multiple parameters supports the interpretation that the detected anomalies reflect real excavation events rather than isolated measurement noise.

FIGURE 10

The TBM time series analysis provided vital diagnostic insights into excavation dynamics, performance degradation, and ground-structure interaction. By combining operational metrics with anomaly detection frameworks, this methodology enabled targeted root cause analysis and supported predictive maintenance strategies. This direction is supported by recent research showing that TBM prediction models can be extended into residual-based risk early-warning frameworks for operational safety management (Zhang et al., 2026). It will further allow integration with AI-based anomaly detectors, such as SHAP or Isolation Forest, to automate and augment tunnel performance evaluation in future work.

4.2.3 SHAP-based feature importance

To gain interpretable insights into the underlying drivers of TBM thrust force, a machine learning model was trained using the XGBoost regression algorithm, a scalable implementation of gradient-boosted decision trees. XGBoost builds an ensemble of decision trees sequentially, with each successive tree optimized to correct the residuals of the previous one, enabling the model to capture nonlinear dependencies and complex interactions across the full set of operational and geometric input variables.

Before model training, all candidate features were systematically audited for two types of problematic dependency. The first type is circular dependency on the target variable. Three features contain thrust force the target within their mathematical formulations. Teale’s Specific Energy is defined as SE = (F·v + 2π·T·N)/(A·v), where F is thrust; the Penetration Index (I_p) is computed as I_p = F/p, where F is thrust; and the QTBM index incorporates thrust as a normalizing factor. Using any of these as a predictor of thrust is equivalent to predicting the target from an expression that already contains it. All three were removed. Their removal reduced the apparent test R² from 0.997 to 0.872 confirming that the original high performance was a mathematical artefact rather than a genuine predictive result.

The second type is internal redundancy, where a feature is mathematically derivable from other predictors already present in the model. Three features were identified through empirical verification. CH Advance Rate (mm/min) equals the product of CH Penetration per Revolution and CH Rotation Speed; the Pearson correlation between Advance Rate and this product was r = 1.000, with a mean absolute difference of 0.018 mm/min. Total Grout Volume equals the sum of Grout Volume Components A and B, verified to within 0.00002 m³. The summary Grout Pressure line equals the mean of the six individual port pressures L01–L06, with a mean absolute difference of 0.004 bar. Retaining both a derived summary and its components simultaneously splits SHAP importance weight between redundant representations of the same signal. All three summary features were removed. The six individual grout port pressures were retained in their place, because inter-port Pearson correlations ranged from r = 0.14 to r = 0.70, confirming that each port carries spatially independent information about grouting conditions at a distinct position around the tail shield annulus information that the mean discards. The final feature set comprised 28 independent operational and geometric parameters, none of which are mathematically dependent on the target variable or on each other.

Hyperparameters were selected through preliminary tuning: maximum tree depth of 6, learning rate of 0.05, and subsampling and column-sampling ratios of 0.8. The subsampling and column-sampling introduce controlled stochasticity into each boosting round, reducing overfitting by preventing any single tree from accessing the full dataset or feature set. Early stopping with a patience of 30 rounds was applied to the held-out validation set, halting training when validation performance ceased to improve. The dataset was split 80/20 into training and testing subsets using a fixed random seed (42) for reproducibility. In addition to RMSE and R², the a20-index was computed to quantify engineering reliability defined as the proportion of predictions for which the ratio of actual to predicted thrust falls within ±20% (Asteris et al., 2024). This metric has been applied in recent TBM performance-prediction studies as a practical complement to statistical accuracy measures (Yu et al., 2023). On the held-out test set, the model achieved RMSE = 381.8 kN, R² = 0.872, MAE = 290.1 kN, and an a20-index of 0.993, indicating that 99.3% of predicted thrust values fell within ±20% of measured values. Five-fold cross-validation yielded a mean R² of 0.312 (standard deviation 0.216). This fold-to-fold variance reflects the geological heterogeneity of the 7 km alignment rings originating from geologically distinct zones exhibit substantively different thrust regimes, and random folding can separate these zones from their training counterparts. This observation suggests zone-stratified and cross-project validation strategies discussed in the Future Work section.

Figure 11 presents the SHAP summary plot for the 28-feature model. Each point represents one ring’s SHAP value for a given feature. Horizontal position indicates the magnitude and direction of influence on predicted thrust, and color encodes the raw feature value for that ring (red = high, blue = low). Features are ranked top to bottom by mean absolute SHAP value.

FIGURE 11

To interpret the model’s predictions, SHAP (SHapley Additive exPlanations) was employed (Lundberg and Lee, 2017). SHAP assigns an importance value to each feature for each individual prediction by computing its marginal contribution across all possible feature orderings, decomposing the model output into feature-wise contributions. This quantifies how much each input parameter pushed the predicted thrust above or below the baseline expectation for a given ring, and in which direction.

The most influential predictors of thrust, in rank order, were: Cutterhead Torque (mean |SHAP| = 341.8) the dominant predictor by a clear margin, with high torque values consistently associated with large positive SHAP contributions, confirming the strong mechanical coupling independently established in the Pearson correlation analysis (r ≈ 0.85); Tunnelmeter (mean |SHAP| = 267.2) encodes position-dependent variability along the 7 km alignment, serving as a spatial proxy for unmeasured geological heterogeneity; Cutterhead Penetration per Revolution (mean |SHAP| = 165.6) a direct measure of cutting efficiency, where low penetration at stable or rising torque corresponds to harder ground and consequently higher thrust demand; Ring Advance Time (mean |SHAP| = 76.6) longer advance times associated with difficult conditions; Earth Pressure Detection #3 and #1 (mean |SHAP| = 72.6 and 70.7) face-pressure sensors confirming that face-pressure regulation is a genuine driver of thrust in EPB mode; Articulation offsets (Articulation Point 1 vertical, Articulation Point 2 vertical and horizontal; mean |SHAP| = 63.6, 44.3, and 32.5) shield-alignment contributions to thrust demand; and Individual Grout Pressure ports L02, L01, L03, L04 (mean |SHAP| = 38.7, 24.5, 22.8, and 22.8) revealing spatial variation in grouting behavior around the tail shield.

Consistent with the findings of the original analysis, no single feature dominates the prediction to the exclusion of all others. Thrust force emerges from the combined contributions of mechanical loading (torque), geological context encoded in alignment position, cutting kinematics (penetration), operational pacing (advance time), face-pressure regulation (earth pressure sensors), shield alignment (articulation offsets), and spatially distributed grouting operations (individual port pressures and volume components). This distributed importance pattern reinforces the core argument of this study: that EPB TBM systems must be monitored and interpreted as integrated mechanical-geotechnical systems, and that single-variable threshold approaches are insufficient to capture the full complexity of thrust behaviors.

4.2.4 Detection of mechanically uncorrelated behavior

To complement SHAP-based interpretation, a scatter plot was constructed (Figure 12) comparing normalized thrust force against penetration per revolution. In the context of EPB TBM operation, a strong mechanical coupling is generally expected between axial thrust force and the penetration rate per cutterhead revolution. As thrust increases, the cutterhead should correspondingly advance further into the ground, reflecting efficient cutting and effective transfer of energy into forward progress. To evaluate adherence to this expected relationship, the dataset was statistically partitioned using percentile-based thresholds. Specifically, the top 10th percentile of normalized thrust values was extracted to represent extreme-force conditions, while the bottom 10th percentile of penetration values was selected to represent minimal advance. The intersection of these two sets, data points where the TBM exerted very high thrust but achieved very low penetration shown as red dots in the below figure, was identified as a critical zone of mechanically uncorrelated behavior. This condition indicated a decoupling between input force and excavation response, which is physically implausible under normal operating conditions.

FIGURE 12

This condition is physically implausible under normal operation, as high thrust is expected to result in proportional penetration. The presence of significant red clusters suggests conditions such as blunt or worn cutter tools, cutterhead jamming or clogging, excavation against highly resistant strata (e.g., boulders, cobbles), or inadequate soil conditioning.

These zones are critical from an operational standpoint, as they represent inefficiency and elevated risk of equipment wear or failure.

5 Results and observations

The Dual-Metric Correlation Analysis, which combined Pearson and Spearman correlation coefficients, revealed clear mechanical and operational interdependencies among key TBM parameters. The strongest correlation was observed between cutterhead torque and thrust force, with a Pearson coefficient of approximately +0.85. As documented by Devarajan et al. (2022) in metro TBM operations, harder ground conditions simultaneously demand higher torque and thrust values for steady advance which is a relationship confirmed in this research by the consistent co-elevation of both variables during high-resistance excavation episodes. This agreement between observed data patterns and established mechanical behavior validated the internal coherence of the dataset.

Additional strong correlations were recorded between advance rate, penetration per revolution, and specific energy, reinforcing their coupled role in reflecting excavation efficiency and energy consumption dynamics. Likewise, the variables ring number, station, and tunnel-meter exhibited near-perfect correlation (r ≈ +0.99), validating their interchangeable use as linear indicators of tunneling progress.

In contrast, parameters such as grout volume, alignment offsets, and segment orientation angles showed only weak correlations with mechanical variables, confirming that these describe orthogonal aspects of TBM operation including segment placement accuracy, annular gap management, and backfilling efficiency rather than direct excavation mechanics. The retention of these weakly correlated variables was methodologically important because, as (Rodríguez et al., 2024) demonstrated in multi-sensor fault detection, orthogonal features consistently capture independent fault modes that mechanically dominant variables alone cannot detect, thereby broadening the diagnostic coverage of the monitoring framework.

The visualization of ring-wise time-series data exposed recurring excavation cycles and localized anomalies within the 4,400-ring dataset. The sequences of thrust, torque, and penetration rate displayed characteristic oscillations corresponding to cut–install–reset phases typical of EPB TBM operation. Beyond these regular cycles, sharp and statistically significant deviations were also identified which were isolated using a ±3 Z-score threshold, a robust outlier-screening criterion validated by (Guo et al., 2025) for high-noise industrial sensor datasets. These observations confirmed the presence of genuine performance anomalies superimposed on otherwise stable excavation patterns.

Several notable outlier patterns emerged. Rings showing high thrust, but abnormally low penetration suggested inefficient cutting, likely arising from cutter wear, material adhesion on the cutterhead, or locally increased ground strength. These conditions were validated as mechanically uncorrelated anomalies in both scatter-plot analyses and SHAP-derived influence distributions.

Furthermore, the grout-volume series exhibited episodic spikes reaching 2 to 3 times the baseline average, signaling potential over-excavation, cavity formation, or soil-loss events. These anomalies typically occurred without corresponding changes in torque or penetration, implying geotechnical disruptions such as soft-pocket intrusion or localized inflow rather than mechanical malfunction.

Overall, the temporal inspection and ring-based outlier screening delineated both mechanical and geotechnical irregularities inherent in real-world TBM operation. Such findings substantiated the framework’s ability to isolate abnormal behavior within otherwise regular excavation sequences, as previously highlighted in EPB-specific analytics research (Xie et al., 2024). Representative examples of ring-indexed time-series patterns and anomaly markers are displayed in Figures 5–10.

The feature importance (SHAP) helped to extend the interpretability beyond descriptive analysis. The XGBoost regression model was trained to estimate thrust force from selected mechanical, geotechnical, and alignment parameters. The model’s predictions were interpreted using SHapley Additive exPlanations (SHAP) to quantify the relative influence of each feature Lundberg and Lee (2017). The resulting SHAP summary plot (Figure 11) ranked features by their mean absolute contribution to predicted thrust.

After excluding all thrust-derived and internally redundant features (Specific Energy, I_p, QTBM, CH Advance Rate, Total Grout Volume, and summary Grout Pressure), the analysis identified Cutterhead Torque (mean |SHAP| = 341.8), Tunnelmeter (267.2), and Cutterhead Penetration per Revolution (165.6) as the dominant independent variables governing thrust demand. Ring Advance Time, Earth Pressure Detection (#3 and #1), Articulation offsets, and individual Grout Pressure ports (L01–L04) followed as secondary contributors. Together, these factors account for the majority of model variance, linking mechanical effort directly to torque coupling, geological context along the alignment, cutting efficiency, and face-pressure regulation.

Interpretation of SHAP magnitudes aligned closely with engineering intuition. Elevated cutterhead torque corresponded proportionally to higher thrust demand, reflecting the physical coupling between axial and rotational loads during excavation an observation consistent with the Pearson correlation result (r ≈ 0.85) and now independently confirmed through the interpretable model. Variations along the tunnelmeter captured position-dependent changes in geology and operational regime, while lower penetration per revolution corresponded to reduced cutting efficiency and consequently increased thrust. Face pressure and grout pressure contributed additional coupling through chamber-stability and backfill-control mechanisms, confirming expected control relationships in EPB mode.

Five-fold cross-validation yielded a mean R² of 0.312 (±0.216). This fold-to-fold variance reflects geological heterogeneity across the 7 km alignment rather than model instability random folding can separate geologically distinct ring zones from their training counterparts. This is reported transparently as a generalization indicator and motivates geology-stratified validation in future work. The a20-index of 0.993 confirms that 99.3% of predicted thrust values fell within ±20% of measured values, demonstrating strong practical reliability.

6 Discussion of results

The strong linear association between cutterhead torque and thrust force (Pearson r ≈ 0.85, Spearman ρ ≈ 0.89) validated the mechanical principle that increased resistance during excavation demanded elevated axial and rotational forces. This trend appeared consistently in the dataset, where both parameters rose in unison across denser ground layers, corroborating earlier findings on TBM behavior Devarajan et al. (2022). Such correspondence also reinforced the interpretability of the SHAP-based model outputs, confirming that torque and thrust reflected the same underlying physical dynamics.

This collinearity implied operational redundancy; either torque or thrust could serve as a surrogate signal for real-time feedback and control systems, minimizing telemetry complexity without compromising monitoring integrity. Meanwhile, features such as grout volume, tail-shield offsets, and articulation deviations displayed weak or negligible correlations with cutting forces, indicating their orthogonal diagnostic relevance. These attributes were retained for their capacity to capture independent fault modes such as grouting irregularities or alignment drift (Benhanifia et al., 2025).

6.1 Correlation insights reinforce TBM mechanics

The torque–thrust relationship confirmed a well-established mechanical interdependence within Earth-Pressure-Balance (EPB) TBM systems. The Pearson correlation of r ≈ 0.85 aligned with theoretical expectations that both axial and rotational loads must rise proportionally to overcome higher cutting resistance. This coupling is consistently reported in TBM performance studies where ground hardness simultaneously elevates both force components (Zhang et al., 2024). This correlation suggested a potential simplification for data acquisition strategies, where either signal could effectively represent mechanical effort, reducing telemetry complexity without compromising monitoring integrity. Conversely, as (Benhanifia et al., 2025) demonstrated in multi-sensor industrial fault detection, weakly correlated parameters such as grout volume and alignment offset capture independent fault modes that mechanically dominant variables cannot represent, validating their complementary diagnostic purpose within the anomaly detection framework.

6.2 Temporal rhythms and seasonality-aware modeling

Ring-wise time-series analysis revealed clear cyclical behavior in thrust, torque, and penetration, corresponding to the cut–install–reset rhythm intrinsic to EPB TBM operation. These cycles reflected natural operational phases driven by excavation, segment installation, and system resets. In addition to cyclicity, local patterns emerged where groups of 10–20 rings exhibited consistent performance drift—likely due to crew shifts, tool reconditioning, or changes in soil strata.

However, the presence of strong seasonality complicated static anomaly detection. Outliers flagged by a ±3 Z-score often coincided with predictable operational pauses or low-load intervals rather than true mechanical anomalies, highlighting the limitation of fixed global thresholds in cyclic tunneling environments. This underscored the need for adaptive, context-sensitive baselines that account for the inherent periodicity of EPB excavation cycles. To improve the reliability of the framework, future iterations should incorporate seasonality-aware methods such as STL (Seasonal-Trend decomposition). This method was identified by (Xiao et al., 2025) as effective for managing temporal dependencies in cyclic industrial data. Integrating these with rolling baseline computations and ring-based autoregressive filtering will help distinguish genuine mechanical faults from routine operational rhythms, thereby reducing false-positive anomaly alerts.

Furthermore, as (Fu et al., 2022) established in ring-based TBM monitoring, high-thrust and low-penetration combinations represent physical boundary conditions that require temporal context for accurate interpretation in machine-learning-based systems. These edge cases provided valuable benchmark instances for supervised model calibration within the present framework.

6.3 SHAP-based interpretability enhances operational trust

The application of SHapley Additive exPlanations (SHAP) introduced transparency to model interpretability, bridging data-driven outputs with engineering intuition (Lundberg and Lee, 2017). After removing all features mathematically derived from thrust or from other predictors, the analysis ranked Cutterhead Torque, Tunnelmeter, and CH Penetration per Revolution as the three most influential features contributing to thrust prediction. Higher torque directly reflects greater rotational resistance, physically coupled to axial thrust demand a result that independently corroborates the Pearson correlation finding (r ≈ 0.85) and unifies the statistical and ML views of the dataset. Tunnelmeter captures geological transitions along the alignment as a spatial proxy, while lower penetration implies reduced cutting efficiency and elevated thrust. Secondary contributors Earth Pressure Detection (both sensors), individual Grout Pressure ports, and Articulation offsets captured face-pressure regulation, spatial grouting variation, and alignment-induced loading respectively. Together, these relationships affirmed the validity of the SHAP framework in interpreting model decisions within an engineering context, now on a methodologically rigorous, non-circular basis.

By translating model predictions into domain-specific language such as, ‘Thrust surge predicted due to elevated specific energy and low penetration efficiency’ the system offered interpretable diagnostics that could enhance operator trust. As Devarajan et al. (2022) demonstrated through decision-tree-based TBM monitoring, model explainability directly determines whether field operators adopt or dismiss data-driven recommendations, making transparency a prerequisite rather than an optional feature.

6.4 Nonlinear dependencies and dimensional reduction

While correlation analysis effectively captured linear relationships, it was inherently limited in identifying nonlinear dependencies, threshold effects, or saturation phenomena typical of mechanical systems.

Similarly, Principal Component Analysis (PCA) would enable clustering of operational states such as normal cutting, over-excavation, and grouting anomaly. The theoretical basis for such dimensionality reduction in high-variance datasets is well established (Hastie et al., 2009), supporting the development of compact and computationally efficient models built upon the EDA foundation developed here.

6.5 Overfitting and generalization considerations

Several precautions were applied to limit overfitting in the XGBoost thrust model. Early stopping with 30 rounds on the held-out validation set was applied during training; tree depth was capped at 6; subsampling and column-subsampling ratios of 0.8 were used to inject stochasticity into each boosting round; and the learning rate was kept moderate (0.05). In addition to the 80/20 held-out evaluation, five-fold cross-validation was performed. The mean cross-validation R² of 0.312 (±0.216) is markedly lower than the held-out test R² of 0.872 and reveals that performance is sensitive to fold composition. This sensitivity arises from the geological structure of the 7 km dataset: adjacent rings within a single stratum are mutually similar, and random folding can place all rings from an unusual zone in the validation set, depressing the score. This is a known limitation of applying random cross-validation to sequential geotechnical data. It motivates the use of zone-stratified or project-aware validation schemes in subsequent work, including cross-project validation on additional EPB datasets once access permits.

6.6 Limitations and mitigation strategies

Collectively, the EDA framework provided a comprehensive depiction of TBM operations, linking physics-based relationships to statistically validated anomaly signals. The combination of correlation mapping, time-series analysis, and SHAP interpretability yielded a structured, interpretable, and operationally relevant understanding of tunneling behavior. These findings established a robust foundation for data-driven predictive maintenance and real-time fault detection in mechanized tunneling, aligning with the broader goal of improving safety, reducing downtime, and enabling intelligent decision support in underground construction projects.

Explainable machine-learning models can be directly integrated with optimization algorithms to enable real-time adjustment of shield tunneling parameters. This will help in extending the role of predictive analytics from passive monitoring toward active operational control, a direction this framework is positioned to support.

6.7 Implications for generative AI in construction engineering

The structured EDA architecture developed in this study provides a basis for new generative AI applications in the Architecture, Construction, and Engineering (ACE) sector, going beyond predictive analytics. Using TBM sensor data, generative models like variational autoencoders, generative adversarial networks, and large language models could develop synthetic operational datasets for uncommon anomaly conditions, simulate excavation scenarios, and support engineers with AI-driven decision-support systems. Future systems may enable real-time scenario exploration and adaptive tunneling techniques by converting previous tunneling data into generative knowledge models. This would be a significant step toward intelligent and autonomous underground construction.

7 Future work

Building upon the systematically completed Exploratory Data Analysis (EDA) modules, this research established a strong foundation for data-driven anomaly detection in Earth-Pressure-Balance (EPB) Tunnel Boring Machine operations. The study successfully combined correlation mapping, time-series anomaly exploration, and SHAP-based interpretability to reveal the underlying mechanical and geotechnical dependencies within TBM systems. The results confirmed that torque–thrust coupling, cyclical excavation patterns, and feature importance hierarchies can be captured through structured EDA, offering actionable insights into excavation performance and machine health.

The next phase of this research will extend these exploratory insights into comprehensive statistical profiling and predictive modeling. Specifically, four major directions are envisioned:

7.1 Descriptive statistics and distributional profiling

A detailed statistical assessment will be conducted for all TBM sensor channels. Key descriptive measures such as mean, median, variance, skewness, kurtosis, and percentiles will quantify central tendency and dispersion. Visual exploration through histograms and multi-facet boxplots will characterize distribution shapes, exposing skewed or multimodal behaviors (e.g., zero-load conveyor phases). Correlating these statistical moments with operational logs will enable differentiation between genuine process variability and data anomalies.

7.2 Low-dimensional visualization and PCA integration

Principal Component Analysis (PCA) will be applied to capture latent operational modes for high-dimensional industrial datasets. Reconstruction-error analysis, which validated as an effective unsupervised anomaly flag, will detect deviations invisible to linear correlation models, exposing outlier clouds and transitional operational states within the ring-based sensor space.

7.3 Anomaly detection model development

To operationalize insights from the EDA stage, a suite of machine-learning algorithms will be deployed, including Isolation Forest, One-Class SVM, and Autoencoder Neural Networks. Training will focus on normal operational segments while excluding known maintenance or fault intervals. Features will be selected based on mutual information rankings, PCA-derived components, and the top SHAP-contributing variables including torque, cutterhead speed, and face pressure. As (Fu et al., 2022) established in ring-based TBM anomaly frameworks, semi-supervised hyperparameter tuning using catalogued fault intervals is critical for balancing precision (minimizing false positives) and recall (maximizing early fault detection) in tunneling anomaly detection systems. In addition (Zhang et al., 2026), recently demonstrated that residual-based warning thresholds and risk grading strategies can be successfully integrated into TBM advance-rate early warning systems, a direction that future iterations of this framework will incorporate.

7.4 Validation, deployment, and broader data integration

Validated models will be compared against operator and maintenance records using standard performance metrics such as precision, recall, and time-to-warning.

Future work will also focus on broadening data horizons through multi-machine and multi-site validation. To address this, incorporating geological inputs such as borehole logs, slurry conditioning parameters, and ground classifications will enrich the feature space, enabling hybrid models that jointly reflect mechanical and subsurface states an integration that (Zhang et al., 2024) demonstrated significantly improves predictive stability across heterogeneous tunneling environments.

7.5 Concluding remarks

In summary, this study demonstrated that a structured EDA framework can bridge engineering intuition with machine-learning interpretability, enhancing the understanding of TBM dynamics and anomaly behavior. The planned extensions incorporating descriptive statistics, PCA, and advanced predictive models will transform these analytical insights into a scalable, real-time anomaly detection system. This integrated approach has the potential to significantly reduce unplanned downtime, improve safety, and contribute to the digital transformation of mechanized tunneling operations.

8 Conclusion

This study established a structured, ring-based Exploratory Data Analysis framework for Earth-Pressure-Balance Tunnel Boring Machine operations, integrating statistical preprocessing, correlation mapping, temporal anomaly detection, and SHAP-based interpretability into a single reproducible pipeline. Applied to more than 4,400 rings from a 7 km EPB TBM project, the framework identified the strong torque–thrust coupling (r ≈ 0.85) as the central mechanical relationship of the dataset, and revealed both mechanical inefficiencies (high thrust with low penetration) and geotechnical disturbances (grout-volume spikes) that conventional threshold-based monitoring would miss.

The XGBoost model with SHAP interpretability constructed after systematic removal of all thrust-derived and internally redundant features achieved R² = 0.872, RMSE = 381.8 kN, and an a20-index of 0.993, and identified Cutterhead Torque, Tunnelmeter (alignment position), and Penetration per Revolution as the principal independent predictors of thrust. By positioning exploratory interpretability as a prerequisite rather than an afterthought, this framework provides a reproducible diagnostic foundation for intelligent TBM monitoring and a methodological basis for subsequent predictive-maintenance and real-time fault-detection applications in mechanized tunneling.

References

• 1

  ApojiD. (2023). Developing AI Systems for EPB TBM Utilizing Sensing Data and Machine Learning. University of California, Berkeley.

  • Google Scholar
  • View reference in article

• 2

  AsterisP. G.TsavdaridisK. D.LemonisM. E.FerreiraF. P. V.FormisanoA.GantesC.et al (2024). AI-powered GUI for prediction of axial compression capacity in concrete-filled steel tube columns. Neural Comput. Appl.36 (35), 22429–22459. 10.1007/s00521-024-10405-w

  • CrossRef
  • Google Scholar
  • View reference in article

• 3

  BaiL.MoD.LiH.HuangW.CaiZ. (2025). Optimized data preprocessing and model selection for TBM cutter wear prediction. Coatings15 (5), 564. 10.3390/coatings15050564

  • CrossRef
  • Google Scholar
  • View reference in article

• 4

  BenhanifiaA.Ben CheikhZ.OliveiraP. M.ValenteA.LimaJ. (2025). Systematic review of predictive maintenance practices in the manufacturing sector. Intelligent Syst. Appl.26, 200501. 10.1016/j.iswa.2025.200501

  • CrossRef
  • Google Scholar
  • View reference in article

• 5

  DevarajanM. V.SacramentoC. S.SambasA. (2022). Data-driven techniques for real-time safety management in tunnel engineering using TBM data. Int. J. Eng. Tech. Mgmt. Res.7 (3), 1–15.

  • Google Scholar
  • View reference in article

• 6

  DongZ. K.TanZ. S.LiX.ZhouZ. L.TianG. S. (2025). Analysis and optimization of key factors in self-supervised lithology recognition from TBM muck images. Rock Mech. Rock Eng., 1–13. 10.1007/s00603-024-04259-w

  • CrossRef
  • Google Scholar
  • View reference in article

• 7

  FlorA.SassiF.La MorgiaM.CerneraF.AmadiniF.MeiA.et al (2023). Artificial intelligence for tunnel boring machine penetration rate prediction. Tunn. Undergr. Space Technol.140, 105249. 10.1016/j.tust.2023.105249

  • CrossRef
  • Google Scholar
  • View reference in article

• 8

  FuT.ZhangT.SongX. (2022). A novel hybrid transfer learning framework for dynamic cutterhead torque prediction of the tunnel boring machine. Energies15 (8), 2907. 10.3390/en15082907

  • CrossRef
  • Google Scholar
  • View reference in article

• 9

  GuoQ.XiaoH.HeR.LiuL.YangH.YuH. (2025). Cutting-edge review of big data preprocessing in TBM tunnel construction. Shock Vib.2025, 8866384. 10.1155/vib/8866384

  • CrossRef
  • Google Scholar
  • View reference in article

• 10

  HarandizadehH.ArmaghaniD. J.AsterisP. G.GandomiA. H. (2021). TBM performance prediction developing a hybrid ANFIS-PNN predictive model optimized by imperialism competitive algorithm. Neural Comput. Appl.33, 16149–16179. 10.1007/s00521-021-06217-x

  • CrossRef
  • Google Scholar
  • View reference in article

• 11

  HastieT.TibshiraniR.FriedmanJ. (2009). The Elements of Statistical Learning: Data Mining, Inference, and Prediction. 2nd ed.Springer. 10.1007/978-0-387-84858-7

  • CrossRef
  • Google Scholar
  • View reference in article

• 12

  HuW.WuK.LiuH.LuoW.LiX.GuanP. (2025). Interpretable machine learning approach for TBM tunnel crown convergence prediction with Bayesian optimization. Front. Earth Sci.13, 1608468. 10.3389/feart.2025.1608468

  • CrossRef
  • Google Scholar
  • View reference in article

• 13

  HuY.AttaZ.RahmanT. U.QiuS.WangJ.WeiW.et al (2026). Shedding light on explainable AI: insights, challenges, and the future of infrastructure management. ISPRS Int. J. Geo-Information15 (3), 100. 10.3390/ijgi15030100

  • CrossRef
  • Google Scholar
  • View reference in article

• 14

  KarahanH.AlkayaD. (2026). Ensemble machine learning for predicting TBM penetration rate with limited geotechnical data. Appl. Sci.16 (7), 3451. 10.3390/app16073451

  • CrossRef
  • Google Scholar
  • View reference in article

• 15

  KuhnM.JohnsonK. (2019). Feature Engineering and Selection: A Practical Approach for Predictive Models. Boca Raton, FL, United States: CRC Press. 10.1201/9781315108230

  • CrossRef
  • Google Scholar
  • View reference in article

• 16

  KwakS. K.KimJ. H. (2017). Statistical data preparation: management of missing values and outliers. Korean J. Anesthesiol.70 (4), 407–411. 10.4097/kjae.2017.70.4.407

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 17

  LeiK.ZhouZ.LiH.TanZ.GongL. (2026). A novel time series classification model with transformer for seamless surrounding rock perception in tunnel boring machine tunneling. Eng. Appl. Artif. Intell.166, 113628. 10.1016/j.engappai.2026.113628

  • CrossRef
  • Google Scholar
  • View reference in article

• 18

  LiZ.Yazdani BejarbanehB.AsterisP. G.KoopialipoorM.ArmaghaniD. J.TahirM. M. (2021). A hybrid GEP and WOA approach to estimate the optimal penetration rate of TBM in granitic rock mass. Soft Comput.25, 11877–11895. 10.1007/s00500-021-06005-8

  • CrossRef
  • Google Scholar
  • View reference in article

• 19

  LiS.ZhangQ.LiuS.MaM. (2024). Robust prediction of thrust for tunnel boring machines with adaptive heavy-tailed error distribution. Adv. Eng. Inf.62, 102619. 10.1016/j.aei.2024.102619

  • CrossRef
  • Google Scholar
  • View reference in article

• 20

  LundbergS. M.LeeS.-I. (2017). A unified approach to interpreting model predictions. Adv. Neural Inf. Process. Syst.30, 4765–4774. 10.5555/3295222.3295230

  • CrossRef
  • Google Scholar
  • View reference in article

• 21

  RodríguezR.BascomptaM.GarcíaH. (2024). Carbon footprint evaluation in tunnels excavated in rock using tunnel boring machine (TBM). Int. J. Civ. Eng.22, 995–1009. 10.1007/s40999-023-00935-0

  • CrossRef
  • Google Scholar
  • View reference in article

• 22

  RousseeuwP. J.LeroyA. M. (2005). Robust Regression and Outlier Detection. Wiley. 10.1002/0471725382

  • CrossRef
  • Google Scholar
  • View reference in article

• 23

  ShanF.HeX.ArmaghaniD. J.ZhangP.ShengD. (2022). Success and challenges in predicting TBM penetration rate using recurrent neural networks. Tunn. Undergr. Space Technol.130, 104728. 10.1016/j.tust.2022.104728

  • CrossRef
  • Google Scholar
  • View reference in article

• 24

  SunW.ShiM.ZhangC.ZhaoJ.SongX. (2018). Dynamic load prediction of tunnel boring machine (TBM) based on heterogeneous in-situ data. Automation Constr.92, 23–34. 10.1016/j.autcon.2018.03.030

  • CrossRef
  • Google Scholar
  • View reference in article

• 25

  WuY.MaC.TanX.YangD.TianH.YangJ. (2019). A new evaluation method for the uniaxial compressive strength ahead of the tunnel face based on the driving data and specification parameters of TBM. Shock Vib.2019, 5309480. 10.1155/2019/5309480

  • CrossRef
  • Google Scholar
  • View reference in article

• 26

  XiaoH.CaoR.ChenZ.HongC.WangJ.YaoM.et al (2025). Handling missing data in large-scale TBM datasets: methods, strategies, and applications. Intell. Geoengin.2 (3), 109–125. 10.1016/j.ige.2025.07.001

  • CrossRef
  • Google Scholar
  • View reference in article

• 27

  XieP.ChenK.YinZ.ZhuY.LuoH.ZhangQ.-B. (2024). A BIM-Based multi-model framework for advancing TBM performance—Part 1: real-Time prediction of thrust force. Tunn. Undergr. Space Technol.151, 105856. 10.1016/j.tust.2024.105856

  • CrossRef
  • Google Scholar
  • View reference in article

• 28

  YuZ.LiC.ZhouJ. (2023). Tunnel boring machine performance prediction using supervised learning method and swarm intelligence algorithm. Mathematics11 (20), 4237. 10.3390/math11204237

  • CrossRef
  • Google Scholar
  • View reference in article

• 29

  ZhangX. P.XieW. Q.LiuQ. S.YangX. M.TangS. H.WuJ. (2021). Development and application of an in-situ indentation testing system for the prediction of tunnel boring machine performance. Int. J. Rock Mech. Min. Sci.147, 104899. 10.1016/j.ijrmms.2021.104899

  • CrossRef
  • Google Scholar
  • View reference in article

• 30

  ZhangY.ZhouJ.QiuY.ArmaghaniD. J.XieQ.YangP.et al (2024). A visual survey of tunnel boring machine (TBM) performance in tunneling excavation: mainstream direction, brief review and future prospects. Appl. Sci.14 (11), 4512. 10.3390/app14114512

  • CrossRef
  • Google Scholar
  • View reference in article

• 31

  ZhangQ.NieY.XueY.DuL.ZhaoX.CaoX.et al (2026). Automated prediction and risk early warning of TBM advance rate. Adv. Eng. Inf.69, 104028. 10.1016/j.aei.2025.104028

  • CrossRef
  • Google Scholar
  • View reference in article

• 32

  ZhouZ.LiZ.TanZ.LeiK.ZhangL. (2025). Optimized decision-making for tunnel boring machine control parameters. Rock Mech. Rock Eng.58, 5711–5731. 10.1007/s00603-025-04200-9

  • CrossRef
  • Google Scholar
  • View reference in article

• 33

  ZhouZ.WangM.YueY.LiuY.LiH. (2026). TBM load time series prediction based on causally interpretable artificial intelligence. Rock Mech. Rock Eng.1–22. 10.1007/s00603-026-05352-y

  • CrossRef
  • Google Scholar
  • View reference in article

• 34

  ZimekA.SchubertE.KriegelH.-P. (2012). A survey on unsupervised outlier detection in high-dimensional numerical data. Stat. Analysis Data Min.5 (5), 363–387. 10.1002/sam.11161

  • CrossRef
  • Google Scholar
  • View reference in article

• 35

  ZouN.ChengH.-Z.DaiT.YangZ.-H. (2026). Real-time surface settlement prediction in shield tunneling using a time-series framework and interpretable machine learning. Results Eng.29, 109478. 10.1016/j.rineng.2026.109478

  • CrossRef
  • Google Scholar
  • View reference in article