Research on Cutter Anomaly Identification in Slightly Weathered Metamorphic Rock Formations Based on BO-Light GBM Model

Abstract

Accurate and timely identification of cutter anomalies is crucial for ensuring the safety and efficiency of shield tunneling. To address the issues of poor timeliness and high costs associated with traditional periodic manual inspection methods, this study establishes a cutter anomaly identification model based on the BO-Light GBM algorithm, focusing on slightly weathered metamorphic rock formations. Six parameters closely related to the tunneling state were selected to construct the feature set, and one-class support vector machines (SVMs) were employed to remove anomalous samples. On this basis, a baseline Light GBM model with preset hyperparameters was developed, achieving a preliminary accuracy of 96.04%. Further hyperparameter tuning using Bayesian optimization boosted the overall accuracy of the final BO-Light GBM model to 99.40% while improving training efficiency by approximately 50% compared to exhaustive grid search. Interpretability analysis conducted via SHAP values revealed that chamber pressure, cutterhead rotation speed, total thrust, and cutterhead torque were the primary contributing features, with patterns consistent with actual tunneling conditions, confirming the accuracy of the model’s predictions. The research outcomes provide valuable theoretical guidance and technical support for similar engineering applications.

1. Introduction

During shield tunneling, the cutter is the core component in direct contact with the excavation face, and its working state and wear condition directly affect engineering safety and construction efficiency [1]. Cutter head wear is mainly divided into normal wear and abnormal wear. Normal wear refers to the uniform consumption of the cutter ring during the rolling rock-breaking process; abnormal wear includes eccentric wear, chipping, cutter ring fracture, and other factors leading to non-uniform wear of the cutter [2]. In strata with relatively good conditions, normal wear accounts for 80% to 90% [3]. In full-section high-strength and composite strata with soft upper and hard lower parts, abnormal wear accounts for more than 30% [4], significantly shortening the service life of the cutter.

Due to the enclosed structure of shield tunneling equipment, the status of cutting tools cannot be directly observed during the tunneling process. Accurate wear conditions of the tools are typically obtained through manual inspections during cutterhead entry maintenance. Non-entry monitoring methods for shield tunneling tools primarily include tool wear threshold monitoring [5], pulsed eddy current monitoring [6], and machine learning model predictions based on tunneling parameters [7]. The tool wear threshold monitoring method involves embedding alarm components inside the tools; when the wear reaches a predetermined threshold, an alert is triggered via odor release or electrical signal transmission. The pulsed eddy current method employs sensors installed on the tool holders to convert the distance from the tool edge to the sensor into a signal output, enabling real-time monitoring of tool wear. These two methods can effectively identify uniform wear under normal tool wear conditions but are relatively limited in addressing abnormal tool wear. The machine learning-based prediction method leverages the fact that the tool wear process alters shield tunneling parameters [8,9,10] to train predictive models for tool wear, achieving non-destructive real-time monitoring without modifying the original tools or equipment.

In the field of abnormal state recognition based on machine learning methods, Sun et al. [11] employed the LSTM (Long Short-Term Memory) algorithm to validate the feasibility of using tunneling parameters to train models for identifying abnormal conditions such as cutterhead cake formation and slurry pipeline blockages. Qin et al. [12] developed a detection model for cutterhead cake formation blockages with an accuracy as high as 98.85% using the LSTM-autoencoder algorithm. Jia et al. [13] explored an improved Elman network for detecting risks such as agglomeration, soil blockage, water inrush, and ground settlement during shield tunneling. Shi Buhai et al. [14] utilized the SMOTE algorithm and Bagging algorithm to identify various fault conditions in shield equipment, resulting in a highly capable anomaly classification model. However, for rock-cutting tools, their structure is relatively independent compared to the shield equipment. The connection between the tools and the cutterhead is purely mechanical, and the large number of tools means the impact of a single tool on tunneling parameters is subtle. This makes predicting and identifying the status of these tools more challenging. To address this, Kui et al. [15] used a BP (Back Propagation) neural network model to explore a wear fault detection method for disc cutters, with input features such as cutterhead rotation speed, cutterhead torque, propulsion speed, soil chamber pressure, and total thrust, preliminarily achieving the identification of disc cutter conditions. Sun Hao et al. [16] employed multiple machine learning models for training with limited samples, achieving a diagnostic model for faulty disc cutters with an accuracy of 96%. Nevertheless, when dealing with high-density shield tunneling data, existing machine learning identification methods still generate a large number of false alarms and fail to reveal the feature relationships and marginal contributions among the input parameters. Additionally, in the application of fault discrimination, more advanced techniques exist, such as those based on attention mechanisms and hybrid methods combining SMOTE (Synthetic Minority Over—sampling Technique) with Trans-CWGAN (Conditional Wasserstein Generative Adversarial Network) [17,18]. However, given the limited computational resources available in engineering projects, the large volume of data, and the difficulty in capturing changes in geological strata, opting for simpler and less time-consuming algorithms is considered more suitable.

Building on this foundation, this paper proposes a tool abnormal state recognition model based on Bayesian Optimization (BO) and the Light Gradient Boosting Machine (Light GBM), named BO-Light GBM. Tunneling parameters from segments where the tool condition is determined are selected as training samples. A dataset is established using total thrust force, cutterhead torque, propulsion speed, cutterhead rotation speed, soil chamber pressure, and friction angle as input features for model training. Furthermore, model interpretability and validation analysis are conducted using SHAP (Shapley Additive Explanations) value theory to clarify the marginal contribution of each feature parameter to the model’s recognition results. This further validates and analyzes the constructed BO-Light GBM model, providing technical support for non-entry condition diagnosis of shield tunneling tools. This method features advantages such as low memory consumption, fast training speed, and support for distributed rapid processing of massive datasets. It aligns well with the characteristics of shield tunneling environments, where hardware performance is often limited and the rapid variation of geological conditions generates large volumes of data that require frequent updates.

The main contributions of this paper are summarized as follows:

• A novel integrated framework: We propose a BO-Light GBM model specifically for cutter anomaly identification, which synergizes Bayesian Optimization for hyperparameter tuning with the Light GBM algorithm for efficient and accurate recognition.
• Enhanced interpretability and validation: We establish a multi-parameter dataset derived from key tunneling parameters and, for the first time in this context, employ SHAP value theory to interpret the model and quantify the marginal contribution of each feature, thereby improving the reliability and providing actionable insights.
• Practical engineering applicability: The developed method is characterized by low memory consumption, fast training speed, and distributed processing capability, making it particularly suitable for deployment in resource-constrained shield tunneling environments with massive and frequently updated data.

2. Model Principle

To achieve high-precision identification of abnormal states of disc cutters during shield tunneling, this study constructs a classification model that integrates unsupervised and supervised machine learning algorithms. The model leverages an unsupervised machine learning algorithm to clean the raw data, eliminating anomalies and noise to enhance the quality of the training samples. Based on the cleaned data, a supervised classification model is established to enable the identification and classification of abnormal states of disc cutters. The model training and parameter optimization process incorporates the Bayesian optimization algorithm to enhance the model’s generalization capability and prediction accuracy.

2.1. Unsupervised Learning Algorithm

One-Class Support Vector Machine (One-Class SVM) is a classification algorithm that efficiently partitions data by non-linearly mapping the original input space into a high-dimensional feature space. It identifies anomalous samples by constructing a hyperplane in this high-dimensional feature space [19].

The objective function:

$${\displaystyle \genfrac{}{}{0pt}{}{min}{w,\rho ,\xi }}{\displaystyle \frac{1}{2}}{‖w‖}^2+{\displaystyle \frac{1}{vn}}\sum _{i=1}^n{\xi }_i-\rho$$

(1)

The constraint conditions:

$$\left(w·\varnothing \left(x_i\right)\right)\ge \rho -{\xi }_i {\xi }_i\ge 0, i=1,\dots ,n$$

(2)

Here: $\varnothing \left(x\right)$ is the feature map; $w$ is the normal vector of the hyperplane; $\rho$ is the bias term; ${\xi }_i$ is a relaxation variable; $v$ is a balancing parameter.

For any sample $x$, its discriminant function is:

$$f\left(x\right)=sign\left(w·\varnothing \left(x\right)\right)-\rho$$

(3)

If $f\left(x\right)<0$, sample x is considered abnormal.

2.2. Supervised Learning Algorithms

(1)

Related model algorithms

(a) KNN

The K-Nearest Neighbors (KNN) algorithm is an instance-based, lazy learning classification algorithm. This algorithm does not involve explicit model training and only utilizes the training data for distance calculations during the prediction phase [20]. Its core process includes calculating the distance between the test sample and all training samples, selecting the K nearest neighbors with the smallest distances, and determining the class of the test sample through a voting mechanism. By adopting Euclidean distance as the metric for measuring the distance between samples, the calculation formula is as follows:

$$d\left(x,x_i\right)=\sqrt{\sum _{j=1}^n{\left(x_j-x_{ij}\right)}^2}$$

(4)

Here: $x$ represents the test sample; n is the feature dimension of the sample; $d\left(x,x_i\right)$ is the Euclidean distance between them.

(b) SVM

Support Vector Machine (SVM) is a classical supervised learning binary classification algorithm. Its fundamental idea is to construct an optimal hyperplane in the feature space that maximizes the geometric margin between two classes of samples [21]. In the linearly separable case, SVM obtains the optimal hyperplane by solving the following problem:

$${\displaystyle \genfrac{}{}{0pt}{}{min}{w,b}}{\displaystyle \frac{1}{2}}{‖w‖}^2$$

(5)

Subject to the constraints:

$$y_i\left(w·x_i+b\right)\ge 1, i=1,\dots ,n$$

(6)

Here, $w$ is the normal vector of the hyperplane; $b$ is the bias term; $x_i$ is the training sample; and $y_i$ is the sample label.

For nonlinear classification problems, the kernel function is employed to map the data into a high-dimensional feature space, making it linearly separable in that space. The core concept is similar to that of One-Class SVM.

(c) Random Forest

Random Forest (RF) is a classification algorithm based on ensemble learning. This method constructs multiple independent decision trees and integrates prediction results through a voting mechanism, thereby enhancing the model’s accuracy and robustness. Random Forest uses the Bootstrap method to perform resampling with replacement on the training samples and randomly selects subsets of features at each node split [22]. For a given test sample $x$, the final prediction of the Random Forest can be expressed as:

$$\widehat{y}=\arg max\sum _{b=1}^{B}I\left[h_b\left(x\right)=c\right]$$

(7)

Here, $h_b\left(x\right)$ is the prediction result of the $b$-th decision tree for sample $x$; $I[\cdot ]$ is the indicator function, which takes the value 1 when the condition is true and 0 otherwise; $B$ represents the total number of trained trees; and $\widehat{y}$ is the predicted class.

(2)

Light GBM

Light GBM is an efficient gradient boosting decision tree model. It features low time and space complexity and supports large-scale data processing, making it suitable for engineering environments with limited resources and requirements for rapid response [23]. Unlike Random Forest, which builds independent decision trees in parallel, Light GBM employs a serial iterative training approach.

The optimization objective function of Light GBM consists of a loss term and a regularization term, expressed as:

$$\mathcal{L}\left(F\right)=\sum _{i=1}^{n}l\left(y_i,F\left(x_i\right)\right)+\mathsf{\Omega }(f_t)$$

(8)

The predicted result for a sample is:

$${\widehat{y}}_i=\sum _{t=1}^T{f}_t\left(x_i\right)$$

(9)

Here, $F\left(x\right)$ is the additive structure composed of trees; $l$ is the loss function; and $\mathsf{\Omega }\left(f\right)$ is the regularization term.

2.3. Bayesian Optimization Algorithm

Bayesian optimization is an efficient global hyperparameter optimization method particularly suitable for scenarios where function evaluations are costly and the objective function lacks an analytical form. This method works by establishing a probability surrogate model of the objective function and using an acquisition function to balance exploration of promising regions and exploitation of the current best parameters, thereby rapidly approximating the optimal solution with a relatively small number of evaluations [24].

The Gaussian Process is commonly used as the surrogate model, fitted to the initial data $D_1$. The model is expressed as:

$$\mathrm{f}\left(\mathrm{x}\right)~\mathcal{G}\mathcal{P}\left(m\left(x\right),k\left(x,x^{\prime }\right)\right)$$

(10)

For any new point $x_{\ast }$, its prediction follows a normal distribution:

$$\mathrm{f}\left({\mathrm{x}}_{\ast }\right)|D~\mathcal{N}\left(\mu \left(x_{\ast }\right),{\sigma }^2\left(x_{\ast }\right)\right)$$

(11)

where $\mu \left(x_{\ast }\right)$ is the predicted mean at the current point, and $\sigma \left(x_{\ast }\right)$ is the predicted standard deviation at the current point.

To select the next evaluation point, the Expected Improvement (EI) is often used as the acquisition function, expressed as:

$$EI\left(x_{\ast }\right)=\left(\left(f^{\ast }-\mu \left(x_{\ast }\right)\right)\right)\mathsf{\Omega }\left(z\right)+\sigma \left(x_{\ast }\right)\omega \left(z\right)$$

(12)

$$z={\displaystyle \frac{f^{\ast }-\mu \left(x_{\ast }\right)}{\sigma \left(x_{\ast }\right)}}$$

(13)

Here, $f^{\ast }$ represents the current best objective value, $\mathsf{\Phi }\left(z\right)$ is the cumulative distribution function of the standard normal distribution, and $\varphi \left(z\right)$ is its probability density function. By maximizing the EI value, effective exploration towards the optimal solution can be achieved.

3. Research Methodology

3.1. Data Collection

This study is based on the second section of the Jiangku Network Raw Water Pipeline Project, which bypasses the Songmushan Reservoir. The project is located in the northeastern area between Lianhuashan and Dalingshan in Dongguan City, Guangdong Province (Figure 1). The main construction of this section employs the full-face earth pressure balance (EPB) shield tunneling method for pipeline excavation, while simultaneously installing tunnel lining rings. The total length of the section is 1987 m, with the tunnel axis buried at a depth of 26.217–64.595 m and a groundwater depth of 26.217–33.320 m. The outer diameter of the shield segments is 6.0 m, with an inner diameter of 5.4 m, and the secondary lining has an inner diameter of 4.6 m. The pipeline operates as a pressure-driven gravity flow system, designed with a pressure rating of 0.45 MPa.

The entire alignment traverses five types of strata: completely weathered metamorphic rock, highly weathered metamorphic rock, moderately weathered metamorphic rock, slightly weathered metamorphic rock, and moderately weathered granite, with the slightly weathered metamorphic rock section being predominant. The metamorphic rock is primarily biotite plagioclase gneiss, composed mainly of plagioclase, biotite, and quartz. It exhibits high hardness, with a saturated uniaxial compressive strength ranging from 6.27 MPa to 106.61 MPa. The images of rock samples before and after uniaxial compression experiment are shown in Figure 2

The shield machine selected for this section is the China Railway Equipment CTE6250 EPB shield (Model 121), with an excavation diameter of 6280 mm, a shield diameter of 6250 mm, a main machine length of 8.427 m, and a total machine length of 80 m. The cutterhead is configured with 6 center double disc cutters and 34 single disc cutters. The equipment is equipped with a fully automatic parameter acquisition system operating at a sampling frequency of 1 Hz, collecting 90 parameters including data ID, date, ring number, main oil tank temperature, gear oil temperature, cutterhead rotation speed, and cutterhead power.

During tunneling in the slightly weathered metamorphic rock section, frequent abnormal tool conditions were recorded on site, with the tool replacement rate significantly higher than in other sections. Various abnormal tool conditions were observed (as shown in Figure 3), including common issues such as eccentric wear, cutter ring detachment, cutter ring fracture, and blade chipping. More severe failures such as cutter shaft damage, bolt fracture, and tools becoming irreversibly deformed—requiring hot work cutting for removal due to inability to disassemble normally—were also encountered.

The tool conditions recorded during opening the chamber for inspection can be considered the actual working states during the shield tunneling process in that section. Therefore, based on the inspection results, the tunneling parameters from several rings excavated prior to opening the chamber are labeled as “abnormal samples.” Following the maintenance after the chamber is sealed, and since all abnormal tools have been replaced, the data from the first excavated ring is labeled as a “normal sample,” as shown in Figure 4.

Depending on the number of abnormal tools found during each inspection when the chamber was opened, the range for extracting abnormal samples is dynamically adjusted. When the number of abnormal tools is high (exceeding 10), data from the three rings excavated immediately before opening the chamber are selected; if the number is low, data from the preceding ring are used. Normal samples are taken from the first ring of tunneling data excavated after the chamber was sealed or from the ring excavated just before the chamber was opened if no abnormalities were detected during the subsequent inspection.

During the trial section, a total of five chamber entry inspections were conducted, obtaining data from 8 rings under abnormal conditions and 6 rings under normal conditions. The original dataset comprises 97,588 records. Detailed information regarding each chamber entry and the corresponding data is presented in

Table 1. Statistics of opening inspection and sample construction information.

Inspection Number	Abnormal Number of Cutting Tools	Number of Rings Marked as Abnormal	Number of Rings Marked as Abnormal	Number of Abnormal Samples	Number of Normal Samples
1	9	2	1	15,991	5692
2	0	0	2	0	16,594
3	5	2	1	13,980	6430
4	3	1	1	6488	6240
5	15	3	1	21,032	5141
total	32	8	6	57,491	40,097

.

Based on existing engineering practices, it is shown that Cutterhead rotation speed, Cutterhead torque, Excavation speed, Total thrust force, and Soil chamber pressure are core parameters closely related to excavation quality [25]. Furthermore, considering excessive wear and damage of cutters, which lead to changes in the reaction force of the rock mass on the equipment [26], this study introduces the rolling angle as an input feature to enhance the comprehensiveness of sample representation regarding cutter anomalies. Specific details are illustrated in Figure 5.

The data collected in this study were all obtained from the slightly weathered metamorphic rock section, and the material properties, structural form, and installation method of the shield cutters remained consistent. Therefore, geological parameters and cutter physical parameters were not considered as variables during the modeling process. The final selected input features and their value ranges are shown in

Table 2. Input Features and Their Range of Values.

Input Features	Maximum	Minimum	Mean	Median	Standard Deviation	Variance
Total thrust/kN	19,255.07	3482.84	12,165.26	12,300.06	2135.07	4,558,537.20
Excavation speed/(mm·min−1)	71.08	0.00	13.00	12.54	6.37	40.59
Cutterhead torque/(kN·m)	4401.86	0.00	2143.93	2173.60	501.15	251,154
Cutterhead rotation speed/(r·min−1)	2.98	0.00	2.17	2.16	0.25	0.06
Soil chamber pressure/(bar)	1.53	0.00	0.52	0.49	0.28	0.08
Roll angle (°)	3.31	−5.00	0.19	0.17	0.86	0.75

. The output target, “whether the cutter is abnormal,” is labeled as “0” for normal and “1” for abnormal.

3.2. Data Processing

The high data density and extensive sampling time span resulted in a significant number of anomalies within the samples, which compromised data quality and reduced the accuracy and stability of model training. Therefore, data cleaning is necessary to enhance the reliability of the training samples and improve data validity.

This paper employs One-Class Support Vector Machine (One-Class SVM) as an automated method for detecting and eliminating abnormal data. This technique learns the distribution characteristics of normal data, possesses strong capability for identifying marginal anomalies, and offers greater adaptability. Using this method, a conservative cleaning of the Excavation speed data was performed by presetting outlier thresholds. After multi-dimensional data cleaning, the Excavation speed was taken as a representative for observation. The results show that this method comprehensively identifies isolated points at the distribution edges of the Excavation speed and gradual anomalous data. The processed sample distribution becomes more balanced, and the Excavation speed sequence demonstrates good continuity and reasonable distribution (Figure 6).

After cleaning, a total of 87,831 valid samples were retained, accounting for 90% of the original data. The removal of outliers significantly enhanced the representativeness and feature stability of the sample set, thereby laying a solid foundation for subsequent model construction.

3.3. Model Construction

(1)

Construction Process

To achieve accurate identification of abnormal states in shield tunnel boring machine disc cutters, this study constructed four types of machine learning algorithms—KNN, SVM, RF, and Light GBM—for comparative analysis. The construction process follows a standard machine learning pipeline, including feature selection, data splitting, model training, and performance evaluation. To ensure the accuracy of the evaluation results, all models utilized the same training and testing set split. The output results of each model were comprehensively evaluated using multiple metrics, with the specific workflow illustrated in Figure 7.

(2)

Evaluation Metrics

To comprehensively evaluate the classification performance of the models in the disc cutter anomaly identification task, the following five evaluation metrics were employed: Accuracy, Precision, Recall, F1-score, and AUC. All metrics are calculated based on the four types of prediction results from the confusion matrix, with details provided in Table 3. The calculations for each metric are as follows.

Accuracy (ACC) represents the overall predictive capability of the model and is defined as:

$$ACC={\displaystyle \frac{TP+TN}{TP+FP+FN+TN}}$$

(14)

Precision (P) for the positive class indicates the proportion of samples predicted as normal that are actually normal:

$$P={\displaystyle \frac{TP}{TP+FP}}$$

(15)

Recall (PR) for the positive class indicates the proportion of actually normal samples that are correctly identified:

$$PR={\displaystyle \frac{TP}{TP+FN}}$$

(16)

The F1-score represents the harmonic mean of precision and recall, reflecting the model’s comprehensive performance in anomaly identification:

$$F1=2{\displaystyle \frac{P·PR}{P+PR}}$$

(17)

The AUC (Area Under the Curve) value is an important metric for evaluating model performance, referring to the area under the ROC curve. It measures the model’s ability to distinguish between positive and negative classes. The closer the AUC is to 1, the better the model performance.

Table 3. Detailed information table of confusion matrix composition.

Designation	Symbol	Meaning
True positive example	TP	The number of samples predicted as normal and actually normal
False positive example	FP	The number of samples predicted as normal but actually abnormal
False negative example	FN	The number of samples predicted as abnormal but actually normal
Ture negative example	TN	The number of samples predicted as abnormal and actually abnormal

Table 3. Detailed information table of confusion matrix composition.

Designation	Symbol	Meaning
True positive example	TP	The number of samples predicted as normal and actually normal
False positive example	FP	The number of samples predicted as normal but actually abnormal
False negative example	FN	The number of samples predicted as abnormal but actually normal
Ture negative example	TN	The number of samples predicted as abnormal and actually abnormal

(3)

Model Training

A total of 87,831 sample data points were used for model training. The dataset was divided into training, validation, and test sets in a 7:2:1 ratio, and all model construction was completed within a unified computational environment.

The hyperparameter settings for each model are as follows:

• For the KNN model, the voting weight (weights) was set to uniform, the number of voting neighbors (n_neighbors) was set to 5, and the distance metric (metric) was set to Minkowski.
• For the SVM model, the penalty coefficient C was set to 1, the kernel function (kernel) was selected as the Gaussian (RBF) kernel, and probability estimation (probability) was enabled.
• For the RF model, the number of trees (n_estimators) was set to 100, the maximum depth of the trees (max_depth) was set to 10, and the minimum number of samples required at a leaf node (min_samples_leaf) was set to 20.
• For the Light GBM model, the objective task (objective) was set to binary classification, the boosting type (boosting_type) was set to Gradient Boosting Decision Tree (gbdt), the learning rate (learning_rate) was set to 0.1, and the number of leaves (num_leaves) was set to 100.

Specific details are provided in

Table 4. Model hyperparameter configuration table.

Model	Parameter Configuration
KNN	n_neighbors = 5; weights = uniform; metric = minkowski
SVM	C = 1; kernel = rbf; probability = true
RF	n_estimators = 100; max_depth = 1; min_samples_leaf = 20
Light GBM	boosting_type = binary; boosting_type = gbdt; learning_rate = 0.1; num_leaves = 30; max_depth = 3; learning_rate = 0.1;subsample = 0.1;coslsample_bytree = 1; reg_alpha = 1; reg_lambd = 1

.

Based on the aforementioned model construction, this study conducted a comparative analysis of the performance of four models—KNN, SVM, RF, and Light GBM—in disc cutter anomaly identification using preset hyperparameters. The confusion matrices for the four models are shown in Figure 8. Figure 9 displays the corresponding ROC curves, and the main performance metrics are detailed in

Table 5. Comparison of classification performance of different models in hob abnormality recognition task.

		ACC	P	PR	F1-Score
KNN	Normal	80.05%	75%	73%	74%
	Abnormal		83%	85%	84%
SVM	Normal	77.02%	78%	56%	65%
	Abnormal		76%	90%	83%
RF	Normal	93.60%	96%	88%	91%
	Abnormal		93%	97%	95%
Light GBM	Normal	96.04%	97%	93%	95%
	Abnormal		96%	98%	97%

.

The SVM model demonstrated the poorest identification performance among the four models (Figure 8b). For the anomaly class, this model achieved a precision of 76%, recall of 90%, and an F1-score of 83%, with an overall accuracy of 77.02% and a corresponding AUC value of only 0.85. The KNN model performed slightly better than SVM, with an accuracy of 80.05%. Its precision, recall, and F1-score for the anomaly class reached 83%, 85%, and 84%, respectively, with an AUC value of 0.86. These results indicate that both algorithms exhibit suboptimal performance in handling high-dimensional features and imbalanced sample identification, resulting in high misjudgment rates.

The identification performance of the RF model was significantly superior to that of both KNN and SVM. It achieved a precision of 93%, recall of 97%, and an F1-score of 95% for anomaly sample identification, with an AUC value of 0.98. The model accurately captured the nonlinear relationship changes in tunneling parameters caused by variations in cutter status. However, its performance in identifying the normal class was moderate, with a certain degree of misjudgment. The precision for normal class samples was 96%, but the recall was 88%, yielding an F1-score of 91%. This indicates a significant performance imbalance in identifying positive and negative class samples.

The Light GBM model delivered the optimal performance across all evaluation metrics. Its overall accuracy reached 96.04%. The identification precision for both normal and anomaly classes exceeded 95% (97% for normal, 96% for anomaly), with recall rates of 93% and 98%, respectively. The F1-scores for both sample types surpassed 95%. Compared to the RF model, Light GBM effectively narrowed the performance gap in discriminating between the positive and negative classes. Its AUC value of 0.99 was higher than RF’s 0.98 and significantly superior to KNN’s 0.86 and SVM’s 0.85. These results demonstrate that the Light GBM model possesses higher discriminative accuracy and robustness when processing high-dimensional, nonlinear, and class-imbalanced tunneling monitoring data.

3.4. Model Optimization

The models with preset hyperparameters adequately demonstrated their applicability in cutter anomaly identification. To further enhance the classification performance and generalization capability of the model, this study introduces the Bayesian optimization algorithm to optimize the hyperparameters of the Light GBM model, constructing an optimized BO-Light GBM model for accurate identification of disc cutter abnormal states. During the optimization process, seven hyperparameters—num_leaves, max_depth, learning_rate, subsample, colsample_bytree, reg_alpha, and reg_lambda—were finely tuned [27]. The parameter tuning ranges are shown in

Table 6. Search range of Bayesian optimized hyperparameters for the BO-Light GBM model.

Hyperparameter Category	Adjustment Range
num_leaves	20–60
max_depth	3–7
learning_rate	0.01–1
subsample	0.5–1
coslsample_bytree	0.5–1
reg_alpha	0–1
reg_lambd	0–1

.

The training results of the BO-Light GBM model are shown in Figure 10. The BO-Light GBM model achieved an identification accuracy of 99.40% for cutter anomalies. Furthermore, the precision, recall, and F1-score for both normal and abnormal samples all exceeded 99%, demonstrating a highly balanced performance in identifying both positive and negative class samples.

To evaluate the computational efficiency of the Bayesian optimization strategy, this study conducted a comparative analysis of the training times required by the BO-Light GBM model and the traditional Grid Search (GS) method under the same data scale. The Grid Search method required 163.91 s to complete parameter optimization by finding the optimal training hyperparameters. In contrast, the Bayesian optimization method required only 83.72 s to identify the optimal hyperparameters, representing a reduction of approximately 50% in computational time overhead. Specific details are shown in

Table 7. Comparison of recognition performance between BO-Light GBM and other hyperparameter optimization methods.

		ACC	P	PR	F1-Score	Time-Consuming
Light GBM	Normal	96.04%	96.87%	92.76%	94.77%	25s
	Abnormal		95.55%	98.11%	96.81%
GR- Light GBM	Normal	99.39%	99.16%	99.26%	99.21%	163.91s
	Abnormal		99.54%	99.47%	99.50%
BO- Light GBM	Normal	99.40%	99.15%	99.29%	99.22%	83.72s
	Abnormal		99.55%	99.46%	99.51%

.

4. Model Interpretability and Validation Analysis

4.1. Theoretical Analysis of SHAP Value

Significant nonlinear coupling relationships exist between tunneling parameters. Traditional experimental methods, which often focus on a single disc cutter, fail to effectively map the relationships between parameters like thrust, torque, and chamber pressure and the status of the disc cutters. In practical engineering, tunneling parameters are influenced by the complex interplay of cutterhead structure, tool arrangement, and the status of multiple cutters, making it difficult to accurately describe these intricate nonlinear relationships through mechanical models. In contrast, machine learning-based methods for cutter status identification can effectively handle these nonlinearities but are typically “black-box” models, making their internal mechanisms difficult to interpret directly. Therefore, introducing model interpretability techniques can reveal the relationships between features and prediction outcomes, enhance model transparency and comprehensibility, and consequently improve their reliability in engineering practice.

SHAP values are a post hoc model output interpretation method based on Shapley value theory from cooperative game theory, widely used to evaluate the contribution of each feature to the prediction result in machine learning models. SHAP values can be used to globally assess the importance of features within the overall model. They can not only explain the influence of individual features but also reveal interactions between features and describe their combined effect on predictions under different conditions. Thus, SHAP values provide transparent interpretability for “black-box” models, effectively enhancing the trustworthiness and understandability of machine learning models in practical applications.

The core idea is to calculate the marginal contribution of each feature $x_i$ in the input feature vector $x=[x_1,x_2,x_3,\dots ,x_n]$ to the model’s prediction. The specific method involves traversing all feature subsets S that do not contain feature i, evaluating the change in the model’s prediction when feature i is added to subset S, and then computing a weighted average of these marginal contributions across all subsets [28]. The mathematical expression for this calculation process is:

$${\varphi }_i=\sum _{S\subseteq F\backslash \left\{i\right\}}{\displaystyle \frac{\left|S\right|!\left(\left|F\right|-\left|S\right|-1\right)!}{\left|F\right|!}}\left[f\left(S\cup \left\{i\right\}\right)-f\left(S\right)\right]$$

(18)

Here: F is the set of all features; $f\left(S\right)$ is the model’s prediction value when using only the features in subset S; ${\varphi }_i$ is the SHAP value for feature $x_i$, representing its marginal contribution to the model’s prediction.

4.2. Global Interpretability Analysis

To gain a deeper understanding of the decision-making process of the highest-accuracy BO-Light GBM model, this study employed SHAP value analysis to quantify the marginal contributions of input features, thereby revealing the relationship between these features and the identification of abnormal cutter states. By analyzing the model output via SHAP value, the contribution of each feature to the cutter anomaly identification task was calculated, and a feature impact distribution plot was generated.

Feature dependency analysis reveals the mode of action of each variable on the model output (Figure 11). The SHAP value for chamber pressure exhibits a clear nonlinear variation (Figure 11a). In the low-value range, it primarily suppresses the anomaly prediction, while exceeding a certain threshold significantly strengthens the model’s tendency towards an abnormal output, indicating its sensitivity under changing working conditions. Both total thrust force and cutterhead rotation speed demonstrate a linear, monotonically positive relationship; as their values increase, the predicted probability of anomaly rises synchronously. Cutterhead torque exhibits a threshold-saturation influence pattern: its contribution is relatively weak below 2000 kN·m, but rapidly strengthens (peaking at 4) once this critical value is exceeded, subsequently tending to decline. Its overall influence capacity is noticeably lower than the top three features. In contrast, the SHAP values for rolling angle and Excavation speed are primarily concentrated within the range [−2, 2], with relatively small overall fluctuation amplitudes, indicating their weaker impact on the model output under the current data conditions.

Figure 12 illustrates the influence of the six input features on cutter anomaly identification. The horizontal axis represents the direction of each feature’s influence, where larger positive values indicate a stronger tendency for the feature to contribute to an “abnormal” prediction, while larger negative values indicate a stronger tendency for a “normal” prediction. The vertical axis ranks the features according to the magnitude of their contribution, with influence capability decreasing from top to bottom. The color bar on the right side of the figure represents the range of values for each feature, transitioning from low (blue) to high (red).

Finally, the analysis of SHAP value indicates that chamber pressure, cutterhead rotation speed, total thrust force, and cutterhead torque are the primary features influencing the model’s predictions. Their marginal contributions are significantly higher than those of the rolling angle and Excavation speed.

4.3. Model Validation Analysis

To validate the practical application value of the trained cutter anomaly identification model in engineering practice, a complete set of tunneling data (totaling 30,000 data points) from the shield tunneling process between two chamber inspections was selected for model validation. By using the model to predict the anomaly probability for each data point, a probability distribution scatter plot (Figure 13) was generated, providing a visual representation of the relationship between tool failure locations and the tunneling process.

The results indicate that during this tunneling drive, when the cutterhead reached approximately sample point 5000, the model’s predicted anomaly probability rapidly increased above the threshold of 0.5. As the operators failed to promptly detect the parameter changes indicative of cutter damage, the shield machine continued operating despite the developing tool abnormality. Between sample points 5000 and 13,000, the anomaly probability for some points scattered within the range of 0.6–1.0, suggesting a phase where a limited number of cutters were failing.

Subsequently, from sample point 15,000 onwards, over 98% of the data points showed an anomaly probability of 100%. Ultimately, by sample point 30,000, the construction personnel recognized the significant changes in the tunneling parameters and organized a chamber inspection. The inspection records revealed extensive abnormal cutter damage and excessive wear on several tools, which strongly corroborates the model’s identification results and further validates its accuracy in practical application.

5. Conclusions

This paper constructed a BO-Light GBM model for shield tunneling cutter anomaly identification using the Bayesian optimization algorithm, comparing it with traditional machine learning models such as KNN, SVM, and RF. Simultaneously, the importance of input features was analyzed for global interpretability, and the model was validated using the SHAP value method. The main conclusions are as follows:

(1)

The Light GBM model demonstrated superior performance in the cutter anomaly identification task compared to KNN, SVM, and RF. Even with constraints of limited depth and fewer decision trees, it achieved an identification accuracy of 96.04%, effectively capturing changes in tunneling parameters caused by alterations in cutter condition. This indicates its stability and applicability under complex working conditions.

(2)

Employing the Bayesian optimization algorithm to optimize the Light GBM model resulted in a significant improvement in model identification accuracy, reaching 99.40%. The optimized model maintained highly balanced performance in identifying both normal and abnormal samples, with a markedly reduced gap in recall between the positive and negative classes. This demonstrates the model’s strong robustness and generalization capability even under conditions of class imbalance.

(3)

The results from the theoretical analysis of SHAP value indicate that chamber pressure, cutterhead rotation speed, total thrust force, and cutterhead torque are the key features influencing the model’s discrimination results, contributing significantly to anomaly identification. The overall contributions of the rolling angle and Excavation speed were limited. The distribution of SHAP value aligns highly consistently with actual tunneling conditions, further validating the identification accuracy and reliability of the BO-Light GBM model.

(4)

The cutter anomaly identification based on the BO-Light GBM model effectively detected the onset time of cutter anomalies in real-time data discrimination. This provides crucial guidance for timely chamber entry, adjustment, or cutter replacement, thereby helping to mitigate negative impacts.

Since the data were collected solely from slightly weathered metamorphic rock strata and the data acquisition equipment was relatively limited, this study lacks application instances in other projects. Future work will focus on conducting identification and analysis across varying geological strata, thereby enhancing the general applicability of the method and its guidance for practical engineering.