Method for determining construction timing of loose soft rock tunnel lining

Abstract

The application provides a method for determining the construction timing of loose soft rock tunnel lining, and belongs to the technical field of loose soft rock tunnel lining construction. The method comprises the following steps: arranging a surrounding rock deformation monitoring section along the axial direction of the tunnel, continuously collecting displacement stress and water seepage data to form a time series data set, inputting the time series data set into a stress release prediction model to output surrounding rock deformation prediction values and stress release stability scores, performing spatial variability risk assessment to calculate the dispersion coefficient of surrounding rock parameters when the stability score is greater than a threshold value and the deformation rate meets the conditions, generating a lining construction permission signal when the dispersion coefficient is less than a threshold value, maintaining real-time monitoring during the pouring process and stopping pouring and re-evaluating when the deformation rate suddenly increases, thereby solving the technical problem of difficulty in accurately determining the construction timing of loose soft rock tunnel lining.

Description

Technical Field

[0001] This invention belongs to the field of construction technology for lining of loose soft rock tunnels, and more specifically, it relates to a method for determining the timing of lining construction for loose soft rock tunnels. Background Technology

[0002] In the construction of tunnels in loose soft rock, the stress release process of the surrounding rock directly affects the stability of the lining structure. Traditional methods for determining the timing of construction mainly rely on the empirical judgment of on-site engineers based on surrounding rock deformation monitoring data. This involves setting a single deformation rate threshold as the initiation condition for lining construction or adopting a static waiting strategy with fixed time periods. In current tunnel construction practice, due to the complex and variable geological conditions of loose soft rock, the stress release process exhibits significant spatiotemporal non-uniformity, making it difficult to accurately grasp the degree of completion of stress release by relying solely on manual experience. Existing methods neglect the prediction of long-term trends in surrounding rock deformation, lack a quantitative assessment mechanism for the stress release process, and cannot identify the risk of stress concentration caused by spatial variations in local rock mass parameters. In other words, existing technologies suffer from the technical problem of accurately determining the timing of lining construction in loose soft rock tunnels. Summary of the Invention

[0003] In view of this, the present invention provides a method for determining the timing of lining construction in loose soft rock tunnels, which can solve the technical problem in the prior art that it is difficult to accurately determine the timing of lining construction in loose soft rock tunnels.

[0004] This invention is implemented as follows: It provides a method for determining the timing of lining construction in loose soft rock tunnels. After tunnel excavation, surrounding rock deformation monitoring sections are set up along the tunnel axis, and convergence deformation monitoring points, surrounding rock stress monitoring sensors, and seepage monitoring devices are installed. Displacement data, stress data, and flow data are continuously collected to form a time-series dataset of surrounding rock deformation. This dataset is input into a stress release prediction model, which outputs predicted future surrounding rock deformation values ​​and stress release stability scores. The stress release prediction model consists of an encoding layer, an attention layer, and a decoding layer. The encoding layer uses long short-term memory units and residual connections to extract temporal features. The attention layer adjusts the number of attention heads according to the burial depth of the monitoring section. The decoding layer outputs predicted surrounding rock deformation values ​​and stress release stability scores. When the stress release stability score and the surrounding rock deformation rate meet threshold conditions, a spatial variability risk assessment is performed to calculate the dispersion coefficient of surrounding rock parameters. When the dispersion coefficient of surrounding rock parameters is less than the threshold, a lining construction permit signal is generated. Real-time monitoring is maintained during the pouring process, and pouring is stopped and reassessed when the surrounding rock deformation rate suddenly increases.

[0005] Among them, the surrounding rock deformation monitoring sections are set up at 5m intervals along the tunnel axis, and convergence deformation monitoring points are installed at the arch crown, arch waist and sidewall positions of each surrounding rock deformation monitoring section.

[0006] Among them, in shallow buried sections with a burial depth of less than 50m, surrounding rock stress monitoring sensors and seepage monitoring devices are densely deployed.

[0007] The data collection frequency is once every 2 hours, and the continuous collection period is no less than 30 days.

[0008] The surrounding rock deformation time series dataset includes displacement data sequences, stress data sequences, and seepage data sequences, which are arranged in chronological order to form a multidimensional time series matrix.

[0009] The input layer of the stress release prediction model receives a 72-dimensional feature vector, which contains the displacement increment stress change rate and seepage fluctuation value for 36 consecutive time steps.

[0010] The coding layer consists of four layers of long short-term memory units, each containing 128 neurons, with gradients transmitted between layers via residual connections.

[0011] The attention layer employs a multi-head attention mechanism, with 8 attention heads when the burial depth of the monitoring section is less than 50m, and 4 attention heads when the burial depth of the monitoring section is greater than or equal to 50m.

[0012] The decoding layer consists of three fully connected layers with 256, 128, and 2 neurons respectively, and outputs 15 time-step predicted values ​​of surrounding rock deformation and 1 stress release stability score.

[0013] The stress release prediction model training adopts a gradient optimization mechanism based on multi-level adaptive weight decay. The variance of the parameter gradient of each layer is calculated. The layer with a variance greater than 1.5 times the average global gradient variance is defined as a high gradient layer and a weight decay coefficient of 0.01 is applied. The layer with a variance less than 0.5 times the average global gradient variance is defined as a low gradient layer and a weight decay coefficient of 0.001 is applied. The intermediate gradient layer is given a weight decay coefficient of 0.005.

[0014] Specifically, the second derivative of the current batch loss function with respect to the parameters is calculated as an approximation of the local curvature information, and the product of the local curvature information and the current parameter value is added to the loss function as a regularization term.

[0015] The sliding window length is set to 10 batches. The standard deviation of the parameter update amplitude within the window is calculated. When the standard deviation is greater than 0.05, the learning rate is decayed to 0.9 times the current value. When the standard deviation is less than 0.01, the learning rate is increased to 1.1 times the current value, but not exceeding the initial learning rate.

[0016] The stress release stability score is based on a weighted sum of three dimensions: the predicted deformation rate stress change trend and the stability of seepage volume, with weighting coefficients of 0.5, 0.3 and 0.2, respectively.

[0017] Among them, spatial variability risk assessment is carried out when the stress release stability score is greater than 85 and the surrounding rock deformation rate is less than 0.1 mm / d for 7 consecutive days.

[0018] The calculation of the dispersion coefficient of the surrounding rock parameters involves extracting the cumulative deformation of all convergent deformation monitoring points within a 10m range before and after the fracture surface of the surrounding rock deformation monitoring section, and calculating the ratio of the standard deviation to the average value of the cumulative deformation.

[0019] Among them, when the deformation rate of the surrounding rock at any convergence deformation monitoring point suddenly increases to more than 0.5 mm / d, the pouring should be stopped immediately and the stress release stability score should be re-evaluated. The sudden increase is determined based on the fact that the daily deformation rate of the surrounding rock increases by more than 5 times compared with the average deformation rate of the surrounding rock in the previous 7 days.

[0020] This invention addresses the shortcomings of traditional methods that lack quantitative evaluation mechanisms by constructing a time-series dataset of surrounding rock deformation and inputting it into a stress release prediction model for future deformation prediction and stability scoring. The stress release prediction model extracts long-term trends and short-term fluctuations in surrounding rock deformation based on long short-term memory units and multi-head attention mechanisms, outputting a stress release stability score that comprehensively reflects the degree of stress release completion and the controllability of subsequent deformation. Spatial variability risk assessment identifies local weak areas by calculating the dispersion coefficient of surrounding rock parameters, avoiding stress imbalance in the lining structure caused by uneven stress release. Real-time monitoring and surge warning mechanisms during the pouring process ensure that the lining structure solidifies in a stable stress field. In summary, this invention solves the technical problem mentioned in the background art of the difficulty in accurately determining the construction timing of lining in loose soft rock tunnels. Attached Figure Description

[0021] Figure 1 This is a flowchart of the method of the present invention.

[0022] Figure 2 This is a schematic diagram of the layout of the tunnel surrounding rock deformation monitoring sections.

[0023] Figure 3 This is a schematic diagram of the stress relief prediction model.

[0024] Figure 4 This is a flowchart for calculating stress release stability score.

[0025] Figure 5 This is a distribution map showing the assessment results of the timing of lining construction at the monitoring sections along the entire line. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0027] like Figure 1The diagram shows a flowchart of a method for determining the construction timing of tunnel lining in loose soft rock, provided by the present invention. This method includes the following steps:

[0028] S01. After the tunnel excavation is completed, a rock deformation monitoring section is set up every 5m along the tunnel axis. Convergence deformation monitoring points are installed at the arch crown, arch waist and sidewall positions of each rock deformation monitoring section. At the same time, rock stress monitoring sensors and seepage monitoring devices are densely deployed in shallow buried sections with a burial depth of less than 50m.

[0029] S02. Continuously collect displacement data of the convergence deformation monitoring point, stress data of the surrounding rock stress monitoring sensor and flow data of the seepage monitoring device. The collection frequency is once every 2 hours, and the continuous collection time is not less than 30 days to form a surrounding rock deformation time series dataset.

[0030] S03. Input the surrounding rock deformation time series dataset into the stress release prediction model, and output the predicted value of surrounding rock deformation and stress release stability score of each surrounding rock deformation monitoring section in the next 15 days. The stress release stability score ranges from 0 to 100.

[0031] S04. When the stress release stability score is greater than 85 and the surrounding rock deformation rate is less than 0.1 mm / d for 7 consecutive days, a spatial variability risk assessment is performed on the corresponding surrounding rock deformation monitoring section, and the dispersion coefficient of the surrounding rock parameters within 10 m in front of and behind the surrounding rock deformation monitoring section is calculated.

[0032] S05. If the dispersion coefficient of the surrounding rock parameters is less than 0.3, it is determined that the corresponding section of the surrounding rock deformation monitoring section meets the conditions for lining construction, a lining construction permit signal is generated and the corresponding station number is recorded as the optimal construction time station number; if the dispersion coefficient of the surrounding rock parameters is greater than or equal to 0.3, the monitoring period is extended by 15 days and steps S02 to S04 are repeated.

[0033] S06. For sections that meet the conditions for lining construction, lining is poured in the order of the optimal construction time and chainage. During the pouring process, the convergence deformation monitoring points are kept working continuously. When the surrounding rock deformation rate at any convergence deformation monitoring point suddenly increases to more than 0.5 mm / d, pouring is stopped immediately and the stress release stability score is re-evaluated.

[0034] The surrounding rock deformation monitoring section is a cross-section perpendicular to the tunnel axis. The deformation characteristics of the surrounding rock are reflected by measuring changes in relative position on this cross-section. The convergence deformation monitoring points use a total station in conjunction with a prism or laser displacement sensor to achieve millimeter-level accuracy measurements. The surrounding rock stress monitoring sensor is a vibrating wire stress gauge or a fiber optic grating stress sensor, embedded at the contact surface between the initial support and the surrounding rock. The seepage monitoring device includes a diversion channel and a flow meter. The diversion channel is arranged along the arch to collect seepage water, and the flow meter records the amount of water passing through per unit time.

[0035] The surrounding rock deformation time-series dataset comprises three subsets: displacement data sequence, stress data sequence, and seepage data sequence. Each subset is arranged in chronological order to form a multidimensional time-series matrix. The displacement data sequence records the changes in the three-dimensional coordinates of each convergent deformation monitoring point. The stress data sequence records the radial and tangential stresses measured by the surrounding rock stress monitoring sensor. The seepage data sequence records the instantaneous and cumulative flow rates of the seepage monitoring device.

[0036] The stress release prediction model has the following structure: the input layer receives a 72-dimensional feature vector, which contains displacement increments, stress change rates, and seepage fluctuations over 36 consecutive time steps; the encoding layer consists of four long short-term memory units, each containing 128 neurons, with gradients transmitted between layers via residual connections; the attention layer employs a multi-head attention mechanism, with the number of attention heads determined by the ratio of the current surrounding rock deformation monitoring section burial depth to 50m; when the burial depth of the surrounding rock deformation monitoring section is less than 50m, the number of attention heads is 8, and when the burial depth is greater than or equal to 50m, the number of attention heads is 4; the decoding layer contains three fully connected layers with 256, 128, and 2 neurons respectively, ultimately outputting the predicted surrounding rock deformation values ​​for 15 time steps and one stress release stability score.

[0037] The displacement increment is the difference in displacement data of the convergent deformation monitoring point between adjacent time steps. The stress change rate is the ratio of the difference in stress data of the surrounding rock stress monitoring sensor between adjacent time steps to the time interval. The seepage fluctuation value is the difference between the flow rate data of the seepage monitoring device at the current time step and the average flow rate data of the previous 10 time steps.

[0038] The steps for establishing the training dataset for the stress release prediction model specifically include: collecting surrounding rock monitoring data from no fewer than 20 completed loose soft rock tunnels, covering the entire process from excavation completion to lining construction; preprocessing the surrounding rock monitoring data to remove outliers during sensor malfunction periods and filling in missing data using linear interpolation; using the surrounding rock monitoring data from 36 consecutive time steps as input samples and the measured deformation data from the subsequent 15 time steps as output labels, while labeling the stress release stability score according to whether lining cracking occurs subsequently: the stress release stability score for uncracked sections is 90 to 100, for slightly cracked sections it is 70 to 89, and for severely cracked sections it is 0 to 69; and dividing the input samples and output labels into training and validation sets in an 8:2 ratio.

[0039] The specific steps for training the stress release prediction model include: initializing model parameters, with the weights of each layer in the encoding layer using an orthogonal initialization method, and the weights of the attention layer and the decoding layer using a Xavier initialization method; training the model using a gradient optimization mechanism based on multi-level adaptive weight decay, wherein the gradient optimization mechanism first calculates the variance of the gradient of each layer's parameters, defining layers with variances greater than 1.5 times the average global gradient variance as high gradient layers, and layers with variances less than 0.5 times the average global gradient variance as low gradient layers; applying a weight decay coefficient of 0.01 to the high gradient layers, a weight decay coefficient of 0.001 to the low gradient layers, and a weight decay coefficient of 0.005 to the intermediate gradient layers; and in each training batch, calculating the approximate second derivative of the loss function of the current batch with respect to the parameters. The value is used as local curvature information, and the product of the local curvature information and the current parameter value is added to the loss function as a regularization term. The sliding window length is set to 10 batches, and the standard deviation of the parameter update amplitude within the window is calculated. When the standard deviation is greater than 0.05, the learning rate is decayed to 0.9 times the current value, and when the standard deviation is less than 0.01, the learning rate is increased to 1.1 times the current value but not exceeding the initial learning rate. During the training process, the mean squared error loss function is used to measure the deformation prediction error, and the cross-entropy loss function is used to measure the stability scoring error. The total loss function is the weighted sum of the two, with weight coefficients of 0.6 and 0.4, respectively. The number of training rounds is set to 200 rounds. Each round iterates through all the training set samples, and the model performance is evaluated on the validation set every 10 rounds. The model parameters with the minimum loss on the validation set are saved as the final model.

[0040] The intermediate gradient layer is a layer whose variance is between 0.5 and 1.5 times the average global gradient variance. The average global gradient variance is the arithmetic mean of the gradient variances of all network layer parameters. The approximate value of the second derivative is calculated using the finite difference method, obtained by applying a small perturbation to the parameters and calculating the ratio of the change in the loss function to the square of the perturbation.

[0041] The gradient optimization mechanism based on multi-level adaptive weight decay achieves refined parameter update control by distinguishing the gradient characteristics of different network layers. In the stress release prediction task, the encoding layer is responsible for extracting temporal features, the attention layer is responsible for identifying key time periods, and the decoding layer is responsible for generating prediction results. The parameter sensitivities of these three layers differ significantly. Traditional uniform weight decay strategies apply the same regularization strength to all layers, leading to over-penalization and underfitting in high-gradient layers, and under-penalization and overfitting in low-gradient layers. The gradient optimization mechanism dynamically adjusts the decay coefficient based on the gradient variance of each layer, giving high-gradient layers stronger constraints to suppress drastic parameter fluctuations, lower-gradient layers weaker constraints to retain subtle feature representation capabilities, and intermediate gradient layers maintaining moderate constraints to balance generalization and fitting. The local curvature information reflects the curvature of the loss function in the parameter space; regions with high curvature correspond to directions sensitive to parameter changes. Constructing the local curvature information into the regularization term guides the optimization process to avoid steep regions, making the parameter update path tend towards a flatter direction and reducing the risk of getting trapped in local extrema. The sliding window statistics capture the short-term fluctuations in parameter updates. When the standard deviation of the update amplitude is too large, it indicates severe oscillations in the optimization process, requiring a reduction in the learning rate to enhance stability. Conversely, when the standard deviation of the update amplitude is too small, it indicates that the optimization process has plateaued, requiring an increase in the learning rate to accelerate convergence. The gradient optimization mechanism, through the synergistic effects of hierarchical differential decay, curvature-guided regularization, and adaptive window adjustment, enables the stress release prediction model to fully learn the long-term trends and short-term fluctuations of surrounding rock deformation during training, while avoiding the decline in generalization ability caused by overfitting. Thus, when facing loose soft rock tunnels with different geological conditions and construction disturbances, it can accurately predict the stress release process and reliably assess the timing of lining construction, providing a quantitative scientific basis for engineering decisions.

[0042] The stress release stability score comprehensively reflects the degree of completion of stress release in the surrounding rock and the controllability of subsequent deformation. The stress release stability score is calculated based on a weighted sum of three dimensions: the predicted deformation rate corresponding to the predicted deformation value of the surrounding rock, the stress change trend corresponding to the stress data sequence, and the seepage stability corresponding to the seepage data sequence. The weighting coefficients are 0.5, 0.3, and 0.2, respectively. A higher stress release stability score indicates more complete stress release in the surrounding rock, and a lower risk of large deformation and lining cracking in the future.

[0043] The predicted deformation rate is the ratio of the difference in predicted displacement between adjacent time steps to the time interval in the predicted deformation values ​​of the surrounding rock. The stress change trend is determined by linearly fitting the stress data sequence over the most recent 30 time steps; a smaller absolute value of the slope of the fitted line indicates that the stress tends to be stable. The stability of seepage volume is characterized by calculating the coefficient of variation of the seepage volume data sequence over the most recent 30 time steps; a smaller coefficient of variation indicates smaller fluctuations in seepage volume.

[0044] The spatial variability risk assessment identifies locally weak areas by quantifying the dispersion of rock mass parameters around the surrounding rock deformation monitoring section. The method for calculating the dispersion coefficient of the surrounding rock parameters is as follows: extract the cumulative deformation of all convergent deformation monitoring points within a 10m radius before and after the surrounding rock deformation monitoring section, and calculate the ratio of the standard deviation to the average value of the cumulative deformation as the dispersion coefficient of the surrounding rock parameters. When the dispersion coefficient of the surrounding rock parameters is less than 0.3, it indicates that the surrounding rock properties are relatively uniform, the stress release process has good spatial consistency, and the stress distribution of the lining structure is balanced. When the dispersion coefficient of the surrounding rock parameters is greater than or equal to 0.3, it indicates the existence of locally abnormal deformation areas, uneven stress release, and the monitoring period needs to be extended until the deformation tends to be uniform before lining construction can proceed.

[0045] The cumulative deformation is the total displacement measured at the convergence deformation monitoring points from the time of excavation completion to the current evaluation time. The surrounding rock deformation rate is the average daily displacement increment at the convergence deformation monitoring points over seven consecutive days.

[0046] The lining construction permit signal is a digital instruction generated by the system and transmitted to the construction management platform to trigger the lining construction preparation process, including formwork installation, rebar tying, and concrete pouring plan preparation. Once generated, the lining construction permit signal is only valid for the corresponding station number section; other sections must independently meet the judgment conditions of steps S04 and S05 to obtain the lining construction permit signal.

[0047] The optimal construction timing station number is the tunnel station number of the surrounding rock deformation monitoring section when the lining construction timing conditions are met. The optimal construction timing station number marks the starting point of lining construction when the surrounding rock stress is fully released and the spatial uniformity is good.

[0048] The sudden increase determination is based on the increase of the daily deformation rate of the surrounding rock relative to the average deformation rate of the surrounding rock over the previous 7 days. When the increase exceeds 5 times and the absolute value reaches 0.5 mm / d, an early warning is triggered, indicating that the stress release of the surrounding rock has accelerated abnormally and the lining construction needs to be suspended for reassessment to avoid the lining structure from solidifying in an unstable stress field, which could lead to cracking and damage.

[0049] The specific implementation methods of the above steps are described in detail below.

[0050] The specific implementation of step S01 involves first determining the time node for tunnel excavation completion, using the advancement of the excavation face to the designed cross-section and the completion of initial support as the criterion for excavation completion. Then, along the tunnel axis from the tunnel entrance to the working face, a surrounding rock deformation monitoring section is set at 5m intervals. This interval is adjusted according to the surrounding rock grade, with closer intervals for lower grades. On each surrounding rock deformation monitoring section, five locations—the arch crown, left arch waist, right arch waist, left side wall, and right side wall—are selected as installation positions for convergence deformation monitoring points. These positions are selected according to a symmetrical distribution principle to comprehensively reflect the cross-sectional deformation characteristics. Expansion bolts are used to fix measuring prisms or laser displacement sensors to the initial support surface, ensuring that the sensors deform synchronously with the surrounding rock. For shallow buried sections with a depth of less than 50m, an intermediate section is added to the existing monitoring sections, shortening the monitoring section spacing to 2.5m. Simultaneously, surrounding rock stress monitoring sensors and seepage monitoring devices are additionally installed on the surrounding rock deformation monitoring sections of these shallow buried sections. The surrounding rock stress monitoring sensor is installed via boreholes, with a drilling depth of 1.5m to 2.0m. After the sensor is buried at the bottom of the hole, it is sealed with cement mortar. The guide channel of the seepage monitoring device is laid longitudinally along the arch, covering the area between two adjacent surrounding rock deformation monitoring sections. A flow meter is connected to the end of the guide channel to achieve continuous measurement. The purpose of these steps is to establish a monitoring network covering the entire tunnel, with a focus on densifying the shallow buried sections, providing a data foundation for subsequent spatiotemporal evolution analysis of the stress release process.

[0051] The specific implementation of step S02 involves setting up a data acquisition system that connects all convergence deformation monitoring points, surrounding rock stress monitoring sensors, and seepage monitoring devices via wired or wireless means. This data acquisition system automatically triggers a measurement command every 2 hours and records the returned data. For convergence deformation monitoring points, three-dimensional coordinates are obtained using a total station in automatic aiming mode or a laser displacement sensor in continuous ranging mode, with a coordinate accuracy requirement of millimeters. For surrounding rock stress monitoring sensors, the frequency of the vibrating wire or the wavelength change of the fiber optic grating is read and converted into stress values ​​through a calibration curve. These stress values ​​include both radial and tangential stress components. For seepage monitoring devices, a flow meter records the flow rate in real time and accumulates it to form instantaneous and cumulative flow data. The data acquisition duration is set to be no less than 30 days, a reference value determined based on engineering experience that the stress release cycle of loose soft rock is typically 20 to 40 days. The collected displacement data is organized into a displacement data sequence according to the monitoring point number and timestamp; the stress data is organized into a stress data sequence according to the sensor number and timestamp; and the flow rate data is organized into a seepage data sequence according to the monitoring device number and timestamp. These three sequences together constitute a time-series dataset of surrounding rock deformation. This step captures the dynamic process of surrounding rock deformation through high-frequency continuous monitoring, providing sufficient time-series samples for stress release prediction models.

[0052] The specific implementation of step S03 involves preprocessing the surrounding rock deformation time-series dataset. First, the displacement increment is calculated by extracting the coordinate difference between adjacent time steps in the displacement data sequence; this difference reflects the deformation per unit time. Then, the stress change rate is calculated by extracting the stress difference between adjacent time steps in the stress data sequence and dividing it by the time interval of 2 hours; this stress change rate characterizes the rate of stress release. Next, the seepage fluctuation value is calculated by extracting the flow rate data of the current time step and subtracting the arithmetic mean of the flow rate data of the previous 10 time steps; this fluctuation value reflects the stability of the seepage state. The displacement increment, stress change rate, and seepage fluctuation value corresponding to 36 consecutive time steps are arranged sequentially to form a 72-dimensional feature vector. This feature vector covers 3 days of monitoring history to capture short-term trends. The feature vector is input into the input layer of the stress release prediction model. The data passes through the Long Short-Term Memory (LSTM) unit of the encoding layer to extract temporal dependencies. The LSM unit selectively retains and updates historical information through forget gates, input gates, and output gates, overcoming the gradient vanishing problem in long sequences of traditional recurrent neural networks. The features output from the encoding layer are weighted by the attention layer. The attention layer determines the number of attention heads based on the burial depth of the current surrounding rock deformation monitoring section. Eight attention heads are used when the burial depth is less than 50m to enhance the ability to focus on multi-source information in shallow sections, while four attention heads are used when the burial depth is greater than or equal to 50m to maintain appropriate computational complexity. The output of the attention layer is passed to the decoding layer, which progressively reduces dimensionality and generates prediction results through three fully connected layers, ultimately outputting 15 time-step predicted values ​​of surrounding rock deformation and one stress release stability score. The predicted surrounding rock deformation values ​​provide the displacement development trend of each monitoring point over the next 7.5 days. The stress release stability score comprehensively evaluates the degree of completion of stress release in the surrounding rock, with a score range of 0 to 100; a higher score indicates more complete stress release. This process utilizes the nonlinear mapping and temporal memory capabilities of the deep learning model to achieve accurate prediction of the stress release process in loose soft rock.

[0053] The specific implementation of step S04 is as follows: First, determine whether the stress release stability score is greater than 85. The threshold of 85 is determined based on historical engineering data statistics. When the score exceeds the threshold, the probability of lining cracking drops below 5%. Then, extract the displacement data of the surrounding rock deformation monitoring section for the most recent 7 days, calculate the daily displacement increment, and take the average value to obtain the surrounding rock deformation rate. Determine whether the surrounding rock deformation rate is less than 0.1 mm / d. The threshold of 0.1 mm / d is the critical rate at which the deformation of loose soft rock tends to stabilize. When both the stress release stability score and the surrounding rock deformation rate meet the conditions, the spatial variability risk assessment procedure is initiated. The spatial variability risk assessment first determines the assessment range, taking the current surrounding rock deformation monitoring section as the center and extending 10m forward and backward along the tunnel axis to form a 20m long assessment section. Within the assessment section, extract the cumulative deformation of all convergent deformation monitoring points from the completion of excavation to the current moment. The cumulative deformation is the sum of the displacement data sequences of each monitoring point. The standard deviation and arithmetic mean of the cumulative deformation are calculated. The standard deviation is divided by the arithmetic mean to obtain the dispersion coefficient of the surrounding rock parameters. The dispersion coefficient quantifies the spatial difference in the deformation of the surrounding rock within the evaluation section. This step, through stability determination in the time dimension and uniformity assessment in the spatial dimension, dually screens sections with sufficient stress release and relatively consistent geological conditions, reducing the risk of cracking due to local stress concentration after lining construction.

[0054] The specific implementation of step S05 involves comparing the dispersion coefficient of the surrounding rock parameters with a threshold of 0.3. This threshold is determined based on the spatial variability characteristics of loose soft rock. When the dispersion coefficient is less than the threshold, it indicates that the mechanical properties of the surrounding rock within the assessment section are relatively uniform. If the dispersion coefficient is less than 0.3, it is determined that the corresponding section of the surrounding rock deformation monitoring section meets the conditions for lining construction. A lining construction permit signal is generated, and the station number of the current surrounding rock deformation monitoring section is recorded as the optimal construction time station number. This optimal construction time station number identifies the optimal starting position for lining construction. If the dispersion coefficient of the surrounding rock parameters is greater than or equal to 0.3, it is determined that there are local weak zones or uneven stress release in the assessment section. In this case, no lining construction permit signal is generated; instead, the monitoring period is extended by 15 days. This extension period is a typical adjustment cycle for stress release in loose soft rock. After extending the monitoring period, the process returns to step S02 to re-collect data and execute subsequent prediction and assessment procedures until the dispersion coefficient of the surrounding rock parameters meets the conditions. The steps described above enable quantitative determination of the timing of lining construction by setting a spatial uniformity threshold, thereby avoiding the uncertainty of subjective experience judgment and ensuring that the lining structure is formed in a relatively uniform stress field.

[0055] The specific implementation of step S06 involves compiling a lining construction plan based on the optimal construction timing station numbers generated from each surrounding rock deformation monitoring section, following an ascending order of station numbers. Priority is given to installing formwork, tying reinforcing bars, and pouring concrete in sections that meet the requirements. During the lining pouring process, the data acquisition system is continuously running, with the convergence deformation monitoring points updating displacement data every 2 hours. The daily surrounding rock deformation rate is calculated in real time by extracting the displacement increment within the most recent 24 hours and dividing it by a 1-day time interval. The daily surrounding rock deformation rate is compared with the arithmetic mean of the surrounding rock deformation rates of the previous 7 days, and the increase ratio is calculated as the daily surrounding rock deformation rate divided by the average of the previous 7 days' surrounding rock deformation rates. When the increase ratio exceeds 5 times and the absolute value of the daily surrounding rock deformation rate reaches 0.5 mm / d, a sudden increase in the surrounding rock deformation rate is identified. This sudden increase indicates an abnormal acceleration in the release of surrounding rock stress or the activation of a new deformation mechanism. Upon triggering a sudden increase warning, the lining pouring operation is immediately stopped, and on-site management personnel are notified to initiate the emergency response procedure. The stress release prediction model is re-initiated to analyze the current surrounding rock deformation time series dataset, outputting an updated stress release stability score. Based on this score, a decision is made on whether to continue waiting for stress release stabilization or to adjust the support parameters. This process, through real-time monitoring and dynamic feedback during construction, promptly detects and addresses abnormal surrounding rock deformation, preventing the lining structure from solidifying under unfavorable stress conditions and causing cracking and damage, thus ensuring the quality of lining construction and the safety of the tunnel structure.

[0056] It should be noted that the key technical ideas of this invention include stress release timing prediction based on long short-term memory units and multi-head attention mechanisms, adaptive weight decay optimization based on gradient variance stratification, and a construction timing determination mechanism that combines temporal stability and spatial uniformity. The stress release timing prediction technology selectively memorizes historical deformation information through the gating mechanism of long short-term memory units, solving the problem that the long-period nonlinear characteristics of stress release processes in loose soft rock are difficult to describe using traditional empirical formulas. The multi-head attention mechanism adaptively allocates computational resources according to burial depth, enabling the model to effectively identify both the multi-source coupling effects in shallow burial sections and the single stress-dominant mode in deep burial sections. Compared with traditional single-factor monitoring and discrimination methods, this prediction technology can integrate multi-dimensional information on displacement, stress, and seepage, and provide deformation trends and stability evaluations 7.5 days in advance, allowing sufficient preparation time for construction decisions. The adaptive weight decay optimization technique dynamically adjusts the regularization strength by calculating the gradient variance of each layer in real time. This synergistically optimizes the feature extraction capability of the encoding layer, the key information recognition capability of the attention layer, and the prediction generation capability of the decoding layer. It overcomes the underfitting or overfitting problems at some levels caused by traditional uniform regularization. Simultaneously, it introduces local curvature information to guide the parameter update path, avoiding steep regions of the loss function. Combined with sliding window statistics to smooth training oscillations, the model maintains stable generalization ability when facing loose soft rock with different geological conditions. Compared to traditional fixed learning rate and uniform weight decay strategies, this optimization technique significantly improves the model's adaptability to geological heterogeneity. The dual-judgment mechanism combines the stress release stability score and surrounding rock deformation rate threshold judgment in the time dimension with the spatial dimension evaluation of the surrounding rock parameter dispersion coefficient. This ensures that the stress release of a single monitoring section meets the standard while also guaranteeing the uniformity of geological conditions in the surrounding area. It avoids the problem of lining cracking caused by localized stress concentration due to spatial differences, which is a problem in traditional methods that only focus on single-point deformation rates. Compared to traditional experience-based judgment methods, this dual-judgment mechanism achieves scientific decision-making on construction timing through quantitative indicators and clear thresholds.

[0057] The synergistic effect of these three key technological approaches lies in constructing a complete closed-loop system from data acquisition to predictive analysis and decision execution. Stress release timing prediction technology transforms discrete multi-source data from on-site monitoring into continuous and predictable deformation trends and stability scores, providing quantitative input for subsequent judgments. Adaptive weight decay optimization technology ensures that the prediction model fully learns the essential laws of stress release in loose soft rock during the training phase without being excessively disturbed by data noise and geological heterogeneity, resulting in highly reliable prediction results. The dual judgment mechanism combines prediction results with spatial analysis, comprehensively evaluating lining construction conditions from both spatiotemporal dimensions, ensuring the scientific and safe nature of decision-making. Compared to the dilemma of traditional methods relying on experience-based judgment, which leads to lining cracking due to premature construction or delays affecting the construction period, this collaborative system, through the organic integration of intelligent prediction, adaptive optimization, and quantitative judgment, achieves precise control of lining construction timing. This ensures that the lining structure forms in a stable stress field, avoiding cracking and damage, and shortens the waiting period and improves construction efficiency through advance prediction and dynamic monitoring, providing a systematic technical solution for loose soft rock tunnel lining construction.

[0058] It should be noted that this invention also solves the following technical problem: the difficulty in coordinating parameter optimization at different network levels during the training of a surrounding rock stress release prediction model. In the task of predicting stress release in loose soft rock tunnels, the parameter sensitivity of the encoding layer extracting temporal features, the attention layer identifying key time periods, and the decoding layer generating prediction results varies significantly. The traditional uniform weight decay strategy applies the same regularization intensity to all levels, leading to over-penalization and underfitting in high-gradient layers, and under-penalization and overfitting in low-gradient layers. This invention calculates the gradient variance of each layer's parameters and dynamically adjusts the weight decay coefficient based on its ratio to the global gradient variance average. Stronger constraints are applied to high-gradient layers to suppress drastic parameter fluctuations, while weaker constraints are applied to low-gradient layers to retain subtle feature expression capabilities. Simultaneously, a regularization term based on the approximation of the second derivative is introduced to guide the optimization process away from steep regions. Combined with the sliding window statistical parameter update amplitude standard deviation, the learning rate is adaptively adjusted, enabling the model to fully learn the long-term trend and short-term fluctuation characteristics of surrounding rock deformation during training while avoiding the decline in generalization ability caused by overfitting.

[0059] Specifically, the principle of this invention is as follows: The stress release prediction model learns the long-term dependencies of surrounding rock deformation time-series data through the long short-term memory units of the encoding layer, capturing the gradual evolution of the stress release process. The attention layer dynamically adjusts the number of attention heads according to the burial depth of the monitoring section, assigning different feature extraction weights to the stress release differences between shallow and deep burial sections. Based on a multi-level adaptive weight decay gradient optimization mechanism, the decay coefficient is dynamically adjusted according to the gradient variance of each layer, so that the high gradient layer obtains stronger constraints to suppress drastic parameter fluctuations, the low gradient layer retains weak feature expression capabilities, and local curvature information guides the optimization process to avoid steep regions and reduce the risk of falling into local extrema. Spatial variability risk assessment quantifies the dispersion of surrounding rock parameters; when the dispersion coefficient is less than a threshold, it indicates good spatial consistency of surrounding rock stress release. By combining deformation prediction, stability scoring, and spatial uniformity judgment, the timing of lining construction is ensured to take into account both the sufficiency of surrounding rock stress release and spatial consistency.

[0060] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0061] The specific implementation methods of steps S01 and S02 are the same as those described above, and will not be repeated in detail here.

[0062] The specific implementation of step S03 involves inputting the surrounding rock deformation time series dataset into the stress release prediction model, with the model input layer receiving a 72-dimensional feature vector. The eigenvector formula is expressed as follows:

[0063] ;

[0064] In the formula, The feature vector contains 36 displacement increment components, 36 stress change rate components, and 36 seepage fluctuation value components, totaling 72 dimensions; For the first The displacement increment per time step, in mm; For the first The rate of change of stress at each time step, in units of ; For the first The fluctuation value of infiltration volume at each time step, in units of ; This is the time step index, with values ​​ranging from 1 to 36. Each feature component is normalized before being input into the model; the normalization formula is as follows:

[0065] ;

[0066] In the formula, For the first Normalized values ​​of each feature component, dimensionless; For the first The original values ​​of each feature component; For the first The minimum value of each feature component in the training set; For the first The maximum value of each feature component in the training set. , and The units are consistent with the corresponding characteristic components. The formula for the displacement increment is as follows:

[0067] ;

[0068] In the formula, For the first Displacement data for each time step, in mm; For the first Displacement data for each time step, in mm. The formula for the rate of change of stress is as follows:

[0069] ;

[0070] In the formula, For the first Stress data at each time step, in units of ; For the first Stress data at each time step, in units of ; The time interval is 2 hours by default. The formula for the fluctuation value of seepage volume is as follows:

[0071] ;

[0072] In the formula, For the first Traffic data at each time step, in units of ; For the first Traffic data at each time step, in units of ; This is the historical time step index, with a value range of [value range missing]. to The model outputs predicted values ​​of surrounding rock deformation at each monitoring section over the next 15 days. and stress release stability score ,in The unit is mm. The range is from 0 to 100, and it is dimensionless.

[0073] The specific implementation of step S04 is when the stress release stability score is... Rock deformation rate greater than 85 and over 7 consecutive days When the deformation is less than 0.1 mm / d, a spatial variability risk assessment is conducted on the corresponding surrounding rock deformation monitoring section, and the dispersion coefficient of the surrounding rock parameters is calculated. The formula for the deformation rate of the surrounding rock is expressed as follows:

[0074] ;

[0075] In the formula, The value represents the deformation rate of the surrounding rock, expressed in mm / d. For the first Displacement data for each day, in mm; For the first Displacement data for each day, in mm; This is the day index, with values ​​ranging from 1 to 7. The formula for the dispersion coefficient of the surrounding rock parameters is as follows:

[0076] ;

[0077] In the formula, represents the dispersion coefficient of the surrounding rock parameters, which is dimensionless; For the first The cumulative deformation at each convergent deformation monitoring point is expressed in mm. This is the average cumulative deformation at all convergent deformation monitoring points, in mm. This represents the total number of convergence deformation monitoring points within 10m before and after the fault face for monitoring surrounding rock deformation. The convergence deformation monitoring points are numbered, with values ​​ranging from 1 to... The formula for the average cumulative deformation is as follows:

[0078] ;

[0079] In the formula, the meanings of each variable are the same as described above. The formula for the cumulative deformation is expressed as follows:

[0080] ;

[0081] In the formula, For the first The cumulative deformation at each convergent deformation monitoring point is expressed in mm. For the first The convergence deformation monitoring point at the first The displacement change at each acquisition time relative to the previous time, in mm; This represents the total number of data collections from the completion of excavation to the current assessment time. This is the data collection time sequence number, with a value range from 1 to... .

[0082] The specific implementation of step S05 is as follows: if the dispersion coefficient of the surrounding rock parameters... If the coefficient of variation is less than 0.3, the corresponding section of the surrounding rock deformation monitoring section is deemed to meet the conditions for lining construction. A lining construction permit signal is generated, and the corresponding station number is recorded as the optimal construction time station number. If the coefficient of variation of the surrounding rock parameters is less than 0.3, the lining construction permit signal is generated, and the corresponding station number is recorded as the optimal construction time station number. If the value is greater than or equal to 0.3, the monitoring period is extended by 15 days and steps S02 to S04 are repeated.

[0083] The specific implementation of step S06 is to pour lining in the section that meets the conditions for lining construction according to the optimal construction timing chainage sequence. During the pouring process, the convergence deformation monitoring point is kept working continuously. When the surrounding rock deformation rate at any convergence deformation monitoring point is... If the stress increase suddenly exceeds 0.5 mm / d, immediately stop pouring and reassess the stress release stability score. The formula for determining the sudden increase is as follows:

[0084] ;

[0085] In the formula, The increase rate of deformation rate of the surrounding rock is dimensionless. This represents the current daily deformation rate of the surrounding rock, expressed in mm / d. For the first 7 days The deformation rate of the surrounding rock per day, expressed in mm / d; This is a historical day index, with values ​​ranging from 1 to 7. More than 5 and An early warning is triggered when the absolute value reaches 0.5 mm / d.

[0086] Stress release stability score The formula is expressed as follows:

[0087] ;

[0088] In the formula, The stress release stability score is dimensionless. The scoring component is dimensionless and is based on the predicted deformation rate. The scoring component is dimensionless and is based on the stress variation trend. These are dimensionless scoring components based on permeability stability. The calculation is based on the predicted deformation rate. The normalization process is expressed by the following formula:

[0089] ;

[0090] In the formula, For reference deformation rate, the empirical value is 0.5 mm / d. The function takes the maximum value to ensure that the result is not less than 0. The function is minimized to ensure the normalized value is no greater than 1. (Rating components) The calculation is based on the stress fitting slope. The normalization process is expressed by the following formula:

[0091] ;

[0092] In the formula, The slope of the linearly fitted line for the stress data sequence, in units of ; For reference slope, the empirical value is 0.01. Rating components The calculation is based on the coefficient of variation of seepage volume. The normalization process is expressed by the following formula:

[0093] ;

[0094] In the formula, The coefficient of variation is used as a reference, with an empirical value of 0.3. Predict the deformation rate. The formula is expressed as follows:

[0095] ;

[0096] In the formula, The deformation rate is predicted in mm / d; For the first Predicted values ​​of surrounding rock deformation for each prediction time step, in mm; For the first Predicted values ​​of surrounding rock deformation for each prediction time step, in mm; To predict the time step interval, a value of 1 day is typically used; The time step index is used for prediction, with a value ranging from 2 to 15. The stress fitting slope is calculated using the least squares method, as expressed in the following formula:

[0097] ;

[0098] In the formula, For the first Stress data at each time step, in units of ; This is the time step index, with values ​​ranging from 1 to 30. The formula for the coefficient of variation of seepage is as follows:

[0099] ;

[0100] In the formula, The coefficient of variation of seepage is dimensionless. For the first Traffic data at each time step, in units of ; This is the average of the traffic data over the most recent 30 time steps, in units of... ; This is the time step index, with values ​​ranging from 1 to 30. The formula for the average traffic data is as follows:

[0101] ;

[0102] In the formula, the meanings of each variable are the same as those described above.

[0103] To better understand and implement this invention, the following is a specific application scenario of this invention, Example 2:

[0104] After the tunnel excavation was completed, the technical team deployed a rock deformation monitoring section every 5 meters along the tunnel axis, totaling 640 monitoring sections, numbered from K0+000 to K3+200. Each monitoring section had convergence deformation monitoring points installed at five locations: the arch crown, left arch waist, right arch waist, left side wall, and right side wall. These monitoring points used laser displacement sensors to achieve millimeter-level precision measurements, with an accuracy of 0.01 mm. In the shallow buried sections (K0+000 to K0+580 and K2+720 to K3+200) with a depth of less than 50 meters, the team densely deployed rock stress monitoring sensors and seepage monitoring devices. The rock stress monitoring sensors were vibrating wire stress gauges, embedded at the contact surface between the initial support and the surrounding rock. Each monitoring section had four stress sensors to measure radial and tangential stresses. The seepage monitoring device consists of a guide channel and a flow meter. The guide channel is laid along the arch to collect seepage water, and the flow meter records the amount of water passing through per unit time with a measurement accuracy of 0.1 liters per hour.

[0105] like Figure 2As shown, the technical team continuously collected data from each monitoring point starting from the moment the tunnel excavation was completed. The collection frequency for displacement, stress, and flow data was uniformly set to once every 2 hours, i.e., 12 times per day. Taking the K1+250 monitoring section as an example, this section has a burial depth of 78 meters, belonging to the deep-buried section. The technical team continuously collected monitoring data for this section for 30 days, forming a time-series dataset of surrounding rock deformation containing 360 time steps. The displacement data sequence recorded the cumulative deformation of the arch crown monitoring point gradually increasing from 0 mm to 15.8 mm over 30 days; the cumulative deformation of the left arch waist monitoring point was 13.2 mm; the cumulative deformation of the right arch waist monitoring point was 14.1 mm; the cumulative deformation of the left wall monitoring point was 9.6 mm; and the cumulative deformation of the right wall monitoring point was 10.3 mm. The stress data sequence recorded the radial stress gradually decreasing from the initial 0.85 MPa to 0.32 MPa, and the tangential stress decreasing from the initial 1.24 MPa to 0.58 MPa. The infiltration data series recorded that the instantaneous flow rate fluctuated from an initial 2.3 liters per hour to 0.8 liters per hour, with a cumulative flow rate of 926 liters.

[0106] The technical team constructed a stress release prediction model to process the collected time-series dataset of surrounding rock deformation. The model's input layer receives a 72-dimensional feature vector, composed of monitoring data from 36 consecutive time steps (3 days). For the K1+250 monitoring section, the team extracted data from time steps 324 to 359 (36 time steps) as the feature vector input. Displacement increment was obtained by calculating the difference in displacement data at the crown monitoring points between adjacent time steps; the displacement increment at time step 359 was 0.08 mm. The stress change rate was obtained by calculating the ratio of the difference in radial stress data between adjacent time steps to the 2-hour time interval; the stress change rate at time step 359 was -0.003 MPa per hour. The seepage fluctuation value was obtained by calculating the difference between the instantaneous flow rate of 0.82 liters per hour at time step 359 and the average flow rate of 0.85 liters per hour for the previous 10 time steps (time steps 349 to 358); the fluctuation value was -0.03 liters per hour. Figure 3 As shown, after the 72-dimensional feature vector is input into the stress release prediction model, it first enters the coding layer for temporal feature extraction.

[0107] The encoding layer consists of four layers of Long Short-Term Memory (LSTM) units. The first layer receives 72-dimensional input and outputs a 128-dimensional hidden state vector. The second layer receives the output of the first layer and superimposes the original input information through residual connections to output a 128-dimensional vector. The third and fourth layers use the same structure to extract higher-level temporal features. The residual connections between layers effectively alleviate the gradient vanishing problem, enabling the model to learn the long-term dependencies of the stress release process. The 128-dimensional feature vector output from the encoding layer is then input into the attention layer for key time period identification. Since the burial depth of the K1+250 monitoring section is 78 meters, which is greater than 50 meters, the number of attention heads in the attention layer is set to 4. The attention layer assigns different weight coefficients to the feature vectors at 36 time steps. The technical team found that the data from time steps 330 to 336, i.e., 27.5 to 28 days after excavation, received higher attention weights, indicating that the deformation features during this period have a significant impact on future deformation prediction. The weighted feature vector output from the attention layer is then fed into the decoding layer to generate the prediction results.

[0108] The decoding layer consists of three fully connected layers with 256, 128, and 2 neurons respectively. The first layer maps the feature vector output from the attention layer to a 256-dimensional space, the second layer further compresses it to a 128-dimensional space, and the third layer outputs a 2-dimensional result. The first 15 values ​​are the predicted values ​​of the surrounding rock deformation over the next 15 time steps (30 hours), and the last value is the stress release stability score. For the K1+250 monitoring section at the 360th time step, the predicted values ​​of the arch deformation over the next 15 time steps are 15.88 mm, 15.95 mm, 16.02 mm, 16.08 mm, 16.14 mm, 16.19 mm, 16.24 mm, 16.29 mm, 16.33 mm, 16.37 mm, 16.41 mm, 16.45 mm, 16.48 mm, 16.51 mm, and 16.54 mm, respectively, with a stress release stability score of 88. The technical team calculated the predicted deformation rate: 0.04 mm / hour (0.96 mm / day) for time steps 360-361, and 0.015 mm / hour (0.36 mm / day) for time steps 374-375. Figure 4 As shown, the stress release stability score is calculated based on a weighted summation of three dimensions.

[0109] The calculation process for a stress release stability score of 88 involves three dimensions: predicted deformation rate, stress change trend, and seepage stability. The predicted deformation rate is calculated by analyzing the predicted deformation values ​​over the next 15 time steps, resulting in an average deformation rate of 0.052 mm / h. This value is less than the threshold of 0.1 mm / day (0.004 mm / h), corresponding to a normalized score of 92. The stress change trend dimension is achieved by linearly fitting radial stress data from the most recent 30 time steps (330th to 359th time steps). The absolute value of the slope of the fitted line is 0.0012 MPa / h, indicating that the stress is stabilizing, resulting in a normalized score of 85. The seepage stability dimension is obtained by calculating the coefficient of variation of instantaneous flow data from the most recent 30 time steps. The standard deviation is 0.06 L / h, the average is 0.84 L / h, and the coefficient of variation is 0.071, corresponding to a normalized score of 86. The three dimensions are weighted and summed with weights of 0.5, 0.3, and 0.2, respectively, ultimately yielding a stress release stability score of 88. The score is greater than the threshold of 85, which preliminarily meets the conditions for the timing of lining construction.

[0110] The technical team further examined the surrounding rock deformation rate at the K1+250 monitoring section over seven consecutive days. From time steps 337 to 360 (days 28 to 30), the daily displacement increments at the crown monitoring point were 0.09 mm, 0.08 mm, 0.08 mm, 0.07 mm, 0.08 mm, 0.09 mm, and 0.08 mm, respectively, with an average surrounding rock deformation rate of 0.081 mm per day, less than the threshold of 0.1 mm per day. After meeting both the stress release stability score and the surrounding rock deformation rate criteria, the technical team conducted a spatial variability risk assessment of the K1+250 monitoring section. The team extracted the cumulative deformation data from all convergent deformation monitoring points within a 10-meter radius before and after the K1+250 section across five monitoring sections (K1+240 to K1+260). The cumulative deformation of the 25 convergent deformation monitoring points across the five monitoring sections is shown in Table 1.

[0111] Table 1. Cumulative Deformation of Monitoring Sections from K1+240 to K1+260

[0112]

[0113] The technical team calculated the standard deviation of 25 cumulative deformation data points to be 2.15 mm, the average to be 12.46 mm, and the coefficient of variation for the surrounding rock parameters to be 0.173. This coefficient of variation is less than the threshold of 0.3, indicating that the surrounding rock properties around the K1+250 monitoring section are relatively uniform, and the stress release process exhibits good spatial consistency. The technical team determined that the corresponding section of the K1+250 monitoring section met the conditions for lining construction, and the system automatically generated a lining construction permit signal and recorded K1+250 as the optimal construction time station. Figure 5As shown, the technical team evaluated 640 monitoring sections along the entire line using the same method, and finally identified 156 sections that met the conditions for lining construction.

[0114] The technical team organized the lining pouring construction according to the optimal construction timing and chainage sequence. During the pouring process in the K1+250 section, the convergence deformation monitoring point remained continuously operational, collecting displacement data every 2 hours. In the first 24 hours after pouring began, the deformation rate of the surrounding rock at the arch crown monitoring point remained stable at approximately 0.09 mm per day. However, in the 28th hour after pouring began, the daily deformation rate of the surrounding rock at the arch crown monitoring point suddenly increased to 0.56 mm per day, a 6.9-fold increase compared to the average deformation rate of 0.081 mm per day over the previous 7 days. The technical team immediately triggered the early warning mechanism, suspending the lining pouring operation in the K1+250 section and reassessing the stress release stability score. After re-collecting the monitoring data for the latest 36 time steps and inputting it into the stress release prediction model, the output stress release stability score decreased to 72, indicating an abnormal acceleration in the stress release of the surrounding rock. The technical team's analysis revealed a hidden groundwater seepage channel above this section. The instantaneous flow rate recorded by the seepage monitoring device suddenly increased from 0.8 liters per hour to 3.2 liters per hour. The groundwater activity caused a decrease in the effective stress of the surrounding rock, leading to accelerated deformation. The technical team decided to extend the monitoring period for the K1+250 section by 15 days, and reassess the timing of lining construction after the groundwater activity stabilizes.

[0115] The stress release prediction model was trained using a gradient optimization mechanism based on multi-level adaptive weight decay. The technical team collected surrounding rock monitoring data from 23 completed loose soft rock tunnels both domestically and internationally, covering the entire process from excavation completion to lining construction. After preprocessing the collected surrounding rock monitoring data, outliers during sensor failure periods were removed, and missing data was filled using linear interpolation, resulting in a dataset containing 8760 training samples and 2190 validation samples. During model training, the team used orthogonal initialization to ensure uniform initial gradient distribution for each layer of the encoding layer, and Xavier initialization to control activation value variance for the attention and decoding layers. During the 50th round of training, the technical team calculated the variance of the gradients of the parameters of each network layer. The gradient variance of the first encoding layer was 0.0089, the second encoding layer was 0.0034, the third encoding layer was 0.0012, the fourth encoding layer was 0.0008, the attention layer was 0.0052, the first decoding layer was 0.0018, the second decoding layer was 0.0009, and the third decoding layer was 0.0006. The global average gradient variance was 0.0033. The gradient variances of the first encoding layer and the attention layer were greater than 0.0050 (1.5 times the global average gradient variance) and were defined as high gradient layers with a weight decay factor of 0.01. The gradient variances of the fourth encoding layer and the third decoding layer were less than 0.0017 (0.5 times the global average gradient variance) and were defined as low gradient layers with a weight decay factor of 0.001. The remaining layers are intermediate gradient layers with a weight decay factor of 0.005.

[0116] The gradient optimization mechanism calculates the approximate second derivative of the loss function with respect to the parameters in each training batch as local curvature information. For a certain weight parameter in the first layer of the encoding layer, its current value is 0.582. After applying a small perturbation of 0.0001 to this parameter, the loss function increases from 1.234 to 1.236, and the approximate second derivative value is 200. The product of the local curvature information 200 and the current parameter value 0.582, 116.4, is added to the loss function as a regularization term to guide the optimization process to avoid steep regions in the parameter space. The technical team sets the sliding window length to 10 batches and calculates the standard deviation of the parameter update magnitude of the first layer of the encoding layer within the window. In batches 80 to 89, the standard deviation of the parameter update magnitude is 0.063, which is greater than the threshold of 0.05. The system automatically decays the learning rate from 0.001 to 0.0009. In batches 140 to 149, the standard deviation of the parameter update magnitude was 0.008, which is less than the threshold of 0.01. The system automatically increased the learning rate from 0.0007 to 0.00077, but not exceeding the initial learning rate of 0.001. During training, the mean squared error loss function was used to measure the deformation prediction error, and the cross-entropy loss function was used to measure the stability scoring error. The total loss function was a weighted sum of the two, with weight coefficients of 0.6 and 0.4, respectively. After 200 rounds of training, the minimum loss on the validation set was 0.0182, and the corresponding model parameters were saved as the final model for practical engineering applications.

[0117] The advancements of this invention compared to traditional methods lie in two aspects: dynamic prediction of the stress release process and quantitative assessment of spatial uniformity. Traditional methods rely solely on deformation monitoring data at a single moment to determine the timing of lining construction, failing to foresee future deformation evolution trends and easily leading to lining cracking due to insufficient stress release. This invention extracts temporal features from 36 consecutive time steps using long short-term memory units and combines this with a multi-head attention mechanism to identify key deformation stages, achieving accurate prediction of surrounding rock deformation over the next 15 time steps. This transforms passive judgment into active prediction, fundamentally solving the problem of difficulty in quantifying the degree of stress release completion. Traditional methods ignore the influence of spatial variability of the surrounding rock on the stress on the lining, focusing only on the deformation state of a single monitoring section, resulting in the failure to identify stress concentrations caused by local weak areas. This invention establishes a quantitative correlation mechanism between the spatial uniformity of the surrounding rock and the timing of lining construction by calculating the dispersion coefficient of the cumulative deformation within 10 meters before and after the monitoring section. This ensures that the lining structure solidifies and forms in a stress-balanced surrounding rock environment, avoiding structural damage caused by uneven stress distribution. The gradient optimization mechanism based on multi-level adaptive weight decay achieves refined parameter update control by distinguishing the gradient characteristics of different network levels. This enables the stress release prediction model to fully learn the long-term trend and short-term fluctuation characteristics of surrounding rock deformation during training, while avoiding the decline in generalization ability caused by overfitting. As a result, it maintains stable prediction performance when facing loose soft rock tunnels with different geological conditions and construction disturbances.

[0118] It should be noted that the variables involved in this invention are explained in detail in Table 2.

[0119] Table 2 Variable Explanation Table

[0120]

[0121] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for determining the timing of lining construction in loose soft rock tunnels, characterized in that, After tunnel excavation, rock deformation monitoring sections are set up along the tunnel axis, and convergence deformation monitoring points, rock stress monitoring sensors, and seepage monitoring devices are installed. Displacement data, stress data, and flow data are continuously collected to form a time series dataset of rock deformation. The time series dataset of rock deformation is input into the stress release prediction model, which outputs the predicted value of future rock deformation and the stress release stability score. The stress release prediction model consists of an encoding layer, an attention layer, and a decoding layer. The encoding layer uses long short-term memory units and residual connections to extract time series features. The attention layer adjusts the number of attention heads according to the burial depth of the monitoring section. The decoding layer outputs the predicted value of rock deformation and the stress release stability score. When the stress release stability score and the rock deformation rate meet the threshold conditions, a spatial variability risk assessment is performed to calculate the dispersion coefficient of the rock parameters. When the dispersion coefficient of the rock parameters is less than the threshold, a lining construction permit signal is generated. Real-time monitoring is maintained during the pouring process, and pouring is stopped and reassessed when the rock deformation rate suddenly increases.

2. The method according to claim 1, characterized in that, The surrounding rock deformation monitoring sections are set up at 5m intervals along the tunnel axis, and convergence deformation monitoring points are installed at the arch crown, arch waist and sidewall positions of each surrounding rock deformation monitoring section.

3. The method according to claim 2, characterized in that, In shallow buried sections with a burial depth of less than 50m, densely deploy surrounding rock stress monitoring sensors and seepage monitoring devices.

4. The method according to claim 3, characterized in that, The data collection frequency is once every 2 hours, and the continuous collection period is no less than 30 days.

5. The method according to claim 4, characterized in that, The surrounding rock deformation time series dataset contains displacement data sequences, stress data sequences, and seepage data sequences, which are arranged in chronological order to form a multidimensional time series matrix.

6. The method according to claim 5, characterized in that, The input layer of the stress relief prediction model receives a 72-dimensional feature vector, which contains the displacement increment stress change rate and seepage fluctuation value for 36 consecutive time steps.

7. The method according to claim 6, characterized in that, The coding layer consists of four layers of long short-term memory units, each containing 128 neurons, with gradients transmitted between layers via residual connections.

8. The method according to claim 7, characterized in that, The attention layer adopts a multi-head attention mechanism. When the burial depth of the monitoring section is less than 50m, the number of attention heads is 8, and when the burial depth of the monitoring section is greater than or equal to 50m, the number of attention heads is 4.

9. The method according to claim 8, characterized in that, The decoding layer consists of three fully connected layers with 256, 128, and 2 neurons respectively, and outputs 15 time-step predictions of surrounding rock deformation and one stress release stability score.

10. The method according to claim 9, characterized in that, The stress release prediction model training adopts a gradient optimization mechanism based on multi-level adaptive weight decay. The variance of the parameter gradient of each layer is calculated. The layers with variance greater than 1.5 times the average global gradient variance are defined as high gradient layers and a weight decay coefficient of 0.01 is applied. The layers with variance less than 0.5 times the average global gradient variance are defined as low gradient layers and a weight decay coefficient of 0.001 is applied. The intermediate gradient layers are defined as middle gradient layers and a weight decay coefficient of 0.005 is applied.

Images