Tunnel surrounding rock deformation rolling time sequence prediction method and system considering excavation disturbance

Abstract

The invention provides a tunnel surrounding rock deformation rolling time sequence prediction method and system considering excavation disturbance, and the method comprises the steps: obtaining a geometrical relationship from a tunnel face to a monitoring point, an excavation area and an excavation footage in a tunnel excavation process, and calculating a construction disturbance factor according to the obtained data; constructing and training a surrounding rock deformation prediction model, wherein the surrounding rock deformation prediction model is obtained by fusing a convolutional neural network, a bidirectional long-short term memory neural network and an attention mechanism layer; acquiring tunnel surrounding rock monitoring data, and preprocessing the monitoring data; and taking the preprocessed monitoring data and the calculated construction disturbance factor as input data of the trained surrounding rock deformation prediction model to obtain a final deformation prediction result. According to the method, the dynamic disturbance characteristics of stress redistribution in the tunnel excavation process are quantified into the construction disturbance factors, the nonlinear coupling relation between surrounding rock deformation and construction disturbance is analyzed through the hybrid model, and the tunnel surrounding rock deformation prediction accuracy is improved.

Description

Technical Field

[0001] This invention belongs to the field of surrounding rock deformation prediction, specifically relating to a method and system for predicting the rolling time sequence of tunnel surrounding rock deformation considering excavation disturbance. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] As a crucial component of modern infrastructure construction, tunnel engineering has always focused on the deformation of surrounding rock in the field of geological engineering. Rock deformation can not only damage the tunnel structure but also significantly impact the surrounding environment and transportation. Therefore, accurately predicting the deformation of the tunnel's surrounding rock, especially during construction, is a critical issue in tunnel engineering management.

[0004] Traditional methods for predicting tunnel surrounding rock deformation can be broadly categorized into three types: mathematical modeling, numerical simulation, and machine learning-based methods. Mathematical modeling methods typically rely on simple physical assumptions and empirical formulas, simplifying complex factors such as geological heterogeneity, groundwater influence, and faults. However, these methods have a narrow scope of application and often fail to provide accurate predictions under complex geological conditions. Numerical simulation demonstrates good performance in predicting tunnel surrounding rock deformation and failure mechanisms. However, tunnel geological conditions are often complex and variable, involving nonlinear rock behavior, interactions between soil and strata, and the influence of groundwater. Furthermore, the chosen constitutive model may not fully capture the stress-strain relationship. Additionally, the complexity of the model is geometrically related to the computational resource requirements. All these factors need to be thoroughly considered during the modeling process.

[0005] With the rapid development of deep learning and machine learning technologies, tunnel surrounding rock deformation prediction methods based on these technologies have gradually become a new research direction. Machine learning methods have powerful data processing capabilities and can automatically learn complex nonlinear relationships extracted from large-scale monitoring data. In tunnel surrounding rock deformation prediction, monitoring data includes static data such as surrounding rock pressure, displacement, and settlement, as well as time-series data. These data are usually high-dimensional and have complex spatiotemporal dependencies. Machine learning models, especially in deep learning methods, can process these complex spatiotemporal features through data self-learning without relying on physical formulas, thus providing more accurate prediction results. However, despite the great potential of machine learning methods in handling tunnel surrounding rock deformation prediction, a significant challenge remains: the prediction results often exhibit time delays, especially the impact of external disturbances on prediction accuracy, causing the model's predictions to fail to respond promptly to real-time disturbances during construction.

[0006] In tunnel construction, construction disturbance is a significant factor leading to surrounding rock deformation. Construction disturbance originates from operations such as face excavation and support structure construction. These operations cause a redistribution of stress in the surrounding rock, thus affecting its deformation. As excavation deepens and support work progresses, the surrounding rock undergoes varying degrees of deformation; this process is highly dependent on time and space. Traditional models often neglect construction disturbance factors, resulting in lag and inaccuracy in predictions. Summary of the Invention

[0007] To address the aforementioned problems, this invention proposes a rolling time-series prediction method and system for tunnel surrounding rock deformation considering excavation disturbance. This invention quantifies the dynamic disturbance characteristics of stress redistribution during tunnel excavation into construction disturbance factors through rolling time-series prediction, and uses a hybrid model to analyze the nonlinear coupling relationship between surrounding rock deformation and construction disturbance, thereby improving the accuracy of tunnel surrounding rock deformation prediction.

[0008] According to some embodiments, the present invention adopts the following technical solution: A method for predicting the deformation and rolling time series of tunnel surrounding rock considering excavation disturbance includes the following steps: Obtain the geometric relationship between the tunnel face and the monitoring point, the excavation area, and the excavation progress during the tunnel excavation process, and calculate the construction disturbance factor based on the obtained data; A surrounding rock deformation prediction model is constructed and trained. The surrounding rock deformation prediction model is obtained by fusing a convolutional neural network, a bidirectional long short-term memory neural network and an attention mechanism layer. Acquire monitoring data of the surrounding rock of the tunnel and preprocess the monitoring data; The preprocessed monitoring data and the calculated construction disturbance factor are used as input data for the trained surrounding rock deformation prediction model. The convolutional neural network performs multi-layer convolution on the input data to extract deep features of the monitoring data and obtain relational feature vectors. The bidirectional long short-term memory neural network extracts the feature relationships in the relational feature vectors. The outputs of the forward long short-term memory neural network and the backward long short-term memory neural network in the bidirectional long short-term memory neural network are concatenated. The attention mechanism layer is used to assign weights to the concatenated result to obtain the final deformation prediction result.

[0009] A time-series prediction system for tunnel surrounding rock deformation rolling considering excavation disturbance includes: The construction disturbance factor calculation module is configured to obtain the geometric relationship between the tunnel face and the monitoring point, the excavation area, and the excavation progress during the tunnel excavation process, and calculate the construction disturbance factor based on the obtained data; The model building and training module is configured to build and train a surrounding rock deformation prediction model, which is obtained by fusing a convolutional neural network, a bidirectional long short-term memory neural network and an attention mechanism layer. The data acquisition module is configured to acquire tunnel surrounding rock monitoring data and preprocess the monitoring data; The deformation prediction module is configured to use preprocessed monitoring data and calculated construction disturbance factors as input data for the trained surrounding rock deformation prediction model. The convolutional neural network performs multi-layer convolution on the input data to extract deep features of the monitoring data and obtain relational feature vectors. The bidirectional long short-term memory neural network extracts the feature relationships in the relational feature vectors. The outputs of the forward long short-term memory neural network and the backward long short-term memory neural network in the bidirectional long short-term memory neural network are concatenated. An attention mechanism layer is used to assign weights to the concatenated result to obtain the final deformation prediction result.

[0010] A computer-readable storage medium for storing computer instructions, which, when executed by a processor, perform the steps in the above method.

[0011] An electronic device includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor, wherein the computer instructions, when executed by the processor, perform the steps in the method described above.

[0012] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention measures the contact pressure of the surrounding rock, the stress of the steel arch frame, and the deformation of the surrounding rock in real time, while recording the area of ​​the bench excavation and the advance of each excavation to obtain complete time-series data. Based on the contact pressure of the surrounding rock, the stress of the steel arch frame, the deformation of the surrounding rock, and the construction disturbance factor, a hybrid model is used to analyze the nonlinear coupling relationship between the deformation of the surrounding rock and the construction disturbance, providing a theoretical basis for dynamic optimization decision-making of the support structure and having important practical value for proactive risk control in the tunnel construction process.

[0013] This invention ensures the accuracy and stability of the model through a time sliding window and an error propagation mechanism. It can respond in real time to changes in the deformation of the surrounding rock during construction and gradually correct the prediction results, thereby improving the accuracy of the prediction.

[0014] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0015] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0016] Figure 1 This is a schematic diagram illustrating the calculation of construction disturbance factors in one embodiment. Figure 2A schematic diagram of the CNN principle structure of one embodiment; Figure 3 This is a schematic diagram of the LSTM principle structure of one embodiment; Figure 4 This is a schematic diagram of the principle structure of BiLSTM in one embodiment; Figure 5 This is a schematic diagram of a tunnel surrounding rock deformation prediction model according to one embodiment; Figure 6 This is a flowchart of an example of advance forecasting; Figure 7 This is an engineering overview diagram of one embodiment; Figure 8 This is a cross-sectional layout diagram of a tunnel deformation monitoring system, as shown in one embodiment. Figure 9 This is a schematic diagram of monitoring data during the initial stage of cross-section construction in one embodiment, where (a) is the deformation of the surrounding rock, (b) is the contact pressure of the surrounding rock, and (c) is the stress of the steel arch frame. Figure 10 This is a construction disturbance factor diagram of one embodiment; Figure 11 This is a monitoring data processing effect of one embodiment, wherein (a) is surrounding rock deformation data, (b) is surrounding rock contact pressure data, (c) is steel arch stress data, and (d) is construction disturbance factor time series data; Figure 12 This is a prediction result of surrounding rock deformation in one embodiment; Figure 13 This is a prediction result of the arch settlement and deformation of one embodiment; Figure 14 This is a prediction result of surrounding rock deformation in one embodiment; Figure 15 This is a comparison diagram of the predicted and measured results of the surrounding rock deformation in one embodiment. (a) is a comparison diagram of the predicted and measured values ​​of the surrounding rock deformation at the left arch shoulder, (b) is a comparison diagram of the predicted and measured values ​​of the surrounding rock deformation at the arch crown, and (c) is a comparison diagram of the predicted and measured values ​​of the surrounding rock deformation at the right arch shoulder. Figure 16 This is a prediction result of the cumulative deformation of the arch in one embodiment. Detailed Implementation

[0017] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0018] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0019] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0020] Where there is no conflict, the embodiments and features described in this application may be combined with each other.

[0021] Example 1 A method for predicting the deformation and rolling time series of tunnel surrounding rock considering excavation disturbance includes the following steps: Obtain the geometric relationship between the tunnel face and the monitoring point, the excavation area, and the excavation progress during the tunnel excavation process, and calculate the construction disturbance factor based on the obtained data; A surrounding rock deformation prediction model is constructed and trained. The surrounding rock deformation prediction model is obtained by fusing a convolutional neural network, a bidirectional long short-term memory neural network and an attention mechanism layer. Acquire monitoring data of the surrounding rock of the tunnel and preprocess the monitoring data; The preprocessed monitoring data and the calculated construction disturbance factor are used as input data for the trained surrounding rock deformation prediction model. The convolutional neural network performs multi-layer convolution on the input data to extract deep features of the monitoring data and obtain relational feature vectors. The bidirectional long short-term memory neural network extracts the feature relationships in the relational feature vectors. The outputs of the forward long short-term memory neural network and the backward long short-term memory neural network in the bidirectional long short-term memory neural network are concatenated. The attention mechanism layer is used to assign weights to the concatenated result to obtain the final deformation prediction result.

[0022] The method provided in this embodiment differs from the simplified treatment of spatiotemporal dynamic coupling effects in traditional models. This work is the first to quantify the dynamic disturbance characteristics of stress redistribution during tunnel excavation as a construction disturbance factor. C d This paper constructs a CNN-BiLSTM-Attention hybrid model to analyze the nonlinear coupling relationship between surrounding rock deformation and construction disturbance. Using the Lehua Tunnel project as a case study, this paper establishes a surrounding rock regression prediction model and a rolling prediction method with an error propagation mechanism, and focuses on verifying the key role of construction disturbance factors in improving the accuracy of time-series predictions. This method not only provides a new technical path for early warning of surrounding rock deformation, but its constructed rolling prediction method also provides a theoretical basis for dynamic optimization decision-making of support structures, and has significant practical value for proactive risk control in tunnel construction.

[0023] Tunnel excavation leads to a redistribution of the stress field, causing induced deformation of the surrounding rock. Specifically, during tunnel construction, as the tunnel face advances, the initial stress balance of the surrounding rock is disrupted. The stress release induced by excavation causes a redistribution of the stress field within the surrounding rock, resulting in changes in the mechanical response of the rock material. This significantly impacts field data, particularly the deformation of the tunnel and the stress changes in the initial support during the bench excavation stage at the tunnel face. Therefore, this embodiment considers the impact of bench excavation during construction and identifies a new characteristic variable: the construction disturbance factor. C d . C d It not only reflects the intensity of disturbance during construction but also directly affects the deformation rate and amplitude of the surrounding rock. This is especially true in weak surrounding rock and layered carbonaceous shale. C d When the changes are significant, the surrounding rock exhibits more pronounced plastic deformation and may even become unstable, such as local collapse and large deformation.

[0024] Construction disturbance factors C d Impact on surrounding rock deformation: The impact of construction disturbance on surrounding rock deformation has significant spatial and temporal dependence. Near the working face, due to the stress concentration effect, construction disturbance often reaches its maximum value, leading to a violent stress release in the surrounding rock in this area, resulting in large-scale displacement and deformation. C d The greater the change, the more obvious the sudden increase in surrounding rock deformation, and the higher the rate of increase in deformation.

[0025] Furthermore, the impact of construction disturbance on surrounding rock deformation exhibits phased characteristics at different construction stages. In bench excavation, each advance triggers a new stress redistribution, causing phased fluctuations in the deformation rate of the surrounding rock. These fluctuations are particularly pronounced at the tunnel crown and shoulders, as these areas are often stress concentration zones and are most directly affected by construction disturbance. As the excavation distance increases, C d The growth rate gradually decreases, and the deformation rate of the surrounding rock gradually stabilizes. However, if C d If not properly controlled or ignored, local deformation of the surrounding rock may exceed the safety threshold, thereby threatening the overall stability of the tunnel.

[0026] therefore, C d It is not only a key parameter for quantifying the impact of stress redistribution during tunnel construction, but also an important indicator for assessing the deformation process and stability of the surrounding rock. Through analysis of... C dIn-depth analysis can reveal the deformation mechanism during construction, providing a scientific basis for optimizing construction plans and rationally adjusting support timing. This, combined with monitoring data and... C d The dynamic changes can better predict the temporal deformation trend of the surrounding rock, thereby improving the safety and reliability of tunnel construction.

[0027] Because the impact of construction disturbance on surrounding rock deformation has significant spatial and temporal dependence, especially the distance between the monitoring section and the working face and the bench excavation, the construction disturbance factor... C d The calculation is based on the geometric relationship between the tunnel face and the monitoring point, the excavation area, and the excavation progress during tunnel excavation. The formula is as follows: Figure 1 As shown: (1) (2) Where, in the formula, Cd,i Indicates the first i Excavation disturbance caused by step excavation; Ai For the tunnel working face i The excavation area is measured in square meters (m2). Li,ki For the first i Step 1 k Within the construction cycle, the first i The distance between the excavation face and the monitoring section, in meters (m), parameters. i For the first i N steps in total. Parameters ki Indicates the first i Number of excavation cycles. As excavation progresses, each new step moves the working face of the roadway further away from the fixed monitoring section, leading to... Li,ki It increases with each cycle. Excavation coefficient Cd Over time, this reflects the increased disturbance to the monitoring section caused by continuous excavation.

[0028] Construction disturbance factors C d The significance lies in considering the impact of bench excavation at the tunnel face during construction. Bench excavation at the deformation monitoring section has the greatest impact on the stability of the surrounding rock mass. As the tunnel face excavation progresses, its impact on the deformation of the surrounding rock at the monitoring section gradually weakens. Therefore, defining a construction disturbance factor is feasible to a certain extent.

[0029] A surrounding rock deformation prediction model is constructed, which is based on the CNN-BiLSTM-Attention structure.

[0030] Convolutional Neural Networks (CNNs) are a type of deep learning network inspired by animal visual mechanisms. They effectively identify local features in images, thereby capturing the information contained within image data, and excel in tasks such as face recognition and image classification. In addition, CNNs are widely used in sequence prediction, particularly in areas such as language text and time series data analysis. A CNN consists of an input layer, convolutional layers, pooling layers connected to a fully connected layer, and an output layer. The basic structure of a CNN is as follows: Figure 2 As shown, its main functions are as follows: Convolutional layers: Convolutional operations effectively extract feature parameters from input data. The convolutional kernel, also called a filter, slides across the input data and performs a dot product operation with the data at each position, resulting in a convolutional value. Different convolutional kernels are used to extract different features from the data.

[0031] Pooling layers: These reduce the dimensionality of feature parameters while retaining the most prominent features. Max pooling is a commonly used pooling operation. It selects the maximum value in each pooling window as the output to reduce the number of parameters, making the model more concise and efficient, and improving its robustness.

[0032] Fully connected layers: These layers transform and combine the feature parameters extracted by convolutional and pooling layers, and then connect them to the output layer. Each neuron in a fully connected layer establishes a connection with neurons in the previous layer to capture the correlations between different features, thereby enabling time-series prediction of the input data.

[0033] The formula for CNN is as follows: (3) (4) (5) In the formula, C This is the output of the convolutional layer; f a For activation functions; w c This is the weight matrix of the convolutional layer; X For input multi-feature data matrix; b c This is the bias matrix of the convolutional layer; P This is the output of the pooling layer; H Output for fully connected layers; w s This is the weight matrix of the fully connected layer; b s This is the bias vector for the fully connected layer.

[0034] Long Short-Term Memory (LSTM) neural networks are a type of recurrent neural network (RNN) variant commonly used for processing time-series data. While RNNs, as traditional neural networks, have certain advantages in processing sequential data, they also suffer from several significant drawbacks. These include gradient vanishing or exploding during backpropagation, insufficient memory capacity to handle long-range dependencies, and difficulty in parallel computation. These shortcomings have consistently limited the performance of RNNs in processing complex sequential data. LSTM neural networks, by introducing gating mechanisms and memory units, can better handle long-term dependencies in complex, long-term sequential data, thus overcoming the limitations of RNNs. LSTM incorporates gating mechanisms including input gates, forget gates, and output gates. These gating units, through the activation functions of their activation functions, control the degree to which information is retained, updated, and output in the memory units. In other words, they determine which parts of the sequence data need to be forgotten, which new information needs to be added to the memory, and how the memory is passed to the next time step. LSTM principle and structure as follows Figure 3 As shown.

[0035] The formula for the forgetting gate: (6) In the formula, w f for h t-1 and x t The weight matrix for the forget gate; b f For bias.

[0036] The function formula for the output gate: (7) (8) In the formula, w o for h t-1 and x t To output gate activation function σ The weight matrix; b o This corresponds to the bias.

[0037] The function formula for the input gate: (9) (10) In the formula, w i , w c They are respectively h t-1 and x t To input gate activation function σ The weight matrix with respect to tanh; b i and b c This is the corresponding bias.

[0038] Cell structure: (11) The flexible control mechanism enables LSTM neural networks to better handle dependencies in long-sequence data and capture long-term memory. BiLSTM neural networks are bidirectional recurrent neural networks containing two independent LSTM layers: one processes the input time-series data forward, and the other processes it backward. This bidirectional structure allows BiLSTM to consider both past and future information simultaneously, better capturing contextual information in the sequence data. Its structure includes an input layer, a forward layer, a backward layer, and an output layer. In BiLSTM, the outputs of the forward and backward LSTMs are merged, such as... Figure 4 As shown.

[0039] Attention mechanisms are computational models that mimic human attention processes and are commonly used in machine learning and artificial intelligence. In deep learning, attention mechanisms help models focus their attention on specific parts of the input data, thereby improving the model's understanding and processing capabilities.

[0040] The core idea of ​​the attention mechanism is that when processing an input sequence, the model can dynamically assign different weights to different positions of the input, thus focusing more on the important parts during output. Through the attention mechanism, the model can adaptively learn the importance of different parts of the input, thereby improving the model's representation ability and generalization ability.

[0041] The principle behind this can be explained through the following steps: Step 1: Input time series data is fed into a neural network or other model for processing.

[0042] Step 2: Query, Key, and Value. Input data is abstracted into queries, keys, and values. Queries, keys, and values ​​can be different representations or characteristics of the input data. Generally, queries are used to find points of interest, while keys and values ​​are the objects being queried.

[0043] Step 3: Similarity calculation. The model calculates the similarity score between the query and the key using different methods, such as dot product, weighted sum, or other similarity measures.

[0044] Step 4: Attention weights. By performing softmax normalization on the similarity scores, the attention weights for each key are obtained.

[0045] Step 5: Weighted summation. The values ​​are summed according to the corresponding attention weights to obtain the final context vector or the weighted value.

[0046] Step 6: Based on the context vector or the weighted value, the model performs further processing, such as generating prediction results or continuing downstream tasks.

[0047] The main calculation formula is: (12) (13) (14) In the formula, h m0 The magnitude of attention; H For learning functions, used to calculate C , h and x The degree of correlation among the three; C This is an attention-weighted feature vector.

[0048] The principle behind attention mechanisms is to calculate the similarity between queries and keys, then transform these similarities into attention weights, and finally use these weights to perform a weighted sum of values, thereby obtaining contextual information or more targeted feature representations. This process enables the model to dynamically focus on different parts of the information while processing input data, thus improving the model's performance on complex tasks.

[0049] This embodiment establishes a surrounding rock deformation prediction model combining a convolutional neural network (CNN), a bidirectional long short-term memory neural network (BiLSTM), and an attention mechanism. The combination of CNN and BiLSTM is used to extract feature information from different dimensions of tunnel monitoring data. The CNN-BiLSTM-Attention model consists of CNN layers, BiLSTM layers, and fully connected layers, as follows... Figure 5As shown, the specific process of the intelligent prediction model's prediction algorithm is as follows: First, the tunnel monitoring data is preprocessed. The Makima method is used to impute missing values ​​in the sequence data, and data smoothing is performed to improve data quality and feature performance. The processed tunnel monitoring dataset is then input into a CNN layer, where multiple convolutions are performed to extract deep features from the tunnel monitoring data and obtain the relational feature vectors of the parameters. Next, the output of the CNN is input into a BiLSTM network for training to further extract feature relationships. Then, the outputs of the forward and backward LSTMs in the BiLSTM are concatenated. Finally, the concatenation is performed on the attention mechanism layer and the fully connected layer to obtain the final tunnel deformation prediction result.

[0050] During neural network model training, prediction results may deviate from actual monitored values ​​due to variations in dataset quality, feature parameters, and model parameters. Therefore, to measure the predictive performance of machine learning algorithms, this embodiment uses mean absolute error (MAE), mean absolute percentage error (MAPE), root mean square error (RMSE), and coefficient of determination (R²). 2 The evaluation metrics are MAE and RMSE, and the residual prediction (RPD). Lower MAE and RMSE values ​​indicate better predictive performance of the model. 2 The closer the value is to 1, the better the model's predictive performance. For RPD, a larger RPD value indicates a better consistency between the actual value and the model's prediction. Conversely, a smaller RPD value indicates a poorer consistency between the actual value and the model's prediction. The relevant evaluation index calculation formulas are as follows: Mean Absolute Error (MAE): (15) Mean Absolute Error (MAPE): (16) Root Mean Square Error (RMSE): (17) Coefficient of determination (R) 2 ): (18) Residual Predicted Residual (RPD): (19) In the formula, n The number of samples; These are actual deformation monitoring values; Deformation prediction value; The mean of the sample; S The standard deviation represents the prediction error.

[0051] The deformation mechanism of tunnel surrounding rock is essentially a dynamic equilibrium process of the surrounding rock-support system under construction disturbance, mainly including three core mechanisms: the interaction between the support structure and the surrounding rock, the mechanical response of the support system itself, and the spatiotemporal propagation effect of construction disturbance. The contact pressure of the surrounding rock, as a direct representation of the support-surrounding rock interaction, reflects the constraint reaction force of the initial support on the deformation of the surrounding rock; its numerical change intuitively reflects the static equilibrium state between the stress release of the surrounding rock and the resistance of the support. The stress of the steel arch frame is a core indicator of the bearing capacity of the support structure; by monitoring the distribution of internal forces in the arch frame, the structural response characteristics of the support system under the action of surrounding rock pressure can be accurately captured. Construction disturbance factor ( C d This study innovatively quantifies the spatiotemporal coupling effect of tunnel face excavation. Its dynamic evolution law, revealed through the correlation function between the step excavation and the excavation sequence, demonstrates the propagation and attenuation characteristics of construction disturbance waves in the surrounding rock medium. The time-series data of surrounding rock deformation fully records the dynamic trajectories such as deformation rate and acceleration stage, providing a time-dimensional observation benchmark for mechanism analysis. These four parameters construct a multi-scale characterization system for surrounding rock deformation from four dimensions: mechanical equilibrium, structural response, disturbance propagation, and dynamic evolution. The first three, as mechanistic characteristic parameters, together with the deformation time-series data, constitute the multi-source input of the deep learning model, realizing the systematic analysis of the spatiotemporal coupling mechanism. Therefore, based on the tunnel surrounding rock deformation mechanism and the needs of deep learning prediction, this embodiment selects four key characteristic parameters as input variables: surrounding rock contact pressure, steel arch stress, surrounding rock deformation time-series data, and construction disturbance factor (…). C d ), and constructed a feature input system that integrates multi-dimensional dynamic features.

[0052] During tunnel construction, on-site monitoring data forms the basis of surrounding rock deformation prediction models. However, due to the complexity of the on-site environment and the limitations of monitoring equipment, monitoring data often suffers from problems such as missing data, noise, discontinuity, and insufficient data points. These problems directly affect the training effect and prediction accuracy of the model. Therefore, this embodiment uses the Makima interpolation completion method to process the monitoring data to ensure data integrity and quality. This embodiment uses the Makima (Modified Akima) interpolation completion method, which is an improvement and extension of the Akima interpolation method. The Akima interpolation method is an interpolation algorithm based on local data points. It interpolates within a small region between adjacent data points, utilizing the slope information of local data points to achieve interpolation. Makima interpolation mainly optimizes the smoothness and stability of the interpolation function compared to Akima interpolation. When calculating the slope, the Makima interpolation algorithm uses more data points to fit the local curve, thereby reducing the fluctuation of the interpolation result. In practical applications, it can effectively solve the oscillation and discontinuity problems that occur during data interpolation and optimize the quality of time-series data.

[0053] The method for predicting surrounding rock deformation is based on a CNN-BiLSTM-Attention hybrid model, which achieves rolling prediction and accuracy control through a time sliding window and a dynamic error feedback mechanism.

[0054] A time-sliding window was used to divide the time-series data into dynamic inputs and outputs. The first n sets of monitoring data (including rock contact pressure, steel arch stress, etc.) were selected. C d The model generates an input sequence (including surrounding rock deformation). n+1 sets of predicted values ​​are used as the model's output, and these predicted values ​​are then fed back into the input sequence to replace the earliest values, thus ensuring the model's rolling operation. This method predicts four sets of features from the n+1 sets, and then inputs the updated sequence into the model to obtain the deformation prediction value for the next time point. By iterating this process, the future deformation value of the surrounding rock can be obtained, enabling the model to achieve time-rolling prediction.

[0055] However, deep learning models have certain uncertainties in actual prediction processes, especially when facing complex geological conditions and excavation disturbances, the prediction results may deviate from the actual values. In multi-step surrounding rock deformation prediction, if the error is not corrected in time, the error will gradually accumulate, eventually leading to a significant difference between the prediction result and the actual result. Therefore, this embodiment proposes an error propagation strategy to avoid the accumulation of errors in rolling prediction. At each time step, the model predicts the following time series data based on the current input data. The model starts prediction from the (n+1)th set of data. After multiple rounds of prediction, when the actual monitoring data arrives, it will be immediately replaced by the corresponding predicted value in the input sequence. The time series data predicted by the model is compared with the actual monitoring data, and the prediction error of the (n+1)th set of data is calculated. The calculated error will be fed back into the model. Specifically, the error is propagated to multiple sets of prediction results that have been predicted, from the (n+1)th set of data to the (n+j)th set of data, to adjust the prediction results. The calculation method of error propagation is shown in equation (20). After error propagation correction, the model continues to predict the time series data of group (n+j+1) based on the latest corrected group (n+j). When the actual monitoring data of group (n+2) arrives, this method is used again for data replacement and error transfer to predict the time series data of group (n+j+2). Through this dynamic error propagation strategy, the model can update the input sequence each time new measurement data is obtained and adjust subsequent time predictions according to the error, thereby significantly reducing the bias amplification effect in multi-step prediction.

[0056] (20) This method ensures the accuracy and stability of the model through a time sliding window and an error propagation mechanism. It can respond in real time to changes in the deformation of the surrounding rock during construction and gradually correct the prediction results, thereby improving the accuracy of the prediction. Figure 6 The method intuitively demonstrates the rolling data updates and error propagation path, highlighting its innovative value in predicting the dynamic stability of tunnel engineering.

[0057] This research on surrounding rock deformation prediction based on a CNN-BiLSTM-Attention model during tunnel construction aims to improve the safety and intelligence of tunnel construction. On-site data acquisition is fundamental to the successful application of this prediction model, including key parameters such as surrounding rock contact pressure, steel arch stress, surrounding rock deformation, and construction disturbance factors. By deploying monitoring elements at the construction cross-section, the contact pressure, steel arch stress, and surrounding rock deformation are measured in real time, while the area of ​​each excavation step and the advance of each excavation are recorded to obtain complete time-series data. After obtaining this data, a multivariate input surrounding rock deformation regression prediction model and a time-series rolling prediction model are constructed based on the CNN-BiLSTM-Attention deep learning model. The accuracy of the deformation prediction is verified by comparing the multi-model regression prediction models, and the time-series rolling model is used to predict the future deformation trend of the surrounding rock, providing a more forward-looking early warning. This method can effectively support intelligent monitoring and safety management in tunnel construction, providing a scientific basis for risk control and decision-making during tunnel construction, and has high application potential and practical significance.

[0058] This embodiment is based on an engineering project described by Li et al. The project involves tunnel construction and monitoring. This project has relatively detailed monitoring records to accurately capture and analyze relevant data, providing a basis for further engineering decisions. The following is a basic description of the project: The Lehua Tunnel is located in Haidong City, Qinghai Province, China. The entrance is in Ledu District, and the exit is in Hualong County. Its geographical location is as follows: Figure 7 As shown. The tunnel is designed with a four-lane independent twin-tube layout, with left and right lanes measuring 5916m and 5933m in length, respectively. The maximum depth of the tunnel is approximately 796m, classifying it as a deep-rock tunnel. The highest point at the tunnel site is located on the Mayin Mountain ridge, at an elevation of approximately 4165 meters.

[0059] The geological conditions along the Lehua Tunnel are complex, with the surrounding rock mainly classified as Class IV and Class V. The strata along the tunnel primarily consist of gravelly soil, strongly to moderately weathered diorite, strongly to moderately weathered argillaceous sandstone, and completely to strongly to moderately weathered granite. The weathered granite strata are locally affected by faults, exhibiting a predominantly fragmented blocky structure with well-developed fissures and uneven rock mass quality. Groundwater in the tunnel site area mainly consists of bedrock fissure water and clastic rock pore water. According to geological surveys, tunnel excavation must cross five fault zones. Groundwater primarily occurs in these fault zones, significantly impacting construction. These fault zones disrupt the integrity and continuity of the rock mass, making it prone to collapse and softening upon contact with water, resulting in poor self-stability.

[0060] This embodiment monitors the deformation of the surrounding rock in the early stage of tunnel construction, the contact pressure of the surrounding rock in the early stage of tunnel construction, and the internal force of the support steel arch frame in the early stage of tunnel construction.

[0061] Tunnel monitoring was conducted at section K41+502. Figure 8 The layout of the measuring elements and monitoring points is shown. To monitor the settlement of the tunnel arch, three reflectors (P1, P2, P3) were installed on the arch, and two measuring lines (Line1 and Line2) were arranged on the upper and lower steps around the perimeter. In addition, pressure units were used to monitor the lateral pressure on the main support caused by soil settlement around the tunnel. A total of seven measuring points were set up, numbered E1, E2, E3, E4, E5, E6, and E7. The pressure units were installed between the surrounding rock and the main support. In the stress test of the steel arch, strain gauges were used and fixed to monitoring positions on a steel frame connected by welds. The strain gauges were securely mounted on the steel arch with bolts. Five strain gauges (S2, S3, S4, S5, and S6) were installed after the upper step of the tunnel was excavated, and subsequently, two more strain gauges (S1 and S7) were installed after the lower step was excavated.

[0062] like Figure 9 As shown, the monitoring data obtained indicates that the tunnel cross-section is approximately... The excavation area of ​​the first and second steps is approximately The third step is approximately The daily excavation progress is approximately Construction disturbance factors such as Figure 10 As shown.

[0063] Due to the limited number of data points, interpolation was performed once between each data point based on the actual deformation monitoring data measured on-site. The temporal curve of the deformation data is shown below. Figure 11 As shown.

[0064] The construction disturbance factor, steel arch stress, and surrounding rock contact pressure at section K41+502 were used as input data for the CNN-BiLSTM-Attention prediction model, with daily cumulative deformation data as output data. The first 70% of the data was selected as the training set, and the last 30% as the test set. The hyperparameters of the CNN-BiLSTM-Attention model were appropriately set. The CNN network consisted of two convolutional layers: convolutional layer 1 had a kernel of (2,1), a stride of (1,1), and 32 channels; convolutional layer 2 had a kernel of (2,1), a stride of (1,1), and 64 channels. The activation function was ReLU, and the pooling layer was global pooling. The BiLSTM neural network consisted of two layers: a forward LSTM and a backward LSTM, with the activation function being tanh. Finally, the surrounding rock deformation prediction results during construction were obtained through the attention layer and the fully connected layer.

[0065] The data from the left shoulder area was input into the CNN-BiLSTM-Attention prediction model for prediction, and compared with the conventional prediction model to further verify the effectiveness of the model.

[0066] The predicted deformation results of the left arch shoulder of the tunnel are as follows: Figure 12 As shown in Table 1, the evaluation metrics for the test set model are as follows.

[0067] Table 1. Evaluation results of the model metrics for the test set.

[0068] The prediction results are analyzed as follows: (1) From Figure 12 The prediction results show that, compared to other models, the CNN-BiLSTM-Attention prediction model performs better in predicting the deformation of the left arch shoulder of the surrounding rock during tunnel construction, and the prediction trends of the models are basically consistent. It can be observed that the predicted deformation values ​​are sometimes smaller than the previous day's predictions. This is because the feature parameters input to the model only map the physical state of the tunnel construction process to the deformation state of the surrounding rock, without fully considering the increasing trend of deformation.

[0069] (2) Analysis of the model evaluation indicators in Table 1 shows that the CNN-BiLSTM-Attention model enhances the prediction ability. Its key indicators, MAE (0.5122), MAPE (0.0162), and RMSE (0.5746), are reduced by 70.2%, 4.8%, and 68.1% respectively compared to the LSTM baseline model. Meanwhile, the average absolute error rate of this model is 1.62%, indicating that the CNN-BiLSTM-Attention prediction model outperforms other models in predicting the deformation of the left arch shoulder. Selecting this model can better extract the features of tunnel time-series monitoring data. This model comprehensively considers the sequential information of time-series data, combining the feature extraction capabilities of convolutional neural networks (CNNs) with the advantages of sequence modeling and attention mechanisms of bidirectional long short-term memory networks. Therefore, it possesses stronger prediction and generalization abilities, and can more accurately describe the relationships between data, making it particularly suitable for the prediction task of tunnel time-series monitoring data.

[0070] Tunnel arch settlement prediction results are as follows Figure 13 As shown in Table 2, the evaluation metrics for the test set model are as follows.

[0071] Table 2. Evaluation Results of Test Set Model Indicators

[0072] The analysis of the predicted settlement results of the arch is as follows: From the deformation prediction results of the model Figure 13 Overall, it can be seen that the CNN-BiLSTM-Attention prediction model can better predict the cumulative settlement and deformation of the surrounding rock arch during tunnel construction compared with other models, and is more in line with the deformation trend.

[0073] From the perspective of error in Table 2, the MAPE of the dome model is only 0.977%, which is 3.8% higher than the LSTM baseline model. Furthermore, it outperforms several other models in all other evaluation metrics, including R... 2 The value is 0.834, which shows that the CNN-BiLSTM-Attention prediction model is far superior to other models in terms of accuracy in predicting tunnel crown settlement. Selecting this model can better extract the features of the time series monitoring data of the tunnel crown, which has important guiding significance for on-site safe construction in special deformation sections of the tunnel.

[0074] The predicted deformation results of the right arch shoulder of the tunnel are as follows: Figure 14 As shown in Table 3, the evaluation metrics for the test set model are as follows.

[0075] Table 3. Evaluation Results of Test Set Model Indicators

[0076] The analysis of the predicted deformation results of the right shoulder is as follows: Based on deformation prediction results Figure 14 As shown in Table 3, both the CNN-LSTM-Attention model and the CNN-BiLSTM-Attention model demonstrate good performance in predicting the deformation of the right shoulder of the surrounding rock. The CNN-BiLSTM-Attention model shows a 2.8% improvement in MAPE compared to the LSTM model, and it outperforms the CNN-LSTM-Attention model in all aspects of prediction. The difference lies in the fact that the CNN-BiLSTM-Attention model can more effectively handle long sequences and important features, thus improving its performance. In regression prediction, the field monitoring data is affected by various factors, and both models exhibit some error.

[0077] The prediction results of the CNN-BiLSTM-Attention regression prediction model demonstrate a strong mapping relationship between the input and output variables, fully illustrating the applicability and rationality of the selected feature parameters for time-series rock rolling prediction. Therefore, based on the regression prediction model, a time-series rock rolling prediction model with multivariate feature input and single output was established to achieve advanced prediction of rock deformation.

[0078] Taking the monitoring data of section K41+502 as an example, 41 sets of time-series data in 4 dimensions were used, including construction disturbance factor, stress of steel arch frame at arch shoulder and arch crown, contact pressure of surrounding rock, and deformation of surrounding rock. A time sliding window was used to divide the time-series data into the model's input and output. n The sequence samples, composed of monitoring data from the group, are used as model input. n The rock deformation monitoring data from group +1 is used as the model's output for rolling training. The predicted value is fed back into the input sequence, replacing the earliest value to ensure continuous model operation. This updated sequence is then input into the model to obtain the deformation prediction value for the next time point. This iterative process yields the future rock deformation value. When predicting the (n+4)th data group, the actual measured data from group n+1 is input in real-time, and the average error between this data and the predicted data is passed to the subsequent two groups. This allows us to continuously obtain the real-time updated prediction data for the last three groups.

[0079] To improve the optimization performance of the model and allow the neural network to find suitable parameters during learning so that the loss function is as small as possible, the optimizer is Adam, the initial learning rate is 0.03, the maximum number of training iterations is 1000, and after training to 700, the learning rate is changed to 0.003.

[0080] Table 4 shows the predicted values ​​for tunnel arch settlement and left and right abutment deformation. The prediction visualization results are shown below. Figure 15As shown.

[0081] Table 4 Model Prediction Results

[0082] The model evaluation metrics are shown in Table 5.

[0083] Table 5. Evaluation Results of Model Indicators

[0084] like Figure 15 As shown in the results, the CNN-BiLSTM-Attention model exhibits excellent predictive capabilities. Specifically, the right abutment deformation prediction metrics (MAE, MAPE, and RMSE) are 0.1644, 0.65%, and 0.1935, all outperforming the predictions for crown settlement and left abutment deformation. 2 The maximum value is 0.9018. Therefore, the overall prediction effect of the right shoulder is better than that of the left shoulder and the crown. This model selects prediction parameters such as construction disturbance factor and internal force of initial support structure, which can effectively overcome the errors caused by many factors such as construction disturbance, geological conditions, and support measures, and more realistically reflect the variation law of displacement of the surrounding rock of gently dipping layered carbonaceous shale tunnels. As the number of predictions increases, we can find that by using the error propagation method, the predicted value gradually matches the actual value, and can better predict the deformation of the tunnel surrounding rock.

[0085] Establishing a rolling time-series prediction model for tunnel surrounding rock deformation can effectively capture key features of tunnel monitoring data and predict future trends of tunnel surrounding rock changes over time in special sections. This model can guide the timing of secondary lining support in tunnels, ensuring safe tunnel construction. Furthermore, it can continuously update the prediction data for the next n days in real time, according to specific needs. This model also serves as a new supplement to traditional data processing methods and is of great significance for improving the level of intelligent tunnel construction.

[0086] This embodiment takes into account the impact of face excavation during construction and determines a new characteristic variable, namely the construction disturbance factor. C d During engineering construction, construction disturbance factors have a significant impact on the stability and deformation control of underground structures. Considering the influence of construction disturbance factors is crucial for improving the accuracy of prediction results when forecasting the deformation of underground structures or tunnels.

[0087] To verify the impact of construction disturbance factors, this embodiment compares the prediction results with and without construction disturbance factors. The specific method is as follows: First, without introducing construction disturbance factors into the prediction model, predictions are made solely based on the surrounding rock contact pressure and the stress state of the steel arch frame to obtain baseline prediction results. Then, construction disturbance factors are incorporated into the model to obtain prediction results with disturbance factors.

[0088] The model evaluation metrics are shown in Table 6.

[0089] Table 6. Evaluation Results of Model Indicators

[0090] Comparison of the two sets of prediction results shows that the predicted values ​​considering construction disturbance factors are in better agreement with the actual measured data, with smaller errors. In contrast, the model that did not consider disturbance factors exhibited significant deviations in its predictions. Specifically, the MAPE of the surrounding rock deformation at the left abutment increased from 7.11% to 1.62%, the MAPE of the surrounding rock deformation at the arch crown increased from 22.61% to 0.98%, and the MAPE of the surrounding rock deformation at the right abutment increased from 10.39% to 1.21%. 2 All values ​​were negative when the disturbance factor was not considered, but approached 1 after considering the construction disturbance factor. The predicted values ​​considering the construction disturbance factor also showed better results for other evaluation parameters than those without. This indicates that the construction disturbance factor has a significant impact on the prediction results of surrounding rock deformation; ignoring these factors may lead to large prediction errors, thereby affecting the formulation of construction plans and risk assessments.

[0091] This comparison leads us to the conclusion that the impact of construction disturbance factors must be fully considered when predicting tunnel surrounding rock deformation in order to improve prediction accuracy and project safety. Therefore, incorporating construction disturbance factors not only effectively improves the accuracy of the prediction model but also provides a more reliable basis for risk control during construction.

[0092] The timing of secondary lining support in tunnels is a critical factor directly affecting the structural integrity and safety of the tunnel. This embodiment utilizes CNN-BiLSTM-Attention to predict tunnel arch settlement in order to determine the optimal timing for secondary lining support. By analyzing measured and predicted displacement data, we aim to provide an evaluation method for this crucial decision-making process.

[0093] According to the "GB50086 Technical Specification for Rock and Soil Anchor and Shotcrete Support Engineering", secondary lining support should be carried out after the horizontal convergence velocity around the tunnel and the vertical displacement velocity of the arch or bottom plate have decreased significantly. Specifically, the horizontal convergence velocity around the tunnel should be less than 0.2 mm / d for 5 consecutive days, and the vertical displacement velocity of the arch or bottom plate should be less than 0.1 mm / d. In this embodiment, the optimal timing for secondary lining support will be determined based on the arch settlement data.

[0094] A CNN-BiLSTM-Attention-based intelligent prediction model for surrounding rock deformation was established by comprehensively utilizing field monitoring data such as the stress of the initial support steel arch, the contact pressure between the initial support and the surrounding rock, construction disturbance factors, and cumulative deformation data of the surrounding rock. This model enables time-series prediction of the cumulative deformation of the surrounding rock. First, the tunnel's on-site deformation monitoring data, construction disturbance factors, steel arch stress, and surrounding rock contact pressure are input into the CNN-BiLSTM-Attention prediction model. The predicted cumulative deformation of the surrounding rock is obtained using a rolling time-series model, and the deformation rate is calculated. Based on construction standard criteria, the timing of secondary lining support is determined.

[0095] Taking section K41+502 as the research object, the monitoring data during the construction period was input into the CNN-BiLSTM-Attention prediction model, with the data from the previous 21 days used for training. Due to the lack of measured deformation monitoring data, the predicted data for the following day was input into the model for rolling prediction. The cumulative deformation prediction result of the arch is as follows: Figure 16 As shown.

[0096] The cumulative data for the predicted cumulative deformation of the arch are shown in Table 7: Table 7. Cumulative data for predicted cumulative deformation of the arch.

[0097] According to relevant standards, the optimal support time is when the vertical displacement velocity of the tunnel arch or floor is less than 0.1 mm / d for five consecutive days. Based on the predicted deformation curve, the optimal support time for the secondary lining at section K41+502, determined by the arch deformation rate, is 32 days. By establishing an intelligent prediction model, the changing trends and anomalies of the tunnel surrounding rock can be detected in a timely manner, providing an important approach and method for determining the support timing of the secondary lining in gently dipping layered carbonaceous shale tunnels.

[0098] Example 2 A time-series prediction system for tunnel surrounding rock deformation rolling considering excavation disturbance includes: The construction disturbance factor calculation module is configured to obtain the geometric relationship between the tunnel face and the monitoring point, the excavation area, and the excavation progress during the tunnel excavation process, and calculate the construction disturbance factor based on the obtained data; The model building and training module is configured to build and train a surrounding rock deformation prediction model, which is obtained by fusing a convolutional neural network, a bidirectional long short-term memory neural network and an attention mechanism layer. The data acquisition module is configured to acquire tunnel surrounding rock monitoring data and preprocess the monitoring data; The deformation prediction module is configured to use preprocessed monitoring data and calculated construction disturbance factors as input data for the trained surrounding rock deformation prediction model. The convolutional neural network performs multi-layer convolution on the input data to extract deep features of the monitoring data and obtain relational feature vectors. The bidirectional long short-term memory neural network extracts the feature relationships in the relational feature vectors. The outputs of the forward long short-term memory neural network and the backward long short-term memory neural network in the bidirectional long short-term memory neural network are concatenated. An attention mechanism layer is used to assign weights to the concatenated result to obtain the final deformation prediction result.

[0099] Example 3 A computer-readable storage medium for storing computer instructions, which, when executed by a processor, perform the steps in the above method.

[0100] Example 4 An electronic device includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor, wherein the computer instructions, when executed by the processor, perform the steps in the method described above.

[0101] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made by those skilled in the art without creative effort within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting the time series deformation and rolling of tunnel surrounding rock considering excavation disturbance, characterized in that, Includes the following steps: Obtain the geometric relationship between the tunnel face and the monitoring point, the excavation area, and the excavation progress during the tunnel excavation process, and calculate the construction disturbance factor based on the obtained data; A surrounding rock deformation prediction model is constructed and trained. The surrounding rock deformation prediction model is obtained by fusing a convolutional neural network, a bidirectional long short-term memory neural network and an attention mechanism layer. Acquire monitoring data of the surrounding rock of the tunnel and preprocess the monitoring data; The preprocessed monitoring data and the calculated construction disturbance factor are used as input data for the trained surrounding rock deformation prediction model. The convolutional neural network performs multi-layer convolution on the input data to extract deep features of the monitoring data and obtain relational feature vectors. The bidirectional long short-term memory neural network extracts the feature relationships in the relational feature vectors. The outputs of the forward long short-term memory neural network and the backward long short-term memory neural network in the bidirectional long short-term memory neural network are concatenated. The attention mechanism layer is used to assign weights to the concatenated result to obtain the final deformation prediction result.

2. The method for predicting the deformation and rolling time series of tunnel surrounding rock considering excavation disturbance as described in claim 1, characterized in that, The construction disturbance factor is: ； ； Where, in the formula, Cd,i Indicates the first i Excavation disturbance caused by step excavation; Ai For the tunnel working face i The area of ​​each excavation step is in square meters. Li,ki For the first i Step 1 k Within the construction cycle, the first i Distance between the excavation face and the monitoring section, parameters i For the first i Steps, N steps in total, parameters ki Indicates the first i The number of excavation cycles per step, as excavation progresses, Li,ki The excavation coefficient increases with each cycle. Cd Over time, this reflects the increased disturbance to the monitoring section caused by continuous excavation.

3. The method for predicting the deformation and rolling time series of tunnel surrounding rock considering excavation disturbance as described in claim 1, characterized in that, The convolutional neural network includes an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer connected in sequence.

4. The method for predicting the deformation and rolling time series of tunnel surrounding rock considering excavation disturbance as described in claim 1, characterized in that, The attention mechanism layer is used to calculate the similarity score between the query and the key based on the input data (query, key, value). By performing softmax normalization on the similarity score, the attention weight of each key is obtained. The values ​​are then weighted and summed according to the corresponding attention weights to obtain the final context vector or weighted value. Based on the context vector or weighted value, a prediction result is generated.

5. The method for predicting the deformation and rolling time series of tunnel surrounding rock considering excavation disturbance as described in claim 1, characterized in that, The tunnel surrounding rock monitoring data includes surrounding rock contact pressure, steel arch stress, and surrounding rock deformation time series data.

6. The method for predicting the deformation and rolling time series of tunnel surrounding rock considering excavation disturbance as described in claim 1, characterized in that, The training process of the surrounding rock deformation prediction model includes selecting the first n sets of monitoring data to form an input sequence, using the (n+1)th set of predicted values ​​as the model output, then feeding the predicted values ​​back into the input sequence to replace the earliest values, predicting four sets of features in the (n+1)th set, and then inputting the updated sequence into the model to obtain the deformation prediction value at the next time point, thus achieving time-rolling prediction.

7. The method for predicting the deformation and rolling time series of tunnel surrounding rock considering excavation disturbance as described in claim 6, characterized in that, The model's predicted time series data is compared with the actual monitoring data. The prediction error of the (n+1)th set of data is calculated, and the calculated error is fed back into the model for error propagation and correction.

8. A time-series prediction system for tunnel surrounding rock deformation and rolling considering excavation disturbance, characterized in that, include: The construction disturbance factor calculation module is configured to obtain the geometric relationship between the tunnel face and the monitoring point, the excavation area, and the excavation progress during the tunnel excavation process, and calculate the construction disturbance factor based on the obtained data; The model building and training module is configured to build and train a surrounding rock deformation prediction model, which is obtained by fusing a convolutional neural network, a bidirectional long short-term memory neural network and an attention mechanism layer. The data acquisition module is configured to acquire tunnel surrounding rock monitoring data and preprocess the monitoring data; The deformation prediction module is configured to use preprocessed monitoring data and calculated construction disturbance factors as input data for the trained surrounding rock deformation prediction model. The convolutional neural network performs multi-layer convolution on the input data to extract deep features of the monitoring data and obtain relational feature vectors. The bidirectional long short-term memory neural network extracts the feature relationships in the relational feature vectors. The outputs of the forward long short-term memory neural network and the backward long short-term memory neural network in the bidirectional long short-term memory neural network are concatenated. An attention mechanism layer is used to assign weights to the concatenated result to obtain the final deformation prediction result.

9. A computer-readable storage medium, characterized in that, Used to store computer instructions, which, when executed by a processor, complete the steps of the method according to any one of claims 1-7.

10. An electronic device, characterized in that, It includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor, which, when executed by the processor, perform the steps of the method according to any one of claims 1-7.

Images