Research on Intelligent Predictions of Surrounding Rock Ahead of the Tunnel Face Based on Neural Network and Longitudinal Deformation Curve

Abstract

Traditional methods for predicting surrounding rock grades ahead of tunnel faces encounter challenges: image-based approaches are susceptible to environmental interference, while parameter-based classification may disrupt construction. This study proposes an intelligent rock grade identification method by integrating longitudinal displacement profile (LDP) evolution patterns with deep learning. First, the numerical model was validated against V-D theoretical curves, and LDP evolution laws were systematically analyzed for three rock types (GSI = 15, 30, 50) under nine geological combinations. The results indicate that (1) homogeneous strata exhibit deformation peaks followed by declines; (2) GSI = 15 strata show significantly larger deformations; and (3) stratified schemes display pre-interface deformation peaks and post-interface deformation controlled by subsequent lithology. A novel hybrid neural network was developed to classify strata using LDP curves as input. The model achieved 93.25% training accuracy and 91.20% validation accuracy. Ablation experiments demonstrated their superiority over the other four models with partial module deletions, achieving improvements in test accuracy of 3.24%, 3.08%, 4.16%, and 6.48%, respectively, compared to those models. This lightweight solution effectively overcomes the limitations of manual expertise dependency in conventional models and environmental sensitivity in visual methods. By synergizing LDP evolution analysis with deep learning, this framework provides a reliable approach for real-time rock grade prediction during tunnel advancement.

1. Introduction

The geological conditions of mountain tunnel engineering are inherently complex, and the information obtained during the investigation phase often fails to comprehensively reflect the actual geological characteristics of the tunnel area. This necessitates the implementation of dynamic design methodologies during construction [1]. Dynamic design imposes stringent requirements on the efficiency and accuracy of surrounding rock classification at tunnel faces during construction phases. Traditional approaches relying on empirical formulas or numerical simulations exhibit limitations in dynamically characterizing the nonlinear deformation features of unexcavated rock masses ahead of tunnel faces [2]. Consequently, algorithm development and deep learning technologies have gained increasing prominence in tunnel and geotechnical engineering research [3,4,5,6,7].

Convolutional neural networks (CNNs), as prevalent deep learning architectures, have demonstrated extensive applications in image feature classification and recognition [8,9,10]. The inherent design of convolutional layers enables CNNs to autonomously extract higher-level abstract features from training datasets. Previous studies have validated the capability of classical models such as AlexNet and ResNet in automatic feature extraction from surrounding rock images [11]. The integration of Principal Component Analysis (PCA) with Probabilistic Neural Networks (PNNs) has shown potential for rock mass stability prediction [12], while Mask R-CNN implementations achieved over 85% accuracy in their mineral content quantification [13]. Although CNN technology effectively replaces manual feature extraction and risk warning tasks, the dynamic identification of rock mass characteristics during construction still faces the following challenges: (1) image acquisition difficulties caused by dust interference, non-uniform illumination, humidity fluctuations, and limited photographic windows [14] and (2) significant heterogeneity in apparent brightness, texture, and grayscale distribution induced by moisture variations and complex fracture patterns, which compromise recognition accuracy [15,16]. Notably, the longitudinal deformation induced by excavation is correlated with intrinsic rock properties, thereby circumventing the environmental interference inherent to image-based methods.

The complete deformation process of rocks surrounding tunnels comprises three distinct phases: pre-excavation for advanced deformation, stabilized deformation following initial support installation, and total deformation prior to secondary lining. Tunnel convergence manifests as a nonlinear process with complex mechanisms, primarily influenced by the surrounding rock stability, excavation-scale effects, and system performance support [17]. Current classification standards for convergence deformation include the Singh [18], Aydan [19], and Hoek [20] schemes based on relative deformation, as well as parameter-based systems utilizing absolute deformation, in situ stress, and stress ratios [21,22]. However, these methods require multiple in situ tests for parameter determination, inevitably disrupting construction operations. Utilizing longitudinal deformation data exclusively could substantially reduce such interference. The Longitudinal Deformation Profile (LDP), as a critical component of the Convergence–Confinement Method (CCM), enables dynamic assessment of support timing and time-dependent self-bearing capacity characteristics, providing practical guidance for construction management [23]. While existing research has extensively analyzed longitudinal deformation data in excavated sections [24,25], the unexcavated portions remain understudied due to challenges in data acquisition. Therefore, a systematic investigation of LDP characteristics ahead of tunnel faces is imperative for the accurate prediction of surrounding rock properties.

To address these challenges, this study aims to assess the following: (1) the comprehensive characterization of LDP curve features under various stratigraphic parameter combinations; (2) the development of a hybrid neural network architecture (SE-CNN-LSTM-Attention) integrating SENet, LSTM, and multiple attention mechanisms; (3) predictive modeling of unexcavated surrounding rock properties; and (4) systematic ablation experiments. This research extends CNN application modalities in rock mass identification and provides practical references for field implementation.

2. Fundamental Theory of Tunnel Surrounding Rock Response

2.1. Hoek–Brown Criterion

The strength criterion of intact rock was initially proposed by E. Hoek and E. T. Brown. Subsequently, by substituting Hoek–Brown parameters, the formulation was generalized to estimate rock mass strength, known as the generalized Hoek–Brown criterion [26]. The mathematical expression is formulated as follows:

$${\sigma }_1={\sigma }_3+{\sigma }_{\mathrm{ci}}{(m_{\mathrm{b}}{\sigma }_3/{\sigma }_{\mathrm{ci}}+s)}^a$$

(1)

where σ₁ and σ₃ denote the major and minor principal stresses at failure of the intact rock, respectively; σ_ci represents the uniaxial compressive strength of the intact rock; and m_b, S, and a are the rock mass parameters characterizing the degree of rock mass fracturing. When the intact rock material constant m_i equals m_b with s = 1 and α = 0.5, Equation (1) reduces the strength expression for intact rock. The Geological Strength Index (GSI) [27] is employed to determine the rock mass parameters m_b, S, and a. E. Hoek et al. [27] introduced the disturbance factor D and proposed an updated methodology for deriving these parameters based on GSI. The calculation formulas are expressed as follows:

$$m_{\mathrm{b}}=m_{\mathrm{i}}\exp ({\displaystyle \frac{GSI-100}{28-14D}})$$

(2)

$$s=\exp ({\displaystyle \frac{GSI-100}{9-3D}})$$

(3)

$$a={\displaystyle \frac{1}{2}}+{\displaystyle \frac{1}{6}}(e^{{\displaystyle \frac{-GSI}{15}}}-e^{{\displaystyle \frac{-20}{3}}})$$

(4)

where D represents the disturbance factor of the surrounding rock. The Geological Strength Index (GSI) is recognized as a quantitative indicator for evaluating rock mass strength and remains the sole method capable of directly determining rock mass mechanical parameters [28]. E. Hoek and M. S. Diederichs employed GSI to estimate the elastic modulus of surrounding rock [29] wherein multiple estimation formulas were systematically compared. Among these, the formulation with reduced empirical error was identified as follows:

$$E_{\mathrm{rm}}={\displaystyle \frac{1-D/2}{1+\exp ((75+25D-GSI)/11)}}\times {10}^5$$

(5)

where E_rm denotes the elastic modulus; D represents the disturbance factor with a value range of 0–1; and GSI indicates the Geological Strength Index ranging from 0 to 100. An estimation formula for Poisson’s ratio of the surrounding rock was proposed by B. Vásárhelyi [30], utilizing rock material parameters and the Geological Strength Index (GSI) under conditions where the Poisson’s ratio for the intact rock remained undetermined:

$${\nu }_{\mathrm{rm}}=-0.002GSI-0.003m_{\mathrm{i}}+0.457$$

(6)

where ν_rm is denoted as Poisson’s ratio of the surrounding rock.

2.2. Longitudinal Deformation Profile of Surrounding Rock

The Longitudinal Deformation Profile (LDP) describes the longitudinal variation in radial displacement at unsupported tunnel walls. A fitting formula for the LDP was initially derived by M. Panet [31] through elastic analysis. Subsequently, T. Unlu and H. Gercek [32] further extended this formulation by incorporating Poisson’s ratio. N. Vlachopoulos and M. S. Diederichs [33] proposed the V-D theoretical curve, which correlates with the plastic zone radius in unsupported tunnels and the longitudinal distance from the excavation face. The theoretical expression is formulated as follows:

$$u^{*}=\left\{\begin{array}{cc}1-(1-u_0^{*}){\mathrm{e}}^{-{\displaystyle \frac{3X^{*}}{2R^{*}}}}& (X^{*}\ge 0)\\ u_0^{*}{\mathrm{e}}^{X^{*}}& (X^{*}<0)\end{array}\right.$$

(7)

$$u_0^{*}={\displaystyle \frac{u_0}{u_{\max }}}={\displaystyle \frac{1}{3}}{\mathrm{e}}^{-0.15R^{*}}$$

(8)

where u* represents the dimensionless ratio of surrounding rock deformation to the maximum deformation at distal locations; X* denotes the dimensionless ratio of the longitudinal distance from the excavation face to tunnel radius; R* indicates the dimensionless ratio of the maximum plastic zone radius to the circular tunnel radius; u_max is defined as the maximum deformation of surrounding rock at distal locations after complete geostress release; and u₀ is characterized as the radial displacement of surrounding rock at the tunnel face.

3. Numerical Modeling and Validation

The numerical model in this study was constructed in COMSOL Multiphysics 6.2 finite element software based on the Hoek–Brown criterion. Post-establishment validation was performed via comparative analysis with the V-D theoretical curve defined in Equations (7) and (8).

3.1. Numerical Model Development

As the Longitudinal Deformation Profile (LDP) characterizes the longitudinal variation in radial displacement at unsupported tunnel walls, a circular tunnel with full-face excavation and no support was selected for modeling.

3.1.1. Fundamental Model Settings

(1) Geometric Parameters

The rock mass was assumed homogeneous and isotropic, with groundwater effects and time-dependent behaviors excluded. Key geometric parameters included the following: tunnel radius: R = 6 m; full-face excavation advance length: l = 1 m; buried depth: H = 50 m; excavated section length: L₁ = 99 m; and unexcavated section length: L₂ = 101 m. The computational domain configuration is illustrated in Figure 1a, with LDP data collected post-excavation.

(2) Meshing Parameters

A hexahedral mesh configuration was adopted in this model, characterized by progressive refinement near the excavation face. Specifically, in the Solid Mechanics module of COMSOL Multiphysics, the “discretization” of the displacement field was set to “quadratic”, and the hexahedral mesh was generated through a sweep operation. The shape functions were 20-node quadratic Serendipity elements. The final mesh consisted of 44,415 hexahedral elements, as shown in Figure 1b.

3.1.2. Parameter Determination for Numerical Modeling

Three distinct cases were established for numerical validation, with key parameters including the uniaxial compressive strength of intact rock (σ_ci), rock density (ρ), the Geological Strength Index (GSI), and intact rock material constant (m_i), as detailed in

Table 1. Parameter value table for numerical simulation.

Categories of Geological Parameters	Ρ (kg/m3)	σci (MPa)	GSI	mi
Condition A	2.5 × 103	35	50	20
Condition B	2.3 × 103	15	30	20
Condition C	2.1 × 103	5	15	20

. The disturbance factor D was assigned a value of 0 according to the criteria specified in

Table 2. Recommended values of the rock mass disturbance coefficient D in tunnel engineering.

Condition	Description of Rock Mass Characteristics	The Recommended Value of D
1	Tunnel boring machines cause minimal disturbance to the surrounding rock during excavation or blasting based on excellent quality control	0
2	Mechanical or manual (non-blasting) excavation of poor-quality rock masses causes minimal disturbance to the surrounding rock	0
3	The squeezing leads to the uplift of the tunnel bottom and requires the installation of an invert	0.5 (when there is no invert)
4	Poor-quality blasting results in severe local damage of 2 to 3 m in the surrounding rock of hard rock tunnels	0.8

[27]. The elastic modulus (E) and Poisson’s ratio (ν) of the surrounding rock were calculated using Equations (5) and (6), respectively.

3.2. Numerical Model Validation

The accuracy of the numerical model was preliminarily verified through comparative analysis with the V-D theoretical curve. The statistical evaluation of the Root Mean Square Error (RMSE) and Coefficient of Determination (R²) between numerical results and theoretical predictions is summarized in

Table 3. Statistical table of RMSE and R² for each condition.

Categories of Geological Parameters	RMSE	R2
Condition A	0.027	0.996
Condition B	0.030	0.996
Condition C	0.033	0.995

. The simulated data were subjected to normalization processing, and the comparative results between numerical simulations and the V-D theoretical method are presented in Figure 2.

4. Analysis of LDP Characteristics Under Different Stratigraphic Parameters

Traditionally, the Longitudinal Deformation Profile (LDP) describes the longitudinal variation in radial displacement at unsupported tunnel walls. However, in scenarios with unfavorable stratigraphic parameters, tunnel support becomes essential. Practical engineering applications typically acquire LDP data from the tunnel’s inner surface in excavated sections, whereas obtaining longitudinal deformation data ahead of the tunnel face remains challenging due to inaccessibility [34]. To enable accurate prediction of surrounding rock properties ahead of the tunnel face, systematic analysis of the LDP within the face advance zone is imperative.

4.1. Numerical Modeling Framework

(1) Geometric and Meshing Parameters

The horseshoe-shaped tunnel cross-section was composed of three circles with distinct radii, as illustrated in Figure 3a. Full-face excavation with an advance length of l = 1 m was implemented, maintaining the computational domain configuration shown in Figure 1a. Key parameters included the following: burial depth: H = 49 m; excavated section length: L₁ = 99 m; unexcavated section length: L₂ = 101 m; unsupported distance behind the face: L₃ = 6 m; and LDP data were collected post-excavation. The mesh configuration (Figure 3c) comprised 21,728 elements.

(2) Stratigraphic and Lining Parameters

The support system was simplified to a single lining layer with the following properties: density: 2500 kg/m³; Young’s modulus: 28.5 GPa; and Poisson’s ratio: 0.25. Stratigraphic parameters remained consistent with those in Section 2.1, encompassing three distinct parameter sets as specified in Table 1. Stratigraphic combination schemes are detailed in

Table 4. Statistics table of different stratum combination schemes.

Combination Scheme Number	Parameters of the Surrounding Rock in the Initial Excavation	Parameters of the Surrounding Rock in Front of the Tunnel Face
1	A	A
2	A	B
3	A	C
4	B	A
5	B	B
6	B	C
7	C	A
8	C	B
9	C	C

.

(3) LDP Monitoring Point Configuration

To simulate practical construction conditions, monitoring points were distributed along a finite-length section ahead of the tunnel face. A total monitoring line length, D = 8 m, was established, with specific positioning shown in Figure 3d. Deformation data were recorded at 0.25 m intervals.

4.2. Characteristic Analysis of Surrounding Rock Longitudinal Deformation Profile

The computed LDP curves for all cases and comparative z-direction displacement contour maps at the z = −50 m cross-section (Cases 3, 6, 9) are presented in Figure 4. Analysis of Figure 4a reveals that Case 1 exhibits significantly smaller longitudinal deformation magnitudes compared to Cases 2 and 3. The longitudinal deformation curves of Cases 2 and 3 predominantly overlap in the pre-interface zone, whereas distinct divergence emerges post-interface. Case 2 demonstrates deformation attenuation, while Case 3 undergoes rapid deformation escalation. As illustrated in Figure 4b, Cases 4–6 exhibit marked numerical discrepancies. Case 6 achieves maximum deformation magnitude of −3.47 mm at 0.5 m ahead of the tunnel face, followed by an abrupt post-interface reduction in deformation. Figure 4c identifies Case 9 to have peak deformation (−4.49 mm) localized at 0.75 m ahead of the face. Cases 7 and 8 demonstrate negligible divergence prior to the interface; however, post-interface deformation magnitudes diverge significantly, with Case 8 exceeding Case 7. The characteristic features of LDP curves across all cases are systematically summarized in

Table 5. A summary of the characteristics of LDP curves for different stratum combination schemes.

Scheme Number	Parameters of the Surrounding Rock	Characterization of the LDP Curve
1	A-A	The deformation amount increases slightly and then decreases slowly, and the overall deformation amount is very small
2	A-B	The deformation amount increases slightly and then decreases slowly. After passing through the stratum interface, the deformation amount decreases rapidly, and the overall deformation amount is relatively small
3	A-C	The deformation amount increases slightly and then decreases slowly. After passing through the stratum interface, it increases rapidly, and the overall deformation amount is moderate
4	B-A	Before reaching the stratum interface, the deformation amount first increases and then decreases. After passing through the interface, it remains basically unchanged, and the overall deformation amount is relatively small
5	B-B	The deformation amount first increases and then decreases slowly, and the overall deformation amount is relatively small
6	B-C	Before reaching the stratum interface, the deformation amount first increases and then decreases. After passing through the interface, it decreases rapidly, and the overall deformation amount is relatively large
7	C-A	Before reaching the stratum interface, the deformation amount first increases and then decreases, and the overall deformation amount is relatively small
8	C-B	Before reaching the stratum interface, the deformation amount first increases and then decreases, and the overall deformation amount is relatively small
9	C-C	The deformation amount first increases and then decreases, and the overall deformation amount is relatively large

.

In summary, homogeneous strata cases generally exhibit deformation patterns characterized by an initial increase followed by gradual attenuation. Conversely, cases involving stratigraphic interfaces exhibit biphasic behavior, where deformation follows an increase–decrease trend prior to the interface, while post-interface deformation magnitudes are governed by lithological parameters of the subsequent stratum.

5. A Prediction Model for Rock Mass Classification Ahead of the Driving Face

The implementation of neural networks to automate the identification of LDP curve characteristics contributes to enhanced construction efficiency and reduced labor costs. To address diverse rock mass combinations, this study proposes a hybrid architecture integrating SENet-enhanced CNN, LSTM, and multiple attention mechanisms to predict and classify rock mass grades ahead of the excavation face. The detailed procedural flowchart is shown in Figure 5.

5.1. Intelligent Prediction Fundamentals for Rock Mass Classification

5.1.1. SE-CNN Architecture

The Squeeze-and-Excitation Network (SENet), proposed by Hu et al. [35], introduces a channel attention mechanism to explicitly model interdependencies between feature channels in convolutional neural networks (CNNs). By embedding SENet modules after CNN convolutional layers, the resulting SE-CNN architecture enhances feature representation by adaptively recalibrating channel-wise feature responses, thereby amplifying critical features while suppressing less relevant ones. This mechanism is particularly effective for extracting instability-related patterns from simulation data in rock mass classification tasks. The SENet module comprises four key components, as shown in Figure 6a.

(1) Transformation

A standard convolutional operation generates uncalibrated feature maps, $\mathit{U}\in R^{H\times W\times C}$, where H, W, and C denote the height, width, and channel dimensions, respectively.

(2) Squeeze Operation

Global Average Pooling (GAP) aggregates spatial information within each channel to produce channel-wise statistics $\mathcal{z}\in R^C$. For the c-th channel, the following is calculated:

$$z_c={\mathit{F}}_{\mathrm{squeeze}}({\mathit{u}}_c)={\displaystyle \frac{1}{H\times W}}\sum _{i=1}^H\sum _{j=1}^W{u}_c(i,j)$$

(9)

This equation eliminates spatial interference while preserving channel-wise relationships, forming a compressed descriptor that captures global contextual information for tunnel face surrounding rock prediction.

(3) Excitation Operation

A gating mechanism with two fully connected (FC) layers learns nonlinear channel dependencies as follows:

$$\mathit{s}={\mathit{F}}_{\mathrm{excitation}}(\mathcal{z},\mathit{W})=\sigma [g(\mathcal{z},\mathit{W})]=\sigma [{\mathit{W}}_2\delta ({\mathit{W}}_1\mathcal{z})]$$

(10)

where ${\mathit{W}}_1\in R^{(C/r)\times C}$ and ${\mathit{W}}_2\in R^{C\times (C/r)}$ are learnable weights; δ(∙) denotes ReLU activation; σ(∙) is the sigmoid function; and r is a reduction ratio controlling model complexity.

(4) Scale Operation

The learned channel-wise weights s were applied to the original feature map U via element-wise multiplication across channels, generating the recalibrated output $\tilde{\mathit{X}}$:

$${\tilde{\mathit{x}}}_c={\mathit{F}}_{\mathrm{scale}}\left({\mathit{u}}_c,s_c\right)=s_c{\mathit{u}}_c$$

(11)

$$\tilde{\mathit{X}}=[{\tilde{\mathit{x}}}_1,\cdots ,{\tilde{\mathit{x}}}_c,\cdots ,{\tilde{\mathit{x}}}_C]$$

(12)

The SENet module was designed to explicitly model inter-channel relationships within neural networks. Through Global Average Pooling (GAP) and fully connected layers, SENet adaptively learns channel-wise weights to amplify informative feature channels critical to target tasks while suppressing less contributive ones, thereby enhancing model performance [35].

5.1.2. Long Short-Term Memory (LSTM)

LSTM demonstrates superior memory retention in processing time-series data. Figure 6d illustrates the fundamental unit structure of an LSTM neural network [36]. The schematic defines the following: X_t: input feature vector at time step t; y_t−1: network output at time step t−1; and c_t−1: cell state at time step t−1 [37]. The LSTM unit integrates the following three gating mechanisms:

(1) Forget gate: This governs the retention/discard of historical cell states;

(2) Input gate: This incorporates dual activation functions for feature normalization;

(3) Output gate: This regulates current state propagation.

In this study, LSTM treated specific positions along stratigraphic Longitudinal Deformation Profiles as sequential time steps, effectively capturing spatiotemporal dependencies.

5.1.3. Attention Mechanism

This study implements spatial attention and multi-head attention mechanisms (Figure 6b,e). The spatial attention module performs geometric transformation on deformation data and autonomously identifies critical spatial regions. For input features of dimension H × W × C (H: height; W: width; C: channels), the spatial attention mechanism generates an H × W weight matrix. Each spatial position in the input feature map is scaled by the corresponding weight through element-wise multiplication as follows:

$$\gamma (x)=\mathrm{Sigmoid}(\mathrm{Conv}1\mathrm{d}(\mathrm{mean}(x);\max (x)))$$

(13)

The multi-head self-attention mechanism is an improvement over the standard self-attention mechanism. It divides the self-attention mechanism into multiple parallel heads for repeated linear projections, performs attention calculations on the different projected results, processes hierarchical information in the input sequence after attention computation, and calculates global attention. In essence, it integrates multiple independent self-attention modules. By dividing the time series into h subspaces, each head performs self-attention calculations within the subspaces to enhance the expressive power of attention. The results from h heads are concatenated and integrated into extracted features. Each head independently focuses on implicit information at different hierarchical levels of the time-series data to capture fine-grained features, synthesize diverse features, and dynamically update the weight matrix to obtain the final output.

5.2. SE-CNN-LSTM-Attention Hybrid Architecture Development

5.2.1. Architectural Configuration

Conventional 1D-CNNs exhibit limited capacity in capturing global temporal dependencies due to constrained receptive fields, whereas the hybrid architecture integrating SENet, LSTM, and attention mechanisms enables the explicit modeling of long-range feature dependencies. For instance, deformation characteristics at stratigraphic interfaces require holistic assessment through sequential data integration across adjacent segments. The proposed architecture is systematically detailed in

Table 6. Specific structure table of the SE-CNN-LSTM-Attention hybrid model.

Module Name	Module Level	Component Name	Input Dimension	Output Dimension	Function Description
Input	1	Raw Input	32	(1, 32)	Input of original features
Convolution module	2	Conv1d	(1, 32)	(64, 32)	Extract local features through convolution
	3	BatchNorm1d	(64, 32)	(64, 32)	Standardize channel features to accelerate training
	4	ReLU	(64, 32)	(64, 32)	Nonlinear activation
	5	SEBlock	(64, 32)	(64, 32)	Channel attention
	6	SpatialAttention1D	(64, 32)	(64, 32)	Spatial attention
	7	MaxPool1d (2)	(64, 32)	(64, 16)	Down-sampling to reduce the computational load
	8	Conv1d	(64, 16)	(128, 16)	Further extract features through convolution
	9	ReLU	(128, 16)	(128, 16)	Nonlinear activation
	10	SEBlock	(128, 16)	(128, 16)	Secondary fine-tuning of channel attention
	11	SpatialAttention1D	(128, 16)	(128, 16)	Secondary fine-tuning of spatial attention
	12	Conv1d	(128, 16)	(256, 16)	Increase in the number of convolution channels
	13	ReLU	(128, 16)	(128, 16)	Nonlinear activation
	14	SEBlock	(256, 16)	(256, 16)	Final fine-tuning of channel attention
	15	SpatialAttention1D	(256, 16)	(256, 16)	Final fine-tuning of spatial attention
	16	AdaptiveMaxPool1d	(256, 16)	(256, 16)	Adaptively fix the output length
LSTM module	17	LSTM	(16, 256)	(16, 256)	Treat positions as time steps to capture sequence dependencies
Multi-head attention module	18	MultiheadAttention	(16, 256)	(16, 256)	Focus on different features to enhance the extraction of key information
Classification module	19	Global Average Pooling	(16, 256)	256	Perform dimensional Average Pooling to generate a global feature vector
	20	Linear	256	128	Reduce the dimension through a fully connected layer
	21	Dropout	128	128	Randomly mask 50% of neurons to prevent overfitting
	22	Linear	128	9	Output classification results

and schematically represented in Figure 6.

5.2.2. Loss Function and Optimizer

The model employs CrossEntropyLoss, which is suitable for multi-class classification tasks. The optimizer uses Adam, with the number of epochs set to 300, the learning rate (lr) set to 0.0001, and the batch size for each training iteration set to 64.

5.3. Model Parameter Determination

Model parameters were determined based on the tabulated methodology summarized by Wang et al. [38] and the Appendix of the Chinese National Standard GB/T 50218-2014 “Standard for Engineering Rock Mass Classification” [39]. The Basic Quality Index (BQ) of the rock mass was calculated using the quantitative parameters R_c (saturated uniaxial compressive strength in MPa) and K_v (rock mass integrity index), following the formula below:

$$\mathrm{BQ}=100+3R_{\mathrm{c}}+250K_{\mathrm{v}}$$

(14)

The saturated uniaxial compressive strength (R_c) is typically lower than the natural-state ultimate strength. As groundwater effects were excluded, the intact rock uniaxial compressive strength (σ_ci) was conservatively equated to R_c. Parameter values (

Table 7. Simple random sampling table of numerical simulation parameters.

Grade of Surrounding Rock		BQ	ρ (kg/m3)	Rc (MPa)	GSI	Kv	mi
Grade III	maximum value	420	2.6 × 103	40	70	0.8	20
	minimum value	390	2.5 × 103	30	50
Grade IV	maximum value	310	2.4 × 103	20	50	0.6	20
	minimum value	295	2.3 × 103	15	30
Grade V	maximum value	230	2.2 × 103	10	30	0.4	20
	minimum value	215	2.1 × 103	5	15

) were assigned via simple random sampling.

Ten parameterized models were established for each label category (

Table 8. Classification table of different labels.

Classification Number	Surrounding Rock Grade at the Initial Excavation	Surrounding Rock Grade in Front of the Tunnel Face
0	III	III
1	III	IV
2	III	V
3	IV	III
4	IV	IV
5	IV	V
6	V	III
7	V	IV
8	V	V

), yielding 90 models and corresponding 1D deformation profiles.

Through basic augmentation, the amount of data becomes four times that of the original data. Then, the original data undergoes 20 rounds of composite augmentation. Eventually, the amount of data obtained is 24 times that of the original data. The specific data augmentation process is shown in Figure 7. This expanded the dataset to 2160 samples (240 per category), divided into training and test sets at a 7:3 ratio. The data distribution is visualized in the scatter plot in Figure 8.

5.4. Results and Analysis

(1) Analysis of Model Loss Function

The training and validation loss curves, along with accuracy metrics, are shown in Figure 9. The results indicate that both training and validation losses converge to low and comparable values, with no evident signs of overfitting or underfitting observed. This demonstrates the robust overall training performance of the model.

(2) Analysis of Model Evaluation Metrics

The key metrics for the machine learning model evaluation typically included precision, recall, F1-Score, and support. To comprehensively assess the hybrid SE-CNN-LSTM-Attention architecture proposed in this study, the model’s performance metrics on the training and test sets are statistically summarized in

Table 9. Statistical table of evaluation indicators for the training set of the SE-CNN-LSTM-Attention model.

Formation Classification	Precision	Recall	F1-Score	Support
0	0.8590	0.7614	0.8072	176
1	1.0000	0.9868	0.9933	151
2	0.9006	0.8735	0.8869	166
3	1.0000	1.0000	1.0000	167
4	0.7778	0.8698	0.8212	169
5	0.8804	0.9153	0.8975	177
6	1.0000	1.0000	1.0000	157
7	1.0000	1.0000	1.0000	172
8	1.0000	1.0000	1.0000	177
Macro avg	0.9353	0.9341	0.9340	1512
Weighted avg	0.9338	0.9325	0.9325	1512
Accuracy	0.9325			1512

and

Table 10. Statistical table of evaluation indicators for the test set of the SE-CNN-LSTM-Attention model.

Formation Classification	Precision	Recall	F1-Score	Support
0	0.9273	0.7969	0.8571	64
1	0.9881	0.9326	0.9565	89
2	0.7317	0.8108	0.7962	74
3	1.0000	1.0000	1.0000	73
4	0.8272	0.9437	0.8816	71
5	0.7288	0.6825	0.7049	63
6	1.0000	1.0000	1.0000	83
7	1.0000	1.0000	1.0000	68
8	1.0000	1.0000	1.0000	63
Macro avg	0.9114	0.9074	0.9080	648
Weighted avg	0.9152	0.9120	0.9123	648
Accuracy	0.9120			648

.

As shown in the table, except for the relatively low data in Class 5 of the test set, the remaining data are generally between 0.8 and 1. This indicates that the model has good performance and high accuracy. Finally, the accuracy of the model on the training set is 93.25%, and the accuracy on the test set is 91.20%. From the confusion matrix in Figure 10, it can be observed that when the true label is from Class 5, the predicted label sometimes appears as Class 2. The reason for this is that after data augmentation, since mirror flipping operations occur along the x-axis in basic augmentation alongside operations that randomly increase and decrease the amplitude in both basic and composite augmentations, the augmented data of Class 2 and Class 5 show highly similar features. Therefore, further research is needed to improve the performance of the model when predicting Class 5 data.

(3) Confusion Matrix

The classification confusion matrix of the model is illustrated in Figure 10. The results demonstrate high predictive accuracy in determining the grade of the surrounding rock ahead of the tunnel face based on the LDP (Load Deformation Profile) curve characteristics, with consistent performance across most rock mass classifications.

6. Discussion

To validate the performance of the proposed hybrid architecture, ablation experiments were conducted on the SE-CNN-LSTM-Attention model. The proposed SE-CNN-LSTM-Attention hybrid architecture was compared with the following variants: SE-CNN-LSTM, CNN-LSTM-Attention, CNN-Attention, and baseline CNN. The architectural configurations of all evaluated models are systematically compared in

Table 11. Statistical table of the structures of various models.

Model Name	Composition of the Core Module	Description of Differences (in Comparison with the Model in this Paper)	Key Parameter Configuration
SE-CNN-LSTM-Attention (Baseline)	SE + SA + CNN3 + LSTM + Attn	none	Conv1d (k = 7, 5, 3), SE (reduction = 16), SA (k = 5, 5, 3), LSTM (h = 256), MultiheadAttn (4 head)
SE-CNN-LSTM	SE + CNN3 + LSTM	-SA, -Attn	Conv1d (k = 7, 5, 3), SE (reduction = 16), LSTM (h = 256)
CNN-LSTM-Attention	CNN3 + LSTM + Attn	-SE, -SA	Conv1d (k = 7, 5, 3), LSTM (h = 256), MultiheadAttn (4 head)
CNN-Attention	CNN3 + Attn	-SE, -SA, -LSTM	Conv1d (k = 7, 5, 3), MultiheadAttn (4 head)
CNN	CNN3	-SE, -SA, -LSTM, -Attn	Conv1d (k = 7, 5, 3)

. This comparative analysis isolates the contributions of SENet channel recalibration, LSTM temporal modeling, and attention-based feature prioritization from the overall predictive performance. Please note that in the table, “SA” refers to spatial attention, “CNN3” represents three convolutional blocks, and “Attn” refers to the multi-head attention module.

The hyperparameters and training data used in the table models were identical to those of the SE-CNN-LSTM-Attention hybrid neural network. The evaluation metrics of each model on the test set are presented in

Table 12. Comparison results of SE-CNN-LSTM-Attention with SE-CNN-LSTM, CNN-LSTM-Attention, CNN-Attention, and CNN on the test set.

Model Name	Accuracy	Precision		Recall		F1-Score
		Macro Avg	Weighted Avg	Macro Avg	Weighted Avg	Macro Avg	Weighted Avg
SE-CNN-LSTM-Attention (Baseline)	91.20%	0.9114	0.9152	0.9074	0.9120	0.9080	0.9123
SE-CNN-LSTM	87.96%	0.8854	0.8897	0.8762	0.8796	0.8732	0.8774
CNN-LSTM-Attention	88.12%	0.8793	0.8862	0.8785	0.8812	0.8755	0.8804
CNN-Attention	87.04%	0.8655	0.8724	0.8655	0.8704	0.8650	0.8709
CNN	84.72%	0.8426	0.8488	0.8455	0.8472	0.8472	0.8432

, and the classification confusion matrices are shown in Figure 11.

As demonstrated in Table 12, the SE-CNN-LSTM-Attention hybrid architecture achieves accuracy improvements of 3.24%, 3.08%, 4.16%, and 6.48% on the test set compared to the SE-CNN-LSTM, CNN-LSTM-Attention, CNN-Attention, and baseline CNN models, respectively. The results indicate that the SE-CNN-LSTM-Attention architecture delivers the best performance on the test set. In conclusion, the proposed SE-CNN-LSTM-Attention hybrid neural network demonstrates robust capabilities for stratigraphic prediction.

7. Conclusions

(1) The Longitudinal Deformation Profile (LDP) characteristics were systematically analyzed for nine stratigraphic combinations across three distinct rock mass types. Homogeneous strata exhibited deformation patterns, characterized by an initial increase followed by gradual attenuation. Rock masses with a Geological Strength Index (GSI) of 15 demonstrated significantly higher deformation magnitudes. For cases involving stratigraphic interfaces, pre-interface deformation followed an increase–decrease trend, while post-interface deformation magnitudes were governed by the lithological parameters of the subsequent stratum.

(2) A hybrid SE-CNN-LSTM-Attention neural network was proposed for predicting the surrounding rock properties ahead of tunnel faces. Utilizing LDP curves as input features, the model achieved classification accuracy rates of 93.25% on the training set and 91.20% on the test set, demonstrating robust predictive capability for stratigraphic characterization.

(3) Comparative ablation studies against four benchmark architectures (SE-CNN-LSTM, CNN-LSTM-Attention, CNN-Attention, and baseline CNN) revealed significant performance enhancements. The proposed model improved the test set accuracy by 3.24%, 3.08%, 4.16%, and 6.48%, respectively, validating the efficacy of the synergistic integration of SENet channel attention, LSTM temporal modeling, and spatial-contextual attention mechanisms.

Notably, the current training data were exclusively derived from numerical simulations, with tunnel burial depth fixed as a constant. Future work will incorporate burial depth as an input variable and validate the model against field-monitored data from practical engineering projects.