Prediction Model for Cutterhead Rotation Speed Based on Dimensional Analysis and Elastic Net Regression

Abstract

The development and maturation of TBM (tunnel boring machine) technology have significantly improved the accuracy and richness of excavation data, driving advancements in intelligent tunneling research. However, challenges remain in managing data noise and parameter coupling, limiting the interpretability of traditional machine learning models regarding TBM parameter relationships. This study proposes a cutterhead rotation speed prediction model based on dimensional analysis. By utilizing boxplot methods and low-pass filtering techniques, excavation data were preprocessed to select appropriate operational and mechanical parameters. A dimensionless model was established and integrated with elastic net regression to quantify parameters. Using TBM cluster data from a water diversion tunnel project in Xinjiang, the accuracy and generalizability of the model were validated. Results indicate that the proposed model achieves high prediction accuracy, effectively capturing trends in cutterhead rotation speed while demonstrating strong generalizability.

1. Introduction

As the development of TBM (tunnel boring machine) technology and the continuous maturation of its core techniques, the application of TBMs in tunnel construction has become increasingly widespread. This has led to a substantial increase in the amount of excavation data collected, with a marked improvement in data accuracy. Effectively leveraging this data to propel TBM technology towards greater automation and intelligence has become a frontier topic in tunneling technology. However, the raw data generated during TBM operation is often highly complex and coupled, posing significant challenges in extracting valuable information while minimizing interference—key hurdles in achieving intelligent tunneling [1,2,3].

To address these challenges, researchers globally have extensively explored TBM data processing and model analysis. Data preprocessing is critical in TBM data analysis because raw data often contains noise and irrelevant information. For instance, Zhang et al. [4] employed the BIRCH (balanced iterative reducing and clustering using hierarchies) algorithm to efficiently compress massive TBM operation data, removing data points where torque, thrust, cutterhead rotation, or penetration equaled zero. Zhao et al. [5] applied K-nearest neighbors (KNN) for outlier detection and combined it with principal component analysis (PCA) for dimensionality reduction, thereby mitigating multicollinearity among features. Rispoli et al. [6] proposed guidelines for efficient data filtering tailored to TBM operations in hard rock, exploring various analytical and modeling techniques, including descriptive statistics, regression analysis, and neural network fitting. Erharter and Marcher [7] utilized methods including standstill data removal, data grouping with mean calculations, outlier elimination, and uniform data point spacing. Li et al. [8] used PCA to identify the parameters most correlated with redundant data, while Wang et al. [9] employed wavelet packet transform to denoise signals, retaining their temporal characteristics. Xu et al. [10] utilized smoothing methods and outlier detection to stabilize four key TBM indicators during the ramp-up phase of complete operational segments. Xiao et al. [11] introduced an automated data processing approach that accurately segmented excavation cycles by analyzing parameters such as spindle rotation speed, advance rate, and downtime. Shan et al. [12] effectively removed noise in time-series data through smoothing techniques, improving data clarity and predictability. Their study indicated that achieving a balance between noise reduction and preserving authentic data characteristics is critical, as excessive smoothing may obscure significant details, while insufficient smoothing could retain excessive noise, reducing model performance. Because label noise in geological data obtained from field investigations presents challenges for accurate predictions. Zhang et al. [13] proposed a confidence learning-based support vector machine (CL-SVM) model to eliminate label noise, enhancing the reliability of geological classification. Despite significant progress in preprocessing methods to improve model performance, limitations persist when addressing complex datasets.

In the realm of model analysis, machine learning, empirical models, and traditional theoretical models are prominent research directions [14,15,16]. In machine learning, Xu et al. [17] applied supervised methods such as KNN, chi-squared automatic interaction detection (CHAID), support vector machines (SVM), classification and regression trees (CART), and neural networks (NN) to predict TBM penetration rate (PR). Their findings identified uniaxial compressive strength as the most critical factor and cutterhead rotation speed as the least significant. Feng et al. [18] employed deep belief networks (DBN) to predict PR, cutterhead speed, torque, and thrust, introducing the field penetration index (FPI) to quantify TBM performance. Xu et al. [10] evaluated statistical and ensemble machine learning methods, including Bayesian ridge regression (BR), nearest neighbor regression, random forests, gradient tree boosting (GTB), and deep neural networks like convolutional neural networks (CNN) and long short-term memory (LSTM). GTB demonstrated the highest prediction accuracy, while BR achieved the shortest computation time. Studies revealed that ensemble methods excel in small datasets, whereas deep learning models show potential when ample data are available. Mahmoodzadeh et al. [19] explored Gaussian process regression (GPR), SVR, decision trees (DT), and KNN for predicting TBM disc cutter life. Yang et al. [20] used evolutionary polynomial regression (EPR) and random forests (RF) to develop models for predicting PR and FPI. Kwon et al. [21] proposed an interpretable machine learning model with data augmentation to predict abnormal disc cutter wear, considering mixed ground conditions. Shin et al. [22] developed a model for predicting cutter wear based on cutter travel distance and CHI intervals. Despite the advancements, machine learning’s “black box” nature limits interpretability. In contrast, empirical models and traditional theoretical models primarily rely on existing empirical data and fundamental theoretical frameworks to explore the dynamic relationships among excavation parameters through data mining and multivariate regression analysis. For instance, Kim et al. [23] developed an excavation efficiency model based on linear cutting tests, unveiling physical relationships among certain parameters. However, these models exhibit limitations when dealing with small datasets and addressing nonlinear relationships, which hinders their ability to accurately describe and predict real-world problems.

To address these challenges, recent studies have sought to integrate physical knowledge with machine learning, among which dimensional analysis has emerged as a widely used approach [24,25]. In the field of TBM tunneling, dimensional analysis has mainly been applied for data preprocessing, while its direct application in model development remains relatively unexplored. Effectively integrating dimensional analysis with TBM excavation data is thus a key research challenge.

TBM excavation involves numerous parameters, but for hard rock TBMs, cutterhead speed is one of the key operational parameters directly specified by operators. It determines the rotational speed of the cutterhead during rock excavation and significantly impacts tunneling costs, construction planning, and risk assessment [26]. Since TBM performance is highly sensitive to geological conditions and rock properties [27], matching the cutterhead speed to the geological conditions is critical. Mismatched speeds can cause severe wear on disc cutters during hard rock excavation, significantly increasing construction costs and reducing efficiency. For example, studies have shown that excessively high cutterhead speeds dramatically accelerate cutter wear, while appropriately increasing the penetration rate can effectively reduce wear rates [28,29].

Therefore, investigating the factors influencing cutterhead speed based on TBM field data and developing reliable predictive models are crucial for optimizing TBM performance, reducing construction risks, and controlling costs.

This study focuses on predicting TBM cutterhead rotation speed using dimensional analysis. By selecting appropriate operational and mechanical parameters, dimensionless variable relationships were constructed and combined with elastic net regression to develop a linear prediction model. Validation on an independent test set demonstrates that the proposed model achieves high prediction accuracy and strong generalizability. These results provide theoretical support for uncovering intrinsic relationships among TBM physical parameters.

2. Dimensional Analysis of Cutterhead Rotation Speed

Cutterhead rotation speed is influenced by multiple factors, including structural and operational parameters, making it a typical multi-variable coupling problem. Developing a prior knowledge model with physical significance is challenging to achieve through theoretical derivation alone. Dimensional analysis offers a robust approach by analyzing numerous physical parameters from the perspective of fundamental dimensions that reflect their intrinsic physical meanings. This method enables the construction of dimensionally homogeneous parameter models that incorporate certain physical mechanisms, thereby facilitating the development of interpretable and generalizable models for cutterhead rotation speed. To enhance the model’s practical applicability and explicit interpretability, this study employs the elastic net regression algorithm for model training.

The modeling process comprises three main steps:

(1)

Parameter Identification. Key parameters significantly influencing cutterhead rotation speed, derived from operational and structural factors, are identified through analysis.

(2)

Model Framework Construction. Based on the principles of the Π-theorem of dimensional analysis and physical constraints, a rational framework for the cutterhead rotation speed prediction model is established.

(3)

Model Training and Prediction. Using specific project data, the constructed framework is employed as input for training the elastic net regression algorithm, which is well-suited for analyzing the statistical characteristics of tunneling data. The algorithm identifies individual working condition features and enables quantitative predictions of cutterhead rotation speed.

By integrating mechanics, dimensional analysis, and engineering data mining, this approach balances model interpretability, generalizability, and accuracy. The workflow is depicted in Figure 1. This workflow combines classic dimensional analysis methods (such as variable selection and π theorem) with modern data-driven techniques (such as elastic network regression analysis and engineering data verification). Through this fusion, not only the scientific nature of the theoretical model is guaranteed, but also the applicability of the model in engineering practice is enhanced.

2.1. Variable Selection

In TBM excavation data, key operational and mechanical factors influencing cutterhead rotation speed (ω) include the cutterhead diameter (D), conveyor speed (Vc), cutterhead torque (T), advance speed (v), total thrust (F), gripper force (Fs), and penetration depth (P). The dimensional expressions of these parameters are summarized in

Table 1. Dimensional Expressions of Parameters.

Parameter	ω	D	Vc	T	v	F	Fs	P
Dimension	[T−1]	[L]	[LT−1]	[ML2T−2]	[LT−1]	[MLT−2]	[MT−2]	[L]

.

The cutterhead rotation speed can be expressed as follows:

$$\omega =f(D,V_c,T,v,F,F_s,p)$$

(1)

Organize the parameters into a column vector q, where:

$$\mathrm{q}={[\omega ,D,V_c,T,v,F,F_s,P]}^{\mathrm{T}}$$

(2)

This can be reformulated into a vector form as:

$$q=f(\omega ,D,V_c,T,v,F,F_s,p)$$

(3)

Equation (3) can be expressed through the dimensional matrix Ω as follows:

$$\Omega =\left[\begin{array}{cccccccc}1& 0& 2& 1& 1& 1& 1& 1\\ 0& 0& 1& 0& 1& 1& 0& 0\\ -1& -1& -2& -1& -2& -2& 0& 0\end{array}\right]$$

(4)

In this paper, Ω represents the parameter dimension matrix, which is used to describe the dimensional decomposition of each parameter in terms of the basic physical quantities (length [L], mass [M], time [T]). Each column of the matrix corresponds to a parameter, such as ω, D, Vc, T, etc., while each row represents the exponent of the respective parameter in the length, mass, and time dimensions. Since the rank of dimensional matrix Ω is 3, three independent variables must be selected as base variables to perform dimensionless analysis on the remaining parameters [30]. In this study, the selection of base variables considers their critical influence on the target variable, dimensional simplicity, and mutual independence [31]. Advance thrust (F), cutterhead diameter (D), and advance speed (v) were selected as the base variables. These variables not only directly impact TBM performance but also minimize potential blind spots in data analysis. The cutterhead diameter (D) represents a fundamental structural parameter with a simple dimensional property, facilitating the extraction of relevant features for rotation speed prediction through dimensional analysis. The advance speed (v), as a key operational parameter in tunneling, is directly correlated with cutterhead rotation speed, making it essential for constructing models that reflect actual operational performance. Selecting F, D, and v adheres to the principles of dimensional analysis while enhancing the interpretability and predictive capability of the model. This selection ensures the development of a robust and physically meaningful model for cutterhead rotation speed prediction.

2.2. π-Theorem Modeling

Using the selected base variables, the remaining variables are combined into dimensionless groups as follows:

$${\pi }_1={\displaystyle \frac{\omega D}{v}},{\pi }_2={\displaystyle \frac{V_c}{v}},{\pi }_3={\displaystyle \frac{P}{D}},{\pi }_4={\displaystyle \frac{T}{FD}},{\pi }_5={\displaystyle \frac{F_s}{F}}$$

(5)

Based on Equation (4), the cutterhead rotation speed can be expressed as: ${\pi }_1=f({\pi }_2,{\pi }_3,{\pi }_4,{\pi }_5)$.

By integrating the dimensionless groups mentioned above:

$${\pi }_1={\displaystyle \frac{\omega D}{v}}=f({\displaystyle \frac{V_c}{v}},{\displaystyle \frac{P}{D}},{\displaystyle \frac{T}{FD}},{\displaystyle \frac{F_s}{F}})$$

(6)

π₄, π₅ primarily represent the influence of load factors on cutterhead rotation speed. These groups are combined through a product transformation into a single dimensionless group, denoted as ${\displaystyle \frac{T{F}_s}{D{F}^2}}$, which captures the impact of load-related operational factors on rotation speed. The other two dimensionless groups, ${\pi }_2={\displaystyle \frac{V_c}{v}},{\pi }_3={\displaystyle \frac{P}{D}}$, represent the effects of speed and length factors, respectively, on the cutterhead rotation speed.

From a mechanical and operational perspective, the interactions among the three dimensionless groups are minimal and can be approximated as linear relationships. Therefore, the model can be expressed as:

$${\displaystyle \frac{\omega D}{v}}=k_1{\displaystyle \frac{V_c}{v}}+k_2{\displaystyle \frac{F_{s}T}{D{F}^2}}+k_3{\displaystyle \frac{P}{D}}$$

(7)

Equation (7) presents a linear model derived from dimensional analysis to describe the relationship between the cutterhead rotation speed ω and key operational and mechanical parameters. Each dimensionless term in the model represents specific contributions: the first term, $k_1{\displaystyle \frac{V_c}{v}}$, indicates the influence of the conveyor speed to advance speed ratio; the second term, $k_2{\displaystyle \frac{F_{s}T}{D{F}^2}}$, captures the combined effects of thrust, torque, and cutterhead diameter as load-related factors; and the third term, $k_3{\displaystyle \frac{P}{D}}$ reflects the impact of the penetration depth to cutterhead diameter ratio. This linear model not only simplifies the analysis of complex mechanical systems but also provides a robust theoretical framework for predicting cutterhead performance under various operational conditions.

3. TBM (Tunnel Boring Machine) Data Processing

3.1. Study Area

The data analyzed in this study originates from a water supply project in Xinjiang, with a total length of approximately 285 km and tunnel diameters ranging from 6.8 to 8.9 m. The construction employed a combination of drill-and-blast methods and open TBM excavation. The TBM used in the project had a cutterhead diameter of 7.8 m. The excavated section primarily consisted of Variscan metamorphic granite, metamorphic biotite granite, and biotite quartz schist, characterized by high integrity, significant rock strength, and minimal joint and fracture development.

The overall tunnel section was intact, with some localized fragmentation. The predominant rock mass classifications were Class II and Class IIIb, with Class IIIa and IIIb accounting for 98.97% of the total rock mass. A 2.5 km long faulted and fractured zone, constituting 0.93% of the total length, was present along the tunnel route. Groundwater in the excavated section was mainly bedrock fissure water, and the tunnel area was located in a region with low groundwater recharge, resulting in minimal seepage. The data acquisition segment spanned terrain with mild undulations, characterized by exposed bedrock and residual aeolian hills.

3.2. Data Preprocessing

This study utilized excavation data from TBM No. 2 as the input dataset. During TBM operation, sensors recorded data at a frequency of seconds, generating massive datasets. These datasets included construction parameters such as total thrust, cutterhead torque, and cutterhead rotation speed, as well as auxiliary system data, such as hydraulic oil tank temperature and main drive cooling water temperature.

Because the sensors operated continuously—even during non-excavation periods such as downtime and maintenance—these datasets contained a significant proportion of non-operational data. These included periods dedicated to surrounding rock support, equipment maintenance, and machine downtime, which were not relevant to the study. Additionally, external disturbances and sensor malfunctions introduced anomalous data points, resulting in datasets that were vast yet sparse in meaningful information.

To improve the accuracy and reliability of the sensor data, reasonable preprocessing techniques were used, including numerical filtering, boxplot analysis, and low-pass filtering. Therefore, data preprocessing was essential to ensure the quality of the input data and the reliability of the analysis results.

3.2.1. Removal of Invalid Data

Due to the operational characteristics of TBMs, the actual excavation work time constituted only a small portion of the total construction time. As a result, the stored data contained large volumes of non-operational data, such as those related to surrounding rock support, equipment maintenance, and downtime. Since these periods were outside the scope of this study, such data were removed. Moreover, the complete excavation cycle was divided into four phases based on TBM excavation data characteristics: idle pushing, ramp-up, steady-state, and ramp-down. Data from the idle pushing and ramp-down phases were classified as invalid rock-breaking data and subsequently removed. As illustrated in Figure 2, gaps between dataset intervals correspond to non-operational segments, which were excluded during preprocessing to focus solely on effective excavation data.

When any operational parameter, such as advance speed, cutterhead rotation speed, or total load, is zero or near zero, it is typically considered as “idle pushing” data or non-operational data. Such data points are removed through direct numerical filtering in this study. The following table shows the statistical values of the processed data.

Table 2. Statistical Distribution of Disc Speed and Torque.

Cutterhead Speed Range	Sample Size	Sample Cutterhead Speed			Sample Cutterhead Torque
		Max	Mean	Std	Max	Mean	Std
0–1	118	0.995	0.506	0.248	452.739	95.290	53.241
1–2	131	1.998	1.563	0.279	268.378	124.274	37.192
2–3	185	2.995	2.511	0.289	329.793	136.090	40.722
3–4	443	3.999	3.614	0.282	861.905	174.697	97.073
4–5	2524	4.999	4.481	0.25	849.932	309.051	172.227
5–6	3738	5.999	5.517	0.29	1018.987	447.003	249.512
6–7	7776	6.999	6.592	0.277	1344.626	736.204	320.109
7–8	8019	7.79	7.29	0.198	1400.759	932.366	295.201

shows the statistical summary of the dataset, partitioned by disc rotational speed. It includes the frequency, maximum, mean, and standard deviation of both rotational speed and torque. The disc speed is mostly concentrated in the 6–8 range, representing the stable operation zone. In contrast, the torque shows greater fluctuations, better reflecting the data variability. Therefore, torque is selected as the key metric for evaluating processing effectiveness.

3.2.2. Outlier Processing

Statistical analysis revealed the presence of outliers in TBM operational data. These anomalies may arise from sensor malfunctions, operator errors, or system glitches, all of which hinder subsequent analysis. Therefore, these outliers must also be removed. As illustrated in Figure 3, the blue boxes highlight the detected outliers.

The boxplot is a simple, efficient, and intuitive data visualization tool commonly used to describe data distributions and identify outliers. It is constructed from several key quartile values, including the first quartile (Q1), median (Q2), and third quartile (Q3), as well as the upper and lower whiskers, which represent the range of typical data values. These values effectively display the central tendency, dispersion, and shape of the data distribution.

A boxplot divides the data into quartiles, where the data inside the box is considered normal. The whiskers reflect the approximate range of the data distribution. The boxplot’s shape indicates whether the data distribution is symmetric and identifies outliers falling outside the whiskers.

The boxplot is particularly advantageous because it is robust to variations in data distribution, making it ideal for summarizing the distribution characteristics of data, comparing differences between datasets, and filtering outliers. In this study, the boxplot method was employed to eliminate outliers. The results of the outlier removal are shown in Figure 4.

3.2.3. Data Denoising

During TBM operations, machine vibrations consistently generate noise in the collected data. Noise in unprocessed data can significantly affect the prediction accuracy of elastic net regression models. Through proper preprocessing, the noise data is removed, which enhances the stability of the model, improves the extraction of the relationship between the cutterhead speed and other parameters, and reduces the overfitting phenomenon. To minimize the impact of noise on the model’s solution speed, this study utilized a low-pass filter for denoising. Low-pass filters smooth the data by restricting high-frequency components while effectively preserving signal variation. Figure 5 compares torque data before and after filtering. The blue box indicates a magnified section, clearly showing that the filtered data is smoother and better suited for subsequent analysis.

4. Elastic Net Regression Analysis

4.1. Elastic Net Regression Setup

When training and regressing the dimensionless model based on engineering data, selecting a suitable method is crucial. Due to challenges such as multicollinearity caused by issues like idle pushing, cutterhead jamming, and parameter regulation delays, along with the presence of anomalous data points, a regression method robust to these issues is required. Elastic net regression excels in handling multicollinearity while balancing variable selection and regularization. By combining L1 and L2 regularization, this method not only selects features but also retains some correlated yet less critical variables, ensuring that significant data is not overlooked. Its applicability and stability make it particularly suitable for dimensionless models predicting TBM cutterhead rotation speed, significantly enhancing the model’s generalizability and predictive accuracy.

To divide the dataset, a 0.2–0.8 split was applied, where 20% of the data was reserved for testing and 80% was used for training. The values of the L1 and L2 regularization parameters were determined using 10-fold cross-validation. This process involves randomly dividing the dataset into 10 subsets. In each iteration, one subset is used as the validation set, while the remaining nine subsets are used as the training set for model training and evaluation. This process is repeated 10 times, with each subset serving as the validation set in turn. The average of the evaluation metrics across all iterations is used to assess model performance for the current hyperparameter combination. This approach identifies the optimal L1 and L2 parameters, enhancing the model’s stability and predictive capability on unseen data.

In this study, the combined parameter L1_ratio was employed to automatically balance the influence of L1 and L2 regularization. The optimal L1_ratio value was determined through cross-validation, ensuring that both regularization effects are effectively considered to achieve the best model performance.

4.2. Regression Analysis and Validation

Through 10-fold cross-validation, the optimal L1_ratio was determined to be 0.1. Using this ratio, the parameters of the dimensionless rotation speed model were calculated as follows: k₁ = 6.25 × 10⁻², k₂ = −2.71 × 10⁻³, k₃ = 1.91 × 10⁻¹.

The resulting dimensionless rotation speed model is expressed as:

$${\displaystyle \frac{\omega D}{v}}=6.25\times {10}^{-2}{\displaystyle \frac{V_c}{V}}-2.71\times {10}^{-3}{\displaystyle \frac{F_{s}T}{D{F}^2}}+1.91\times {10}^{-1}{\displaystyle \frac{P}{D}}$$

(8)

To evaluate the predictive ability and effectiveness of the model, it was applied to the test set for cutterhead rotation speed prediction and compared to the actual values. The results demonstrated a coefficient of determination (R²) of 0.954, indicating a high goodness-of-fit and the model’s ability to effectively capture the relationship between input and output variables.

The mean squared error (MSE), a key metric for assessing model error, was calculated as 0.00286. This low MSE value confirms the minimal discrepancy between the predicted and actual values, further validating the model’s accuracy.

Figure 6 shows the residual plot of the model, which is based on the results of the test set. The residuals predominantly range between −0.1 and 0.1, indicating small prediction errors and excellent model fit. The residuals exhibit uniform fluctuations around zero, suggesting no significant systematic bias in the model. Furthermore, the concentration of residuals within a narrow range reflects a low noise level in the data, validating the effectiveness of the preprocessing steps.

To further validate the model’s reliability, the predicted dimensionless rotation speed values were compared with the actual values, and an error percentage plot was generated. As shown in Figure 7 and Figure 8, the predicted values closely align with the observed values, demonstrating the model’s ability to accurately predict cutterhead rotation speed. Most prediction errors fall within 15%, with 90% of the data showing errors below 10%. This strongly indicates that the constructed rotation speed prediction model effectively captures variations in cutterhead rotation speed observed in real-world engineering scenarios, highlighting its practicality and reliability.

5. Discussion

This study, based on dimensional analysis and elastic net regression, developed a prediction model for the cutterhead rotation speed of tunnel boring machines (TBMs). Using real-world engineering data and excavation parameters, the model aims to enhance construction efficiency and safety. The following points are discussed:

• In this study, we explored the factors influencing the cutterhead rotation speed of tunnel boring machines (TBMs). By identifying the key parameters affecting cutterhead speed, operators are empowered to quickly diagnose issues and adjust parameters accordingly when problems arise. This ability to quickly pinpoint and address the factors impacting cutterhead performance not only enhances operational safety but also improves overall construction efficiency. For example, when encountering unexpected ground conditions or equipment malfunctions, operators can refer to the model’s predictions to identify whether the issue stems from inappropriate rotation speed or other operational parameters. By adjusting the cutterhead speed based on real-time feedback, operators can prevent potential cutter fatigue, minimize rock disturbances, and avoid machine downtime, ultimately optimizing the tunneling process and ensuring timely project completion.
• High-quality data are essential for accurate predictions. This study utilized numerical filtering, boxplot analysis, and low-pass filtering to effectively reduce noise and outliers, improving data reliability. Future work could further enhance data quality by improving sensor precision and integrating anomaly detection and data imputation techniques.
• Despite the promising results, the study has the following limitations: (1) The dataset was sourced from a single project, requiring validation under diverse geological conditions. (2) The model’s ability to dynamically adjust parameters in real time has not been tested.
• Despite the encouraging results of this study, future research should focus on the following aspects: (1) Dynamic real-time prediction: Develop a model that can dynamically adjust the cutterhead speed in real time according to tunnel conditions, combined with a real-time monitoring system to cope with unforeseen operational challenges. (2) Predictive maintenance: Integrate machine learning models to predict TBM downtime and maintenance needs, proactively schedule maintenance and reduce downtime by identifying abnormal patterns in the data. (3) Cross-project and geological validation: Apply the model to different projects and geological conditions for validation to enhance the versatility and accuracy of the model.

6. Conclusions

This study developed a TBM cutterhead rotation speed prediction model based on dimensional analysis, integrating physical knowledge with data analysis to address the traditional limitations in interpreting parameter relationships. The key findings are as follows:

(1)

The TBM cutterhead rotation speed prediction model, established using dimensional analysis and elastic net regression, achieved accurate predictions after comprehensive data preprocessing. Validation on an independent test set demonstrated the proposed model’s strong predictive ability and broad applicability.

(2)

By employing numerical filtering, boxplot analysis, and low-pass filtering techniques, the noise and interference from outliers in the excavation data were significantly reduced. This improved the model’s stability and prediction accuracy, providing crucial theoretical support for advancing intelligent tunneling with shield TBMs.

This study developed a tunnel boring machine (TBM) cutterhead rotation speed prediction model based on dimensional analysis and elastic net regression. By integrating physical knowledge with data analytics, it addresses the limitations of traditional methods in explaining parameter relationships. The main conclusions are as follows:

(1)

This study successfully developed a tunnel boring machine (TBM) cutterhead rotational speed prediction model based on dimensional analysis and elastic net regression. The validation results demonstrated a high degree of consistency between the predicted and actual values, with the majority of prediction errors within 15%, and 90% of the errors below 10%. These results indicate that the model can effectively and accurately predict the cutterhead rotational speed.

(2)

Through data preprocessing techniques such as numerical filtering, boxplot analysis, and low-pass filtering, significant reductions in noise and outliers in the excavation data were achieved. This further enhanced the model’s stability and prediction accuracy, providing reliable support for data-driven modeling in complex operating conditions.

(3)

By integrating dimensional analysis with elastic net regression optimization, the study effectively balanced the L1 and L2 regularization terms, ensuring both the physical interpretability of the model and its adaptability. This provides a theoretical foundation for capturing the complex relationships between the TBM cutterhead rotational speed and input parameters.

(4)

This research offers theoretical support for the intelligent tunneling of TBMs. Future work could focus on improving data quality, expanding the range of model parameters, and combining machine learning techniques for predictive maintenance, further enhancing the operational efficiency and stability of TBM operations.

In summary, this study successfully constructed a TBM cutterhead rotational speed prediction model based on dimensional analysis and elastic net regression, validating its high accuracy and broad applicability under various working conditions. By optimizing data processing methods and parameter selection, the model not only improved the stability and prediction accuracy of TBM operations but also laid the foundation for the further development of intelligent tunneling technology. Future research should aim to further enhance data quality, expand the model’s applicability, and incorporate intelligent maintenance technologies to improve the efficiency and safety of TBM operations.