A bidirectional LSTM-based underground tunnel support timing analysis method

Abstract

The application relates to the technical field of deep learning applied to underground tunnel construction, and discloses a method for analyzing underground tunnel support timing based on a bidirectional LSTM, which comprises the following steps: obtaining surrounding rock geology condition data of actual underground tunnel construction as to-be-analyzed condition data; loading an underground tunnel support timing analysis model; the underground tunnel support timing analysis model adopts a network structure combining a bidirectional LSTM and an attention mechanism, is trained and generated by using underground tunnel environment samples constructed by tunnel surrounding rock displacement time sequence characteristics under typical working conditions, and is used for realizing accurate prediction of the best support distance of the underground tunnel; and the to-be-analyzed condition data is inputted, and the minimum distance corresponding to the best support timing under the surrounding rock geology condition is obtained. According to the technical scheme, the best support timing of the underground tunnel can be determined under multiple surrounding rock categories and different ground stress conditions, construction safety is ensured, efficiency is improved, and cost is reduced.

Description

Technical Field

[0001] This invention relates to the technical field of applying deep learning to underground tunnel construction, and more specifically, to a method for analyzing the timing of underground tunnel support based on bidirectional LSTM. Background Technology

[0002] In underground tunnel construction, accurately determining the timing of support is crucial for ensuring rock mass stability, mitigating geological hazards such as collapses, and controlling project costs. Supporting too early increases construction interference and costs, while supporting too late leads to excessive rock displacement, decreased stability, and a higher risk of rock bursts and collapses, severely threatening construction safety and delaying the project schedule. Currently, the analysis of underground tunnel support timing mainly relies on two methods: numerical simulation (such as FLAC3D and ABAQUS software) and on-site monitoring. Numerical simulation requires building specific numerical models for different rock types, geostress conditions, and geological structures, resulting in complex and time-consuming calculations that demand highly skilled operators, making it unsuitable for rapid on-site construction. On-site monitoring is affected by factors such as monitoring point placement and environmental interference, leading to unstable data accuracy. Furthermore, it can only passively acquire rock displacement data and cannot predict the optimal support timing in advance, hindering proactive and precise control of support timing.

[0003] While there are some attempts to apply machine learning / deep learning to tunnel construction analysis, most focus on predicting the stability level of surrounding rock, early warning of tunnel face instability, or recommending support schemes. They haven't studied the core indicator for support timing—the minimum distance from the tunnel face. Furthermore, existing deep learning applications cannot accurately capture the long-distance temporal dependencies of surrounding rock deformation in their network structures, limiting model accuracy and generalization ability. There has also been a lack of accumulation and application of data on different surrounding rock types (II to V) and different ground stresses (low to high ground stress), making it impossible to quickly and accurately output the optimal support timing under different geological conditions.

[0004] Therefore, there is an urgent need for a technical solution to overcome the shortcomings of existing technologies, such as the time-consuming nature of underground tunnel support timing analysis, low accuracy, poor adaptability, and inability to quickly respond to on-site construction needs. This solution should fully utilize dedicated data from multi-condition numerical simulations and combine it with intelligent optimization algorithms to determine the optimal support timing for underground tunnels under various rock types and different ground stress conditions. This would provide scientific guidance for on-site construction, ensure construction safety, improve efficiency, and reduce costs. Summary of the Invention

[0005] To achieve the above objectives, this application provides a method for analyzing the timing of underground tunnel support based on bidirectional LSTM, comprising the following steps: Obtain the geological conditions of the surrounding rock during actual underground tunnel construction as the data to be analyzed; Loading an underground tunnel support timing analysis model; the underground tunnel support timing analysis model adopts a bidirectional LSTM network structure combined with an attention mechanism. It is generated by training underground tunnel environment samples constructed from the temporal characteristics of tunnel surrounding rock displacement under typical working conditions, and is used to achieve accurate prediction of the optimal support distance of underground tunnels; typical working conditions refer to: surrounding rock grade supporting Class II to V, and ground stress conditions supporting high, medium and low levels. Input the data of the conditions to be analyzed into the underground tunnel support timing analysis model to obtain the minimum distance corresponding to the optimal support timing under the surrounding rock geological conditions.

[0006] Among them, before loading the underground tunnel support timing analysis model, the underground tunnel support timing analysis model is generated. The steps involved in generating the underground tunnel support timing analysis model are as follows: Construct multi-dimensional boundary conditions for typical working conditions to generate simulation scenarios; Sample pairs are constructed based on time-series features collected from simulated scenarios; The sample pairs are preprocessed to generate underground tunnel environment samples; the underground tunnel environment samples are then divided into training set, validation set and test set. Define the analysis model; the analysis model adopts an end-to-end network structure combining bidirectional LSTM with an attention mechanism; The analysis model was trained using environmental samples from underground tunnels to construct an analysis model for the timing of underground tunnel support.

[0007] When constructing multi-dimensional boundary conditions for typical working conditions, geological working conditions are matrixed to obtain orthogonal parameter matrices; the orthogonal parameter matrices reflect various surrounding rock grades, different geostress levels, and different cross-sectional shapes and sizes. The simulation scenario is generated using 3D simulation software based on an orthogonal parameter matrix, and is used to simulate the step-by-step excavation process of actual construction.

[0008] Furthermore, constructing sample pairs includes: Monitoring points were set up to collect time series data D(t) of the surrounding rock displacement as the working face advanced; The process of gradually excavating the tunnel face is simulated, and the displacement release amount at each monitoring point is collected in real time. The displacement release rate η at each monitoring point is calculated, and the spatial distance L between the tunnel face and the monitoring point corresponding to the displacement release rate η at a preset threshold is defined as the target label. Y ; Define a sliding window sequence formed by the time series data of the current step and the previous n historical steps as the input feature X, combined with the target label. Y Construct sample pairs ( X , Y ).

[0009] The generated underground tunnel environment samples include: Gaussian noise perturbation and moving average smoothing are introduced into the sample pairs to construct an enhanced sample set covering multiple operating conditions; Based on the enhanced sample set, displacement rate and displacement acceleration are extracted, and displacement release rate is calculated as a surrounding rock stability discrimination index to construct a multi-dimensional feature vector. The multidimensional feature vectors are scaled to generate a unified sample set.

[0010] Furthermore, the bidirectional LSTM network structure combined with the attention mechanism is set up with an input layer, hidden layer, regularization layer, fully connected layer and output layer in sequence along the data transmission direction. The features are transferred and transformed step by step between each layer through tensor dimension matching. By capturing the long-distance dependency of the surrounding rock displacement release time series data and screening the key stage features of the surrounding rock deformation, the accurate prediction of the optimal support distance of the underground tunnel is achieved. The trainable parameters of the overall network structure are optimized globally through backpropagation. The input layer receives time-series data with a structure of [Batch_size, Time_steps, Features], including displacement, velocity, surrounding rock grade coding, and in-situ stress value; the output layer outputs the optimal support distance predicted by a single node.

[0011] Furthermore, the hidden layers include bidirectional LSTM layers and attention layers; Among them, the bidirectional LSTM layer is used to simultaneously capture the impact of historical information on the current state and the inverse inference of future trends on the current state, generating a bidirectional fused feature tensor to achieve complete extraction of full temporal features. The Attention layer introduces an attention mechanism to automatically calculate the weight of time nodes; the time nodes include the moment of excavation and the stage of ground stress redistribution.

[0012] In this process, after the time-series data enters the bidirectional LSTM layer, it is divided into two parallel data streams: a forward data stream and a reverse data stream. The forward data stream is a data stream transmitted in ascending time sequence, according to the excavation step sequence t1→t2→…→t n Forward propagation extracts historical deformation features; the reverse data stream is a time-reverse data transmission stream, based on the excavation step sequence t. n →t n-1 →…→t1 backpropagation, extracting future deformation features; The two data streams output feature vectors at each time step, which are then concatenated and fused into a bidirectional fused feature tensor.

[0013] Furthermore, before training the analysis model, the hyperparameters of the analysis model are globally optimized using the particle swarm optimization algorithm. During training, the training set is input into the analysis model, and the model parameters are locally iteratively optimized using the adaptive momentum optimization algorithm. The number of training iterations is set to 500-1000, the initial learning rate is 0.001, and the learning rate is adaptively adjusted according to the changes in the loss function during training. The mean squared error is used as the loss function. During training, the analysis model is validated in real time using a validation set, and the structure and parameters of the analysis model are optimized based on the validation results. After training, the analysis model is tested and cross-validated using a test set to ensure that the analysis model is within the boundary range of typical working conditions.

[0014] If the data under the conditions to be analyzed exceeds the range of typical working conditions, supplement the numerical model and data for the corresponding working conditions to fine-tune and train the underground tunnel support timing analysis model.

[0015] According to the present invention, the optimal support timing can be accurately determined, avoiding construction interference and cost waste caused by supporting too early, and avoiding geological disasters such as surrounding rock collapse and rock burst caused by supporting too late, thereby reducing construction costs, reducing the incidence of safety accidents and rework costs, and facilitating efficient on-site construction. Attached Figure Description

[0016] Figure 1 This is a step diagram of the underground tunnel support timing analysis method based on bidirectional LSTM provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the dataset structure of the orthogonal parameter matrix provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the layout of monitoring points in an underground tunnel according to an embodiment of the present invention. Detailed Implementation

[0017] Tunnel surrounding rock displacement data has strong temporal sequence, nonlinearity and multi-factor coupling characteristics, and displacement release has lag and long-term dependence. Long Short-Term Memory (LSTM) network has specialized temporal modeling and long-range memory capabilities. Through gating mechanism, it can effectively capture deep correlations between sequences. Therefore, it is particularly suitable for processing data with complex temporal dynamic characteristics such as tunnel surrounding rock displacement release.

[0018] This invention provides a technical solution that constructs a dedicated dataset for multi-condition numerical simulation, builds a bidirectional LSTM-attention mechanism model, combines intelligent optimization algorithms to optimize hyperparameters, and trains and generates an analysis model for the timing of underground tunnel support. This enables the rapid and accurate calculation of the minimum length of the optimal support from the tunnel face under various surrounding rock and geostress conditions.

[0019] The specific implementation of the present invention will now be described in detail with reference to the accompanying drawings.

[0020] Step S100: Obtain the geological conditions of the surrounding rock during actual underground tunnel construction as the data to be analyzed; The geological conditions data of the surrounding rock include the displacement release characteristics of the surrounding rock, the type of surrounding rock, and the parameters of the geostress level.

[0021] This invention provides a specific implementation example: an underground tunnel project in a water conservancy and hydropower project, with a tunnel cross-section of 8m × 9.5m, an excavation speed of 1.0m / d, and an excavation step distance of 0.8m / step. It involves two types of surrounding rock, Class II and Class IV, and two levels of ground stress, low and high. To determine the minimum length of the optimal support distance from the tunnel face, displacement release characteristic data of the surrounding rock under Class II surrounding rock and low ground stress conditions, and Class IV surrounding rock and high ground stress conditions, were collected as two sets of data to be analyzed.

[0022] Step S110: Load the underground tunnel support timing analysis model; Before loading the underground tunnel support timing analysis model, the underground tunnel support timing analysis model is generated, including the following steps: Step S111: Construct multi-dimensional boundary conditions for typical working conditions to generate simulation scenarios; When constructing multi-dimensional boundary conditions for typical working conditions, geological working condition matrixing is performed. During geological working condition matrixing, an orthogonal parameter matrix is ​​obtained; the dataset structure of the orthogonal parameter matrix is ​​as follows: Figure 2 As shown, this reflects various surrounding rock grades, different in-situ stress levels, and different cross-sectional shapes and dimensions. Typical working conditions refer to surrounding rock grades supporting Class II to V, and in-situ stress conditions supporting high, medium, and low levels.

[0023] Furthermore, based on the orthogonal parameter matrix, a numerical model is established using 3D simulation software such as FLAC3D or ABAQUS, based on the Mohr-Coulomb or Hawke-Brown criteria. The model generates a simulation scene by simulating the step-by-step excavation process of actual construction with a step size of 0.5~1.0m, and accurately records the geostress-strain response of the surrounding rock under each excavation step.

[0024] To support the examples provided in this invention, the multi-dimensional boundary conditions can cover Class II and Class IV surrounding rock, and scenarios with low ground stress (8MPa, <10MPa) and high ground stress (28MPa, >20MPa); the numerical model simulates the actual tunnel excavation conditions in a way that is consistent with actual construction.

[0025] Step S112: Based on the temporal features collected from the simulated scene, construct sample pairs; Before collecting time-series data, monitoring points are set up to collect time-series data D(t) of surrounding rock displacement as the tunnel face advances. The monitoring point layout is as follows: Figure 3As shown, virtual monitoring points are arranged in key parts of the numerical model (arch top, arch waist, sidewalls). For example, monitoring points are evenly arranged at the top, sidewalls, and bottom of the tunnel, with a spacing of 1.2m. 20 monitoring points are arranged for each model.

[0026] The process of gradually excavating the tunnel face was simulated, and the displacement release at each monitoring point was collected in real time. The displacement release data at each time point constituted a time series data D(t). Based on the displacement back analysis theory, the displacement release rate η at each monitoring point was calculated. The focus was on recording the distance from the tunnel face corresponding to the displacement release rate η reaching a preset threshold (35%~40%). The spatial distance L from the monitoring point at the tunnel face corresponding to the preset threshold was defined as the target label. Y ; Define a sliding window sequence formed by the time series data of the current step and the previous n historical steps as the input feature X, combined with the target label. Y Construct sample pairs ( X , Y ).

[0027] Step S113: Preprocess the sample pairs to generate underground tunnel environment samples; The preprocessing process includes the following steps: 1) Considering the systematic differences between the data collected in the simulated scenario and the measured data in the field in terms of noise level, sampling interval and boundary constraints, Gaussian noise disturbance and moving average smoothing are introduced into the sample pairs to construct an enhanced sample set covering multiple working conditions, so as to improve the robustness and generalization ability to complex field environments.

[0028] 2) Based on the enhanced sample set, displacement rate is extracted by first-order difference, displacement acceleration is extracted by second-order difference, and displacement release rate is calculated as the stability index of surrounding rock. A multi-dimensional feature vector containing original displacement, deformation rate, acceleration and release rate is constructed to enhance the characterization ability of the dynamic law of deformation evolution of surrounding rock.

[0029] 3) Use Z-score standardization or Min-Max normalization to scale the multidimensional feature vectors to generate a unified sample set, so as to eliminate the difference in dimensions and the order of magnitude; and divide the unified sample set into training set (70%), validation set (20%) and test set (10%) according to the time series sequence.

[0030] Step S114: Define the analysis model. The analysis model adopts an end-to-end network structure of bidirectional LSTM combined with attention mechanism. The structure is set up with input layer, hidden layer, regularization layer, fully connected layer and output layer in sequence along the data transmission direction. The features are transferred and transformed step by step between layers through tensor dimension matching. By capturing the long-distance dependency of the surrounding rock displacement release time series data and screening the key stage features of surrounding rock deformation, the accurate prediction of the optimal support distance of underground tunnels is achieved. The trainable parameters of the overall network structure are optimized globally through backpropagation.

[0031] The specific structure, parameters, and functional settings for each layer are as follows: 1) Input layer: Receives time-series data with the structure [Batch_size, Time_steps, Features]. Features include displacement, velocity, surrounding rock grade code, and geostress value.

[0032] 2) The hidden layer includes a bidirectional LSTM layer and an Attention layer; Among them, the bidirectional LSTM layer is different from the ordinary LSTM. It can simultaneously capture the influence of historical information on the current state and the inverse inference of future trends on the current state, and better fit the nonlinear process of surrounding rock deformation. Specifically, after the input time series data enters the bidirectional LSTM layer, it is divided into two paths for parallel transmission: forward data stream and reverse data stream: (1) The forward data stream is a data stream transmitted in the forward time sequence, according to the excavation step sequence t1→t2→…→t n Forward propagation to extract historical deformation features; (2) Reverse data stream: Transmit data stream in reverse time order, according to the excavation step sequence t n →t n-1 →…→t1 backpropagation extracts future deformation features. The two data streams output feature vectors at each time step, which are concatenated and fused into a bidirectional fused feature tensor. Unlike unidirectional LSTM which only propagates along a single path, this achieves complete extraction of full-time-series features.

[0033] The Attention layer introduces an attention mechanism that automatically calculates the weights of time nodes such as the moment of excavation and the stress redistribution stage, giving higher attention to key deformation stages, solving the gradient vanishing problem in long sequence prediction, and effectively improving the model fitting accuracy.

[0034] 3) Regularization layer: Add a Dropout layer (with a ratio of 0.2 to 0.3) before the fully connected layer to prevent overfitting.

[0035] 4) Output layer: Outputs the optimal support distance predicted by a single node. L pred That is, the minimum distance from the working face.

[0036] Step S115: Train the analysis model to construct an analysis model for the timing of underground tunnel support.

[0037] Before training, intelligent hyperparameter optimization is adopted, that is, the particle swarm optimization algorithm (PSO) is used to globally optimize the hyperparameters of the analysis model, and the mean squared error (MSE) of the validation set is used as the fitness function to search for the optimal combination of hyperparameters (number of neurons 16-64, sliding window size 5-20, dropout ratio 0.2-0.3, initial learning rate 0.0005-0.001).

[0038] During training, the training set is input into the analysis model, and the adaptive momentum optimization algorithm (Adam) is used to locally iteratively optimize the model parameters. The number of training iterations is set to 500-1000, the initial learning rate is 0.001, and the learning rate is adaptively adjusted according to the changes in the loss function during training. The mean squared error (MSE) is used as the loss function.

[0039] During training, the analysis model is validated in real time using a validation set, and the structure and parameters of the analysis model are optimized based on the validation results. After training, the analysis model is tested and cross-validated using a test set to ensure that the minimum length error of the output from the tunnel face is controlled within 5% and the mean absolute error (MAE) is ≤0.1m within the boundary range of 12 geological conditions, thus obtaining a qualified underground tunnel support timing analysis model.

[0040] At this point, the generated underground tunnel support timing analysis model adopts a bidirectional LSTM network structure combined with an attention mechanism. It is generated by training underground tunnel environment samples constructed from the temporal characteristics of tunnel surrounding rock displacement under typical working conditions, and is used to achieve accurate prediction of the optimal support distance for underground tunnels.

[0041] Step S120: Input the data of the conditions to be analyzed into the underground tunnel support timing analysis model to obtain the minimum distance corresponding to the optimal support timing under the surrounding rock geological conditions; At this point, the underground tunnel support timing analysis model can quickly output the minimum distance from the tunnel face corresponding to a rock displacement release rate of 35% to 40% under the corresponding geological conditions. This length is the minimum distance corresponding to the optimal support timing under the current working conditions, guiding the precise implementation of support construction on site.

[0042] In the case provided by this invention, the data of the conditions to be analyzed under Class II surrounding rock and low ground stress conditions are input into the underground tunnel support timing analysis model. Within 8 seconds, the minimum length from the tunnel face is obtained as 2.8m, which can guide the on-site support construction to be carried out when the tunnel face is advanced to 2.8m. When the data of the conditions to be analyzed under Class IV surrounding rock and high ground stress conditions are input into the underground tunnel support timing analysis model, the minimum length from the tunnel face is output as 1.5m within 10 seconds. After construction is carried out on-site according to these parameters, the displacement release rate of the surrounding rock is controlled between 25% and 35%, the stability is good, and no collapse risk occurs.

[0043] Step S121: Fine-tune and train the underground tunnel support timing analysis model based on the data of the conditions to be analyzed.

[0044] If the data to be analyzed exceeds the following conditions... Figure 2 The 12 combination ranges shown are supplemented with corresponding numerical models and data for fine-tuning the model. The supplemented data accounts for no less than 5% of the original dataset, and the number of fine-tuning iterations is between 200 and 500, in order to expand the adaptability of the underground tunnel support timing analysis model.

[0045] The underground tunnel support timing analysis model constructed in this invention is based on a dedicated dataset built from 12+ geological conditions. It is generated through intelligent optimization algorithms using a bidirectional LSTM-attention mechanism model, enabling accurate prediction of the minimum distance from the tunnel face with low error. It covers various surrounding rock types and multiple geostress conditions, and can be directly adapted to most underground tunnel construction scenarios. Furthermore, the underground tunnel support timing analysis model supports fine-tuning training with newly added data, eliminating the need for repeated model development for different projects, making it highly practical.

[0046] The underground tunnel support timing analysis method provided by this invention can accurately determine the optimal support timing, avoid construction interference and cost waste caused by supporting too early, and avoid geological disasters such as surrounding rock collapse and rock burst caused by supporting too late, thereby reducing construction costs, reducing the incidence of safety accidents and rework costs, and helping to promote efficient on-site construction.

[0047] This invention deeply integrates deep learning, intelligent optimization algorithms, and underground tunnel construction technology to construct a proxy numerical calculation model for support timing analysis. It breaks through the limitations of traditional support timing analysis that relies on numerical simulation and passive monitoring, and provides a brand-new technical path for underground tunnel support timing analysis. This promotes the development of tunnel construction towards intelligence and precision, and has significant technological innovation and industry promotion value.

[0048] The above-disclosed embodiments are merely a few specific examples of the present invention. However, the present invention is not limited thereto, and any variations that can be conceived by those skilled in the art should fall within the protection scope of the present invention.

Claims

1. A method for analyzing the timing of underground tunnel support based on bidirectional LSTM, characterized in that, Includes the following steps: Obtain the geological conditions of the surrounding rock during actual underground tunnel construction as the data to be analyzed; Load the underground tunnel support timing analysis model; the underground tunnel support timing analysis model adopts a bidirectional LSTM network structure combined with attention mechanism, and is generated by training the underground tunnel environment sample constructed from the displacement time sequence characteristics of the tunnel surrounding rock under typical working conditions, so as to achieve accurate prediction of the optimal support distance of the underground tunnel. The typical working conditions refer to: surrounding rock grade supporting Class II to V, and in-situ stress conditions supporting high, medium, and low levels; Input the data of the conditions to be analyzed into the underground tunnel support timing analysis model to obtain the minimum distance corresponding to the optimal support timing under the surrounding rock geological conditions.

2. The method for analyzing the timing of underground tunnel support based on bidirectional LSTM according to claim 1, characterized in that, Before loading the underground tunnel support timing analysis model, the underground tunnel support timing analysis model is generated. The generation of the underground tunnel support timing analysis model includes the following steps: Construct multi-dimensional boundary conditions for typical working conditions to generate simulation scenarios; Based on the time-series features collected from the simulated scenario, sample pairs are constructed; The sample pairs are preprocessed to generate underground tunnel environment samples; the underground tunnel environment samples are then divided into training set, validation set and test set. Define the analysis model; the analysis model adopts an end-to-end network structure of bidirectional LSTM combined with attention mechanism; The analysis model was trained using environmental samples from underground tunnels to construct an analysis model for the timing of underground tunnel support.

3. The method for analyzing the timing of underground tunnel support based on bidirectional LSTM according to claim 2, characterized in that, When constructing multi-dimensional boundary conditions for typical working conditions, geological working conditions are matrixed to obtain orthogonal parameter matrices; the orthogonal parameter matrices reflect various surrounding rock grades, different geostress levels, and different cross-sectional shapes and sizes. The simulation scenario is generated using 3D simulation software based on an orthogonal parameter matrix, and is used to simulate the step-by-step excavation process of actual construction.

4. The method for analyzing the timing of underground tunnel support based on bidirectional LSTM according to claim 2, characterized in that, The constructed sample pairs include: Monitoring points were set up to collect time series data D(t) of the surrounding rock displacement as the working face advanced; The process of gradually excavating the tunnel face is simulated, and the displacement release amount at each monitoring point is collected in real time. The displacement release rate η at each monitoring point is calculated, and the spatial distance L between the tunnel face and the monitoring point corresponding to the displacement release rate η at a preset threshold is defined as the target label. Y ; Define a sliding window sequence formed by the time series data of the current step and the previous n historical steps as the input feature X, combined with the target label. Y Construct sample pairs ( X , Y ).

5. The method for analyzing the timing of underground tunnel support based on bidirectional LSTM according to claim 2, characterized in that, The generated underground tunnel environment samples include: Gaussian noise perturbation and moving average smoothing are introduced into the sample pairs to construct an enhanced sample set covering multiple operating conditions; Based on the enhanced sample set, displacement rate and displacement acceleration are extracted, and displacement release rate is calculated as a surrounding rock stability discrimination index to construct a multi-dimensional feature vector. The multidimensional feature vectors are scaled to generate a unified sample set.

6. The method for analyzing the timing of underground tunnel support based on bidirectional LSTM according to claim 1, characterized in that, The bidirectional LSTM network structure combined with the attention mechanism is set up with an input layer, a hidden layer, a regularization layer, a fully connected layer and an output layer in sequence along the data transmission direction. The features are transferred and transformed step by step between the layers through tensor dimension matching. By capturing the long-distance dependency of the surrounding rock displacement release time series data and screening the key stage features of the surrounding rock deformation, the network structure can accurately predict the optimal support distance of the underground tunnel. The trainable parameters of the overall network structure are optimized globally through backpropagation. The input layer receives time-series data with a structure of [Batch_size, Time_steps, Features], including displacement, velocity, surrounding rock grade coding, and in-situ stress value; the output layer outputs the optimal support distance predicted by a single node.

7. The method for analyzing the timing of underground tunnel support based on bidirectional LSTM according to claim 1, characterized in that, The hidden layer includes a bidirectional LSTM layer and an Attention layer; Among them, the bidirectional LSTM layer is used to simultaneously capture the impact of historical information on the current state and the inverse inference of future trends on the current state, generating a bidirectional fused feature tensor to achieve complete extraction of full temporal features. The Attention layer introduces an attention mechanism to automatically calculate the weight of time nodes; the time nodes include the moment of excavation and the stress redistribution stage.

8. The method for analyzing the timing of underground tunnel support based on bidirectional LSTM according to claim 7, characterized in that, After the time-series data enters the bidirectional LSTM layer, it is divided into two parallel data streams: a forward data stream and a reverse data stream. The forward data stream is transmitted in ascending time sequence, according to the excavation step sequence t1→t2→…→t. n Forward propagation extracts historical deformation features; the reverse data stream is a time-reverse transmission data stream, transmitted according to the excavation step sequence t. n →t n-1 →…→t1 backpropagation, extracting future deformation features; The two data streams output feature vectors at each time step, which are then concatenated and fused into a bidirectional fused feature tensor.

9. The method for analyzing the timing of underground tunnel support based on bidirectional LSTM according to claim 2, characterized in that, Before training the analysis model, the hyperparameters of the analysis model are globally optimized using the particle swarm optimization algorithm. During training, the training set is input into the analysis model, and the adaptive momentum optimization algorithm is used to locally iteratively optimize the model parameters. The number of training iterations is set to 500-1000, and the initial learning rate is set to... 0.001, the learning rate is adaptively adjusted according to the changes in the loss function during training, with mean squared error as the loss function; During training, the analysis model is validated in real time using a validation set, and the structure and parameters of the analysis model are optimized based on the validation results. After training, the analysis model is tested and cross-validated using a test set to ensure that the analysis model is within the boundary range of typical working conditions.

10. The method for analyzing the timing of underground tunnel support based on bidirectional LSTM according to claim 2, characterized in that, If the data under the conditions to be analyzed exceeds the range of typical working conditions, supplement the numerical model and data for the corresponding working conditions to fine-tune and train the underground tunnel support timing analysis model.

Images