Tunnel rock-soil deformation identification method and system

Abstract

The invention discloses a tunnel rock-soil deformation identification method and system, relates to the technical field of tunnel and underground engineering monitoring and early warning, and remarkably improves the reliability and timeliness of tunnel weak surrounding rock deformation risk identification. In the abrupt change response link, a time domain convolutional network and a displacement acceleration self-adaptive strengthening mechanism are introduced, so that the capturing capability of the system on nonlinear displacement surge is improved in a leap-over manner; according to the method, the acceleration component is focused and the physical constraint weight is applied, so that the sudden change event is locked in real time, and accidents such as unstability of the tunnel face are prevented from being missed in report; in order to solve the problem of point cloud registration distortion of a joint development area, the geological law of a rock mass structural plane is fused into a registration algorithm, and overall deformation misjudgment caused by rock mass local dislocation is inhibited from the source through a Gaussian mixture model coupling normal elastic deformation threshold and a tangential coulomb friction criterion; the measurement precision of laser scanning in a fracture zone and a fault zone breaks through environmental limitation, and a reliable basis is provided for supporting decision.

Description

Technical Field

[0001] The present invention relates to the technical field of tunnel and underground engineering monitoring and early warning, and in particular to a tunnel rock and soil deformation identification method and system. Background Art

[0002] In tunnel engineering, deformation monitoring of weak surrounding rock (such as mudstone and fractured zones) is directly related to construction safety. Such rock masses are prone to nonlinear mutations after being disturbed by excavation, such as face instability or vault collapse. Current projects generally rely on real-time monitoring systems (such as 3D laser scanning and fiber optic sensing) and intelligent early warning models to achieve dynamic identification of deformation risks.

[0003] Existing identification schemes are mainly based on two types of technologies. One is to use long short-term memory networks (LSTMs) to analyze displacement sensor data and predict surrounding rock deformation trends. The other is to calculate rock mass shape variables through multi-period laser scanning point cloud registration algorithms, such as ICP. However, in weak surrounding rock scenarios, the LSTM model relies on the continuity of historical data and has a delayed response to sudden deformations with a single-day displacement surge of more than 200%. Point cloud registration is easily disturbed by local rock block dislocation in rock masses with developed joints, leading to misjudgment of the overall deformation. Although fiber optic monitoring systems can capture microstrain, they cannot associate them with the physical yield mechanism of rock and soil to predict critical risks.

[0004] Recently, some solutions have optimized data input or integrated multi-source sensors, such as increasing the laser scanning frequency to improve the point cloud density, or combining vibrating string sensors to assist in judging rock loosening. Although such solutions can alleviate the noise interference of a single data source, they do not solve the core problem. The extraction of temporal features of mutation events and the analysis of non-rigid deformation of jointed rock masses still rely on a general algorithm framework, which is difficult to adapt to the complex mechanical behavior of weak surrounding rock. Attempts such as incremental learning and transfer learning for mutation omissions have not yet been implemented, but the false alarm rate remains high because they do not embed the physical boundary conditions of rock and soil failure. Summary of the Invention

[0005] In view of the above existing problems, the present invention is proposed.

[0006] The present invention provides a tunnel rock and soil deformation identification method and system to solve the problem of high underreporting rate of soft surrounding rock deformation caused by delayed mutation response, distorted joint registration and missing microstrain analysis in existing solutions.

[0007] In order to solve the above technical problems, the present invention provides the following technical solutions: In a first aspect, an embodiment of the present invention provides a method for identifying tunnel rock and soil deformation, which includes: Step S1: Acquire a multi-period 3D point cloud of the tunnel cross section using a multispectral laser scanner, acquire settlement / convergence time series data using a laser displacement meter, and acquire microstrain time series data using a distributed optical fiber; time-align all data using a unified clock, and complete coordinate unification and outlier removal; Step S2: Input the displacement time series into the time domain convolutional network to obtain the displacement acceleration features including the second-order time difference; Step S3, based on the displacement acceleration amplitude and noise scale estimation within the window, the adaptive enhancement weight is calculated to obtain the mutation feature enhancement sequence; Step S4, performing normal estimation and plane segmentation on the three-dimensional point cloud, and determining the main direction of the joint surface in combination with on-site structural surface mapping / survey line statistics; using rock mass quality indicators including RQD / Jv as parameter inputs of constraint strength or prior confidence; In step S5, a non-rigid point cloud registration based on a Gaussian mixture model is used to register and solve the multi-period point cloud. A joint constraint term is introduced into the energy function so that the deformation field is dominated by tangential slip along the main direction of the joint, while satisfying: The normal deformation does not exceed the threshold determined by the elastic parameters; The amount of tangential slip is limited by the Coulomb friction criterion; Step S6, inputting the mutation feature enhancement sequence, the point cloud deformation obtained based on registration, and the optical fiber microstrain time series into a fusion decision module; Step S7, analyzing the state transition of the microstrain feature through a hidden Markov state machine, and outputting a risk score or a critical risk signal; Step S8: Compare the risk score or critical risk signal with the yield critical threshold calculated based on the Mohr-Coulomb parameter, trigger a graded warning, and output the result to the construction management system.

[0008] As a preferred solution of the tunnel rock and soil deformation identification method described in the present invention, the multispectral laser scanner simultaneously emits near-infrared and visible light bands and records the echo intensity, and performs dust point identification and point cloud completion based on the band intensity ratio and spatial neighborhood consistency.

[0009] As a preferred solution of the tunnel rock and soil deformation identification method described in the present invention, the adaptive enhancement weight is calculated based on the absolute amplitude of the window second-order time difference and the robust noise scale, and time continuity constraints are performed to suppress isolated abnormal points. The robust noise scale includes the median absolute deviation.

[0010] As a preferred solution of the tunnel rock and soil deformation identification method described in the present invention, the main direction of the joint is determined by the main plane set obtained by point cloud normal clustering and the strike / dip statistics of on-site structural surface mapping, and RQD is used to adjust the normal constraint threshold or constraint weight.

[0011] As a preferred solution of the tunnel rock and soil deformation identification method described in the present invention, the non-rigid point cloud registration adopts collaborative point drift or a non-rigid method based on a Gaussian mixture model, and adds a normal elastic threshold term and a tangential Coulomb friction term to the objective function to constrain the normal deformation and tangential slip of the joint surface.

[0012] As a preferred solution of the tunnel rock and soil deformation identification method described in the present invention, the hidden Markov state machine includes three types of hidden states: elasticity, damage evolution, and instability. The observed quantities are the amplitude, growth rate, and frequency domain energy of microstrain, and the state transition probability can be updated online.

[0013] As a preferred solution of the tunnel rock and soil deformation identification method described in the present invention, the graded warning includes at least three levels, and each level corresponds to a different response suggestion and construction control parameter range.

[0014] In a second aspect, the present invention provides a tunnel rock and soil deformation identification system, comprising: Multi-source acquisition unit, used to obtain multi-spectral 3D point clouds, displacement time series and optical fiber micro-strain time series, and complete time synchronization; Timing analysis module, used for runtime convolutional networks and mutation feature enhancement calculations; Point cloud processing module, used to perform non-rigid registration based on Gaussian mixture model with joint constraints and output deformation; Fusion decision module, with built-in hidden Markov state machine and physical critical interface, used to output risk score or critical risk signal; The early warning and visualization unit is used to dynamically mark and display the graded early warning results and joint slip zones in the BIM / construction management model.

[0015] As a preferred solution of the tunnel rock and soil deformation identification system described in the present invention, the physical critical interface receives the yield critical threshold calculated by the Mohr-Coulomb parameters obtained by indoor experiments or numerical inversion, and supports online update.

[0016] As a preferred solution of the tunnel rock and soil deformation identification system described in the present invention, the early warning and visualization unit is configured to: when a high risk level is reached, generate suggestions for grouting reinforcement range and process parameters based on a preset rule library for manual confirmation, and highlight the displacement mutation area and joint sliding zone in the BIM model.

[0017] The beneficial effects of the present invention are as follows: the present invention significantly improves the reliability and timeliness of deformation risk identification of weak surrounding rock in tunnels. In the mutation response link, the time-domain convolutional network and the displacement acceleration adaptive enhancement mechanism are introduced to greatly improve the system's ability to capture the surge in nonlinear displacement. The present invention ensures that mutation events are locked in real time by focusing on acceleration components and applying physical constraint weights, thereby avoiding missed reports of accidents such as face instability. In order to solve the problem of point cloud registration distortion in joint-developed areas, the geological laws of the rock structure surface are integrated into the registration algorithm, and the Gaussian mixture model is coupled with the normal elastic deformation threshold and the tangential Coulomb friction criterion to suppress the misjudgment of overall deformation caused by local dislocation of rock blocks from the root. The measurement accuracy of laser scanning in broken zones and fault zones breaks through environmental limitations, providing a reliable basis for support decisions.

[0018] In terms of identifying hidden risks, this solution transcends the static threshold limitations of traditional fiber-optic monitoring and constructs a geotechnical damage evolution chain based on a hidden Markov state machine. By mapping microstrain characteristics to a three-state transition model of elasticity, damage, and instability, the system can capture the cumulative effects of continuous creep from seemingly smooth data streams, enabling risk prediction hours in advance. This approach requires no additional hardware; it unlocks the value of microstrain data through algorithmic reconstruction alone.

[0019] Furthermore, the entire process utilizes a multi-source data fusion architecture: The outputs of the time series analysis module and the point cloud processing module, after physical logic arbitration, drive a graded early warning system to dynamically mark high-risk areas within the BIM model and automatically generate process parameter recommendations such as grouting range and anchor density. This not only reduces reliance on expert experience but also reduces construction response time to minutes, providing all-weather safety assurance for tunnel construction in complex geological conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following is a brief introduction to the drawings required for use in the embodiments. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope of the present application.

[0021] Figure 1 Schematic diagram of the process of the tunnel rock and soil deformation identification method of the present invention.

[0022] Figure 2 Schematic diagram of the framework of the tunnel rock and soil deformation identification system of the present invention. DETAILED DESCRIPTION

[0023] In order to make the purpose, technical solutions and advantages of this application more clearly understood, the present application is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0024] All terms (including technical and scientific terms) used in this application have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein should be interpreted as having a meaning consistent with the context of this specification and should not be interpreted in an idealized or overly rigid manner.

[0025] For example, the terms "first," "second," etc. used in this application are only used to distinguish and describe similar objects, to differentiate a first object from another object, rather than to describe a specific order or sequence, and should not be understood as indicating or implying relative importance.

[0026] This application proposes a tunnel rock and soil deformation identification method, combined with Figure 1 and Figure 2 As shown, the method includes: Step S1: Acquire a multi-period 3D point cloud of the tunnel cross section using a multispectral laser scanner, acquire settlement / convergence time series data using a laser displacement meter, and acquire microstrain time series data using a distributed optical fiber; time-align all data using a unified clock, and complete coordinate unification and outlier removal; Specifically, clock unification means that acquisition devices record timestamps using the same timing source and verify relative drift. Coordinate unification refers to establishing external parameter correspondences for laser scanners, displacement meters, and fiber optic sensors based on the construction reference coordinate system. Outlier rejection robustly removes dust noise, reflection saturation points, and isolated outlier periods. Furthermore, the point cloud to project coordinates conversion uses on-site control points or fixed targets to determine pose. The installation mileage and orientation of the displacement meters and fiber optic lines are based on construction measurement records and mapped to a table. Regarding numerical caliber, the default scanning period is 15 minutes and can be adjusted from 10 to 30 minutes; the default sampling period for the displacement meter is 10 seconds and can be adjusted from 1 to 60 seconds; the default sampling frequency for the fiber optic line is 20 Hz and can be adjusted from 5 to 50 Hz. The allowed residual error for time alignment is ≤200 ms by default and can be adjusted to ≤500 ms, based on the cross-correlation peak width of the commissioning data and the timing accuracy of the equipment. Optionally, PPS hardware pulses or network time synchronization can be used for timing. Coordinate unification can be performed once for initial deployment and once after equipment relocation. Necessary exception and boundary processing includes: when a data stream is missing for more than 5 consecutive sampling cycles, the minimum executable caliber is to linearly extrapolate its latest available fragment and mark it as "degraded fusion" while maintaining normal processing of the remaining data streams.

[0027] Step S2: Input the displacement time series into the time domain convolutional network to obtain the displacement acceleration features including the second-order time difference; Specifically, the time-domain convolutional network (TCN) is a one-dimensional convolutional architecture that takes as input a sequence of chronologically ordered displacement scalars and outputs multi-scale time series features, including first-order and second-order differences. Second-order time differences are computed within the network using a fixed difference kernel or during preprocessing and used as channel input. Similarly, the receptive field is determined by a combination of dilated convolution and layer stacking, covering the time span from short-term perturbations to intraday trends. Regarding numerical caliber, the feature extraction window length defaults to 10 minutes and is adjustable from 2 to 30 minutes; the number of convolutional layers defaults to 3 to 5 with a stride of 1; and the inference sliding stride defaults to 1 sample point. These settings are based on a trade-off between recall of sudden transitions and false positives for steady-state segments on the validation set. Alternatively, in scenarios with limited computational resources, an equivalent one-dimensional residual convolutional network can be substituted, or a lightweight implementation using only a fixed difference kernel and sliding statistics can be used, preserving the dimensionality and order of the output features. When the input sequence length is shorter than the minimum window length, it degenerates to directly output finite difference features and marks them as "degraded features" Step S3, based on the displacement acceleration amplitude and noise scale estimation within the window, the adaptive enhancement weight is calculated to obtain the mutation feature enhancement sequence; Step S4: normal estimation and plane segmentation are performed on the 3D point cloud, and the main direction of the joint surface is determined by combining on-site structural surface mapping / survey line statistics; rock mass quality indicators including RQD / Jv are used as parameter inputs of constraint strength or prior confidence; Similarly, normal estimation involves obtaining the normal direction by least-squares plane fitting of a local neighborhood point set. Plane segmentation involves extracting a set of approximately coplanar points using region growing or random consistency strategies. In-situ structural surface mapping / survey line statistics involves using strike, dip, and spacing information recorded at the tunnel or excavation face as orientation priors. Optionally, the neighborhood selection radius defaults to 0.2-0.5 m or is adaptive based on the number of points per neighborhood (20-50). The plane fitting residual threshold defaults to 1-2 times the point spacing. The orientation consistency threshold defaults to a range of 10-20°, set based on point cloud density and in-situ joint roughness. Regarding parameter interpretation, when RQD and Jv are used as weights or threshold adjustment factors, their values ​​are derived from core records and field survey line statistics. Their purpose is to moderately relax normal constraints or increase prior uncertainty in areas of poor quality. The necessary anomaly and boundary processing is as follows: when the local point density is insufficient or dust obscuration leads to normal instability, the main direction of the joint is output only in the high confidence area and the results of the previous period are used in the low confidence area to maintain time series continuity.

[0028] In step S5, a non-rigid point cloud registration based on a Gaussian mixture model is used to register and solve the multi-period point cloud. A joint constraint term is introduced into the energy function so that the deformation field is dominated by tangential slip along the main direction of the joint, while satisfying: The normal deformation does not exceed the threshold determined by the elastic parameters; The amount of tangential slip is limited by the Coulomb friction criterion; In step S6, the mutation feature enhancement sequence, the point cloud deformation obtained based on registration, and the fiber microstrain time series are input into the fusion decision module. Specifically, the input vectors of the fusion decision module are aligned by timestamp and then dimensionally normalized and checked for abnormal duty cycle. When a single source is missing, fusion continues in a weighted attenuation manner to maintain output stability. Optionally, the input update cycle defaults to 5 minutes and can be adjusted from 1 to 10 minutes. The time alignment tolerance defaults to 1 sampling step. The setting is based on the construction response timeliness and computing resource constraints. Regarding parameter interpretation, the setting of the fusion threshold or weight is derived from the distribution fitting of historical samples or expert rules and can be dynamically adjusted according to the construction stage. When multiple sources are missing simultaneously for more than one update cycle, the output enters the "hold" state and records the alarm "data cannot be judged" to avoid false triggering.

[0029] In step S7, the state transition of the microstrain characteristics is analyzed through a hidden Markov state machine, and a risk score or critical risk signal is output. Furthermore, the observation quantity of the hidden Markov state machine includes a normalized combination of the microstrain amplitude, slope and frequency band energy. The initial state probability and the transition matrix are obtained by statistically analyzing the samples of the historical stable segment and the unstable pre-segment, and are fine-tuned online with a sliding window during operation. Similarly, the minimum dwell time is used to suppress frequent state switching, and the default is 2-5 observation steps. Smoothing is achieved by two equivalent methods: forward-backward probability averaging or Viterbi path backtracking. When the proportion of missing observations exceeds 30%, the transfer update is suspended and only the posterior most likely state of the previous moment is output, and a one-time reassessment is performed after recovery.

[0030] Step S8: Compare the risk score or critical risk signal with the yield critical threshold calculated based on the Mohr-Coulomb parameter, trigger a graded warning, and output the result to the construction management system; Specifically, the thresholds for graded warnings are derived from a combination of indoor triaxial test parameters and on-site inversion results and are reviewed by experts. Risk levels and response recommendations are issued via a data interface within the construction management system. Optionally, to avoid jitter, a time hysteresis mechanism and event merging rules are used: the maintenance of the same risk level must meet a minimum duration, and two triggers are merged into one event when the interval is less than the set window. The necessary boundary processing is: when the Mohr-Coulomb parameter interface is temporarily unavailable, a temporary comparison is performed according to the most recent valid threshold, and after recovery, it is automatically rechecked to see whether an upgrade or downgrade is required.

[0031] In one embodiment, a multispectral laser scanner simultaneously emits near-infrared and visible light bands and records the echo intensity, and performs dust point identification and point cloud completion based on the band intensity ratio and spatial neighborhood consistency; In one embodiment, the adaptive enhancement weight is calculated based on the absolute amplitude of the windowed second-order time difference and a robust noise scale, and a temporal continuity constraint is performed to suppress isolated outliers. The robust noise scale includes the median absolute deviation. Specifically, in step S3, the adaptive reinforcement weight calculation process is as follows: At each sampling moment In the sliding window, the second-order time difference from step S2 is used as the mutation sensitivity, and the robust noise scale is used as the normalization benchmark. The standardized significance is first calculated and mapped to the weight. Then, the time continuity and morphological constraints are imposed on the weight sequence. Finally, the weight is multiplied back by the second-order time difference to obtain the mutation feature enhancement sequence for use in the subsequent fusion decision module. The calculation formula includes: The dimensionless adaptive reinforcement weight is obtained by using a monotone S-shaped mapping: , in, Indicates time The adaptive reinforcement weight is dimensionless and has a value of , is the mapping slope coefficient, dimensionless, used to adjust the transition steepness, For the moment The second-order time difference amplitude has the same dimension as the displacement, is the robust noise scale within the window, and its dimension is consistent with the displacement. To prevent the denominator from approaching zero, the dimension is consistent with the displacement and the value is taken as the acquisition range. times, is the significance center threshold, dimensionless, used to control the reporting level; In the formula, the second-order time difference adopts the central difference approximation: , in, Same as above. 、 、 The displacement timing is respectively 、 、 The value of is consistent with the displacement dimension. The robust noise scale is estimated using MAD: , , in, is the Gaussian approximation calibration constant , For is the median operation of the index set, For The sliding window index set centered on Display window The index variable in the same domain as i, is the first-order difference of displacement, and its dimension is consistent with that of displacement. For the displacement timing at index The value of , the dimension is consistent with the displacement; In getting After obtaining the original sequence, a first-order recursive time continuity constraint is first applied to suppress isolated spikes, and then grayscale morphological closing-opening processing is performed to fill in gaps and filter out short responses, resulting in a smooth and structurally consistent weight sequence. This sequence is then point-by-point multiplied by the second-order time difference to form a mutation feature enhancement sequence, which enters the subsequent fusion judgment and early warning link. Specifically, this step uses the second-order time difference as the mutation sensitivity, and constructs dimensionless significance in conjunction with the robust noise scale. A monotonic S-shaped mapping is then used to obtain weights between zero and one, thereby maintaining comparability under different noise levels and dimensional conditions. Temporal continuity processing allows the response to appear naturally adhered on the time axis, avoiding overreaction to occasional spikes. Morphological closing-opening further suppresses fragmented short segments and connects real events cut off by noise, making the obtained weight interval closer to the actual duration of surrounding rock deformation. The dot product of the weight and the second-order time difference outputs an enhanced sequence, whose amplitude and direction simultaneously retain the mutation information and change sign, facilitating temporal alignment and thresholding with the displacement field increments and fiber microstrain sequences obtained by registration. At the parameter level, the slope and center threshold dominate the sensitivity and alarm level, small positive numbers stabilize numerical calculations, and the robust scale derived from the MAD is insensitive to outliers and adapts to complex working conditions such as dust and vibration. Overall, this step constitutes a closed loop from detection to calculation to weight adjustment, providing stable and available mutation feature input for fusion judgment. For example, the sliding window refers to a set of symmetric or half-open interval indices around the current sampling point, the morphological closed-open processing is performed with a structural element of fixed sample length, and the time continuity constraint is implemented in a first-order recursive manner to ensure that the weight transitions smoothly within the event segment. Furthermore, the default window span is 3-10 minutes and adapts to the sampling period, the default length of the morphological structural element is 3-7 sampling points, the default minimum event duration is 2-5 minutes, and the default update period is consistent with the inference stride. The above values ​​are based on the statistical distribution of offline annotated mutation fragments and the false alarm control target. Regarding the interpretation of the formula parameters, the mapping slope coefficient and the center threshold are selected based on the ROC optimal point of the validation set or the fixed false positive rate principle. The small positive number gives the order of magnitude according to the sensor quantization bit and full scale. The robust noise scale is estimated by the median absolute deviation of the first-order difference in the window and quantile correction is used when the event has a high duty cycle. Optionally, the monotone S-shaped mapping can be replaced by a piecewise linear approximation when resources are limited without changing the weight normalization property; the morphological order can also be switched between closed-open and open-closed according to the false alarm preference. Necessary anomaly and boundary processing includes: using mirror filling or nearest neighbor preservation when the window is at the edge of the sequence or there are scattered missing measurements; when the robust scale estimate approaches zero, enabling an alternative scale based on quantile difference and limiting the rate of change of the upper bound of the weight.

[0032] In one embodiment, the main direction of the joint is determined by combining the main plane set obtained by point cloud normal clustering and the strike / dip statistics of on-site structural surface mapping, and RQD is used to adjust the normal constraint threshold or constraint weight.

[0033] In one embodiment, the non-rigid registration adopts a cooperative point drift or a non-rigid method based on a Gaussian mixture model, and adds a normal elastic threshold term and a tangential Coulomb friction term to the objective function to constrain the normal deformation and tangential slip of the joint surface.

[0034] Specifically, in step S5, during the registration and solution of multi-phase point clouds, GMM / CPD is used as the backbone for data consistency, and physical constraints are imposed on the displacement field in the main direction of the joint: the normal component is limited by the elastic threshold, and the tangential component satisfies the displacement equivalent of the Coulomb criterion. The pose and non-rigid deformation coefficients are jointly solved through a unified objective function to obtain a physically interpretable deformation field, which is then output to the subsequent fusion judgment. The solution process formula is as follows: , in, is the total energy, is the GMM / CPD data item, is a non-rigid transformation, and are the set of points to be registered and the set of reference points, N and M are the number of points to be registered and the number of reference points, is the GMM variance term, the unit is the square of the displacement dimension, is the displacement field smoothing regularization coefficient, dimensionless, is a smoothing operator (such as the thin plate spline Laplacian matrix), is the kernel coefficient splicing vector, are the normal and tangential constraint weights, are the normal and tangential constraint energies, respectively; Where: , , in, is the kernel function, A set of control points The control points, is the number of control points, is the Gaussian kernel width, For the The displacement coefficient vector of the control points; , Among them, the sum index Corresponding points Number, is the two-norm; Where: Define displacement increment , define the joint set The unit normal is , obtained from step S4; let each point be mapped to the nearest joint set index , and the distance threshold Forming a constrained index set , The point-to-surface distance threshold from a point to its joint surface, in meters, is calibrated on-site according to the minimum distance threshold from a point to a joint surface; The normal threshold is derived using elastic parameters and monitoring span and combined with RQD adjustment: , , in, is the normal displacement threshold, is the unadjusted baseline threshold, is the RQD adjustment coefficient, is the rock mass quality index, 0-100, is the geometric coefficient, dimensionless, To monitor span length, is Poisson's ratio, is the elastic modulus, is the characteristic normal stress scale, unit: Pa; The normal constraint energy uses a hinge penalty: , in, for point The component of displacement in the joint normal direction; The tangential component is given by the projection matrix have to: , , in, is a 3×3 unit matrix (2×2 in the case of 2D cutting), is the tangential slip, is the normal compression displacement; contact stiffness is introduced Shift the Coulomb criterion: , in, Joint set The friction coefficient is dimensionless and can be set by the joint set. is the cohesive strength, unit is Pa, are the normal and tangential contact stiffness, respectively, in Pa / m; Weights and solution process: 1) Initialization and detection data: Use the data given in step S4 With RQD, monitoring span calculate , and set Baseline value, dimensionless, empirical range , calibrated with the validation set; 2) Formula calculation: In each outer iteration In the calculate , build , the GMM / CPD step adopts closed form or iterative update corresponding to the expectation, and solves and ; 3) Parameter adjustment: If or If the corresponding violation ratio increases, fine-tune the linear strategy. or moderate relaxation / tightening (by and RQD upper and lower bounds), until the residual is stable and the output is terminated and displacement field ; Specifically, this step superimposes joint-guided physical constraints within the probabilistic registration framework, ensuring that the resulting deformation field not only fits point cloud observations but also complies with the rock mass's force and structural surface motion characteristics. Data terms are modeled through mixed Gaussian pair correspondences, regularization terms restrict violent oscillations in the displacement field, and normal constraints are adaptively relaxed or tightened based on elastic modulus, Poisson's ratio, and the displacement threshold of the monitoring span, combined with rock mass quality indicators, to avoid unreasonable out-of-plane compression or expansion. The tangential constraint translates the Coulomb friction criterion into a displacement-domain inequality, introduces contact stiffness and cohesion, and sets separate slip tolerances for different joint sets. During the optimization process, constraint violations are fed back into subtle adjustments to weights and thresholds, forming a closed loop of observation-calculation-parameter adjustment. This design reduces the non-physical deformations that occur in crack-rich areas due to pure geometric alignment. The obtained displacement is decomposed into normal and tangential components, which can be directly used for subsequent fusion and early warning. In this embodiment, the control point set is selected on the reference point cloud by voxel downsampling or uniform grid, and the kernel width and smoothing operator are matched accordingly to take into account both details and overall continuity. The iteration stop condition is defined by the energy drop threshold and the maximum number of external iterations. Furthermore, the control point spacing defaults to 0.5-2.0 m, the maximum external iteration defaults to 30-80 times, and the energy drop threshold defaults to a relatively low 1e-4-1e-3; these settings are based on a compromise between offline alignment error and convergence speed. Regarding the interpretation of the formula parameters, the normal and tangential constraint weights are determined by weighting the loss function of the normal limit and tangential misjudgment in the validation set. The elastic modulus and Poisson's ratio in the normal threshold are derived from indoor tests or on-site tests. The monitoring span is taken from the construction section scale. The friction coefficient and cohesion are derived from joint shear tests or inversion. The normal and tangential contact stiffness are estimated based on the contact model and the structural surface roughness. Optionally, in cases of severe occlusion or sharp deformation gradients, multi-resolution registration (coarse-to-fine) or block-based solving can be employed to maintain solution stability and an unchanged output interface. Necessary exception and boundary handling includes: When data terms and constraint terms alternately dominate, resulting in oscillation and non-convergence, temporarily reducing constraint weights and limiting the tightening rate of the normal threshold; if the outer iteration reaches the upper limit but still fails to meet the stopping condition, the current optimal solution is output and marked as "registration degraded."

[0035] In one embodiment, the hidden Markov state machine includes three types of hidden states: elasticity, damage evolution, and instability. The observed quantities are the amplitude, growth rate, and frequency domain energy of microstrain, and the state transition probability can be updated online.

[0036] In one embodiment, the graded warning includes at least three levels, and each level corresponds to a different response suggestion and construction control parameter range.

[0037] This application also proposes a tunnel surrounding rock deformation identification system, including: Multi-source acquisition unit, used to obtain multi-spectral 3D point clouds, displacement time series and optical fiber micro-strain time series, and complete time synchronization; Timing analysis module, used for runtime convolutional networks and mutation feature enhancement calculations; Point cloud processing module, used to perform non-rigid registration based on Gaussian mixture model with joint constraints and output deformation; Fusion decision module, with built-in hidden Markov state machine and physical critical interface, used to output risk score or critical risk signal; Early warning and visualization unit, used to dynamically mark and display graded early warning results and joint slip zones in the BIM / construction management model; Specifically, the system receives the risk level, main joint direction, and location attributes of the tangential slip zone through a unified data interface of engineering coordinates, and expresses them in the construction management model as sections, mileage, and arch positions. Furthermore, the default data display refresh cycle is 5 minutes, adjustable from 1 to 10 minutes, and the single rendering delay does not exceed 30 seconds, based on the construction scheduling response requirements. Optionally, when computing resources are limited, it can be downgraded to outputting only text and section numbers and omitting continuous surface rendering, keeping the linkage logic unchanged. When the model interface is unavailable, the system locally caches the most recent results and retransmits them in batches after the interface is restored.

[0038] In one embodiment, the physical critical interface receives the yield critical threshold calculated from the Mohr-Coulomb parameters obtained by indoor experiments or numerical inversion, and supports online updates. Similarly, the online update of the physical critical interface refers to pulling the Mohr-Coulomb parameters or inversion results from the parameter management library at a set period and performing consistency verification. If the parameter change exceeds the set amplitude, the comparison threshold is adjusted synchronously and the version number is recorded. Optionally, the update cycle defaults to 24 hours and is adjustable from 2 to 72 hours. The verification rules are determined based on the historical parameter fluctuation range and the upper limit of the experimental error. When the new parameters fail to pass the verification or the interface times out, the previous version of the parameters will continue to be used and an internally visible alarm will be triggered to remind manual review.

[0039] In one embodiment, the early warning and visualization unit is configured to: when a high risk level is reached, generate suggestions for grouting reinforcement scope and process parameters based on a preset rule base for manual confirmation, and highlight displacement mutation areas and joint slip zones in the BIM model; For example, visualization of displacement mutation zones and joint slip zones uses segment highlighting and text annotation consistent with project coordinates. Color and transparency are mapped according to risk level, and clicking on the corresponding time window and key values ​​is displayed. Furthermore, event browsing and export functions are provided with a default retention period of 30 days, and the export format and fields comply with the data specifications of the construction management system. Optionally, during low-load periods at night, the system can perform batch replay of historical events for review without changing online operating parameters. If three consecutive rendering failures occur or the rendering time exceeds the threshold, the system automatically downgrades to displaying only the segment number and risk level text and recording the performance event.

[0040] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some or all of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present application.

[0041] Furthermore, those skilled in the art will appreciate that although some embodiments herein include certain features included in other embodiments but not others, combinations of features from different embodiments are intended to be within the scope of this application and to form different embodiments. For example, all of the above embodiments may be used in any combination. The information disclosed in this background section is intended solely to enhance understanding of the overall background of this application and should not be construed as an admission or any form of implication that such information constitutes prior art known to those skilled in the art.

Claims

1. A method for identifying tunnel rock and soil deformation, characterized in that: include, Step S1: Acquire a multi-period 3D point cloud of the tunnel cross section using a multispectral laser scanner, acquire settlement / convergence time series data using a laser displacement meter, and acquire microstrain time series data using a distributed optical fiber; time-align all data using a unified clock, and complete coordinate unification and outlier removal; Step S2: Input the displacement time series into the time domain convolutional network to obtain the displacement acceleration features including the second-order time difference; Step S3, based on the displacement acceleration amplitude and noise scale estimation within the window, the adaptive enhancement weight is calculated to obtain the mutation feature enhancement sequence; Step S4, performing normal estimation and plane segmentation on the three-dimensional point cloud, and determining the main direction of the joint surface in combination with on-site structural surface mapping / survey line statistics; using rock mass quality indicators including RQD / Jv as parameter inputs of constraint strength or prior confidence; In step S5, a non-rigid point cloud registration based on a Gaussian mixture model is used to register and solve the multi-period point cloud. A joint constraint term is introduced into the energy function so that the deformation field is dominated by tangential slip along the main direction of the joint, while satisfying: The normal deformation does not exceed the threshold determined by the elastic parameters; The amount of tangential slip is limited by the Coulomb friction criterion; Step S6, inputting the mutation feature enhancement sequence, the point cloud deformation obtained based on registration, and the optical fiber microstrain time series into a fusion decision module; Step S7, analyzing the state transition of the microstrain feature through a hidden Markov state machine, and outputting a risk score or a critical risk signal; Step S8: Compare the risk score or critical risk signal with the yield critical threshold calculated based on the Mohr-Coulomb parameter, trigger a graded warning, and output the result to the construction management system.

2. A tunnel rock and soil deformation identification method according to claim 1, characterized in that: The multispectral laser scanner simultaneously emits near-infrared and visible light bands and records the echo intensity, and performs dust point discrimination and point cloud completion based on the band intensity ratio and spatial neighborhood consistency.

3. The method for identifying tunnel rock and soil deformation according to claim 1, wherein: The adaptive enhancement weight is calculated based on the absolute amplitude of the windowed second-order time difference and a robust noise scale, and a time continuity constraint is performed to suppress isolated outliers. The robust noise scale includes the median absolute deviation.

4. The method for identifying tunnel rock and soil deformation according to claim 1, wherein: The main direction of the joint is determined by the main plane set obtained by point cloud normal clustering and the trend / dip statistics of on-site structural surface mapping, and RQD is used to adjust the normal constraint threshold or constraint weight.

5. The method for identifying tunnel rock and soil deformation according to claim 1, wherein: The non-rigid point cloud registration adopts collaborative point drift or a non-rigid method based on a Gaussian mixture model, and adds a normal elastic threshold term and a tangential Coulomb friction term to the objective function to constrain the normal deformation and tangential slip of the joint surface.

6. The method for identifying tunnel rock and soil deformation according to claim 1, wherein: The hidden Markov state machine contains three types of hidden states: elasticity, damage evolution, and instability. The observed quantities are the amplitude, growth rate, and frequency domain energy of microstrain, and the state transition probability can be updated online.

7. The method for identifying tunnel rock and soil deformation according to claim 1, wherein: The hierarchical warning includes at least three levels, and each level corresponds to different response suggestions and construction control parameter ranges.

8. A tunnel rock and soil deformation identification system, based on a tunnel rock and soil deformation identification method according to any one of claims 1 to 7, characterized in that: include: Multi-source acquisition unit, used to obtain multi-spectral 3D point clouds, displacement time series and optical fiber micro-strain time series, and complete time synchronization; Timing analysis module, used for runtime convolutional networks and mutation feature enhancement calculations; Point cloud processing module, used to perform non-rigid registration based on Gaussian mixture model with joint constraints and output deformation; Fusion decision module, with built-in hidden Markov state machine and physical critical interface, used to output risk score or critical risk signal; The early warning and visualization unit is used to dynamically mark and display the graded early warning results and joint slip zones in the BIM / construction management model.

9. The tunnel rock and soil deformation identification system according to claim 8, characterized in that: The physical critical interface receives the yield critical threshold value calculated from the Mohr-Coulomb parameters obtained from indoor experiments or numerical inversion, and supports online update.

10. The tunnel rock and soil deformation identification system according to claim 8, characterized in that: The early warning and visualization unit is configured to generate recommendations on grouting reinforcement scope and process parameters based on a preset rule base for manual confirmation when a high risk level is reached, and highlight displacement mutation areas and joint sliding zones in the BIM model.

Images