Clogging Risk Early Warning for Slurry Shield Tunneling in Mixed Mudstone–Gravel Ground: A Real-Time Self-Updating Machine Learning Approach

Abstract

Clogging constitutes a significant obstacle to shield tunneling in mudstone soils. Previous research has focused on investigating the influence of soils and slurry properties on clogging, although little attention has been paid to the impact of tunneling parameters on clogging, and particularly early clogging warning during tunneling. This paper contributes to developing a real-time clogging early-warning approach, based on a self-updating machine learning method. The clogging judgment criteria are based on the statistical characteristics of whole-ring tunneling parameters. The paper proposes the use of random forest (RF) for a real-time self-updating early warning strategy for clogging. The performance of this approach is illustrated through its application to a slurry-pressure-balanced shield tunneling construction of Nanning metro line 1. Results show that the RF-based approach can predict clogging during a ring construction with only four minutes of tunneling data, with an accuracy of 95%. The RF model provided the best performance compared with the other machine learning methods. Furthermore, the RF model can realize an accurate clogging prediction in one ring, using less tunneling data with the self-updating mechanism.

1. Introduction

Clogging constitutes a serious challenge in shield tunneling in clayey soils for earth-pressure-balanced (EPB) tunnel boring machines (TBMs) and slurry-pressure-balanced (SPB) TBMs [1,2]. This may lead to a severe disturbance in tunneling construction, such as tunnel face collapse [3], severe cutterhead wear [4], and poor construction efficiency [5]. The SPB shield tunneling is a fast and sustainable construction method in urban areas, due to its low environmental disturbance. However, clogging has made the SPB shield tunneling unsustainable, due to the overflowing slurry which has caused environmental pollution. As a typical clayey soil, mudstone has a high potential to cause clogging [6,7]. Several projects in China have encountered clogging problems in the mudstone-rich area, for example, the Wuhan Sanyang cross-river road tunnel [8], the metro tunnel line 1 and 2 in Nanning city [9], Nanchang Metro Line 1 [10], and the Nanjing Yangze river tunnel [11]. This field experience indicates that it is difficult to maintain the normal tunneling state in mudstone-rich areas, especially for the mixed ground condition of round gravel and mudstone.

There are two pressurized chambers in the modern SPB machine. The excavation chamber is near the tunnel face. It is filled with excavated material (usually soil and slurry) to support the tunnel face stability. The second is the working chamber, which uses an air cushion at the top to realize the pressure adjustment [7]. A submerged wall is used to separate these chambers. Still, the hydraulic connection is through an opening with two communicating pipes at the bottom and the middle of the submerged wall. This is essential for maintaining the slurry pressure in the excavation chamber (SPE), with the help of an air cushion and slurry circulation systems. During tunneling in the mixed ground of round gravel and mudstone, the excavated soil may adhere to the opening and the communicating pipes on the submerged wall, and obstruct the hydraulic connection between the two chambers, as shown in Figure 1. When clogging occurs, the slurry feed lines can still work well, but the slurry return lines will be unable to transport the excavated material, leading to the high-pressure fluctuation of SPE with an increase in the cutterhead wear and a decrease in the shield machine advance rate (AR). Unfortunately, the criteria for clogging are still missing using tunneling parameters; in particular, with regard to how we can use the few available minutes of tunneling data for clogging early warnings.

The literature review shows that previous research has focused on the influence of soil properties on clogging, with little attention on the use of tunneling data for clogging early warnings. This research proposes a real-time self-updating machine learning (ML) system for clogging early warnings. Real-time means using minimal tunneling data to predict clogging during the construction of a ring. The self-updating mechanism means the hyper-parameters of the ML model can be updated automatically as the shield machine advances with the tunneling data of newly finished rings. The research contributions can be summarized as follows: firstly, we investigated the statistical characteristics of the tunneling parameters in SPB tunneling in the mudstone-rich areas, and identified the judgment criteria for clogging based on the variation of tunneling parameters. Secondly, we presented a real-time self-updating approach for clogging early warnings during the tunnel construction process, which achieved a 95% prediction accuracy using four minutes of tunneling data in one ring. Thirdly, the research showed that the cutterhead torque should be considered first in clogging early warnings. Our approach is suitable for abnormal situation prediction in soft-soil shield tunneling and hard-rock TBM construction. The self-updating mechanism helps obtain more accurate prediction results with the data-driven model.

This paper is organized as follows. Section 2 provides a literature review about clogging judgment and tunneling data variation due to clogging, and applications of random forest (RF)-based ML models. Section 3 presents the research methodology and the tunnel project used to support this research. Section 4 shows a statistical analysis of the tunneling parameters of Nanning Metro and an elaboration of the clogging criteria. Section 4 illustrates the development and use of a clogging early warning model based on the RF method. The final section includes recommendations for elaborating and applying the clogging early warning model in tunneling.

2. Literature Review

2.1. Clogging Judgment Criterion and Tunneling Parameter Variation Due to Clogging

As clogging results from the adherence of the excavated clayey soil, it is influenced by the properties of clay soil and the slurry flow behavior [2,12]. Experiments have assessed the possibilities of clogging in one specific kind of soil and slurry flow behavior [2,12,13,14,15,16]. They allowed the establishment of some direct and indirect methods for clogging judgment. These methods rely on the soil properties. Since most of these methods focus on sand or clayey soils, they cannot be used in mixed ground conditions [2,17]. The criteria to evaluate clogging potential, as raised by Hollmann and Thewes [13] and Feinendegenet al. [14], mainly focused on soil properties but paid little attention to slurry or foam properties.

Moreover, we still do not know enough about the shield tunneling parameter variation rules in the clogging situation. Shirlaw [17] reported the clogging problem when SPB tunneling was used in completely decomposed granite (CDG) with high clay (clayey CDG). It was found that the values for the Penetration Index (Pind, related to penetration rate (PR) and thrust (THR)) and specific energy (SE, related to rotation speed (RS) of cutterhead, PR, THR, and torque (TOR)) were large in the clayey CDG, much larger than in traditional CDG. The SE values in the clayey CDG were almost as large or larger than that in the granite rock. Although valuable experience about THR, TOR, and AR in the clogging situation was reported in Shirlaw’s paper, the variation of slurry pressure in different geological conditions was not discussed. This variation is crucial for settlement control in urban shield tunneling. Xie et al. [3] proposed a parametric analysis of the SPB shield tunneling parameters in a mixed ground of round gravel and mudstone. In this research, a k-means clustering method was applied to study the relationship between the mixed ratio $\lambda$ (the height ratio of mudstone in the excavation face) and the difference between the slurry pressure in the working chamber (SPW) and SPE. The average values of SPE, TOR, THR, and PR were analyzed with different $\lambda$. However, the statistical characteristics of the instantaneous value of tunneling parameters was not considered. This issue will be explored in this paper. Avunduk and Copur [18] investigated the influences of soil properties on the excavation performance of an EPB shield employed in the Istanbul Ayvali Waste Water Tunnel. Clogging occurred in the high-plasticity clay (CH). The values of TOR and SE obtained in the CH were nearly double that obtained in brown-colored silty sand and dark-gray-colored lean clay. They used multivariable linear regression to predict the TOR, SE, and instantaneous AR using geotechnical characteristics, such as vane shear strength, fall cone penetration depth, plastic limit, plasticity index. They achieved a coefficient of determination ($R^2$) around 0.8.

The literature review shows that (i) it is challenging to predict clogging occurrence using only geological survey information, and (ii) since clogging significantly influences tunneling, it is essential to establish clogging judgment criterion using tunneling data.

2.2. Application of RF-Based Prediction Model

The modern shield machine can generate tunneling data every ten seconds, recorded by the programmable logic controller (PLC) system. Tunneling data contains information about the interaction between the TBM and geological environment, which is helpful for clogging judgment and early warnings [19]. The ML-based approaches, such as the RF method, are appropriate for analyzing tunneling data during the tunneling process. They have been employed in several tunnel projects [20,21,22,23]. Since the clogging state can be regarded as an abnormal tunneling situation, it can be considered a binary classification problem.

RF algorithm has been widely used in developing predictive engineering data-mining models [24,25]. Nitsche et al. [26] proposed an RF model to predict the range and standard deviation of the weighted longitudinal profile to evaluate the pavement roughness using vehicle response. When reducing the number of features, the RF was more stable than the multilayer perceptron (MLP). Kohestani et al. [27] presented an RF-based soil seismic liquefaction potential prediction model, which used the cone penetration test data as input. The proposed RF models provided more accurate results than the artificial neural network (ANN) and the support vector machine (SVM) models. Zhou et al. [28] employed eleven algorithms to predict the rockburst classification. He showed that the RF provided the best results. Alipour et al. [29] established a bridge load-capacity prediction model based on RF for bridge maintenance decisions. Shaikhina et al. [30] presented the use of RF for predicting early transplant rejection using only 80 samples. They achieved a prediction accuracy of 85%. They confirmed the advantages of using an RF-based model with low samples.

The RF-based model has been used in shield tunnel construction to predict soil settlement and tunneling parameters. Zhang et al. [31,32] successfully used the RF-based model for the ground settlement in several cases. Zhou et al. [33] investigated the feasibility of an RF-based model in predicting shield tunnel-induced ground movements and found that the RF approach provided promising results, and can be regarded as an alternative method for settlement prediction. Luo et al. [34] used the RF-based model to analyze the relationship between the slurry pressure and some tunneling factors and optimize the slurry pressure in SPB tunneling and the particle swarm optimization. Zhu et al. [35] proposed a cloud RF model to assess the structure losses of metro tunnels in the operational period. They also introduced a self-training framework to achieve higher predictive accuracy.

The above-mentioned applications of the RF-based classification model are static models. They employed well-trained RF models for all test data sets, unsuitable for the shield tunneling process. The required model should be trained with a few rings and realize self-updating for a real-time early warning model for clogging. To achieve this objective, a training data set updating method is proposed in this paper to realize an early clogging warning during shield tunneling construction.

3. Methodology and Material

3.1. Methodology

3.1.1. Random Forest (RF) Classifier

The random forest method was used for classification. It is based on constructing a set of decision trees [36]. As illustrated in Figure 2, samples and features are randomly selected from the data set via the bootstrap aggregating approach. First, sub-samples were created by choosing random features with replacements. Then, every decision tree in the forest was trained on the sub-sample, and the final class (clogging or normal) was determined by averaging the probabilistic prediction of all the trees, as shown in Equation (1):

$$P(c)={\displaystyle \sum _{j=1}^n{P}_j(c)}$$

(1)

where $P(c)$ is the final prediction from a total number of n trees, and $P_j(c)$ is the individual prediction result from tree #j for a sub-sample of the input data set.

Several hyper-parameters in the RF method were determined by cross-validation, such as the number of trees in the forest ($n\_estimators$), the maximum depth of the tree ($\max \_depth$), etc. The randomized search strategy improved training efficiency during shield tunneling construction. This strategy used a random search over the hyper-parameters, where each candidate set was generated from a distribution over possible hyper-parameter values.

3.1.2. Real-Time Clogging Early Warning Process

Figure 3 shows the process for the real-time early warning of clogging risks. Suppose there are L rings in total for a tunnel section, and the shield machine will excavate at ring #i ($i\le L$). Firstly, we selected N minutes of tunneling data samples at the start of each i-1 ring as the early warning input feature ($N=0.5,1,2,3\dots$), and took the i-1 rings clogging state as output, composed as the training set. Secondly, the ith clogging prediction model was trained with the above training set, whose hyper-parameters were calculated by the 5-fold cross-validation and randomized search [37]. Then, the proposed prediction model was applied to predict the clogging state of ring #i via the N minutes tunneling data samples at the start of the ith ring. Finally, when ring #i was finished, we determined the real clogging state for the ith ring based on the clogging judgment criteria, which will be introduced in Section 4. The training data set was updated by adding the input feature and clogging state of ring #i. This process continued until the tunneling section was achieved.

3.2. Data Collection—Nanning Metro Line 1 Project

The Nanning metro line 1 in Guangxi province supports this research. It connects the Bai Cang Ling Station and Railway Station (BR section). Its length equals 1209 m with 806 rings. The width of the segment is equal to 1.5 m. The tunnel section was excavated using an SPB shield machine, 6.28 m.

Figure 4 shows the geological conditions and the adjacent buildings. The ground is composed of a mixture of mudstone and round gravel. Thirty-six boreholes were used for soil investigation (

Table 1. Soil Properties of Round Gravel and Mudstone in The Nanning Metro Line 1.

Soil	$\mathit{\rho }$	$\mathit{\upsilon }$	$\mathit{c}$	$\mathit{\phi }$	${\mathit{E}}_0$	$\mathit{K}$	${\mathit{k}}_0$
	g/cm3		kPa	°	MPa	m/d
Round gravel	2.05	0.27	0.0	35.0	25.0	90	0.37
Mudstone	2.15	0.20	90.0	21.0	30.0	0.01	0.25

Note: $\rho$ is the natural soil density, $\upsilon$ is the Poisson’s ratio, $c$ is the cohesion, $\phi$ is the friction angle, $E_0$ is the deformation modulus, $K$ is the permeability coefficient, $k_0$ is the coefficient of lateral earth pressure.

). The buried depth of the tunnel varies between 14 m to 22 m, while the groundwater level of the tunnel ranges from 1.5 m to 9.0 m. The gravel is saturated in the BR section. It mainly consists of gravel with a small number of pebbles (Figure 5a). The average content of particles with a grain size of 2–20mm is 53%, while the average content of particles with a grain size greater than 20 mm is 25%. The inter-particle filling is mainly medium and coarse sand. The gravel has a high permeability coefficient (K = 90 m/d), while the mudstone (Figure 5b) has a low permeability (K = 0.01 m/d). Figure 5c illustrates that the grain size of the mudstone is almost smaller than 0.1 mm, and about 64% of the particles have a grain size smaller than 0.005 mm.

The SPB shield was employed to control the soil settlement and reduce damage risk to the adjacent building. The SPB was equipped with 16 cutters with a diameter of 17 inches, 46 scrapers, and 36 tearing cutters to deal with pebbles of a large size. To convey excavated material from the tunnel face, the cutterhead opening ratio was 35%. Sixteen pairs of double cylinders were employed in the shield machine. They provided a maximum THR of 42.575 MN (350 bar). The maximum values of the RS and AR of the cutterhead were 3 rpm and 50 mm/min, respectively. The rated and the breakout values of the TOR were 5.488 MN·m and 6.619 MN·m, respectively.

The satisfactory performance of the SPB shield was observed in the round gravel area in the last section. The ground settlement was about 3 mm on average, while the average value of the AR was around 30 mm/min. Besides, there was minimal wear of the cutterhead after driving 1468 m. Observations in the BR section containing mudstone showed clogging (

Table 2. Consistency Limits of Mudstone for Given Sampling Locations.

Sample	${\mathit{\omega }}_{\mathit{n}}$	${\mathit{\omega }}_{\mathit{L}}$	${\mathit{\omega }}_{\mathit{P}}$	${\mathit{I}}_{\mathit{P}}$	${\mathit{\omega }}_{\mathit{L}}-{\mathit{\omega }}_{\mathit{n}}$	${\mathit{\omega }}_{\mathit{p}}-{\mathit{\omega }}_{\mathit{n}}$	${\mathit{I}}_{\mathit{c}}$
	%	%	%	%	%	%	%
Sample 1	14.9	35.6	19.6	16.0	20.7	4.7	1.29
Sample 2	15.2	38.4	22.0	16.4	23.2	6.8	1.41

Note: ${\omega }_n$ is the natural water content, ${\omega }_L$ is the liquid limit, ${\omega }_P$ is the plastic limit, $I_P$ is the plasticity index, $I_C$ is the consistency index.

) due to the mixture of the clayey mineral component and high viscosity slurry (Figure 6a). This resulted in severe disturbances, including environmental pollution due to the slurry oozing (Figure 6b). The tunneling efficiency was poor with an AR smaller than 10 mm/min. High wear of cutterhead happened as the cutterhead jammed when the SPB machine was in the arrival process. The consistency limits of the mudstone samples section were measured to evaluate the mudstone clogging potential. Figure 7 summarizes the obtained results. According to the SPB classification diagram [13] (Figure 7), the mudstone was in the form of lumps. This could result in extensive clogging as large lumps of the mudstone could form a blockage in the slurry circulation system of the SPB shield machine [18]. The tunnel face stability of the water-rich round gravel required a slurry with Marsh funnel viscosity in the range of 24–32 s. This process made the excavated material more easily clogged at the submerged wall opening between the excavation and working chambers, as illustrated in Figure 6c,d.

The PLC system was used to automatically record the tunneling data at an interval of 10 s, including the time, location, and sensors data in the SPB shield machine. For example, the SPE was measured by a pressure sensor around the middle line of the cutterhead in the excavation chamber. In addition, the operational tunneling data were collected for 666 rings. These data will be presented and analyzed in the following section.

4. Statistical Analysis of The Tunneling Parameters in The BR Section

The tunneling parameters used in the research included SPE, SPW, TOR, THR, RS, and THR. Tunneling data were denoised by removing the obvious measurement errors, such as values that exceeded the sensor’s range. The following mixed ratio (λ) was used to evaluate the effect of the mudstone on clogging Equation (2) [3]:

$$\lambda =H_m/D$$

(2)

where $H_m$ is the mudstone thickness at the excavation face, as determined from the geological report. $D$ is the excavated diameter of the SPB.

The average, standard deviation, maximum, and range values were calculated for each ring to investigate the relationship between these parameters and the mixed ratio (λ). The analysis will focus on the period when AR > 0.

4.1. Statistical Analysis of Tunneling Parameters

Figure 8 illustrates the variation of the statistical indexes of the SPE, together with the mixed ratio ($\lambda$). It shows a strong positive correlation between these parameters, in particular between λ and the SPE standard deviation (Figure 8b), the maximum (Figure 8c), and the range (Figure 8d). This result indicates that the mudstone clogging resulted in a hydraulic obstruction between the two chambers; therefore, the excavated material cannot be transported via the slurry return line to the slurry treatment plant. This obstruction caused an increase in the SPE up to 480 kPa, with a risk of slurry backflow to the shield tail and severe slurry spouting at the shield tail. Besides, the slurry oozing problem easily occurred due to the significant fluctuation of the SPE, resulting in passive failure of the tunnel face and environmental pollution.

Figure 9 shows the variation of the SPW statistical parameters. We observed a more regular change than that of the SPE parameters. Figure 9a shows that the mean values of the SPW were not influenced by clogging. ${\rho }_{\mathrm{mean},\lambda }$ was smaller than 0.2. This weak influence was mainly due to the air cushion system. However, Figure 9b,d show fluctuations in SPW values in the mudstone-rich area, which are confirmed by the high positive values of ${\rho }_{\mathrm{std},\lambda }$ and ρ_rangeλ. Clogging increased the maximum values of the SPW, particularly around ring #400 (Figure 9c).

Figure 10 shows the TOR statistical parameters together with the mudstone distribution. The correlation coefficient between TOR and $\lambda$ is around 0.5. Hysteresis phenomenon was observed between the TOR statistical indexes, and λ. We take the mean values of the TOR as an example: when $\lambda$ increases, the mean value of the TOR increases. When $\lambda$ begins to decrease, the mean value of the TOR remains high during the construction of several rings. This observation could be attributed to the difficulty in clean-up of the cutterhead. Clogging causes a significant increase in the TOR, which varies from 1.5 MN·m to 3.8 MN·m. The TOR induced by clogging accounts for 27% to 70% of the TOR in normal conditions. (Figure 10c) shows that the maximum value of the TOR can reach 4.8 MN·m (87% of the rated TOR), which indicates an overloaded operation condition of the cutterhead motor.

Figure 11 shows that the statistical parameters of the THR are similar to those of the TOR. There was a positive relationship between the statistical indexes of the THR and the mixed ratio (correlation coefficients = 0.5), except for the standard deviation of the THR (correlation coefficient = 0.14). The mean values of the THR increased with the increase in $\lambda$, but were affected by the buried depth. The hysteresis phenomenon between the statistical indexes of the THR and $\lambda$ was observed.

Figure 12 shows the variation of the statistical indexes of the RS together with that of the mixed ratio. A high RS was observed in the mudstone-rich area (Figure 12a). Therefore, there was a moderate positive relationship between the mean values, maximum values of RS, and $\lambda$. As the RS was an active control parameter chosen by the operator, a higher cutterhead RS was employed to mitigate the clogging problem in some cases, but with low efficiency. Figure 12b shows a slight variation in the standard deviation of RS in the different ground conditions.

Figure 13 illustrates the variation of the statistical indexes of the AR. It shows a strong negative correlation between these indexes and $\lambda$. In the mudstone-rich area, the mean value of the AR was less than 10mm/min (Figure 13a). This low value indicates low tunneling efficiency because of clogging. Other statistical parameters of the AR confirm this result in the mudstone-rich area, which was smaller than those observed in the round gravel area (Figure 13b–d).

4.2. Determination of Clogging Based on Tunneling Parameters

According to the above analysis, SPE, TOR, and AR were selected to evaluate the clogging risk. The difference between SPE and SPW is used to eliminate the influence of the buried depth and water table on the SPE (Equation (3)).

$$\Delta SP=SPW-SPE$$

(3)

Figure 14 illustrates some examples of the variation of ΔSP. In normal conditions, $\Delta SP$ it varies in the range 0–20 kPa (Figure 14d), but in clogging conditions, it varies between 150 and 50 kPa. Therefore, a slurry pressure fluctuation index (SPFI) $\theta$ will be used subsequently. This parameter is determined according to Equation (4) and Figure 14a.

$$\left\{\begin{array}{l}{S}_{positive}=\frac{1}{2}({\displaystyle \int \Delta SPdt}+{\displaystyle \int \left|\Delta SP\right|dt})\\ S_{negative}=\frac{1}{2}({\displaystyle \int \Delta SPdt}-{\displaystyle \int \left|\Delta SP\right|dt})\\ \theta =\frac{abs(S_{negative})}{S_{positive}+abs(S_{negative})}\end{array}\right.$$

(4)

The conditions of clogging occurrence were established from field observations and statistical parameters of the tunneling parameters. Clogging in a ring was expected when two of the following conditions were satisfied:

• The SPFI of the whole ring was larger than 0.15;
• The average value of the TOR was larger than 1 MN·m;
• The average value of the AR was smaller than 15 mm/min.

Figure 15 illustrates the variation of the slurry pressure fluctuation index $\theta$ and the mixed ratio $\lambda$, as well as the clogging state (green color). It shows a dramatic fluctuation of the SPE in the clogging rings. The values of $\theta$ were close to one for clogging states. The amount of the mixed ratio $\lambda$ was not always consistent with the clogging state. This observation could be related to the uncertainties in the geological survey, and the mudstone stuck on the cutterhead.

Table 3. Mean Values of Statistical Indexes of The Tunneling Parameters in The Normal and Clogging State.

Parameters\Indexes	Mean		Std		Max		Range
	Normal	Clogging	Normal	Clogging	Normal	Clogging	Normal	Clogging
SPE (kPa)	163.7	203.5	5.4	22.2	179.1	310.9	30.0	160.0
SPW (kPa)	180.2	181.7	3.6	7.0	191.9	210.2	25.1	53.9
TOR (MN·m)	0.9	2.2	0.2	0.5	1.3	3.5	1.1	3.1
THR (MN)	12.1	16.5	1.2	1.4	14.1	18.4	7.9	11.8
RS (rpm)	1.0	1.2	0.03	0.05	1.1	1.3	0.2	0.2
AR (mm/min)	26.9	10.1	8.0	4.4	37.5	25.2	37.4	25.2

summarizes the average values of the statistical parameters of the tunneling parameters for rings in normal and clogging states. It shows remarkable differences for SPE, TOR, and AR between the normal and clogging states. The mean values of SPE in the clogging state were about 25% higher than those in the normal state. A slight difference was observed for the mean values of SPW between the normal and clogging states. Moreover, the standard deviation of the SPE in the clogging state was about four times that in the normal state. The range value of the SPE in the clogging state was about 5.3 times that in the normal state. This observation confirms the sharp fluctuations in the SPE in clogging conditions. For the cutterhead torque, the mean values in the clogging state were about 2.5 times than in the normal state. The maximum and range values of the TOR in the clogging state were about 2.7 times that in the normal state. For the cutterhead thrust and rotation speed, the statistical indexes in the clogging state were larger than those in the normal state, but with low differences. The mean values of the AR in the normal state were about 2.7 times those in the clogging state.

5. Clogging Prediction Results in the BR Section

5.1. Prediction Results

Figure 16 shows an example of the six-minute tunneling data used as an input feature for clogging prediction. The PR is determined from the AR and RS values as follows in Equation (5):

$$PR=\frac{AR}{RS}$$

(5)

This represents the penetration distance per rotation of the cutterhead.

First, analysis was conducted with 17 kinds of features as input data, including the statistical indexes (mean values, standard deviations, maximum values, and range values) of the $\Delta SP$, THR, TOR, PR, and $\theta$. This analysis aims at the identification of the most influential parameters. This will help us investigate the influence of the parameter N.

A confusion matrix was used to assess the RF method performances, as shown in

Table 4. Confusion Matrix Used for Clogging Prediction Assessment.

		Actual Class
		P: Clogging	N: Normal
Predicted class	P: Clogging	TP	FP
	N: Normal	FN	TN

. The positive (P) and negative (N) values present the clogging state and the normal state, respectively. The true positive (TP) indicates a clogging state for both the current and predicted states. False positive (FP) indicates a current normal state but a clogging predicted state. True negative (TN) refers to a current clogging but a normal predicted state. Finally, false negative (FN) indicates the normal for the current and predicted states.

The model performance was evaluated via the following indexes: error rate, precision, recall, and the F1 score (Equation (6)).

$$\left\{\begin{array}{l}error rate=\frac{FP+FN}{TP+TN+FP+FN}\times 100\%\\ precision=\frac{TP}{TP+FP}\times 100\%\\ reacall=\frac{TP}{TP+FN}\times 100\%\\ F1=\frac{2\times precision\times reacall}{precision+reacall}\end{array}\right.$$

(6)

The data presented in Section 3.2 are used in the following analysis. They include data for 666 rings. According to the clogging criteria, 229 rings experienced clogging (Figure 17). The first 90 rings were selected for the initial training data and the analysis of the tunneling data length ($N=0.5,1,2,3,4,5,6,8$). The candidate hyper-parameters ($n\_estimators$ and $\max \_depth$) used for the RF model was a random number between 0 to 1000. We also conducted three other models to evaluate the model performances, including the K-nearest neighbor (KNN) model [38], the support vector classification (SVC) model [39], and the MLP model [40]. The hyper-parameters ($n\_neighbors$) distribution for the KNN model was a number that was selected randomly from 0 to 10. The SVC kernel function used in the SVC model was the RBF function. The C and gamma values were all logarithmically spaced in the range 1–1000, and 0.001–1000, respectively. The vector length was 20 and 7 for the C and gamma values, respectively. The MLP model had one hidden layer with neurons in the range 5–25, with an increase of 5. The learning rate of the MLP model was set as logarithmically spaced values ranging from 1 to 1000, whose vector length was 4. All the hyperparameters used in the KNN, SVC, and MLP models were determined via the 5-fold cross-validation and randomized search, which was the same as the RF model. The ranges of these hyperparameters mentioned above were determined according to previous experience.

Figure 17 shows the influence of the tunneling data length on the performances of the capacity of four machine learning models (RF, KNN, SVC, and MLP) to predict clogging in the 666 rings. It was found that there was a general improvement in the models’ performance with the increase in the data length. This also shows that the RF model provided the best performance, followed by the KNN and SVC models. With four minutes of data samples, the RF achieved good performance (error state = 5%, precision, recall, and F1 score around 94%) (Figure 17), which is acceptable in tunnel construction.

5.2. Discussion

5.2.1. Model Performance without Training Data Updating

Table 5. Metrics Comparison of RF Clogging Prediction Model Considering Different Training Strategies.

Tunneling Data Length (Minutes)	Error Rate (%)		Precision (%)		Recall (%)		F1 (%)
	Updating	No Updating	Updating	No Updating	Updating	No Updating	Updating	No Updating
0.5	16.5	20.7	80.5	80.2	77.3	62.8	78.9	70.4
1	11.6	18.6	88.6	77.5	81.2	75.2	84.7	76.3
2	9.7	24.5	89.5	64.3	85.6	84.5	87.5	73.0
3	6.4	12.5	82.5	80.1	91.3	91.2	86.7	85.3
4	5.2	9.7	93.8	84.6	93.4	92.7	93.6	88.5
5	4.2	5.4	95.6	94.1	93.9	92.8	94.7	93.4
6	4.3	5.5	95.1	95.2	93.9	92.6	94.5	93.9
8	4.3	5.2	95.1	94.2	94.0	92.8	94.6	93.5

Note: updating means the training data set is updated via the process in Figure 3, and no updating means the training data set is always the initial data set (the first 90 rings tunneling data).

summarizes the RF clogging predictions according to two training strategies. The first strategy uses the updated training data set, while the second one employs the initial training data without updating. It can be observed that the model with updated data provided better performance than the model with static data. Table 5 also shows that the RF model with static data required five to eight minutes of tunneling data to achieve the performance of the updated data model that took 4 minutes (error data = 5.2%). Figure 18 clearly shows this result.

5.2.2. Influence of Input Features on Clogging Prediction

This section analyzes the importance of the 17 features used in clogging prediction with the RF model. The permutation feature importance technology [41] was adopted in this investigation with four minutes of tunneling data. Figure 19 shows the obtained results with clogging, normal, and all states. It can be observed that only the mean and maximum values of the TOR had an impact exceeding 10%. The effect of the PR on the normal state rings exceeded 10% for the maximum values, the range values, and the standard deviations. Finally, the importance of the mean and maximum values of the TOR on clogging rings exceeded 15%, which means that this factor should be considered in clogging prediction.

6. Conclusions and Future Work

This paper presented an approach for the use of the RF method for the real-time early warning of clogging, and an evaluation of the performance of this approach. The main results of the conducted analyses are as follows:

• The SPB shield tunneling in the mudstone-rich area was frequently subjected to clogging, making the statistical characteristics of the tunneling parameters quite different from the normal conditions. Furthermore, clogging was generally accompanied by high fluctuations of SPE and TOR, and low values of AR, which was harmful to tunneling safety and efficiency;
• Data allowed us to establish the following three clogging criteria for tunnels constructed in mixed ground of round gravel and mudstone: (i) SPFI of the whole ring larger than 0.15 ($\theta >0.15$); (ii) average value of the TOR higher than 1 MN·m; and (iii) an average value of the AR smaller than 15 mm/min. Clogging occurred when two of these conditions were satisfied;
• The RF model provided a good prediction for the early warning of clogging using four minutes of tunneling data with an accuracy of 95%. In addition, the RF model yielded the best performance compared with the KNN model, the SVC model, and the MLP model;
• Feature importance analysis showed the crucial role of the TOR and PR in clogging. Furthermore, with the self-updating mechanism, the RF model can make a good prediction of clogging from the early beginning of one ring, which is beneficial for clogging early warnings.

Although we have realized some achievements in the early prediction of clogging using tunneling parameters, we still need to develop measures to mitigate clogging in mixed grounds. For example, more appropriate slurry properties designed for mixed ground containing mudstone is vital for tunnel face stability, and for conveying the excavated material smoothly. Developing intelligent models to control the tunneling parameters in the areas of high clogging risk is also necessary.