Monitoring Methods for Progressive Failure and Time-Related Cracking of Surrounding Rock in Ultra-Long and Ultra-Deep TBM Tunnels

Abstract

This invention relates to the field of tunnel and underground engineering safety monitoring technology, and discloses a method for monitoring the progressive failure and time-dependent cracking of surrounding rock in ultra-long and ultra-deep TBM tunnels. This method enables real-time monitoring, precise location, time-dependent quantification, and rapid feedback of surrounding rock failure, thereby improving the construction safety of ultra-long and ultra-deep TBM tunnel projects. The invention employs a combination of fixed deployment and dynamic inspection to construct a multi-modal sensor network, collecting data from multiple sources. Through multi-source data preprocessing and spatiotemporal fusion to unify the data benchmark, noise and spatiotemporal asynchrony issues are eliminated. A CNN-LSTM attention hybrid model is used to accurately predict the location and probability of progressive rock failure. Time-series analysis is used to extract cracking characteristics and construct quantitative indicators to assess the degree of time-dependent cracking and predict its long-term evolution. Finally, based on the failure probability and cracking quantitative indicators, graded early warning is implemented, and TBM execution parameter adjustments and emergency response are linked. This invention is applicable to ultra-long and ultra-deep TBM tunnel projects.

Description

Technical Field

[0001] This invention relates to the field of tunnel and underground engineering safety monitoring technology, specifically to a method for monitoring the progressive damage and time-related cracking of surrounding rock in ultra-long and ultra-deep TBM tunnels. Background Technology

[0002] With the development of underground engineering towards ultra-long and ultra-deep tunnels, TBMs (tunnel boring machines) have been widely used in the construction of ultra-long and ultra-deep tunnels (over 50 km in length and over 2000 m in depth) due to their high efficiency. However, these projects are generally located in environments with high ground stress, high seepage pressure, and complex geological structures, resulting in significantly higher construction risks than conventional tunnels. Under these conditions, the surrounding rock is prone to progressive failure and time-dependent cracking under long-term high ground stress, inducing geological hazards such as delayed rockbursts and long-term creep deformation. These hazards are characterized by strong concealment, significant time lag, large destructive range, and strong suddenness, easily causing major safety accidents such as damage to tunneling equipment and tunnel collapses. They have become a core hidden danger restricting the safe construction and long-term stable operation of ultra-long and ultra-deep TBM tunnels.

[0003] Existing methods for monitoring surrounding rock mainly focus on detecting single physical signals or assessing local risks, which has significant limitations. For example, some technologies rely on microseismic monitoring to capture acoustic signals of crack propagation, or on acoustic emission tests and ultrasonic testing to assess rock mass damage. However, these methods can only acquire single-dimensional data and cannot achieve collaborative analysis of multi-source data, resulting in insufficient accuracy in locating the damage site, with errors often exceeding 10 meters, making it difficult to meet the precise monitoring needs of ultra-long and ultra-deep tunnels. Other technologies analyze the evolution of surrounding rock damage through numerical simulation or combine TBM tunneling parameters to assess rockburst risk, but they ignore the dynamic development process of time-dependent cracking, cannot predict the evolution path of delayed failure, and rely on manual judgment of data trends, resulting in response times of up to several hours, making it difficult to achieve real-time early warning.

[0004] Internationally, research on the aging cracking of surrounding rock has largely focused on ultrasonic testing and creep behavior analysis, obtaining rock mass aging parameters through laboratory tests. However, laboratory test conditions differ significantly from the high ground stress and complex geological environments in the field, resulting in poor applicability of the test results. Furthermore, existing technologies lack a data lifecycle management mechanism, making monitoring data susceptible to tampering and unable to provide reliable evidence for long-term tunnel stability assessments. Fixed sensor deployment methods also have monitoring blind spots, making it difficult to detect crack propagation in the face of the tunnel face and in the dead corners of the tunnel sidewalls.

[0005] In summary, existing monitoring methods cannot simultaneously meet the needs of full coverage, real-time response, precise positioning, and long-term assessment for ultra-long and ultra-deep TBM tunnels. There is an urgent need for a monitoring method that integrates multimodal sensing, intelligent analysis, and end-to-end assurance to achieve precise positioning of the progressive failure location of the surrounding rock and quantitative analysis of the time-dependent cracking characteristics, thereby providing technical support for safe construction and long-term operation and maintenance of the project. Summary of the Invention

[0006] The technical problem to be solved by this invention is to provide a method for monitoring the progressive failure and time-dependent cracking of surrounding rock in ultra-long and ultra-deep TBM tunnels, so as to realize real-time monitoring, precise location, time-dependent quantification and rapid feedback of surrounding rock failure, thereby improving the construction safety of ultra-long and ultra-deep TBM tunnel projects.

[0007] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows:

[0008] A method for monitoring progressive failure and time-related cracking of surrounding rock in ultra-long and ultra-deep TBM tunnels includes the following steps:

[0009] S1. Construct a multimodal sensor network that combines fixed deployment with dynamic inspection to collect surrounding rock strain data, microseismic data, TBM tunneling parameters, and inspection data of blind spots in fixed sensor monitoring;

[0010] S2. Preprocess and spatiotemporally fuse the collected multi-source data to obtain spatiotemporally fused data;

[0011] S3. Based on the pre-trained CNN-LSTM attention hybrid model, using the spatiotemporal fusion data as input, predict the location and probability of progressive failure of the surrounding rock;

[0012] S4. Extract time-dependent cracking characteristics based on spatiotemporal fusion data, construct time-dependent cracking quantitative indicators, and complete the assessment of the degree of time-dependent cracking of surrounding rock and the prediction of its long-term evolution.

[0013] S5. Based on the probability of damage and the quantitative indicators of time-dependent cracking, a graded early warning system is implemented, which is linked to the TBM execution parameter adjustment and emergency response, and the monitoring and early warning results are output through a visual interface.

[0014] This solution, through a closed-loop process of sensing and acquisition, data fusion, intelligent prediction, cracking assessment, and early warning and disposal, can fully reflect the changes in the surrounding rock from deformation accumulation and crack expansion to time-related cracking. It overcomes the problems of existing technologies that can only monitor locally, have isolated data, cannot simultaneously realize damage location and cracking quantification, and are difficult to form a closed-loop disposal system.

[0015] Furthermore, in step S1, the construction of a multimodal sensor network combining fixed deployment and dynamic inspection includes:

[0016] Distributed fiber optic sensors are deployed along the tunnel axis and radial direction to collect surrounding rock strain data;

[0017] A microseismic sensor array is arranged in front of the tunnel face and on the tunnel sidewall to collect microseismic data generated by the propagation of microcracks in the surrounding rock.

[0018] Parameter sensors are integrated into the TBM cutterhead, propulsion cylinder, and torque sensor to collect TBM tunneling parameters;

[0019] By deploying autonomous, mobile IoT micro-robots, inspection data can be collected from blind spots in fixed sensor monitoring.

[0020] In this scheme, distributed optical fibers can continuously acquire the strain field of the surrounding rock across the entire area, reflecting the stress distribution and deformation accumulation law of the surrounding rock; microseismic arrays can capture the elastic wave signals released by rock fractures, corresponding to the physical nature of crack initiation and propagation; TBM tunneling parameters directly reflect the disturbance intensity of construction on the surrounding rock; and IoT robots fill the structural dead angles and geometric blind spots that fixed sensors cannot cover, thereby improving the integrity and reliability of monitoring through a multimodal sensor network.

[0021] Furthermore, step S1 also includes:

[0022] The surrounding rock strain rate is calculated based on the surrounding rock strain data acquired by distributed fiber optic sensors. And based on the surrounding rock strain rate Dynamically adjust the sampling frequency of the multimodal sensor network:

[0023] when At that time, the sampling frequency is maintained at 1Hz;

[0024] when At that time, the sampling frequency was increased to 10Hz;

[0025] when At that time, the sampling frequency was adjusted to 100Hz.

[0026] In this scheme, the strain rate It directly reflects the severity of surrounding rock deformation and is a precursor characteristic of surrounding rock entering an unstable evolution. The sampling frequency of the multi-modal sensor network is dynamically adjusted based on the strain rate as the trigger condition. In the stable stage, the sampling frequency is low to reduce data redundancy and transmission pressure. In the abnormal accelerated deformation stage, the frequency is automatically increased to capture high-frequency abrupt change signals, providing a reliable data source for the real-time monitoring and quantification of subsequent surrounding rock failure.

[0027] Furthermore, in step S2, the preprocessing includes: outlier removal based on the 3σ criterion, noise filtering of the data, and data standardization.

[0028] The spatiotemporal fusion includes: using an improved Kalman filter algorithm, by introducing an adaptive weighting factor, to perform spatiotemporal fusion on the preprocessed surrounding rock strain data, microseismic data, TBM tunneling parameters and inspection data, and to uniformly map data with different timestamps and different spatial locations to the three-dimensional tunnel coordinate system.

[0029] In this scheme, data preprocessing can preserve the true surrounding rock response and eliminate the dimensional differences of multi-source data. In spatiotemporal fusion, adaptive weights can be used to dynamically allocate contribution based on signal credibility, thereby unifying multi-source data to the same spatiotemporal benchmark and laying a good foundation for accurate data processing in the future.

[0030] Furthermore, in step S3, the CNN-LSTM attention hybrid model includes:

[0031] The CNN feature extraction layer is configured to perform spatial feature extraction on spatiotemporal fusion data and output a spatial feature map.

[0032] The LSTM temporal analysis layer is configured to take the spatial feature map as input, extract temporal evolution features, and output a temporal feature sequence.

[0033] The self-attention layer is configured to assign feature weights to the temporal feature sequence and output weighted fusion features.

[0034] The output layer is configured to output the three-dimensional coordinates and failure probability of the progressive failure location of the surrounding rock based on the weighted fusion features.

[0035] In this scheme, the CNN feature extraction layer spatially identifies the spatial morphology of precursors to failure, such as stress concentration, strain anomalies, and dense microseismic areas; the LSTM temporal analysis layer uses a gating mechanism to memorize long-term temporal dependencies, characterizing the lag pattern of surrounding rock deformation accumulation and crack propagation over time; the self-attention layer strengthens weak abrupt change signals in high-risk areas through dynamic weights, suppressing background noise interference; the coupling of these three layers can simultaneously capture the "spatial location characteristics" and "temporal development characteristics" of failure, matching the spatiotemporal evolution mechanism of progressive surrounding rock failure, thereby improving positioning accuracy and prediction reliability.

[0036] Furthermore, the pre-training methods for the CNN-LSTM attention hybrid model include:

[0037] A training dataset was constructed by collecting historical monitoring data, geological disaster case data, and indoor simulation test data of ultra-long and ultra-deep TBM tunnel projects.

[0038] The training dataset is input into the CNN-LSTM attention hybrid model, and the prediction results are output after spatial feature extraction, temporal feature extraction and attention weighting.

[0039] Based on the predicted results and the loss calculated from the actual labels, an adaptive optimizer is used to iteratively optimize the model parameters until the model accuracy meets the requirements.

[0040] In this scheme, real engineering data and high-stress indoor test data are used as training sets to enable the model to learn failure modes that conform to the mechanical properties of deep rock masses. Through real-label supervised training, the model establishes a mapping relationship between multi-source signals, failure locations, and failure probabilities, enabling the model to predict progressive failure locations and failure probabilities.

[0041] Furthermore, in step S4, the method for extracting time-dependent cracking features based on spatiotemporal fusion data includes:

[0042] The ARIMA model combined with wavelet transform is used to perform temporal decomposition and feature extraction on the spatiotemporal fusion data to obtain the time-dependent cracking characteristics of the surrounding rock. The time-dependent cracking characteristics include crack propagation rate, cracking lag, creep index and crack frequency.

[0043] In this scheme, wavelet transform has multi-scale resolution characteristics and can extract weak cracking signals from strong noise; the ARIMA model separates the slow development trend of time-dependent cracking from the time-series signal; the crack propagation rate, time delay, creep index and crack frequency extracted therefrom are all intrinsic features of time-dependent cracking, which can objectively reflect the cracking process of rock mass under long-term load and solve the problem that traditional methods cannot quantify time effects.

[0044] Furthermore, in step S4, the method for constructing the time-dependent cracking quantification index includes:

[0045] Based on the extracted time-dependent cracking characteristics, the Time-dependent Cracking Index (TCI) is constructed as a quantitative indicator of the degree of time-dependent cracking in the surrounding rock.

[0046] The aging cracking index is calculated using a weighted summation method, and the calculation formula is as follows:

[0047] ;

[0048] in, , , , These are the weight coefficients predetermined using the analytic hierarchy process (AHP). Surrounding rock strain rate The mapping score, ; The maximum principal stress of the surrounding rock The mapping score, ; Crack frequency The mapping score, ; Crack propagation rate The mapping score, ;

[0049] The age-related cracking levels of the surrounding rock are classified according to the TCI numerical range:

[0050] Corresponding to a stable state;

[0051] Corresponding to a slightly cracked state;

[0052] This corresponds to a severely cracked state.

[0053] In this scheme, the TCI index normalizes and weights four core time-dependent cracking parameters—surrounding rock strain rate, maximum principal stress, crack frequency, and crack propagation rate—to objectively and quantitatively characterize the degree of cracking development of surrounding rock under long-term high ground stress. By setting graded thresholds, the standardization of surrounding rock cracking status is achieved, providing a quantitative basis for subsequent graded early warning.

[0054] Furthermore, in step S4, the methods for assessing the degree of age-related cracking of the surrounding rock and predicting its long-term evolution include:

[0055] The current age-related cracking level of the surrounding rock is determined based on the current age-related cracking index (TCI) value.

[0056] Blockchain technology is used to record time-sensitive data, with each data block containing a timestamp, monitoring location, feature parameters, and TCI value information;

[0057] Based on historical data stored on the blockchain, a time series prediction model is used to extrapolate the time-dependent cracking trend of the surrounding rock in the next 1–12 months, and obtain the time-dependent cracking evolution path.

[0058] A visual heat map is generated by combining the cracking evolution path, marking high-risk cracking areas at different time periods.

[0059] In this scheme, the blockchain adopts chain-hash encryption and distributed notarization to ensure data traceability and provide a reliable data source for long-term cracking analysis; the time series prediction is based on the extrapolation of historical cracking trends, which conforms to the time evolution law of time-sensitive phenomena such as creep and delayed failure; the heat map uses spatial color to intuitively reflect the degree of cracking and development direction, which is convenient for identifying potential delayed rockbursts and long-term instability areas.

[0060] Furthermore, in step S5, the method of classifying and issuing early warnings based on the probability of damage and the quantitative index of time-dependent cracking, and linking the TBM execution parameters for adjustment and emergency response, includes:

[0061] When the probability of failure is between 60% and 80% or the time-dependent cracking index (TCI) meets the requirements... When this occurs, it is determined to be a Level 1 warning, the original tunneling parameters of the TBM are maintained, and the monitoring sampling frequency is increased;

[0062] When the probability of failure is between 80% and 90% or the time-dependent cracking index (TCI) meets the requirements... When the situation is deemed a Level II warning, the TBM is controlled to reduce its tunneling speed and thrust.

[0063] When the probability of failure is ≥90% or the time-dependent cracking index (TCI) meets the requirements... When the situation is assessed as a Level 3 warning, the TBM is immediately shut down, and the emergency support procedure is initiated.

[0064] In this scheme, the probability of failure reflects the risk of instantaneous failure, while the TCI reflects the degree of time-dependent cracking. Using these two indicators as the basis for early warning judgment can avoid false alarms and missed alarms. The early warning is directly linked to TBM control, which can quickly suppress the development of surrounding rock cracking according to the risk level and reduce the probability of accidents such as collapse and rock burst.

[0065] The beneficial effects of this invention are:

[0066] (1) The monitoring coverage is more comprehensive and the data integrity is higher:

[0067] This invention employs a multimodal sensor network that combines fixed deployment with dynamic inspection. Through the collaborative acquisition of data via distributed optical fibers, micro-seismic arrays, TBM tunneling parameter sensors, and IoT micro-robots, it achieves full-area, multi-dimensional, and blind-spot-free monitoring of the surrounding rock, fundamentally solving the problems of blind spots, single data dimensions, and inability to fully reflect the true state of the surrounding rock caused by traditional fixed sensors.

[0068] (2) Data processing is more accurate and reliable, providing a high-quality foundation for subsequent analysis:

[0069] This invention performs spatiotemporal fusion through outlier removal, noise filtering, standardization, and improved Kalman filtering with adaptive weights. This effectively eliminates noise interference, spatiotemporal asynchrony, and dimensional differences from multi-source data, reduces data fusion errors, improves data consistency and credibility, and ensures the authenticity and reliability of damage localization and cracking assessment results.

[0070] (3) High accuracy in locating the damaged location and stronger prediction accuracy:

[0071] This invention employs a CNN-LSTM attention hybrid model, which can simultaneously extract the spatial distribution features and temporal evolution features of the surrounding rock state. It also enhances weak signals in high-risk areas through a self-attention mechanism, thereby achieving accurate output of progressive failure locations and quantitative assessment of failure probability, significantly improving positioning accuracy and prediction reliability.

[0072] (4) It can realize quantitative assessment of age-related cracking and prediction of long-term evolution:

[0073] This invention extracts the time-dependent cracking characteristics of surrounding rock using ARIMA combined with wavelet transform, and constructs a time-dependent cracking index (TCI) that is mapped and weighted by a score of 0 to 10, thereby achieving a standardized determination of the degree of cracking in surrounding rock and enabling an objective and quantitative characterization of the time-dependent cracking state of surrounding rock. At the same time, relying on the time-series prediction model and blockchain-based trusted evidence data, it can predict future cracking trends and achieve early identification of hidden risks such as delayed failure and long-term creep.

[0074] (5) The early warning is scientific and reasonable, and the risk handling is more timely and efficient:

[0075] This invention uses a dual-indicator approach of probability of failure and time-dependent cracking index (TCI) for graded early warning, resulting in more accurate early warning judgments and effectively reducing false alarms and missed alarms. The early warning results directly link the TBM to adjust tunneling parameters or implement automated emergency support, rapidly suppressing the development of surrounding rock cracking, reducing the risk of accidents such as rock bursts and collapses, and improving construction safety. Attached Figure Description

[0076] Figure 1 This is a flowchart of the monitoring method for progressive failure and time-related cracking of surrounding rock in ultra-long and ultra-deep TBM tunnels in this invention. Detailed Implementation

[0077] This invention aims to provide a method for monitoring the progressive failure and time-dependent cracking of surrounding rock in ultra-long and ultra-deep TBM tunnels, enabling real-time monitoring, precise location, time-dependent quantification, and rapid feedback of surrounding rock damage, thereby improving the construction safety of ultra-long and ultra-deep TBM tunnel projects. Its core idea is to construct an integrated monitoring system based on the spatiotemporal evolution of progressive failure and time-dependent cracking of surrounding rock in ultra-long and ultra-deep TBM tunnels, encompassing full-domain multimodal perception, spatiotemporal fusion of multi-source data, accurate prediction of failure location and probability, quantitative characterization of time-dependent cracking, and graded early warning and closed-loop response. Specifically, a multimodal sensor network is constructed by combining fixed deployment with dynamic inspection to achieve full-area, multi-dimensional, and blind-spot-free data acquisition of the surrounding rock. Multi-source data preprocessing and spatiotemporal fusion are used to unify the data benchmark, eliminating noise and spatiotemporal asynchrony issues. A CNN-LSTM attention hybrid model is used to accurately predict the location and probability of progressive rock failure. Time-series analysis is used to extract cracking characteristics and construct quantitative indicators to assess the degree of time-related cracking and predict its long-term evolution. Finally, based on the probability of failure and the quantitative cracking indicators, graded early warning is achieved, and the TBM (Total Damage Machine) is linked to perform parameter adjustment and emergency response, forming a complete closed loop of monitoring, analysis, early warning, and response. This addresses the shortcomings of existing technologies, such as insufficient monitoring coverage, isolated data, low positioning accuracy, inability to quantify time-related cracking, and difficulty in providing real-time early warning.

[0078] For specific implementation details, see [link / reference]. Figure 1 The present invention provides a method for monitoring progressive failure and time-related cracking of surrounding rock in ultra-long and ultra-deep TBM tunnels, comprising the following steps:

[0079] S1. Construct a multimodal sensor network to collect data from multiple sources:

[0080] In this step, based on the geological characteristics and monitoring needs of ultra-long and ultra-deep TBM tunnels, a multi-modal sensor network combining fixed deployment and dynamic inspection is constructed to collect surrounding rock strain data, microseismic data, TBM tunneling parameters, and inspection data of fixed sensor monitoring blind spots.

[0081] In one exemplary implementation, the various sensor deployment schemes in a multimodal sensor network are as follows:

[0082] The specific deployment scheme for the multimodal sensor network is as follows:

[0083] Deployment of Distributed Fiber Optic Sensors (DFOS): Distributed optical fibers are laid along the tunnel axis and radial direction. The axial fibers are fixed in pre-embedded slots on the tunnel wall every 0.5m, and the radial fibers are implanted into the surrounding rock through boreholes to a depth of 1.5-2 times the tunnel radius, covering the stress concentration area of ​​the surrounding rock. Phase-sensitive optical time-domain reflectometry is used to achieve full-domain monitoring of strain field and temperature changes, with a monitoring accuracy of 1με and a temperature resolution of ±0.1℃.

[0084] Deployment of microseismic sensor array: A set of microseismic sensors is arranged within a range of 10-15m in front of the tunnel face and every 5m on the tunnel sidewall. Each set contains 3 three-component sensors to form a three-dimensional monitoring array. The sensor sensitivity is ≥100V / m / s and the frequency response range is 1-1000Hz. It can capture the weak acoustic wave signals generated by the propagation of micro-cracks in the surrounding rock and realize the early detection of crack initiation and propagation.

[0085] TBM tunneling parameter sensor integration: Parameter sensors are integrated into key components such as the TBM cutterhead, propulsion cylinder, and torque sensor to collect real-time data on propulsion force (measurement range 0-5000kN), torque (0-2000kN·m), tunneling speed (0-5m / h), and machine vibration acceleration (0-10g).

[0086] To balance monitoring accuracy with system energy consumption and storage pressure, this invention also incorporates an adaptive sampling mechanism for the acquisition parameters of the multimodal sensor network, providing a reliable data source for subsequent real-time monitoring and quantification of surrounding rock damage.

[0087] Based on the real-time stress level of the surrounding rock, the sampling frequency is dynamically adjusted, and the strain rate of the surrounding rock is calculated using DFOS data. :

[0088] when At that time, the deformation rate of the surrounding rock was at a low level, the overall rock mass structure was stable, and high-frequency acquisition was not required; the sampling frequency was maintained at 1Hz.

[0089] when At that time, the deformation rate of the surrounding rock increased significantly, and the rock mass entered a slight time-dependent cracking stage. It is necessary to increase the sampling frequency to capture subtle changes, and the sampling frequency was increased to 10Hz.

[0090] when At this time, the deformation of the surrounding rock is developing rapidly, the rock mass cracking is accelerating and showing signs of gradual failure. High-frequency sampling is required to capture the abnormal evolution process, and the sampling frequency is adjusted to 100Hz.

[0091] Deployment of wearable IoT microrobot sensors: Configure no fewer than 5 wearable IoT microrobots. The robots are equipped with micro strain sensors, acoustic sensors and high-definition cameras, weigh ≤2kg, have a battery life of ≥8h, and can move autonomously along the tunnel wall to perform dynamic inspections of blind spots of fixed sensors (such as the corners of the tunnel face and the tunnel joints). The robots communicate with the main control system through a wireless mesh network to achieve dynamic coverage of the monitoring range and blind spot filling.

[0092] In terms of data transmission, a combination of wireless transmission (5G+LoRa dual-mode) and wired transmission is adopted. DFOS and micro-vibration sensor data are transmitted to the edge computing node via fiber optic wired transmission. TBM tunneling parameters are uploaded in real time through the device bus, and IoT robot data is transmitted through the 5G network to ensure that the data transmission latency is ≤100ms. The edge node is equipped with redundant storage modules to prevent data loss.

[0093] S2. Multi-source data preprocessing and spatiotemporal fusion:

[0094] In this step, the collected multi-source data is preprocessed and spatiotemporally fused to obtain spatiotemporally fused data.

[0095] In one exemplary implementation, the edge computing device preprocesses the collected data, including outlier removal (based on the 3σ criterion), noise filtering (wavelet thresholding is used for DFOS data and adaptive notch filtering is used for microseismic data), and data standardization (normalizing data of different dimensions to the [0,1] interval) to reduce data redundancy and interference.

[0096] Multi-source data fusion: An improved Kalman filter algorithm is adopted, and an adaptive weighting factor is introduced to perform spatiotemporal fusion on the preprocessed DFOS strain data, microseismic acoustic data, TBM tunneling parameters and robot inspection data. Data with different timestamps and spatial locations are uniformly mapped to the three-dimensional tunnel coordinate system to form a standardized spatiotemporal dataset with a data fusion error of ≤2%.

[0097] S3. Predicting the location and probability of progressive rock failure based on a CNN-LSTM attention hybrid model:

[0098] In this step, based on the pre-trained CNN-LSTM attention hybrid model, the spatiotemporal fusion data is used as input to predict the location and probability of progressive rock failure.

[0099] In one exemplary implementation, the CNN-LSTM attention hybrid model structure includes a feature extraction layer, a temporal analysis layer, an attention layer, and an output layer. The feature extraction layer uses a CNN network (containing 3 convolutional layers and 2 pooling layers) to extract spatial features (such as stress concentration areas and crack propagation directions) from the fused data. The temporal analysis layer uses an LSTM network (containing 2 hidden layers, each with 128 neurons) to capture the temporal evolution of the data. The attention layer, through a self-attention mechanism, focuses on weak signals in high-risk areas (such as sudden changes in microseismic signal amplitude and abnormal strain accumulation), strengthening the weights of key features. The output layer outputs the three-dimensional coordinates (x, y, z) of the failure location and the failure probability (0-100%).

[0100] Model Training and Optimization: Historical monitoring data (including rockburst and collapse case data) and indoor simulation test data (simulating the surrounding rock failure process under high ground stress) from more than 30 ultra-long and ultra-deep TBM tunnel projects at home and abroad were collected to form the training dataset. The dataset size is ≥100,000 sets. The Adam optimizer is used with adaptive adjustment of the learning rate (initial learning rate 0.001, decaying by 10% every 100 rounds) and overfitting is prevented through early stopping strategy. The model supports online learning, absorbing new monitoring data and engineering cases in real time, iteratively optimizing model parameters, and adapting to different geological conditions (such as granite, sandstone, shale, etc.).

[0101] Damage location localization: The fused spatiotemporal dataset is input into the trained model, and the model outputs the three-dimensional coordinates of the damage location with a localization accuracy of sub-meter level (error ≤ 0.8m); at the same time, it outputs the damage probability distribution, which clarifies the scope and development trend of high-risk areas and provides a basis for early warning classification.

[0102] S4. Conduct quantitative analysis of aging cracking characteristics:

[0103] In this step, time-dependent cracking characteristics are extracted based on spatiotemporal fusion data, time-dependent cracking quantitative indicators are constructed, and the assessment of the degree of time-dependent cracking of the surrounding rock and the prediction of its long-term evolution are completed.

[0104] In one exemplary implementation scheme, the methods for extracting time-dependent cracking characteristics, constructing quantitative indicators, assessing the degree of time-dependent cracking in surrounding rock, and predicting long-term evolution are as follows:

[0105] Cracking Feature Extraction: An ARIMA model combined with wavelet transform was used to perform time-series decomposition on the fused data, extracting core features of time-dependent cracking, including: crack propagation rate (…). ), cracking time lag effect (the time difference from stress loading to crack initiation), surrounding rock creep index (based on the Nishihara model fitting), and crack frequency (the number of new cracks per unit time).

[0106] Among the above characteristics, crack frequency and crack propagation rate together characterize the degree of cracking activity; cracking time lag effect helps to determine the type of disaster (short time lag indicates brittle failure, long time lag indicates creep deformation); the surrounding rock creep index is used to predict long-term deformation trend (>1.2 indicates accelerated creep).

[0107] Construction of the Time-Based Cracking Index (TCI): The TCI is defined as a quantitative indicator of the degree of time-based cracking. It comprehensively considers the surrounding rock strain rate, stress level, and crack development characteristics. To eliminate differences in the dimensions and numerical ranges of each parameter, the original monitoring data is first normalized and mapped to a unified scoring range of 0-10. Then, a weighted sum is calculated based on the mapped scores of each parameter. The calculation formula is as follows:

[0108] ;

[0109] in, , , , These are the weight coefficients predetermined using the analytic hierarchy process (AHP). Surrounding rock strain rate The mapping score, ; The maximum principal stress of the surrounding rock The mapping score, ; Crack frequency The mapping score, ; Crack propagation rate The mapping score, ;

[0110] In the calculation of the above-mentioned mapping fraction values, the surrounding rock strain rate Maximum principal stress of surrounding rock Crack frequency Crack propagation rate The base value (denominator) is taken respectively. , , , This is based on statistical data from measured ultra-deep TBM tunnel projects and typical engineering cases at home and abroad, covering extreme working conditions of age-induced cracking of surrounding rock under high ground stress.

[0111] The classification method for age-related cracking of surrounding rock is as follows: It is in a stable state; It is in a slightly cracked state; It is in a severely cracked state.

[0112] Assessment of the degree of aging cracking of surrounding rock: Based on the current value of the aging cracking index (TCI) and the above classification of the aging cracking level of surrounding rock, the current aging cracking level of surrounding rock is determined.

[0113] Long-term evolution prediction: Blockchain technology is used to record timely fission data. Each data block contains information such as timestamp, monitoring location, characteristic parameters, and TCI value, so as to achieve data immutability and traceability. Based on the historical data stored on the blockchain, the fission evolution path in the next 1-12 months can be predicted through time series prediction models, generating a visual heat map and marking high-risk fission areas at different time points.

[0114] S5. Tiered early warning and linkage with TBM execution parameter adjustment and emergency response:

[0115] In this step, graded early warnings are issued based on the probability of damage and time-dependent cracking quantification indicators, and the TBM execution parameters are adjusted and emergency response is carried out in conjunction with these measures. The monitoring and early warning results are then output through a visual interface.

[0116] In one exemplary implementation scheme, the methods for tiered early warning, intelligent feedback adjustment, and visualized decision-making are as follows:

[0117] Tiered early warning system: Three levels of early warning are set based on the probability of damage and the TCI value.

[0118] Level 1 Warning (Blue): Probability of damage 60%-80% or This suggests increasing the monitoring frequency;

[0119] Level II Warning (Yellow): Probability of damage 80%-90% or The system triggers an on-site inspection command, and the IoT robot focuses on the warning area to strengthen the inspection, while controlling the TBM to reduce the tunneling speed and reduce the thrust.

[0120] Level 3 Warning (Red): Probability of damage ≥ 90% or The emergency alarm will be triggered immediately, the TBM will be shut down immediately, the emergency support procedure will be initiated, and on-site personnel will be notified to evacuate.

[0121] Intelligent feedback adjustment: After the alarm is triggered, the system automatically feeds back the warning information and adjustment suggestions to the TBM control system. In the case of a first-level warning, the tunneling parameters are maintained and the sensor sampling frequency is optimized. In the case of a second-level warning, the tunneling speed is reduced (to 50% of the original speed) and the thrust is reduced. In the case of a third-level warning, the TBM is shut down and the emergency support process is initiated.

[0122] AR Visualization Decision Making: Integrating an augmented reality (AR) interface, field engineers can view real-time monitoring data, 3D models of the damage location, cracking thermal maps, and evolution prediction results through mobile terminals (tablets, AR devices on safety helmets), gaining an intuitive understanding of the surrounding rock condition; the AR interface supports gesture interaction, allowing for zooming in to view local details, assisting engineers in quickly formulating treatment plans and reducing human error.

[0123] Example:

[0124] This embodiment uses a TBM tunnel project in granite strata with a burial depth of 2500m and a length of 60km as an application scenario. The method of this invention is used to monitor the location of progressive failure and time-dependent cracking characteristics of the surrounding rock. The specific implementation steps are as follows:

[0125] Multimodal sensor network deployment: Slots are pre-embedded in the tunnel wall every 0.5m along the tunnel axis to deploy distributed fiber optic sensors (Φ-OTDR type, strain accuracy 1με). Fiber optics are radially implanted through boreholes to a depth of 8m (tunnel radius 5m). A set of micro-seismic sensors (three-component, frequency response 1-1000Hz) is deployed every 5m in front of the tunnel face and on the tunnel sidewall, for a total of 60 sets. The existing propulsion and torque sensors of the TBM are integrated, and vibration acceleration sensors are added. Six wearable IoT micro-robots are configured to autonomously inspect the tunnel sidewall blind spots and tunnel face corners.

[0126] Real-time data acquisition and fusion: 5G+LoRa dual-mode data transmission is adopted. DFOS data and microseismic data are transmitted to edge computing nodes via fiber optic cable. TBM parameters are uploaded via device bus, and robot data is transmitted via 5G. Edge nodes use the 3σ criterion to remove outliers, wavelet thresholding for denoising DFOS data, and adaptive notch filtering for microseismic data. The sampling frequency of the multi-modal sensors is adjusted based on strain rate, and initial surrounding rock stability (…) ), sampling frequency 1Hz; when stress concentration occurs in front of the working face ( The sampling frequency was increased to 10Hz; an improved Kalman filter algorithm was used to fuse multi-source data to form a three-dimensional spatiotemporal dataset.

[0127] Progressive damage location prediction: Import the trained CNN-LSTM attention model (training dataset contains 200,000 sets of historical data), input the fused data, and the model outputs the three-dimensional coordinates of the damage location (x=1250m, y=2.3m, z=4.8m), with a damage probability of 82% and a positioning error of 0.6m; the model absorbs the monitoring data through online learning and optimizes the parameters to adapt to the characteristics of the granite strata.

[0128] Quantitative analysis of cracking characteristics over time: The ARIMA model combined with wavelet transform was used to extract features, and the crack propagation rate was found to be 0.3 mm / d, the time delay effect was 24 h, and the creep index was 0.8.

[0129] Based on DFOS data, strain values ​​at adjacent times from the same monitoring location are selected for differential calculation to obtain the strain change per unit time, thus obtaining the current surrounding rock strain rate. After calculation By combining the rock mass elastic parameters with the initial in-situ stress conditions, the maximum principal stress of the surrounding rock was obtained through elasticity inversion. .

[0130] By capturing microseismic signals generated by the initiation and propagation of microcracks in the surrounding rock in real time using a microseismic sensor array, and counting the number of effective microseismic events per unit time, the crack frequency is obtained. .

[0131] Then the strain rate of the surrounding rock can be calculated. Mapping score Maximum principal stress of surrounding rock Mapping score Crack frequency Mapping score Crack propagation rate Mapping score ; and then through Calculate the TCI value ( , , , Based on the Analytic Hierarchy Process (AHP), the values ​​are 0.35, 0.25, 0.25, and 0.15 respectively. According to the preset judgment rules, it is judged to be in a state of slight cracking. By recording data through blockchain, the cracking evolution path in the next 7 days is predicted, and heat maps are generated to mark high-risk areas.

[0132] Early warning and feedback: Due to TCI=4.345, but the CNN-LSTM attention model predicted a failure probability of 82%, a level 2 warning (yellow) was triggered. The system automatically fed back to the TBM control system, reducing the tunneling speed from 3m / h to 1.5m / h and decreasing the thrust by 30%. On-site engineers viewed the 3D model and heat map of the failure location through an AR tablet and instructed the IoT robot to focus on the warning area and strengthen inspections, submitting an inspection report every 2 hours. In subsequent monitoring, the crack propagation rate dropped to 0.08mm / d, the failure probability dropped to 75%, the level 2 warning was lifted, and normal tunneling parameters were restored.

[0133] Although embodiments of the present invention have been described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the present invention, and all such changes and alterations shall not depart from the protection scope of the present invention.

Claims

1. A method for monitoring progressive failure and time-related cracking of surrounding rock in ultra-long and ultra-deep TBM tunnels, characterized in that, Includes the following steps: S1. Construct a multimodal sensor network that combines fixed deployment with dynamic inspection to collect surrounding rock strain data, microseismic data, TBM tunneling parameters, and inspection data of blind spots in fixed sensor monitoring; S2. Preprocess and spatiotemporally fuse the collected multi-source data to obtain spatiotemporally fused data; S3. Based on the pre-trained CNN-LSTM attention hybrid model, using the spatiotemporal fusion data as input, predict the location and probability of progressive failure of the surrounding rock; S4. Extract time-dependent cracking characteristics based on spatiotemporal fusion data, construct time-dependent cracking quantitative indicators, and complete the assessment of the degree of time-dependent cracking of surrounding rock and the prediction of its long-term evolution. S5. Based on the probability of damage and the quantitative indicators of time-dependent cracking, a graded early warning is issued, which is linked to the TBM execution parameter adjustment and emergency response, and the monitoring and early warning results are output through a visual interface; In step S4, the method for extracting time-dependent cracking features based on spatiotemporal fusion data includes: The ARIMA model combined with wavelet transform is used to perform temporal decomposition and feature extraction on the spatiotemporal fusion data to obtain the time-dependent cracking characteristics of the surrounding rock. The time-dependent cracking characteristics include crack propagation rate, cracking lag, creep index and crack frequency. The method for constructing the time-sensitive cracking quantitative index includes: Based on the extracted time-dependent cracking characteristics, the Time-dependent Cracking Index (TCI) is constructed as a quantitative indicator of the degree of time-dependent cracking in the surrounding rock. The aging cracking index is calculated using a weighted summation method, and the calculation formula is as follows: ； in, , , , These are the weight coefficients predetermined using the analytic hierarchy process (AHP). Surrounding rock strain rate The mapping score, ; The maximum principal stress of the surrounding rock The mapping score, ; Crack frequency The mapping score, ; Crack propagation rate The mapping score, ; The age-related cracking levels of the surrounding rock are classified according to the TCI numerical range: Corresponding to a stable state; Corresponding to a slightly cracked state; This corresponds to a severely cracked state.

2. The method for monitoring progressive failure and time-related cracking of surrounding rock in ultra-long and ultra-deep TBM tunnels as described in claim 1, characterized in that, In step S1, the construction of a multimodal sensor network combining fixed deployment and dynamic inspection includes: Distributed fiber optic sensors are deployed along the tunnel axis and radial direction to collect surrounding rock strain data; A microseismic sensor array is arranged in front of the tunnel face and on the tunnel sidewall to collect microseismic data generated by the propagation of microcracks in the surrounding rock. Parameter sensors are integrated into the TBM cutterhead, propulsion cylinder, and torque sensor to collect TBM tunneling parameters; By deploying autonomous, mobile IoT micro-robots, inspection data can be collected from blind spots in fixed sensor monitoring.

3. The method for monitoring progressive failure and time-related cracking of surrounding rock in ultra-long and ultra-deep TBM tunnels as described in claim 2, characterized in that, Step S1 also includes: The surrounding rock strain rate is calculated based on the surrounding rock strain data acquired by distributed fiber optic sensors. And based on the surrounding rock strain rate Dynamically adjust the sampling frequency of the multimodal sensor network: when At that time, the sampling frequency is maintained at 1Hz; when At that time, the sampling frequency was increased to 10Hz; when At that time, the sampling frequency was adjusted to 100Hz.

4. The method for monitoring progressive failure and time-related cracking of surrounding rock in ultra-long and ultra-deep TBM tunnels as described in claim 1, characterized in that, In step S2, the preprocessing includes: outlier removal based on the 3σ criterion, noise filtering of the data, and data standardization. The spatiotemporal fusion includes: using an improved Kalman filter algorithm, by introducing an adaptive weighting factor, to perform spatiotemporal fusion on the preprocessed surrounding rock strain data, microseismic data, TBM tunneling parameters and inspection data, and to uniformly map data with different timestamps and different spatial locations to the three-dimensional tunnel coordinate system.

5. The method for monitoring progressive failure and time-related cracking of surrounding rock in ultra-long and ultra-deep TBM tunnels as described in claim 1, characterized in that, In step S3, the CNN-LSTM attention hybrid model includes: The CNN feature extraction layer is configured to perform spatial feature extraction on spatiotemporal fusion data and output a spatial feature map. The LSTM temporal analysis layer is configured to take the spatial feature map as input, extract temporal evolution features, and output a temporal feature sequence. The self-attention layer is configured to assign feature weights to the temporal feature sequence and output weighted fusion features. The output layer is configured to output the three-dimensional coordinates and failure probability of the progressive failure location of the surrounding rock based on the weighted fusion features.

6. The method for monitoring progressive failure and time-related cracking of surrounding rock in ultra-long and ultra-deep TBM tunnels as described in claim 5, characterized in that, The pre-training methods for CNN-LSTM attention hybrid models include: A training dataset was constructed by collecting historical monitoring data, geological disaster case data, and indoor simulation test data of ultra-long and ultra-deep TBM tunnel projects. The training dataset is input into the CNN-LSTM attention hybrid model, and the prediction results are output after spatial feature extraction, temporal feature extraction and attention weighting. Based on the predicted results and the loss calculated from the actual labels, an adaptive optimizer is used to iteratively optimize the model parameters until the model accuracy meets the requirements.

7. The method for monitoring progressive failure and time-related cracking of surrounding rock in ultra-long and ultra-deep TBM tunnels as described in claim 1, characterized in that, In step S4, the methods for assessing the degree of age-related cracking of the surrounding rock and predicting its long-term evolution include: The current age-related cracking level of the surrounding rock is determined based on the current age-related cracking index (TCI) value. Blockchain technology is used to record time-sensitive data, with each data block containing a timestamp, monitoring location, feature parameters, and TCI value information; Based on historical data stored on the blockchain, a time series prediction model is used to extrapolate the time-dependent cracking trend of the surrounding rock in the next 1–12 months, and obtain the time-dependent cracking evolution path. A visual heat map is generated by combining the cracking evolution path, marking high-risk cracking areas at different time periods.

8. The method for monitoring progressive failure and time-related cracking of surrounding rock in ultra-long and ultra-deep TBM tunnels as described in claim 1, characterized in that, In step S5, the method of classifying and issuing early warnings based on the probability of damage and the quantitative index of time-dependent cracking, and linking the TBM execution parameters and emergency response, includes: When the probability of failure is between 60% and 80% or the time-dependent cracking index (TCI) meets the requirements... When this occurs, it is determined to be a Level 1 warning, the original tunneling parameters of the TBM are maintained, and the monitoring sampling frequency is increased; When the probability of failure is between 80% and 90% or the time-dependent cracking index (TCI) meets the requirements... When the situation is deemed a Level II warning, the TBM is controlled to reduce its tunneling speed and thrust. When the probability of failure is ≥90% or the time-dependent cracking index (TCI) meets the requirements... When the situation is assessed as a Level 3 warning, the TBM is immediately shut down, and the emergency support procedure is initiated.

Images