Surrounding Rock Squeezing Classification in Underground Engineering Using a Hybrid Paradigm of Generative Artificial Intelligence and Deep Ensemble Learning

Abstract

Surrounding rock squeezing is a common geological disaster in underground excavation projects (e.g., TBM tunneling and deep mining), which has adverse effects on construction safety, schedule, and property. To predict the squeezing of the surrounding rock accurately and quickly, this study proposes a hybrid machine learning paradigm that integrates generative artificial intelligence and deep ensemble learning. Specifically, conditional tabular generative adversarial network is devised to solve the problems of data shortage and class imbalance for data augmentation at the data level, and the deep random forest is built based on the augmented data for subsequent squeezing classification. A total of 139 historical squeezing cases are collected worldwide to validate the efficacy of the proposed modeling paradigm. The results reveal that this paradigm achieves a prediction accuracy of 92.86% and a macro F₁-score of 0.9292. In particular, the individual F₁-scores on strong squeezing and extremely strong squeezing are more than 0.9, with excellent prediction reliability for high-intensity squeezing. Finally, a comparative analysis with traditional machine learning techniques is conducted and the superiority of this paradigm is further verified. This study provides a valuable reference for surrounding rock squeezing classification under a limited data environment.

1. Introduction

With the rapid development of social economy and the accelerating advancement of urbanization, the demand for all kinds of infrastructure has presented a significant growth trend, which undoubtedly brings unprecedented challenges to infrastructure construction [1,2]. In particular, tunnel engineering is an indispensable part of water conservancy, transportation, mining, and other fields, and its construction scale and technical difficulty are rising [3]. Today’s tunnels are not only designed to be deeper and longer but also increasingly complex in shape and compact in layout, which means that their construction needs to face harsher geological conditions and construction environments [4]. Weak, fractured, or high-stress areas are all important safety factors that cannot be ignored in tunnel construction, which greatly increase the construction risk [5,6].

Surrounding rock squeezing is a geological disaster frequently occurring in tunnel construction [7,8]. The squeezing of the surrounding rock causes the deformation of the tunnel structure, such as arch subsidence and waist convergence around the tunnel [9]. At the same time, the supporting structure may also be subjected to excessive pressure and fail as a result of the pressure, which will make the tunnel structure lose stability, further aggravate the deformation of the tunnel structure and even lead to its collapse [10].

To predict the risk of surrounding rock squeezing, many researchers have proposed various empirical criteria derived from engineering experience and case studies. For instance, Aydan et al. [11] suggested using the ratio of uniaxial compressive strength of intact rock to vertical ground stress, with a threshold value of two to differentiate between squeezing and non-squeezing conditions. Similarly, Barla [12] introduced a criterion based on the ratio of uniaxial compressive strength of rock mass to vertical ground stress. These criteria mentioned above essentially fall under the category of strength-stress ratios. Additionally, Farhadian et al. [13], Singh et al. [14], and Jimenez et al. [15] developed squeezing criteria that consider extra factors such as burial depth and surrounding rock quality. These criteria have been widely applied to numerous tunnel projects, and their effectiveness has been validated.

As research has progressed, it has become evident that surrounding rock squeezing is a complicated nonlinear problem controlled by multiple trigging agents [7,16,17,18]. Consequently, some researchers have turned to machine learning techniques to predict the squeezing potential of the surrounding rock, given their strong capabilities to handle complex nonlinear problems [19,20,21]. For example, Zhang et al. [22] developed a back-propagation neural network model for squeezing prediction, using burial depth, excavation diameter, strength-stress ratio, rock mass quality index, and support stiffness as input parameters. Similarly, Bo et al. [23] built a logistic regression model to achieve squeezing classification based on the above five parameters, while Feng et al. [24] and Azizi et al. [25] proposed Bayesian network models for the same purpose. Additionally, support vector machine [26,27,28,29], decision tree [30,31], and probabilistic neural network [8] have also been popular methods for predicting the squeezing potential of the surrounding rock. These practises mentioned above fully demonstrate the effectiveness of machine learning techniques in the prediction of surrounding rock squeezing.

However, the performance of machine learning models is highly sensitive to data quality [32,33]. Class imbalance and data shortage are two common data issues. In the context of surrounding rock squeezing, the collection of squeezing cases is difficult, so the available data are usually limited, especially for strong and extremely strong squeezing cases [34]. On the one hand, it exacerbates the data shortage problem. On the other hand, it also aggravates the data imbalance of different squeezing grades. When data are scarce, it is difficult for machine learning models to effectively capture the mapping relationships contained in the data. Additionally, when the classes are imbalanced, machine learning models have a low generalization level for the minority.

On the basis of considering data shortage and class imbalance, this study proposes a hybrid intelligent modeling paradigm for surrounding rock squeezing classification. Firstly, conditional tabular generative adversarial network is designed to cope with data shortage and class imbalance at the data level. After that, the deep random forest is built based on the augmented data to capture the mapping relationship between the trigging agents and the squeezing grades. Finally, the generalization performance of the devised model is evaluated quantitatively using an independent test set and two professional metrics. The rest of this paper is organized as follows: Section 2 introduces the proposed hybrid intelligent modeling paradigm in detail; Section 3 describes the collected database and conducts data augmentation; Section 4 analyzes the generalization performance of the model and makes some comparisons with other common models; and Section 5 summarizes the main conclusions.

2. Proposed Hybrid Modeling Paradigm

This study aims to propose a hybrid modeling paradigm to solve the problems of data shortage and class imbalance and build a model for surrounding rock squeezing classification. Particularly, the task of data augmentation is undertaken by conditional tabular generative adversarial network, and the deep random forest is designed to learn the mapping relationship from the augmented data to achieve squeezing classification. In order to clearly demonstrate the proposed hybrid modeling paradigm, Section 2.1 describes conditional tabular generative adversarial network, Section 2.2 introduces deep random forest, Section 2.3 depicts the comparison scheme used to verify the advantages of the proposed hybrid modeling paradigm, and Section 2.4 illustrates how to evaluate the generalization performance of the model quantitatively. Figure 1 gives the research framework of this study.

2.1. Conditional Tabular Generative Adversarial Network

Conditional tabular generative adversarial network (CTGAN), as the representative algorithm of generative artificial intelligence, is a variant of generative adversarial network (GAN) [35,36]. The CTGAN has two characteristic properties. One is data preprocessing; that is, all continuous features are processed through model (i.e., variational Gaussian mixture model)-specific normalization. The other is conditional training, in which category-level unbalance of categorical features is addressed.

Resembling the GAN, the CTGAN includes the following two modules (Figure 2): generator and discriminator. The aim of the generator is to learn the conditional distribution of real data and then generate synthetic data. The purpose of the discriminator is to distinguish the difference between real data and synthetic data. When synthetic data are very close to real data, the discriminator will output a large score. Instead, when synthesized data differs significantly from real data, it will output a small score. In general, the generator is optimized for deceiving the discriminator as much as possible. If the discriminator cannot tell the difference, then the CTGAN is considered as a finely trained model that can be used to generate synthetic data.

To reserve the essential information of real data and generate effective synthetic data, the loss functions of the generator and discriminator are designed separately. For the generator, the loss function $L_G$ is expressed by the following equation:

$$L_G=-{\displaystyle \frac{1}{n/N}}{\displaystyle {\sum }_{i=1}^{n/N}D\left({\widehat{\nu }}_i,{\kappa }_i\right)}+{\displaystyle \frac{1}{n}}{\displaystyle {\sum }_{j=1}^{n}CE\left({\tau }_j,{\epsilon }_j\right)}$$

(1)

For the discriminator, the loss function $L_D$ is expressed by the following equation:

$$L_D={\displaystyle \frac{1}{n/N}}{\displaystyle {\sum }_{i=1}^{n/N}D\left({\widehat{\nu }}_i,{\kappa }_i\right)}-{\displaystyle \frac{1}{n/N}}{\displaystyle {\sum }_{i=1}^{n/N}D\left({\nu }_i,{\kappa }_i\right)}$$

(2)

where n is the batch size; N is the packing size used to handle model collapse; $\nu$ is the real data; $\widehat{\nu }$ is the synthetic data; $\kappa$ is the conditional vector; $\tau$ is the one-hot conditional vector; $\epsilon$ is the mask vector associated with $\tau$; L is the loss function; G is the generator; D is the discriminator; and CE is the cross entropy.

In the CTGAN’s training process, the generator and discriminator are optimized and updated by an Adam optimizer. Based on real data and synthetic data, the $L_D$ can be calculated, based on which the discriminator is updated first. Then, synthetic data are delivered to the updated discriminator and the $L_G$ can be calculated, based on which the generator is updated. Thus, when building the CTGAN, the discriminator is updated followed by the generator.

2.2. Deep Random Forest

The random forest (RF) is an ensemble learning method for classification and regression tasks [37]. The RF combines multiple single predictors into one integrated predictor, so as to attain excellent generalization performance. These single predictors can be generated in parallel and allow for diversity among each other. Generally, the RF uses binary decision trees, such as classification and regression tree (CART), as single predictors, due to the low parameters required, strong robustness, and low training cost [38]. Figure 3 illustrates the topology of the RF.

Inspired by the layered structure of deep learning, the deep random forest (DRF) is proposed [39]. As illustrated in Figure 4, the DRF creates multiple layers. For each layer, multiple RFs are assembled. Through cascading connections between different layers, the potential features in the input data are represented and learned. For the first layer of the DRF, it receives the original input data as the input, and correspondingly, each RF in this layer will generate a class probability vector, denoted as $\mathit{p}=\left(p_1,p_2,\cdots ,p_n\right)$. In this class probability vector, $p_n$ represents the probability that the sample belongs to the n-th class, and the dimension is equal to the number of classes set by the classification tasks. For subsequent layers, the input is composed by two parts: the original input data and the class probability vectors output by random forests in the previous layer. Thus, the dimension of the input is characterized by the following formula:

$$di{m}_{input}^i=di{m}_{data}+n\times m^{i-1}$$

(3)

where $di{m}_{input}^i$ is the input dimension of the current layer; $di{m}_{data}$ is the dimension of the original input data; n is the number of classes set by the classification tasks; $m^{i-1}$ is the number of RFs contained in the previous layer; and i and i − 1 are the serial numbers of the layers.

In the last layer of the DRF, the class probability vectors of all RFs here will be averaged to obtain the final class probability vector, as shown in Equation (4). The index corresponding to the maximum in the final class probability vector is the class of the sample predicted by the DRF.

$$ve{c}_0={\displaystyle \frac{1}{U}}{\displaystyle {\sum }_{k=1}^{U}ve{c}_k}$$

(4)

where $ve{c}_0$ is the final class probability vector, output generated by the last layer of the DRF; $ve{c}_k$ is the k-th class probability vector, output generated by the k-th RF of the last layer; and U is the number of RFs contained in the last layer.

It is well known that the DRF has a strong ability of representation learning [40]. By cascading RFs between multiple layers, the DRF enables fine-grained modelling of data. On the one hand, it maximizes data value. On the other hand, the separability of samples in the mapping space is improved by mining deep-level features. Therefore, the DRF is favourably popular in solving practical engineering problems.

2.3. Benchmark Comparison

To demonstrate the superiority of the proposed model, referred to as the CTGAN-DRF, five frequently used machine learning algorithms, including the back-propagation neural network (BPNN) [41], support vector machine (SVM) [42], Gaussian process classifier (GPC) [43], Naïve Bayesian classifier (NBC) [44], and k-nearest neighbor (KNN) [45], are chosen for benchmark comparison.

The BPNN is an algorithm for forward propagation and back propagation by graph computation. In forward propagation, the input data are passed layer by layer through the network, and the output result is finally obtained. In back propagation, according to the error between the output result and the desired target, the gradient of the parameters of each layer is calculated using the chain rule, and the network parameters are updated to reduce the error.

The SVM is a supervised learning algorithm mainly used for classification and regression tasks. It achieves efficient data classification by finding the best separated hyperplane in the feature space. Moreover, it performs well when dealing with high-dimensional data, nonlinear problems, and few training samples.

The GPC is a supervised classification algorithm based on Bayesian framework. It assumes that there is an implicit function for a Gaussian process, which corresponds monotonically to the probability that the sample belongs to a certain class. By selecting an appropriate likelihood function to describe this mapping relationship, the GPC can calculate the posterior probability of the sample and achieve classification prediction. The GPC has the characteristics of automatic parameter optimization and can give the classification result in probabilistic form.

The NBC is a classification algorithm based on Bayes theorem. It assumes that the sample features are independent of each other, calculates the posterior probability of the sample class using Bayesian formula according to the known prior probability and the probability distribution of the sample class, and then takes the class of the sample with the largest posterior probability as the output of the classifier. The NBC has the advantages of being a stable mathematical model, with good classification effect, and needing few training parameters.

The KNN is a supervised learning algorithm for classification and regression tasks. The classification principle states that if most of the K nearest neighbors near the sample belong to a certain class, then this sample also belongs to that class. The KNN is simple and flexible, without requiring training process, and can be applied to various data types.

2.4. Evaluation Metric

In order to quantitatively evaluate the generalization performance of the model, we select two evaluation metrics, namely accuracy and macro F₁-score. The accuracy describes the ability of the model to correctly classify the samples without considering class differences, calculated as follows:

$$\mathrm{accuracy}={\displaystyle \frac{{\displaystyle {\sum }_{k=1}^K{u}_{k,k}}}{{\displaystyle {\sum }_{i=1}^K{\displaystyle {\sum }_{j=1}^K{u}_{i,j}}}}}$$

(5)

where K is the number of sample classes and $u_{i,j}$ is the number of samples belonging to the i-th class but predicted as the j-th class.

Different from the accuracy, the macro F₁-score takes into account the model’s local generalization performance on each class. The macro F₁-score is defined as the average of all individual F₁-scores on various classes, calculated as follows:

$$\mathrm{macro}\hspace{0.33em}{F}_1\text{-}\mathrm{score}={\displaystyle \frac{1}{K}}{\displaystyle {\sum }_{k=1}^K{F}_1\text{-}{\mathrm{score}}_k}$$

(6)

where $F_1\text{-}{\mathrm{score}}_k$ is the F₁-score on the k-th class. It is calculated as follows:

$$F_1\text{-}{\mathrm{score}}_k={\displaystyle \frac{2\cdot Pr{e}_k\cdot Re{c}_k}{Pr{e}_k+Re{c}_k}}$$

(7)

where $Pr{e}_k$ and $Re{c}_k$ are the precision and the recall on the k-th class, respectively. They are calculated as follows:

$$Pr{e}_k={\displaystyle \frac{u_{k,k}}{{\displaystyle {\sum }_{i=1}^K{u}_{i,k}}}}$$

(8)

$$Re{c}_k={\displaystyle \frac{u_{k,k}}{{\displaystyle {\sum }_{j=1}^K{u}_{k,j}}}}$$

(9)

3. Data Collection and Augmentation

3.1. Data Collection

To verify the effectiveness of the proposed approach, we collected 139 cases from 42 tunnel projects worldwide to build the database (see Supplementary Materials) [46]. Each case contains five input parameters, including burial depth, excavation diameter, strength-stress ratio, rock mass quality index, and support stiffness. Regarding their acquisition, the detailed description can be found in [46]. In essence, the occurrence of surrounding rock squeezing is closely related to excavation scale, ground stress, rock strength, rock mass structure, and support conditions [29]. Among the input parameters, burial depth is an indicator associated with ground stress, excavation diameter reflects the excavation scale, strength-stress ratio is a composite index integrating rock strength and ground stress, rock mass quality index emphasizes rock mass structure, and support stiffness characterizes applied support conditions. Figure 5 presents the visual statistical distribution of various parameters in this database.

A personalized label, namely squeezing strength of surrounding rock, is used as the output parameter of each case. Based on the relative strain of the surrounding rock, the squeezing strength is divided into five grades (

Table 1. Classification criteria of surrounding rock squeezing.

Squeezing Grade of the Surrounding Rock	Relative Strain of the Surrounding Rock
No	<1.0%
Mild	1.0%~2.5%
Moderate	2.5%~5.0%
Strong	5.0%~10.0%
Extremely strong	>10.0%

) [8,47]. When the relative strain of the surrounding rock is <1.0%, 1.0%~2.5%, 2.5%~5.0%, 5.0%~10.0%, and >10.0%, it indicates no squeezing, mild squeezing, moderate squeezing, strong squeezing, and extremely strong squeezing, respectively. In the established database, there are 28 cases with no squeezing (accounting for 20.14%), 37 cases with mild squeezing (accounting for 26.62%), 33 cases with moderate squeezing (accounting for 23.74%), 26 cases with strong squeezing (accounting for 18.71%), and 15 cases with extremely strong squeezing (accounting for 10.79%), as shown in Figure 6.

In order to ensure the representativeness of the test set, the ratio of training set to test set is taken as 7:3; that is, the training set has 97 cases, and the test set has 42 cases. In this way, there are enough test samples to reliably test the generalization performance of the model. In addition, the samples of different classes in the test set should be as balanced as possible. Based on these guidelines, we randomly select 8 cases with no squeezing, 8 cases with mild squeezing, 8 cases with moderate squeezing, 9 cases with strong squeezing, and 9 cases with extremely strong squeezing from the database to construct the test set. Correspondingly, the rest cases in the database are used as the training set, including 20, 29, 25, 17, 6 cases with no squeezing, mild squeezing, moderate squeezing, strong squeezing, and extremely strong squeezing, respectively.

3.2. Data Augmentation

In order to avoid disturbance to the test set and ensure the independence of the test set, data augmentation is only implemented on the training set. Firstly, the CTGAN is trained based on the true data (i.e., the training set). After that, the trained CTGAN is used to generate the synthetic data. On the one hand, the class imbalance is eliminated. On the other hand, the data shortage issue is remedied. After data augmentation, the number of cases with no squeezing is increased from 20 to 200. Similarly, the number of cases with mild squeezing, moderate squeezing, strong squeezing, and extremely strong squeezing is also increased to 200, respectively.

Table 2. Data augmentation overview.

Squeezing Intensity	Sample Size
	Before Data Augmentation	After Data Augmentation
No	20	200
Mild	29	200
Moderate	25	200
Strong	17	200
Extremely strong	6	200

summarizes the changes in the training set after data augmentation. Correspondingly, Figure 7, Figure 8, Figure 9, Figure 10 and Figure 11 display the kernel density distribution of each parameter before and after data augmentation under varying squeezing grades.

Correlation coefficient matrix is used to evaluate the quality of the synthetic data. The higher the similarity between the correlation coefficient matrices of the synthetic data and the true data, the better the quality of the synthetic data, which represents that the synthetic data have more consistent underlying topology with the true data. For the “no squeezing” class, Figure 12a,b illustrate the correlation coefficient matrices of the true data and the synthetic data, respectively, and Figure 12c calculates the absolute difference between the correlation coefficient matrices of the true data and the synthetic data. In particular, the value in each box of Figure 12a,b is characterized by the Pearson correlation coefficient, which is a commonly used indictor to measure correlation. Figure 13, Figure 14, Figure 15 and Figure 16 present the relevant analysis results for the “mild squeezing” class, “moderate squeezing” class, “strong squeezing” class, and “extremely strong squeezing” class, respectively. As shown in Figure 12c, all absolute differences are less than 0.2, which mean that the synthetic data have a similar internal topology to the real data, and the CTGAN attains a high-level capability to learn from the real data and is capable of producing high-quality synthetic data. In addition to the no-squeezing samples, the consistent pattern can be found for both the moderate-squeezing samples and strong-squeezing samples when utilizing the CTGAN to generate the synthetic data, as presented in Figure 14c and Figure 15c, respectively. However, there are some anomalies for the mild-squeezing samples and extremely strong-squeezing samples. For example, as demonstrated in Figure 13c, the absolute difference in correlation coefficients between burial depth and excavation diameter for the real data and the synthetic data slightly exceeds 2, reaching 0.2071; and as indicated in Figure 16c, the absolute difference in correlation coefficients between burial depth and strength-stress ratio for the real data and the synthetic data is 0.3099. Nevertheless, most of the differences are within 0.2. Overall, the CTGAN has a superior data augmentation ability for various squeezing intensities.

4. Results and Discussion

4.1. Prediction Results of CTGAN-DRF Model

After data augmentation using the CTGAN, the DRF is trained using the augmented data. Correspondingly, the built model is referred to as the CTGAN-DRF. Based on the independent test set, the generalization performance of the CTGAN-DRF is tested quantitatively. In particular, the training and testing of the model are executed on a personal computer with Windows 10, i7-9750H CPU, 16GB RAM, and RTX 2060. Figure 17 depicts the confusion matrix of the CTGAN-DRF on the test set. In this confusion matrix, the sum of the values in each row is equal to the number of test samples of a specific class. From the bottom up, the sum of each row is eight, eight, eight, nine, and nine, which indicates that the number of test samples belonging to no squeezing, mild squeezing, moderate squeezing, strong squeezing, and extremely strong squeezing is eight, eight, eight, nine, and nine, respectively.

Particularly, the values on the diagonal of the confusion matrix denote the number of correctly predicted test samples, and those on the non-diagonal represent the number of wrongly predicted test samples. In terms of no squeezing, mild squeezing, moderate squeezing, strong squeezing, and extremely strong squeezing, the number of correctly predicted test samples by the CTGAN-DRF is eight, seven, eight, eight, and eight, respectively, accounting for 100.00%, 87.50%, 100.00%, 88.89%, and 88.89% of each corresponding class of test samples. For mild squeezing, there is one wrongly predicted test sample out of eight test samples, and the predicted squeezing grade is no squeezing. For both strong squeezing and extremely strong squeezing, there is also one incorrect prediction. Regrading to both the former and the latter, the misclassified label is moderate squeezing.

Relying on the confusion matrix, the accuracy and macro F₁-score of the CTGAN-DRF are calculated using Equation (5) and Equation (6), respectively. The results reveal that the accuracy of the CTGAN-DRF is 92.86%, and the macro F₁-score reaches a score of 0.9292, as shown in Figure 18a. Furthermore, the F₁-score of the CTGAN-DRF on each individual class is analyzed, as illustrated in Figure 18b. It can be observed that the F₁-scores on strong squeezing and extremely strong squeezing exceed 0.9, which means that the CTGAN-DRF attains superior prediction reliability for high-grade squeezing disasters and makes significant contribution to squeezing management.

4.2. Comparison Analysis with Benchmark Models

In this section, we compared the proposed CTGAN-DRF with the BPNN, SVM, GPC, NBC, and KNN in terms of the prediction performance of squeezing disaster. The confusion matrices of the BPNN, SVM, GPC, NBC, and KNN are shown in Figure 19, based on which the accuracy and macro F₁-score of the BPNN, SVM, GPC, NBC, and KNN are calculated according to Equation (5) and Equation (6), respectively. Referring to Figure 20, in comparison with the BPNN, SVM, GPC, NBC, and KNN, the accuracy of the CTGAN-DRF is increased by 45.24%, 40.48%, 47.62%, 50.00%, and 57.14%, respectively, and the macro F₁-score is raised by 0.4579, 0.4075, 0.4812, 0.5003, and 0.5834, respectively. Based on these strong evidences, the superiority of the CTGAN-DRF is demonstrated.

5. Conclusions

Surrounding rock squeezing is a kind of geological disaster that occurs frequently in underground construction, posing a serious threat to safety, schedule, and property. The physical mechanism of surrounding rock squeezing is complex, whose occurrence is influenced by multiple factors and exhibits high nonlinearity with them. In order to establish the mapping relationship between the squeezing and the influencing factors, machine learning is introduced. Aiming at data shortage, class imbalance, and nonlinear topology, this study proposes a hybrid machine learning modeling paradigm. In this paradigm, conditional tabular generative adversarial network (CTGAN) is designed to achieve data augmentation, and the deep random forest (DRF) is integrated to construct the nonlinear topology from the augmented data. Consequently, the CTGAN-DRF model is developed for surrounding rock squeezing classification. To validate the efficacy of the developed model, a total of 139 historical squeezing cases are collected. The results indicate that the CTGAN-DRF model achieves a prediction accuracy of 92.86% and macro F₁-score of 0.9292. In particular, the individual F₁-scores on strong squeezing and extremely strong squeezing are more than 0.9, which means that the CTGAN-DRF model has excellent prediction reliability for high-intensity squeezing. Additionally, in comparison with other common machine learning models, such as back-propagation neural network, support vector machine, Gaussian process classifier, Naïve Bayesian classifier, and k-nearest neighbor, the CTGAN-DRF model exhibits an advantage in its generalization ability.

Overall, this study leverages the geological, support, and construction data under a specific environment to evaluate the potential surrounding rock squeezing tendency. The developed model is effective and meets the construction requirement, and it can provide the prediction results of squeezing tendency to assist engineers in taking advanced measures so as to avoid squeezing disaster. In particular, it also considers data shortage and class imbalance in the modelling, providing a meaningful reference for surrounding rock squeezing classification and other similar engineering problems with limited data quality. However, this study still has some limitations. First, due to the difficulty of real-time acquisition of input parameters, the model developed here is biased towards long-term predictions. If we plan real-time predictions, the real-time monitoring surrounding rock deformation data can be considered as the input to establish the model. Second, both the training and validation of the current model are executed based on the collected historical squeezing cases. In future research, we will deploy this model in real projects to further examine its performance under new engineering environments. If necessary, incremental learning can be used to update the current model. Third, exploring more data mining and feature characterization techniques is still an important task in future research, which can improve data utilization and decrease data loss in the modeling, especially when there is limited data available. Finally, it should be noted that surrounding rock squeezing is a very complex, nonlinear process, and its occurrence is controlled by many factors. In the future, we need to pay more attention to establish the input system and take into account as many squeezing-inducing factors as possible. At the same time, the investigation on novel squeezing prediction models, such as fuzzy logic predictor, is also necessary for squeezing management.