Risk classification assessment and early warning of large deformation of soft rock in tunnels based on CNN-LSTM model Authors: Xianmeng Zhang, Wang Minghao, Dan Feng, Wang Jingchun Published: 2024 ABSTRACT AbstractTunnels in the Western Plateau region of China often face numerous problems of high ground stress, as well as highly fractured and weak rock masses in deep sections of the tunnel, which results in an increased risk of large soft rock deformation in tunnels, posing substantial challenges to safe and efficient construction. Hence, risk assessment and establishment of an early safety warning system are essential. By examining factors affecting the large soft rock deformation in tunnels, a set of index system consisting of three primary indicators (i.e., hydrogeological conditions, design factors, and construction factors) and eight secondary indicators (i.e., strength-to-stress ratio, rock mass index, groundwater influence, groundwater influence, etc.) for assessing these risks is constructed, while the risk classification and identification criteria are also established. As an example, a tunnel on the Xicheng Railway is adopted for regression prediction of rock mass deformation and risk warning classification based on CNN-LSTM model. The results indicate that the model is effective in predicting rock mass deformation with minimal fluctuations in settlement data, a MAPE of 0.29525, and a RMSE of 6.1311. The model demonstrates a high accuracy rate of 91.67% for early risk warning prediction of soft rock deformation in the tunnel, moreover, when applied to 10 cross-sections of the Xicheng Railway Tunnel for deformation risk early warning, the model demonstrated a high prediction accuracy of 90%. which serves as a reference for the assessment and prediction of large deformation risks in similarly complex and challenging tunnels in the western region of China.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-024-81816-x.Keywords: Tunneling, Large soft rock deformation, Convolutional neural network (CNN), Long short-term memory neural network (LSTM), Risk assessmentSubject terms: Civil engineering, Information technology Introduction China’s western mountainous regions are characterized by complex terrain and frequent geological hazards, which significantly affects the construction safety of tunnels. Due to high geopathic stress, weak rock mass, groundwater, fault fracture zones, and other complex geological conditions, the construction of tunnels is prone to large rock mass deformation hazards, which is defined as “rock mass deformation that is significantly greater than the normal level and may lead to adverse outcomes”1. Large deformations in soft rock tunnels can result in severe accidents such as vault collapse, face instability, and support system failure, which potentially causes casualties and property damage during construction and poses an extremely high safety risk2. Therefore, conducting risk assessment on high geo-stress soft rock tunnels is essential to make risk predictions and early warnings for potential large soft rock deformations. Researchers worldwide have conducted comprehensive studies on the large deformation of rock mass. Yang et al.3 employed the discrete element simulation method to simulate the large deformation occurring during tunnel excavation and to characterize the damage modes in both unsupported and primary-support-equipped tunnels. They devised an improved backstop support control scheme based on the simulation results. Similarly, Yang et al.4 utilized the UDEC software to simulate the substantial deformation associated with ventilation tunnel construction through discrete element simulation and to propose a novel “bolt-cable-net-sprayed concrete + shell” composite support system for the ventilation tunnel stabilization. Liu et al.5,6conducted long-term, real-time monitoring of structural stress in tunnels with severe deformation, developed four stress-time regression models and compared the cross diaphragm (CRD) method with the three-bench annular excavation method with reserved core soil in extremely soft and fractured rock, which advanced methods for controlling large deformations. Lei and Chao7 explored the principle of circumferential yielding support through the lens of energy conversion during the tunnel excavation-support process. They analyzed the primary factors influencing structural deformation in soft rock tunnels with significant deformations and their bearing structures using a structural mechanics analytical methodology. Additionally, they employed finite element analysis software to compare the mechanical characteristics of a traditional bearing system with those of the circumferential yielding bearing system. Zhang et al.8 analyzed factors influencing tunnel rock mass deformation by performing numerical simulations with discrete element software, and then identified key factors affecting thin-layer slate tunnel deformation such as joint inclination, depth, water absorption, and rock mass softening, etc. Deng et al.9 investigated the geology and deformation properties of the large deformation zones in Da-ping Mountain Tunnel, and determined the creep parameters of rock mass in large deformation zones by conducting creep tests. Also, they assessed the geo-stress conditions within these zones through displacement inversion analysis, and confirmed the effectiveness of the proposed optimization strategies for tunnel construction. Han et al.10 examined the squeezing deformation mechanism of soft rock tunnel from the perspective of stress relaxation and plastic rheology. They proposed a deformation control principle centered on “multi-stage and hierarchical collaborative control”, in which the degrees of rock mass deformation were categorized into five levels: slight, moderate, substantial, severe, and extremely severe. He et al.11identified 15 risk factors influencing large deformations in soft rock tunnels through case studies, and devised a risk assessment model for these deformations by utilizing the game theory-cloud model. Kang and Wang12 outlined the principles of large deformation in rock masses and the mechanical prediction methods related to tunneling projects, and introduced a Support Vector Machine (SVM) based prediction model for tunnel rock mass deformations. Yang et al.13 developed a prediction model for large deformations in laminated soft rock tunnels using the Particle Swarm Optimization- Support Vector Machine (PSO-SVM) algorithm based on the analysis of deformations in such strata structures. Thereafter, the model was implemented in real tunnel scenarios to enhance project expertise by selecting key factors such as rock and soil compressive strength, laminae inclination, and initial geo-stress conditions among the six indicators as parameters. Ma et al.14 implemented an innovative MIC-LSTM machine learning algorithm designed to control deformation in soft rock tunnels subject to extremely high geo-stress, which included its successful deployment in real-world scenarios involving high and extremely high geo-stress during construction. Xu et al.15 combined the LSTM algorithm with numerical simulation and, after training with parameters including uniaxial compressive strength, ground stress, and joint spacing among others, they introduced a novel deformation prediction approach specifically designed for large cross-section tunnel engineering. Liu et al.16 evaluated the strengths and weaknesses of three algorithms—Gaussian process regression, Support Vector Machines, and Long Short-Term Memory Networks—in predicting perimeter rock deformation in hydraulic tunnels, and determined that the SVM model outperformed the others, thus laying a theoretical foundation for predicting perimeter rock deformation. He et al.17 introduced a nonlinear deformation prediction method for surrounding rock utilizing machine learning, based on studies of the surrounding rock deformation from two actual tunnels. Following comparative analysis of the GP, SVM, and WNN models, they concluded that the SVM model exhibited superior performance in predicting settlements and convergence deformations. Qiu et al.18 used the Delphi method and cloud model theory to study the deformation of rocks around tunnels, established a grade standard for the deformation of surrounding rock in high-speed railway tunnels, and analyzed the key factors affecting the deformation grade of surrounding rock. He et al.19 analyzed a case of large deformation in soft rock tunnels and proposed 15 factors affecting the risk of large deformation in soft rock tunnels. Bai et al.20 employed Data Envelopment Analysis (DEA) to evaluate the efficiency of 12 factors influencing the deformation of soft rock tunnels. The results indicate that moisture content of the soft rock is the most critical factor, followed by tunnel depth, uniaxial compressive strength, excavation method, strength-stress ratio, and groundwater seepage conditions. Current research on large deformations of rock masses primarily focuses on the damage mechanisms of soft rock and the efficacy of support measures. Risk assessments are typically performed using traditional safety evaluation methods, while machine learning algorithms are most often utilized for deformation prediction of soft rock tunnels21. Research on risk prediction for these deformations is limited. Evidences demonstrate the significant time-dependence of soft rock deformation in western mountain tunnels. Therefore, this study employs an integrated approach of the convolutional neural networks (CNNs) and the long short-term memory networks (LSTMs) to develop the machine learning model. The comparison of the effects of specific risk prediction methods is described in Table 1. An assessment system for large deformations in soft rock tunnels is developed, and its effectiveness in predicting rock mass deformations is verified by actual tunnel monitoring data, thereby enabling risk classification and early warning for significant deformations in soft rock tunnels. Table 1Comparison of risk prediction methods for large deformation of soft rock.Prediction methodIntroduction descriptionTraditional methods:cloud model, etc.The method is outdated, requires extensive computation time, involves significant manpower, and the model’s prediction accuracy is poor.Machine learning:SVMThe method features fast operation time, excellent linear data processing capability, higher prediction accuracy compared to traditional methods, and can be integrated with monitoring and early warning modules.Machine learning:CNN-LSTMThe method demonstrates good processing effect for nonlinear time series data and achieves high prediction accuracy. However, its operation time is long. It can also be integrated with monitoring and early warning modules. Comparison of risk prediction methods for large deformation of soft rock. Traditional methods: cloud model, etc. Machine learning: SVM Machine learning: CNN-LSTM Machine learning models Convolutional neural network A Convolutional Neural Network (CNN) is a specialized type of artificial neural network that distinguishes itself from other models through its signature feature of the convolution operation22–25. Whilst the accuracy of CNN models is undermined by their limited generalization capability when dealing with time series data, they excel at feature extraction. Assessing the risk of significant deformations requires an appropriate set of indicators, each of which represents a distinct feature. Given that data of indicators influencing significant deformations in soft rock tunnels are one-dimensional and lack spatial context, this paper employs a one-dimensional (1D) convolutional approach for the CNN model. Figure 1 illustrates the structure of the 1D CNN model. CNN model flowchart. Long short-term memory network Long Short-Term Memory networks (LSTMs) are a specialized form of Recurrent Neural Networks (RNNs). Since a standard RNN can be considered as a multi-layered feedforward neural network with feedback connections, the extensive historical information carried by long sequences can lead to the vanishing and exploding gradients during training26. LSTMs are one of the most extensively utilized deep learning models, which are renowned for their robustness in handling long-term dependencies and their superior performance in processing time series data27. The rock mass deformation in tunnels exhibits typical time series characteristics, allowing the LSTM model to be well suited for predicting such deformations. An LSTM network mainly consists of a forget gate, an input gate, and an output gate, where the cell state of the previous step is denoted as (C_{t-1}), the hidden state is denoted as (), and the current input is denoted as (). The cell state updates its values after passing through the forget gate, and the tanh function is used at the input gate to filter information relevant to the current context. Then () is combined with the modulation of the output gate via a tanh function to produce the next hidden state, where () represents the sigmoid activation function28. Figure 2 depicts the specified LSTM model. LSTM model diagram. Convolutional neural networks-long short-term memory The convolutional neural networks-long short-term memory (CNN-LSTM) model is effective in both regression prediction and classification tasks. The CNN layer is responsible for extracting and selecting features from the risk assessment data of large soft rock deformation, while LSTM layers are tasked with the prediction of time series data, which then estimate the final extent of rock mass deformation. Based on the risk criteria for rock mass deformation, the model assesses the risk level of large soft rock deformation in tunnels. The model consists of two components. One is the CNN convolutional network, which consists of convolutional layers and pooling layers that perform feature extraction on the training data. The other one is the LSTM model, which processes the features extracted by the CNN to forecast the rock mass deformation. The number of convolutional layers and the number of convolutional kernels within the CNN affect both computational time and complexity. Therefore, it adopts a dual-layer configuration with kernel dimensions of 2 × 1, 32 convolutional kernels, and a maximum pooling strategy for pooling layer. The LSTM utilizes a single layer with 128 neurons with a loss rate of 0.25, and employs ‘adam’ as the optimizer. Also, it uses a regularization coefficient with a value of 1e-3, where the loss function adopts Mean Squared Error (MSE) and the activation function uses Softmax. The detailed framework of the CNN-LSTM model illustrated in Fig. 3. Structural diagram of CNN-LSTM. The developed CNN-LSTM model is utilized to perform regression analysis for predicting tunnel rock mass deformation and for conducting early warning classification for large deformation risks in soft rock tunnels. The prediction results are compared with those from independent CNN and LSTM models to validate the effectiveness of integrated model. Model evaluation metrics Regarding the performance metrics of the model, to quantify the prediction accuracy of the model, the performance indicators of Mean Absolute Percentage Error (MAPE) and Root Mean Square Error (RMSE) commonly utilized in machine learning for regression analysis of rock mass deformation are employed. Their respective formulas are presented in Eq. (1) and Eq. (2). For early warning prediction of large soft rock deformation, the standard performance metrics include Accuracy (A), the formulation of which is detailed in Eq. (3). In Eq. (1) and Eq. (2), is the i-th actual value; is the i-th predicted value; is the sample size. In Eq. (3), is the number of correct predictions by the model that align with the actual classifications; is the number of incorrect predictions by the model that are actually correct; is the number of times the model correctly predicts the level, but the actual classification is incorrect; and indicates that both the model’s prediction and the actual classification are incorrect. Risk assessment system for large deformation in soft rock tunnels Risk classification of large deformation in soft rock tunnels Combined with the Code for Tunnelling in Squeezing Rocks (Q/CR 9512 − 2019)29, Technical Specification for Monitoring Measurement of Railway Tunnel (Q/CR 9218 − 2024)30, and Guidelines related to the design and construction of high ground stress soft rock tunnels, the classification of large deformations for actual projects is investigated. The large deformation classification for soft rock tunnels mainly includes the strength-to-stress ratio, the relative deformation amount, and the deformation potential31,32.The risk classification criteria for large deformation in soft rock tunnels based on the deformation potential and the relative deformation amount are developed. Based on deformation potential and relative deformation amount, the large deformation is classified into grades I-IV. The relative deformation and the deformation potential are more frequently employed in field construction. For risks in levels III-IV, it is recommended that construction teams perform additional ground stress assessment, calculate strength-to-stress ratio, and ascertain large deformation level. Quantitative criteria for identifying and controlling deformation are presented in Table 2, and the proposed classification method for large deformation serves as a reference for the risk assessment of similar tunnels in the future. Table 2Classification of large deformations by deformation potential and relative deformation amount.Risk levelIIIIIIIVRelative deformation1–2.5%2.5–5%5–10%> 10%Metamorphic potentialMinimalMediumIntenseExtremely intenseDeformation rateVp (mm/d)10 ~ 3030 ~ 5050 ~ 80≥ 80Deformation rate classificationLow rateMedium rateHigh rateExtremely high rate Classification of large deformations by deformation potential and relative deformation amount. Deformation rate Vp (mm/d) Indicators for assessing large deformations in soft rock tunnels High ground stress fields and presence of soft rock masses are the primary factors contributing to large soft rock deformations. However, ground stress measurements in projects are often rarely performed as they are time-consuming and not mandatory. The relevant rock mass parameters can be found in the geological investigation report33. Therefore, review of collected 120 typical tunnel cases of large soft rock deformation worldwide identifies that hydrogeological conditions, design factors, and construction factors are the main contributors to these large soft rock deformation hazards. Therefore, a risk assessment index system for soft rock large deformation in tunnels is developed, including three main indicators of hydrogeological conditions, design factors, and construction factors as well as eight sub-indicators of strength-to-stress ratio, rock quality, groundwater conditions, tunnel depth, equivalent diameter, support strength, monitoring and measurement practices, and site management, as depicted in Fig. 4. Risk assessment system for large deformation in soft rock tunnels. The strength-to-stress ratio (GN) is defined as the ratio of the rock mass strength to its self-gravity stress, which serves as a key parameter in classifying the degree of soft rock large deformation in tunnels. It illustrates the relationship between the rock mass strength and the surrounding complex environment. Therefore, it’s a principal indicator for assessing large soft rock deformations. Tunnel depth (H) is a critical factor, especially in the western mountainous regions where tunnels are often deeply buried, and increasing tunnel depth corresponds to high stress in the rock mass. The equivalent diameter (D) of a tunnel accounts for non-circular tunnel sections commonly employed in soft rock excavations, i.e., , where A is the cross sectional area of the tunnel. The rock strength index (BQ) that indicates the rock strength is a major factor influencing large deformations in soft rock tunnels. Monitoring, management, and additional factors during tunnel construction are closely related to the risk of large deformation in soft rock tunnels, thus they are integral to the indicators selected for a large deformation assessment system. The support strength (K) characterizes how effectively the implementation of tunnel support measures reduce deformation, making it a viable indicator for assessing large deformation in soft rock tunnels. The support strength (K) is composed of the sum of the stiffness of concrete lining , the stiffness of steel arch and the anchorage stiffness . The specific calculation formula is shown in Eqs. (4)-(7). The monitoring and measurement intensity(M1) and site management intensity(M2) are mainly determined according to the scoring and assignment of site experts. In Eq. (5), is the elastic modulus of shotcrete layer; is the Poisson’s ratio of shotcrete layer; is the tunnel radius; is the thickness of the shotcrete layer. In Eq. (6), is the radial support pressure of steel arch, is the radial deformation. In Eq. (7), and are the circumferential and longitudinal spacing of the anchor rod respectively, is the diameter of the anchor rod, is the elastic modulus of the anchor rod, is the force displacement constant, and is the length of the anchor rod. Based on the indicator factor analysis of the aforementioned assessment system, the aforementioned studies on indicator affecting the risk of large deformation, and an extensive review of the related literatur34,35, the classification criteria for the risk of large soft rock deformation are established, as detailed in Table 3. Table 3Risk classification indicators for large deformation.Risk classificationIIIIIIIVGN>0.200.20 ≥ GN>0.150.15 ≥ GN>0.10≤ 0.10[BQ]> 550400–550250–400< 250Underground water W(Setting value)Dry75–100Wet50–75Spout25–50Drizzle0–25Tunnel depth H< 100 m100–300 m300–500 m> 500 mEquivalent diameter D< 6 m6–8 m8–10 m> 10 mSupport strength K> 500 MPa200–500 MPa20–200 MPa< 20 MPaMonitor and measure intensityM1 (Setting value)Well75–100Normal50–75Mediocre25–50Err0–25Site management intensityM2 (Setting value)Well75–100Normal50–75Mediocre25–50Err0–25 Risk classification indicators for large deformation. Underground water W (Setting value) Dry 75–100 Wet 50–75 Spout 25–50 Drizzle 0–25 Monitor and measure intensity M1 (Setting value) Well 75–100 Normal 50–75 Mediocre 25–50 Err 0–25 Site management intensity M2 (Setting value) Well 75–100 Normal 50–75 Mediocre 25–50 Err 0–25 Prediction of rock mass deformation in tunnels Engineering overview The elevation of the Xicheng railway tunnel is about 3220 m. The tunnel portal section consists of a 2293-m long main hole and a pilot tunnel. The lengths of grade III, IV, V, and VI rock masses are 220 m, 400 m, 1450 m, and 203 m respectively, with grade V and VI rock masses accounting for 72.1% of the section. The tunnel portal section is located within a highly active tectonic zone with strong block difference activity. The presence of weak, fragmented, water-rich rock mass, high geo-stress, significant deformation, elevated temperatures, and hot underground water in the site increases engineering complexity as well as poses considerable engineering and construction challenges with extremely high risks. Time series prediction of rock mass deformation Twenty-two sets of 40-day monitoring data on tunnel vault settlement are collected using the automatic monitoring information system of rock mass in Xicheng Railway tunnel. Among them, Twenty data sets are used as the training sets, while the remaining two sets that correspond to typical tunnel sections are used as the test sets. The CNN, LSTM, and CNN-LSTM models for predicting rock mass deformation are developed based on the Matlab software platform by utilizing the settlement values during the initial 40 days. After Standardization processing the collected data according to Eq. (8), the learning rate parameter for the models is designated as 0.005, and the number of iterations is fixed at 300. The prediction results for rock mass deformation during the subsequent 40 days after model convergence are depicted in Fig. 5, whereas the performance of the model in terms of warning error is summarized in Table 4. The detailed prediction result data can be found as Supplementary Table S1 online. Comparison of vault settlements at typical sections. Table 4Comparison of prediction errors with different models.ModelMAPERMSEMaximum error (mm)Minimum error (mm)Time (s)CNN0.6781744.976713.60531.04878.028LSTM0.4216628.60519.46740.94796.611SVM0.5835914.838221.2250.2581.545PSO-SVM0.745215.367210.11190.86864.448LSTM-SVM0.5835912.312119.16610.95718.745CNN-LSTM0.295256.13115.73520.50613.719 Comparison of prediction errors with different models. In Eq. (8), is the original data, is the mean of the data, is the standard deviation of the data, and is the normalized data As depicted in Fig. 5(a), the prediction curve of the CNN-LSTM model demonstrates convergence, with a mean absolute percentage error (MAPE) of 0.29525 and a corresponding root-mean-square error (RMSE) of 6.1311 for the section AEYDK0 + 437. The MAPE and RMSE for the section AEYDK0 + 508 are 0.2218 and 12.5755, respectively. The prediction curves indicate that the CNN-LSTM model exhibits greater efficacy in predicting settlement deformation of rock mass. Compared with the independent CNN and LSTM models, the integrated CNN-LSTM model demonstrates superior improvements in prediction accuracy and computational efficiency. To validate the feasibility of the CNN-LSTM model, a comparison was conducted with SVM, PSO-SVM, and other models. As shown in Fig. 5; Table 3, the CNN-LSTM model demonstrates superior accuracy compared to other models; however, its computation time is 13.719s, which is longer than that of SVM and other models. To verify the model’s accuracy in predicting peripheral rock deformation, monitoring data from two distinct phases—deformation rise and deformation fluctuation—were selected. Figure 5(a) shows that the peripheral rock vault settlement increased steadily and stabilized after 15 days, with the vault settlement rate approximate at 10 mm per day, and a deformation value of 1%, corresponding to the tunnel’s relative radius of 5 m, indicating a grade I risk level for large deformation. The model’s predictions for peripheral rock deformation at the tunnel’s end align with the observations in Fig. 5(b). In Fig. 5(b), the surrounding rock vault settlement displays a pattern of initial increase followed by decrease. Despite the substantial deformation fluctuations in this data, the model’s predictive accuracy diminishes relative to Fig. 5(a), yet the general trend is consistent, with the settlement rate maintained at 10 mm/d and the risk level remaining at grade I for large deformation. Early warning for risk classification of large deformation in soft rock tunnels Model comparison On the basis of the favorable outcomes of the above regression prediction of tunnel rock mass deformation, the risk of large deformation is quantified by relative deformation, so that the CNN-LSTM model can be utilized for early detection and prediction of the risk of large deformation in soft rock tunnels. The collected data on rock mass properties, rock body strength index [BQ], ground stress levels, tunnel depth (H), diameter (D), support strength (K), and other key indices from various tunnel sections prone to soft rock deformation, as well as computed key indicators such as strength-to-stress ratio (GN) and relative deformation (ξ) yield a large deformation database for soft rock tunnels, which consists of 120 data sets. Due to space constraints, only a subset of the data is presented in Table 5. The entire list of 120 cases can be found as Supplementary Table S2 online. Given that only ten cases of large deformation in shallow-buried tunnels were collected—a sample size deemed insufficient—and subsequently excluded from the database, this study primarily focuses on assessing the risk of large deformations in deep-buried tunnels. Table 5Database of large deformation cases in soft rock tunnels.NumberGNH/m[BQ]D/mK/MPaWM1M2ξ/%Level10.04880016711.21052.676560753.57II20.16235019011.2845.196560753.13II30.16242516012.671027.844070657.26III40.1546514811.5627.364070659.57III……1160.03630755.511719080850.75I1170.01400805.511716565753.59II1180.90545.12437.8974.978580801.28I1190.3071332037.8974.978380801.76I1200.34120.12097.8974.9710405011.34IV Database of large deformation cases in soft rock tunnels. 70% of the tunnel soft rock deformation database is utilized for training, while the remaining 30% serves as the validation set, and the dataset is shuffled during the prediction. The risk level of tunnel soft rock deformation is classified based on relative deformation (ξ). A many-to-one prediction structure is developed by using the remaining eight indices as the input layer and the risk level as the output layer. When the CNN-LSTM neural network model processes predictions, the batch size is designated as 24, the number of epochs is determined as 300, the learning rate is adjusted to 0.001, and the input data is normalized according to the specified normalization formula. After training, the relative deformation values calculated by the CNN-LSTM network are benchmarked against established classification criteria to determine the respective deformation risk level. The resulting confusion matrix of the deformation risk classification model is depicted in Fig. 6. Models Predict Confusion Matrix. Figure 6 reveals that the model provides superior predictions of the risk level of large soft rock deformation in the Xicheng railway tunnel. Furthermore, the CNN-LSTM model yields a reduced prediction error for relative deformation, with an accuracy of 91.67%, while the accuracy of the CNN model and LSTM model were 83.33% and 77.78%, respectively. In the test samples with risk level I, the model predicts 17 samples, while the actual results indicate 18 samples. In the test samples with risk level III, the model predicts 2 samples, one predicts as level 2 and the other predicts as level 4. The model achieves an accuracy of 100% in classifying samples belonging to risk levels II and IV. The CNN-LSTM model yields an overall accuracy exceeding 90% for classification assessment and prediction of risk levels of large soft rock deformation in tunnels while the prediction error is within the acceptable range, which provides a reliable basis for risk assessment and prediction of large soft rock deformation in practical projects. Example analyses Ten cross-sections were selected from a section of the Xicheng railway tunnel, where field research and the expert scoring method were utilized to assign values and scores to the indicators in the large deformation assessment system. The resulting data are presented in Table 6. This data serves as the prediction set, while the 120 sets of data in Sect. 4.1 constitute the training set for the machine learning prediction model. The prediction results of the CNN-LSTM model are presented in Fig. 7. The table reveals that the model’s prediction accuracy is 90%, with the exception of section number 5, whose predicted risk level differs from the actual level. Table 6Table of risk section indicators for large deformation.NumberGNH/m[BQ]D/mK/MPaWM1M211.37126248446186808520.33234031113.111239.6248606530.43723429012.721266.178808540.39216025713.111239.6270808550.24571523612.841303.2546808460.24671023912.741287.8278808470.258233.91805.21827.1365707580.24671023912.741287.8278808490.38945830212.841303.25828084100.355284230141239.62858080 Table of risk section indicators for large deformation. CNN-LSTM Models Predict Results and Confusion Matrix. Therefore, for tunnels in high-stress soft rock with risk level I and II, it is recommended to strengthen the initial support as the primary deformation control measure. Strategies such as resistance-based support, enhanced reinforcement, increased reserved deformation capacity, anchor reinforcement, and optimization of cross-sectional form are advised. For sections with risk level III and IV, it is recommended to transition from a single support form to a joint support system. This approach forms a diversified deformation control strategy, incorporating measures such as moderate release, long, medium, and short anchor reinforcement, multi-support systems, and circular sections. The use of machine learning models to predict the risk level of large soft rock deformation in tunnels enables rapid and precise outcomes. These models can be incorporated into risk monitoring and warning platforms to enhance the efficient deployment of human resources and provide a strong safeguard for the risk assessment and early warning systems in tunnel engineering. Discussion A machine learning model was utilized to predict the trend of perimeter rock deformation and to establish a large deformation dataset for risk classification warnings, with prediction results falling within the acceptable range. The CNN-LSTM model is employed for predicting the risk of large deformations, capable of reducing the risk by compiling historical data on large deformations into a database and pre-establishing relevant index parameters for predictive analyses. Prior to the onset of large deformations, appropriate preventive measures can be taken, such as implementing anti-support structures and overrun grouting, among others. Although this paper employs a machine learning model to facilitate hierarchical early warnings for large deformation risks, it does not address the weight of indicators within the large deformation assessment system, a consideration that can be pursued in future research. Concurrently, due to inherent limitations in the dataset, this study did not examine large deformations in shallow-buried tunnels, nor did it address cases of significant tunnel deformations in arid regions where groundwater is scarce. Groundwater levels and tunnel depth were treated as linear variables, without consideration for their potential nonlinear relationships. This aspect warrants further investigation in future research. Conclusions In this study, a risk assessment system specifically for large soft rock deformation in western mountain tunnels is established with reference to the Xicheng railway tunnel, and a CNN-LSTM-based risk classification assessment model for large deformations is developed. This framework aims to provide a basis for safety risk assessment of large soft rock deformation in complex and hazardous mountain tunnels prevalent in the western region. The following main conclusions are drawn: A risk assessment system is developed for large soft rock deformations in western mountain tunnels, comprising three core indicators: hydrogeological conditions, design parameters, and management aspects. This is further complemented by eight subsidiary indicators: strength-to-stress ratio, rock quality designation, groundwater influence, tunnel depth, equivalent diameter, support robustness, monitoring intensity, and on-site management consistency, with classification criteria developed for each indicator. A risk assessment system is developed for large soft rock deformations in western mountain tunnels, comprising three core indicators: hydrogeological conditions, design parameters, and management aspects. This is further complemented by eight subsidiary indicators: strength-to-stress ratio, rock quality designation, groundwater influence, tunnel depth, equivalent diameter, support robustness, monitoring intensity, and on-site management consistency, with classification criteria developed for each indicator. A prediction model for rock mass deformation based on CNN-LSTM framework is developed, which is validated by the automatic monitoring data of rock mass deformation. The results demonstrates that the prediction performance for rock mass deformation is enhanced, with a Mean Absolute Percentage Error (MAPE) of 0.29525 and a Root Mean Square Error (RMSE) of 6.1311 in test scenarios. A prediction model for rock mass deformation based on CNN-LSTM framework is developed, which is validated by the automatic monitoring data of rock mass deformation. The results demonstrates that the prediction performance for rock mass deformation is enhanced, with a Mean Absolute Percentage Error (MAPE) of 0.29525 and a Root Mean Square Error (RMSE) of 6.1311 in test scenarios. A prediction model for rock mass deformation based on CNN-LSTM framework is developed, which is validated by the automatic monitoring data of rock mass deformation. The results demonstrates that the prediction performance for rock mass deformation is enhanced, with a Mean Absolute Percentage Error (MAPE) of 0.29525 and a Root Mean Square Error (RMSE) of 6.1311 in test scenarios. A prediction model for rock mass deformation based on CNN-LSTM framework is developed, which is validated by the automatic monitoring data of rock mass deformation. The results demonstrates that the prediction performance for rock mass deformation is enhanced, with a Mean Absolute Percentage Error (MAPE) of 0.29525 and a Root Mean Square Error (RMSE) of 6.1311 in test scenarios. Electronic supplementary material Below is the link to the electronic supplementary material. Supplementary Material 1 Acknowledgements The authors would like to extend their sincere appreciation to Key Laboratory of Large Structure Health Monitoring and Control, Shijiazhuang, 050043 (Grant Number: KLLSHMC2113). Author contributions Methodology, X.Z.; Formal analysis, X.Z.; Investigation, D.F.; Data curation, M.W.; Writing—original draft, X.Z.; Writing—review & editing, J.W.; Supervision, J.W. All authors have read and agreed to the published version of the manuscript. Data availability The Supplementary Materials include all data. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1 Data Availability Statement The Supplementary Materials include all data.