Tunnel excavation footage and support parameter optimization method based on machine learning

Abstract

The application discloses a tunnel excavation footage and support parameter optimization method based on machine learning, relates to the technical field of tunnel excavation, and comprises the following steps: establishing a multi-scale stress monitoring baseline in a deep tunnel construction environment, continuously collecting surrounding rock low-frequency stress fluctuation data by using a high-precision sensing array, calculating surrounding rock energy distribution based on the collected data, and obtaining a dynamic energy distribution factor for representing dynamic energy changes of the surrounding rock; and constructing a force field disturbance evolution mechanism based on the dynamic energy distribution factor, performing time sequence expansion on the dynamic energy distribution factor, and extracting a potential trigger time window of the surrounding rock low-frequency stress mutation. Through energy feature migration and counterfactual closed-loop regulation, the application realizes dynamic perception of the surrounding rock energy field and self-adaptive optimization of the support parameters, improves the discrimination accuracy and real-time response capability of the model under low-frequency mutation, and guarantees the stability and intelligent level of the tunnel construction.

Description

Technical Field

[0001] This invention relates to the field of tunnel excavation technology, and more specifically, to a method for optimizing tunnel excavation advance and support parameters based on machine learning. Background Technology

[0002] Machine learning-based optimization of tunnel excavation advance and support parameters refers to the use of machine learning models to mine and model a large amount of historical engineering data, geological parameters, surrounding rock deformation monitoring data, and construction feedback information during tunnel construction. This establishes a nonlinear mapping relationship between excavation advance and support parameters. Traditional methods typically rely on engineers' experience or static design specifications to determine the excavation length and support strength, but under complex geological conditions, this can easily lead to engineering problems such as over-excavation, insufficient support, or excessive support. In contrast, machine learning methods can learn implicit patterns from data, enabling dynamic prediction of surrounding rock stability and real-time assessment of construction risks. This allows for the automatic recommendation of optimal excavation advance and appropriate support methods during construction, such as shotcrete thickness, anchor bolt spacing, and steel frame strength. This method not only helps improve the safety and economy of tunnel construction but also continuously self-corrects and optimizes based on real-time monitoring data, forming a closed-loop control mechanism of prediction, feedback, and adjustment, thereby significantly improving the intelligence level of the construction process.

[0003] However, existing technologies still have certain shortcomings. Current machine learning-based methods for optimizing tunnel excavation progress and support parameters largely rely on the continuous stability of surrounding rock monitoring data for feature extraction and model extrapolation. However, in deep tunnel construction, low-frequency abrupt changes caused by sudden releases of regional ground stress are frequently encountered. These abrupt changes typically manifest as a slow but significant dynamic adjustment of the overall energy field of the surrounding rock within a short period. Due to the strong concealment and obvious nonlinear superposition characteristics of these low-frequency disturbances, existing models struggle to effectively cover such scenarios during training, causing their original feature weights to quickly become invalid when stress abrupt changes occur, leading to a sharp decline in feature discrimination accuracy. More seriously, machine learning models struggle to capture changes in the stability state of the surrounding rock in a timely manner under low-frequency abrupt changes, resulting in a significant lag in the output of support parameters. This leads to a mismatch between the support structure and the actual stress state of the surrounding rock, potentially triggering large-scale rockfalls or even cascading collapses, severely threatening construction safety and overall structural stability.

[0004] The information disclosed in the background section is intended to further enhance the understanding of the background of this disclosure, and therefore may contain some content that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] To address the aforementioned technical problems in related technologies, this invention proposes a machine learning-based method for optimizing tunnel excavation progress and support parameters, which can overcome the shortcomings of existing technologies.

[0006] To achieve the above-mentioned technical objectives, the technical solution of the present invention is implemented as follows:

[0007] A machine learning-based method for optimizing tunnel excavation progress and support parameters;

[0008] The machine learning-based method for optimizing tunnel excavation progress and support parameters includes the following steps:

[0009] Step 1: Establish a multi-scale stress monitoring baseline in the deep tunnel construction environment, continuously collect low-frequency stress fluctuation data of the surrounding rock using a high-precision sensor array, and calculate the energy distribution of the surrounding rock based on the collected data to obtain the dynamic energy distribution factor.

[0010] Step 2: Construct a force field disturbance calculation mechanism based on the dynamic energy distribution factor, perform time-series expansion on the dynamic energy distribution factor, extract the potential triggering time window of low-frequency stress change in the surrounding rock, and generate a characteristic transition map based on the energy change trajectory within the triggering time window;

[0011] Step 3: Train a non-stationary feature transfer network based on the feature transition map, and guide the dynamic redistribution of feature weights within the non-stationary feature transfer network through the energy patterns in the feature transition map to obtain a model with adaptively adjusted feature weights;

[0012] Step 4: Based on the model after adaptive adjustment of feature weights, construct a counterfactual replay chain, replay the historical low-frequency mutation trajectories in the feature transition map using counterfactual methods, compare and analyze the replay feature response results with historical collapse event samples, and generate a correction instruction stream with causal consistency.

[0013] Step 5: Based on the correction command stream, construct a closed-loop support optimization and control unit, use the correction command stream to dynamically correct the support parameters output in real time, and synchronize the correction results to the non-stationary feature migration network through the parameter write-back mechanism to realize closed-loop adaptive optimization between support parameters and dynamic energy distribution of surrounding rock.

[0014] Furthermore, step 1 specifically includes:

[0015] Based on the tunnel burial depth, geological structure and distribution of surrounding rock joints, the surrounding rock space is divided into three characteristic sections: the main control stress zone, the secondary control stress zone and the local disturbance zone. High-precision stress sensor arrays are arranged in each section to form the multi-scale stress monitoring baseline.

[0016] The high-precision stress sensor array is used to continuously collect low-frequency stress fluctuation data of the surrounding rock, and the sampling error is corrected by a baseline drift self-correction mechanism.

[0017] Based on the corrected low-frequency fluctuation data, the energy distribution of the surrounding rock is calculated using mechanical energy field theory and wave energy spectrum analysis method, and a three-dimensional energy distribution field is established.

[0018] The dynamic energy distribution factor is constructed based on the temporal and spatial gradient variation characteristics of the three-dimensional energy distribution field.

[0019] Furthermore, the high-precision stress sensor array includes a combination structure of a triaxial stress sensor and a distributed fiber optic strain sensor; and the baseline drift self-correction mechanism uses the average stress value in a short-period stable range to correct the sampled data.

[0020] Furthermore, step 2 specifically includes:

[0021] Based on the dynamic energy distribution factor, a force field disturbance response model for surrounding rock is established. The force field disturbance response matrix is ​​calculated by coupling the stress-strain energy density relationship with the energy release rate.

[0022] The dynamic energy distribution factor is subjected to time-series expansion, and the time-frequency response surface of energy fluctuation is constructed using multi-scale time window analysis and Hilbert transform.

[0023] Based on the time series unfolding results, the energy change rate is calculated and combined with the force field disturbance response matrix to determine the spatial concentration region of energy anomalies, thereby extracting the potential triggering time window;

[0024] Energy change trajectory features are extracted within the potential triggering time window, and the feature transition spectrum is generated using principal component analysis and nonlinear embedding.

[0025] Further, generating the feature transition map includes:

[0026] Cluster analysis was performed on the energy change trajectories within the potential triggering time window, and they were divided into several energy transition clusters based on similarity.

[0027] Calculate the center trajectory of each energy transition cluster;

[0028] The central trajectories are connected sequentially in the time series to form a continuous feature transition map.

[0029] Furthermore, step 3 specifically includes:

[0030] Based on the aforementioned feature transition map, a feature mapping set is constructed, and the energy density change rate, energy gradient direction, stress response sensitivity, and energy release rate are decomposed in multiple dimensions. The weight mapping function of the energy mode is obtained through the eigenvalue decomposition of the covariance matrix.

[0031] The weight mapping function is used to dynamically redistribute the features within the non-stationary feature transfer network, and the feature weight gradient is corrected according to the fluctuation trend of the energy pattern.

[0032] The dynamically redistributed feature set is input into the non-stationary feature transfer network, and a feature alignment relationship is established between stable geological samples and low-frequency mutation samples through a transfer learning strategy.

[0033] Furthermore, when dynamically redistributing the internal features of the non-stationary feature migration network, a force field perturbation regularization term is introduced to constrain the spatial continuity of the energy pattern.

[0034] Furthermore, step 4 specifically includes:

[0035] The characteristic transition map is subjected to time-series inversion, and the energy transition segment with the most drastic change in energy gradient is selected as the counterfactual replay sample to reconstruct the historical energy evolution path;

[0036] A time series analysis is performed on the characteristic response results generated by the model during the counterfactual replay process. The phase difference and correlation coefficient between the model output and the energy trajectory are calculated to form a model response deviation distribution map.

[0037] The model response deviation distribution map is compared with historical collapse event samples to establish a causal consistency matrix between model behavior and geological events;

[0038] The correction instruction stream is generated based on the causal consistency matrix to synchronously correct the model feature weights and support parameter outputs.

[0039] Furthermore, when generating the correction command stream, the adjustment parameters of the correction command include the feature weight correction coefficient, the time delay compensation amount, and the support parameter adjustment factor; and the correction amplitude is adaptively adjusted through the dynamic gain function corresponding to the energy change rate.

[0040] Furthermore, step 5 specifically includes:

[0041] Based on the correction command flow, a parameter correction matrix is ​​established, and the feature weight correction coefficient, time delay compensation amount and support adjustment factor are mapped to the support parameters output by the model. The shotcrete thickness, anchor bolt spacing, steel arch stiffness and lock foot anchorage depth are dynamically corrected step by step according to the gain factor of the correction command flow.

[0042] The corrected support parameters are verified by energy response. The correction results are input into the surrounding rock energy distribution calculation framework to calculate the energy release rate and stress concentration factor, so as to judge the correction effect and optimize the parameter correction matrix.

[0043] The energy-verified support parameter correction values, energy response change rate, and dynamic energy distribution factor are encoded as feature feedback vectors and input into the non-stationary feature transfer network. The model parameters are dynamically synchronized through the parameter write-back mechanism.

[0044] The stability of the closed-loop support optimization and control process is evaluated, and the construction process is kept stable in the low-frequency sudden change scenario through multidimensional statistical analysis and force field inertial constraints.

[0045] The beneficial effects of this invention are as follows: By establishing a multi-scale stress monitoring and dynamic energy distribution factor construction mechanism, high-precision dynamic perception of the non-stationary evolution of the surrounding rock energy field is achieved. This enables the machine learning model to adaptively adjust its internal weights based on the feature transition spectrum to respond to low-frequency abrupt changes. Furthermore, by combining counterfactual playback and causal correction mechanisms, precise support parameter correction instructions are generated. Finally, through a closed-loop optimization and control unit, real-time matching and adaptive optimization of support parameters and the dynamic energy distribution of the surrounding rock are achieved, significantly improving the stability safety margin and overall intelligent control level of tunnel construction under complex geological conditions. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0047] Figure 1 This is a flowchart of a method for optimizing tunnel excavation progress and support parameters based on machine learning, according to an embodiment of the present invention. Detailed Implementation

[0048] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the description of this disclosure will be more complete and fully convey the concept of the exemplary embodiments to those skilled in the art.

[0049] This invention provides, for example Figure 1 The machine learning-based method for optimizing tunnel excavation progress and support parameters includes the following steps:

[0050] Step 1: Establish a multi-scale stress monitoring baseline in the deep tunnel construction environment, continuously collect low-frequency stress fluctuation data of the surrounding rock using a high-precision sensor array, and calculate the energy distribution of the surrounding rock based on the collected data to obtain the dynamic energy distribution factor.

[0051] Step 2: Construct a force field disturbance calculation mechanism based on the dynamic energy distribution factor, perform time-series expansion on the dynamic energy distribution factor, extract the potential triggering time window of low-frequency stress change in the surrounding rock, and generate a characteristic transition map based on the energy change trajectory within the triggering time window;

[0052] Step 3: Train a non-stationary feature transfer network based on the feature transition map, and guide the dynamic redistribution of feature weights within the non-stationary feature transfer network through the energy patterns in the feature transition map to obtain a model with adaptively adjusted feature weights;

[0053] Step 4: Based on the model after adaptive adjustment of feature weights, construct a counterfactual replay chain, replay the historical low-frequency mutation trajectories in the feature transition map using counterfactual methods, compare and analyze the replay feature response results with historical collapse event samples, and generate a correction instruction stream with causal consistency.

[0054] Step 5: Based on the correction command stream, construct a closed-loop support optimization and control unit, use the correction command stream to dynamically correct the support parameters output in real time, and synchronize the correction results to the non-stationary feature migration network through the parameter write-back mechanism to realize closed-loop adaptive optimization between support parameters and dynamic energy distribution of surrounding rock.

[0055] According to the machine learning-based tunnel excavation advance and support parameter optimization method of the present invention, in a specific embodiment, step 1 specifically includes:

[0056] Based on the tunnel burial depth, geological structure and distribution of surrounding rock joints, the surrounding rock space is divided into three characteristic sections: the main control stress zone, the secondary control stress zone and the local disturbance zone. High-precision stress sensor arrays are arranged in each section to form the multi-scale stress monitoring baseline.

[0057] The high-precision stress sensor array is used to continuously collect low-frequency stress fluctuation data of the surrounding rock, and the sampling error is corrected by a baseline drift self-correction mechanism.

[0058] Based on the corrected low-frequency fluctuation data, the energy distribution of the surrounding rock is calculated using mechanical energy field theory and wave energy spectrum analysis method, and a three-dimensional energy distribution field is established.

[0059] The dynamic energy distribution factor is constructed based on the temporal and spatial gradient variation characteristics of the three-dimensional energy distribution field.

[0060] In a specific embodiment of the machine learning-based tunnel excavation advance and support parameter optimization method of the present invention, the high-precision stress sensor array includes a combination structure of triaxial stress sensor and distributed fiber optic strain sensor; and the baseline drift self-correction mechanism uses the stress mean value of short-period stable interval to correct the sampled data.

[0061] According to the machine learning-based tunnel excavation advance and support parameter optimization method of the present invention, in a specific embodiment, step 2 specifically includes:

[0062] Based on the dynamic energy distribution factor, a force field disturbance response model for surrounding rock is established. The force field disturbance response matrix is ​​calculated by coupling the stress-strain energy density relationship with the energy release rate.

[0063] The dynamic energy distribution factor is subjected to time-series expansion, and the time-frequency response surface of energy fluctuation is constructed using multi-scale time window analysis and Hilbert transform.

[0064] Based on the time series unfolding results, the energy change rate is calculated and combined with the force field disturbance response matrix to determine the spatial concentration region of energy anomalies, thereby extracting the potential triggering time window;

[0065] Energy change trajectory features are extracted within the potential triggering time window, and the feature transition spectrum is generated using principal component analysis and nonlinear embedding.

[0066] According to the machine learning-based tunnel excavation advance and support parameter optimization method of the present invention, in a specific embodiment, generating the feature transition map includes:

[0067] Cluster analysis was performed on the energy change trajectories within the potential triggering time window, and they were divided into several energy transition clusters based on similarity.

[0068] Calculate the center trajectory of each energy transition cluster;

[0069] The central trajectories are connected sequentially in the time series to form a continuous feature transition map.

[0070] According to the machine learning-based tunnel excavation advance and support parameter optimization method of the present invention, in a specific embodiment, step 3 specifically includes:

[0071] Based on the aforementioned feature transition map, a feature mapping set is constructed, and the energy density change rate, energy gradient direction, stress response sensitivity, and energy release rate are decomposed in multiple dimensions. The weight mapping function of the energy mode is obtained through the eigenvalue decomposition of the covariance matrix.

[0072] The weight mapping function is used to dynamically redistribute the features within the non-stationary feature transfer network, and the feature weight gradient is corrected according to the fluctuation trend of the energy pattern.

[0073] The dynamically redistributed feature set is input into the non-stationary feature transfer network, and a feature alignment relationship is established between stable geological samples and low-frequency mutation samples through a transfer learning strategy.

[0074] In a specific embodiment of the machine learning-based tunnel excavation advance and support parameter optimization method of the present invention, when dynamically redistributing the internal features of the non-stationary feature migration network, a force field perturbation regularization term is introduced to constrain the spatial continuity of the energy pattern.

[0075] According to the machine learning-based tunnel excavation advance and support parameter optimization method of the present invention, in a specific embodiment, step 4 specifically includes:

[0076] The characteristic transition map is subjected to time-series inversion, and the energy transition segment with the most drastic change in energy gradient is selected as the counterfactual replay sample to reconstruct the historical energy evolution path;

[0077] A time series analysis is performed on the characteristic response results generated by the model during the counterfactual replay process. The phase difference and correlation coefficient between the model output and the energy trajectory are calculated to form a model response deviation distribution map.

[0078] The model response deviation distribution map is compared with historical collapse event samples to establish a causal consistency matrix between model behavior and geological events;

[0079] The correction instruction stream is generated based on the causal consistency matrix to synchronously correct the model feature weights and support parameter outputs.

[0080] According to the machine learning-based tunnel excavation advance and support parameter optimization method of the present invention, in a specific embodiment, when generating the correction command stream, the adjustment parameters of the correction command include feature weight correction coefficient, time delay compensation amount and support parameter adjustment factor; and the correction amplitude is adaptively adjusted by the dynamic gain function corresponding to the energy change rate.

[0081] According to the machine learning-based tunnel excavation advance and support parameter optimization method of the present invention, in a specific embodiment, step 5 specifically includes:

[0082] Based on the correction command flow, a parameter correction matrix is ​​established, and the feature weight correction coefficient, time delay compensation amount and support adjustment factor are mapped to the support parameters output by the model. The shotcrete thickness, anchor bolt spacing, steel arch stiffness and lock foot anchorage depth are dynamically corrected step by step according to the gain factor of the correction command flow.

[0083] The corrected support parameters are verified by energy response. The correction results are input into the surrounding rock energy distribution calculation framework to calculate the energy release rate and stress concentration factor, so as to judge the correction effect and optimize the parameter correction matrix.

[0084] The energy-verified support parameter correction values, energy response change rate, and dynamic energy distribution factor are encoded as feature feedback vectors and input into the non-stationary feature transfer network. The model parameters are dynamically synchronized through the parameter write-back mechanism.

[0085] The stability of the closed-loop support optimization and control process is evaluated, and the construction process is kept stable in the low-frequency sudden change scenario through multidimensional statistical analysis and force field inertial constraints.

[0086] To facilitate understanding of the above technical solutions of the present invention, the following detailed description of the above technical solutions of the present invention will be provided through specific embodiments.

[0087] In practical application, the tunnel excavation advance and support parameter optimization method based on machine learning according to the present invention includes the following steps:

[0088] In the deep tunnel construction environment, a multi-scale stress monitoring baseline is established. High-precision sensor arrays are used to continuously collect low-frequency stress fluctuation data of the surrounding rock. Based on the collected data, the energy distribution of the surrounding rock is calculated to obtain a dynamic energy distribution factor that characterizes the dynamic energy change of the surrounding rock, which serves as a benchmark for subsequent anomaly identification and model calibration.

[0089] The specific steps to achieve this process are as follows:

[0090] Establishing a multi-scale stress monitoring baseline in the deep tunnel construction environment is crucial for characterizing the stress evolution of the surrounding rock within different spatial levels and geological structural units. The establishment of this baseline involves classifying the spatial distribution characteristics of the tunnel's surrounding rock and non-uniformly distributing monitoring points. Specifically, based on tunnel depth, geological structure, surrounding rock integrity, and rock joint distribution, the surrounding rock space is divided into three characteristic zones: the primary stress zone, the secondary stress zone, and the local disturbance zone. Within each characteristic zone, a high-precision stress sensor array is deployed along the tunnel axis and radial direction. The spacing between sensors is dynamically calibrated based on rock wave impedance and energy attenuation patterns to ensure the comparability and continuity of the collected data across spatial scales. This deployment method enables the formation of a high- and low-frequency compatible stress sampling network across different spatial levels, thereby achieving high-resolution reconstruction of the overall stress field of the deep surrounding rock.

[0091] Based on the established multi-scale stress monitoring baseline, a high-precision sensor array is used to continuously acquire low-frequency stress fluctuation data of the surrounding rock to capture its true response under dynamic stress disturbance conditions. The high-precision sensor array is designed with long-term stability and measurement accuracy in deep, high-temperature, and high-pressure environments in mind. It achieves simultaneous monitoring of principal stress, shear stress, and volumetric strain of the surrounding rock by combining triaxial stress sensors with distributed fiber optic strain sensors. To ensure the continuity and anti-interference of the sampling results, a baseline drift self-correction mechanism is introduced during signal acquisition. This mechanism uses the average stress value within a short-period stable range as a real-time correction baseline to correct drift in subsequent sampling data, thereby eliminating offset errors caused by long-term sensor operation. Simultaneously, by performing multi-scale decomposition on the acquired signal, high-frequency noise components are separated from low-frequency effective fluctuations, retaining the low-frequency stress fluctuation trajectory that reflects the overall energy release characteristics of the surrounding rock, providing high-confidence raw input for subsequent energy distribution calculations.

[0092] Based on the collected low-frequency stress fluctuation data of the surrounding rock, the energy distribution of the surrounding rock is calculated to reveal the energy accumulation and release laws of deep surrounding rock under dynamic stress disturbance. This process is achieved by combining mechanical energy field theory with wave energy spectrum analysis. First, the collected low-frequency stress fluctuation signal is converted into a strain energy density time series curve. The total strain energy per unit volume of surrounding rock at any time is calculated using an integral form, thus obtaining the energy accumulation function over time. Then, the energy accumulation function is combined with the viscoelastic parameters of the surrounding rock medium to solve for the energy release rate distribution, thereby describing the spatiotemporal evolution characteristics of energy at different locations. To further improve the physical accuracy of the energy distribution calculation, the influence of the heterogeneity of the surrounding rock on wave propagation needs to be considered. By constructing a multi-point inversion model in the numerical domain, the energy distribution is corrected and smoothed so that the obtained energy distribution can truly reflect the intrinsic energy state of deep surrounding rock. After the above calculations, a high-resolution three-dimensional energy distribution field can be obtained throughout the monitoring range, providing a reliable data foundation for the extraction of dynamic energy factors.

[0093] After calculating the energy distribution of the surrounding rock, a dynamic energy distribution factor is extracted based on the temporal characteristics and spatial gradient variation patterns of energy changes at each monitoring location. This dynamic energy distribution factor serves as a benchmark for subsequent anomaly identification and model calibration. The dynamic energy distribution factor is obtained through normalization and feature reconstruction of the energy distribution field. Its core function is to reflect the dynamic migration trend of energy within the surrounding rock through the temporal fluctuation characteristics of the energy field. Specifically, firstly, continuous gradient information of the energy distribution field in the time dimension is extracted, and the energy release rate per unit time is calculated and used as the dynamic energy response coefficient. Secondly, combined with the energy density change rate in the spatial direction, an energy gradient vector field is constructed to describe the migration direction of energy from high-potential areas to low-potential areas. Thirdly, covariance analysis is performed on the coupling characteristics of the temporal and spatial gradients to form the feature matrix of the dynamic energy distribution factor, which comprehensively reflects the energy fluctuation intensity and stability of the surrounding rock. Through the above processing, the dynamic energy distribution factor can not only characterize the energy evolution state of the surrounding rock in real time, but also show significant response characteristics when low-frequency stress disturbances occur, thus providing a unified, stable and physically interpretable quantitative reference for subsequent low-frequency mutation identification and machine learning model calibration.

[0094] Through the above implementation steps, a complete process from multi-scale stress monitoring, low-frequency stress fluctuation acquisition, energy distribution calculation to dynamic energy distribution factor construction was achieved in the deep tunnel construction environment. The entire process is interdependent and progressive, which not only solves the problems of inconsistent data scales, feature distortion, and dynamic response lag in traditional stress monitoring, but also achieves high-precision dynamic characterization of the evolution law of the surrounding rock energy field, laying the physical and data foundation for subsequent low-frequency mutation identification and intelligent optimization of support parameters.

[0095] Based on the dynamic energy distribution factor, a force field disturbance calculation mechanism is constructed. The dynamic energy distribution factor is temporally expanded to extract the potential triggering time window of low-frequency stress change in the surrounding rock. Based on the energy change trajectory within the triggering window, a characteristic transition spectrum is generated to characterize the non-stationary change mode of the surrounding rock energy field.

[0096] The specific steps to achieve this process are as follows:

[0097] Based on the obtained dynamic energy distribution factor, a disturbance response model of the surrounding rock force field is established to reveal the correspondence between energy fluctuations and stress evolution. Specifically, multiple energy response points are first selected within the monitoring area. Based on the spatial and temporal gradients of the dynamic energy distribution factor, the surrounding rock is considered as a nonlinear viscoelastic continuous medium, and a force field response function is introduced to analyze the disturbance of energy distribution. The force field response function is coupled with the energy release rate through the stress-strain energy density relationship, thereby establishing a transfer equation between energy disturbance and stress field change in the time domain. By simulating small disturbances of the dynamic energy distribution factor at each monitoring point, a force field disturbance response matrix is ​​obtained. This matrix characterizes the energy absorption, conversion, and release characteristics of the surrounding rock under external disturbances. By solving and normalizing the disturbance response matrix, the response sensitivity distribution of the surrounding rock stress field at different time scales is obtained, providing dynamic constraint boundaries for subsequent time-series derivation.

[0098] Under the constructed force field perturbation calculation mechanism, the dynamic energy distribution factor is subjected to time-series expansion to reveal the non-stationary fluctuation characteristics in the energy evolution process. This process introduces a multi-scale time window analysis method to decompose the dynamic energy distribution factor into multiple overlapping intervals in the time dimension. Within each time interval, the first and second derivatives of the energy change rate are calculated to characterize the rate characteristics of energy accumulation and release. Simultaneously, by performing a Hilbert transform on the time series, the instantaneous energy phase and envelope curves are extracted, thereby constructing the time-frequency response surface of energy fluctuations. Subsequently, the energy phase difference and energy amplitude changes of adjacent time intervals are superimposed to form a time-series coupling map to describe the transmission and interference relationship of energy in the time dimension. By continuously integrating this time-series coupling map, the non-stationary evolution trajectory of the surrounding rock energy distribution in the time domain can be obtained, providing a basis for the temporal localization of potential abrupt events.

[0099] Based on time-series unfolded energy response data, this study extracts potential triggering time windows for low-frequency stress mutations in the surrounding rock to achieve a quantitative correlation between energy disturbances and mutation events. Specifically, firstly, the energy change rate curves within each time interval are calculated, and the variance of energy fluctuations is used as the mutation criterion. A dynamic threshold function is set to identify time periods with abnormally increased energy change rates. Secondly, after detecting abnormal energy intervals, the spatial concentration areas of energy mutations are located by combining the stress response sensitivity at the corresponding time in the force field disturbance response matrix, forming a time-space coupled mutation triggering window. To eliminate the interference of random noise on mutation identification, a weighted moving average method is used to smooth the energy change curves, and the start and end times of mutations are iteratively corrected using Kalman filtering, thus obtaining a physically stable and statistically significant potential triggering time window. The determination of this potential triggering time window not only reveals the critical point of energy release in the surrounding rock under low-frequency disturbances but also provides a time reference for the generation of characteristic transition maps.

[0100] After obtaining the potential triggering time window, a characteristic transition map is generated based on the energy change trajectory within the triggering window to characterize the non-stationary change mode of the surrounding rock energy field. Specifically, within the triggering time window, multidimensional features of the energy change trajectory are extracted, including the rate of change of energy density, energy gradient direction, energy release rate, and stress response sensitivity. These high-dimensional features are then mapped to a low-dimensional feature space using principal component analysis and nonlinear embedding techniques to form a visualized energy evolution trajectory within the feature space. Subsequently, cluster analysis is performed on the mapped energy trajectories, dividing them into several energy transition clusters based on similarity, and the central trajectory of each energy cluster is calculated. By sequentially connecting the central trajectories of the energy clusters in the time series, a continuous characteristic transition map is formed. This characteristic transition map reflects the evolution process of the surrounding rock energy field from a stationary state to a non-stationary state under low-frequency disturbances, where energy abrupt changes correspond to high-gradient transition regions in the map, while energy slow release corresponds to low-gradient smooth regions. By analyzing the feature transition spectrum, the energy transfer pattern, abrupt change intensity, and duration of the surrounding rock under specific geostress conditions can be identified, providing interpretable input for the feature redistribution and dynamic adaptation of subsequent machine learning models.

[0101] Through the above implementation steps, a force field perturbation calculation mechanism was constructed and applied based on the dynamic energy distribution factor. The entire process starts with the dynamic analysis of energy distribution, proceeds through force field response calculation, time series expansion, abrupt change window identification, and characteristic transition map generation, forming a complete chain of energy perturbation, response, and characteristic mapping. By introducing the force field perturbation calculation mechanism and the characteristic transition map generation process, not only can the formation mechanism of low-frequency abrupt changes be quantitatively revealed, but also the continuous tracking and spatial visualization of energy evolution characteristics under non-stationary conditions can be achieved.

[0102] The non-stationary feature transfer network is trained based on the feature transition graph. The energy patterns in the feature transition graph guide the dynamic redistribution of feature weights within the model. This enables the machine learning model to adaptively adjust feature weights when it detects low-frequency abrupt changes, thereby maintaining the model's real-time discrimination ability for surrounding rock stability and reducing prediction lag.

[0103] The specific steps to achieve this process are as follows:

[0104] Based on the generated feature transition map, a feature mapping set is constructed to extract energy pattern information during the feature transition process. The feature transition map contains the non-stationary evolution trajectory of the surrounding rock energy field in different time intervals, where energy abrupt changes correspond to high gradient regions and energy slow release corresponds to low gradient regions. To extract this spatiotemporal information, the feature transition map is first decomposed into multiple dimensions, separating quantitative indicators such as energy density change rate, energy gradient direction, stress response sensitivity, and energy release rate to form a joint feature vector set of time and space dimensions. Then, the joint feature vector set is normalized, and the covariance matrix between features is calculated to characterize the correlation between energy features. By decomposing the eigenvalues ​​of the covariance matrix, the contribution distribution of energy features under different perturbation states can be obtained, thus obtaining the weight mapping function of the energy pattern. This weight mapping function reflects the relative importance of different features in the stress abrupt change process, providing a basis for subsequent dynamic feature weight adjustment.

[0105] After constructing the feature mapping set, the feature weights within the machine learning model are dynamically redistributed based on the energy patterns of the feature transition map. Specifically, firstly, the input features in the model are reparameterized using a weight mapping function, that is, the weight factors of the feature inputs are dynamically adjusted according to the importance of each feature in the energy pattern, so that the features with the most significant energy changes receive higher gradient contributions during model training. To achieve temporal continuity in weight redistribution, a temporal smoothing constraint of the energy pattern is introduced, limiting the changes in feature weights within adjacent time windows to a set range, thereby avoiding model oscillations caused by weight fluctuations. Secondly, during training, the weight distribution is updated in real time using the backpropagation algorithm, and the weight gradient is corrected according to the fluctuation trend of the energy pattern, so that the model's feature response changes synchronously with the energy mutation process. When a sudden transition occurs in the energy pattern, the model weights will quickly converge to a new equilibrium state to ensure that the model still has discriminative stability under non-stationary energy fields. This dynamic feature weight redistribution process realizes the transformation of the machine learning model from static parameter mapping to dynamic adaptive learning, enabling the model to maintain high response sensitivity to energy disturbances caused by sudden changes in geostress.

[0106] After dynamically redistributing feature weights, energy patterns guide the training process of the non-stationary feature transfer network to achieve knowledge transfer and adaptive generalization under different geological conditions and stress disturbance environments. Specifically, firstly, the feature set after dynamic weight adjustment is input into the non-stationary feature transfer network, and a feature alignment relationship is established between the source and target domains through a transfer learning strategy. The source domain consists of energy distribution samples under stable geological conditions, and the target domain consists of energy disturbance samples under low-frequency abrupt change environments. By minimizing the dynamic difference function between the feature distributions of the two domains, the network can identify and transfer core feature patterns that remain stable under different disturbance conditions. Secondly, a force field perturbation regularization term is introduced during the transfer training process to constrain the spatial continuity of the energy patterns, ensuring that the transferred feature distribution is consistent with the real surrounding rock energy field. Through repeated training and parameter convergence, the network can form an adaptive mapping mechanism, that is, when low-frequency abrupt change features appear in the energy patterns, the network can automatically adjust the parameter structure of the feature layer and the decision layer, thereby quickly adapting to the new stress state. This energy pattern-based transfer learning method significantly enhances the robustness and adaptability of the model in complex construction environments, enabling it not only to cope with the sudden impact of ground stress release, but also to maintain stable predictive capabilities under multi-field coupling conditions.

[0107] After training the non-stationary feature transfer network, the model's real-time discrimination capability and lag response need to be verified and corrected. Specifically, by comparing the model's discrimination results under different energy perturbation intensities with actual monitoring data, the model's prediction error and response delay are calculated, and reverse calibration is performed based on the energy trajectory of the feature transition spectrum. When the model exhibits discrimination lag within a certain abrupt change interval, the feature weight update rate is adjusted according to the corresponding energy abrupt change slope in the spectrum to accelerate the model's adaptive response; when the model exhibits over-response, the error amplification effect is suppressed by reducing the weight gradient gain. Through the above closed-loop calibration process, the model achieves self-stabilization and self-adjustment during continuous learning, enabling the feature transfer network to dynamically maintain the optimal feature allocation state during long-term operation. Ultimately, when the network detects low-frequency abrupt change features, it can instantly adjust the internal feature weights, keeping the model's prediction results synchronized with the actual stability changes of the surrounding rock, thereby significantly reducing the deviation of support parameters and construction risks caused by abrupt change lag.

[0108] Through the above implementation steps, adaptive training and dynamic weight allocation of a non-stationary feature transfer network were achieved based on the feature transition map. The entire process uses energy patterns as the core logical thread, forming a complete energy-driven adaptive learning closed loop from feature extraction, weight reallocation, transfer learning to discrimination calibration. This implementation introduces an energy pattern-based feature transfer mechanism, enabling the machine learning model to possess self-learning and stable response capabilities under low-frequency stress mutation scenarios, thereby maintaining high-precision discrimination of surrounding rock stability in actual tunnel construction.

[0109] After completing the adaptive adjustment of feature weights, a counterfactual replay chain is constructed to replay the historical low-frequency mutation trajectories in the feature transition map. The feature response results obtained from the replay are compared and analyzed with historical collapse event samples to generate a correction instruction stream with causal consistency, which is used to correct the support parameter prediction results output by the model.

[0110] The specific steps to achieve this process are as follows:

[0111] Based on the model state after adaptive adjustment of feature weights, a time-series inversion of the feature transition map is performed to reconstruct the historical evolution path of the surrounding rock energy field. Specifically, firstly, representative energy transition segments in low-frequency abrupt change scenarios are selected as counterfactual playback samples, corresponding to the regions in the feature transition map where energy gradient changes most drastically. Then, the time series of these energy transition segments is reverse-ordered, allowing the model to receive the same energy pattern input as before the abrupt change, simulating the energy disturbance environment at that time. Through inversion processing, the model can regenerate the feature response sequence under this abrupt change environment, thereby achieving dynamic reconstruction of the historical stress state. In this process, to ensure the physical rationality of the energy inversion process, energy conservation constraints and temporal consistency constraints need to be introduced. That is, the total energy and time step are kept consistent during the reverse playback process to ensure that the energy trajectory received by the model remains equivalent to the original scene, thus avoiding deviation of the playback results from the actual physical process.

[0112] After replaying the energy trajectory counterfactually, the characteristic response results generated by the model during the replay process are extracted and analyzed temporally to reveal the model's response lag and discrimination bias during historical mutation processes. Specifically, firstly, the output response intensity and feature weight distribution of the model at each time step are calculated, and a time-aligned coordinate system is established with the energy mutation point as the center to synchronously map the model output with the actual energy change. Then, the response bias of the model at different stages is quantified by calculating the phase difference and correlation coefficient between the model response sequence and the original energy trajectory. When the phase difference between the response curve and the energy trajectory exceeds a preset threshold, it is determined that the model has a feature lag phenomenon in that period; when the correlation of the response curve decreases significantly, it indicates that the model has experienced feature mismatch at the mutation moment. By aggregating and analyzing these bias indicators, a response bias distribution map of the model under historical mutation scenarios can be formed to reveal the dynamic drift characteristics of the model's internal weights under mutation conditions.

[0113] Based on the comparative analysis between the model response deviation distribution map and real engineering event samples, a causal consistency mapping between model behavior and geological events is established. Specifically, firstly, historical collapse event samples corresponding to the feature transition maps are selected, including observational data such as surrounding rock collapse volume, peak stress of the support structure, strain rate, and energy release rate, as a validation reference set. Subsequently, the feature response results generated by the model's counterfactual playback are compared with the aforementioned historical samples over time. By calculating the deviation metric function between the model's predicted energy change curve and the actual collapse energy release curve, the causal deviation interval between the two is identified. Furthermore, to reveal the stability of the causal relationship, the coupling relationship between the change trend of the contrast metric function and the feature weight adjustment rate is compared to determine whether the model's response remains consistent before and after energy mutations. When the model exhibits similar response shift patterns in multiple mutation events, it indicates that the deviation has structural characteristics and can be systematically compensated through causal correction methods. The causal consistency matrix formed through comparative analysis can be used as the logical basis for generating correction command flows, providing an interpretable causal basis for subsequent support parameter correction.

[0114] Finally, based on the aforementioned causal consistency matrix, a correction command stream with causal consistency is generated and applied to correct the model's predicted output. Specifically, firstly, the adjustment parameters of the correction command are determined according to the direction and magnitude of the model response deviation, including the feature weight correction coefficient, time delay compensation, and support parameter adjustment factor. Then, these correction parameters are input into the model in time series form to synchronously correct the current feature weight distribution and predicted output. To ensure the stability of the correction command, a dynamic gain function based on the energy change rate is introduced, allowing the correction magnitude to adaptively adjust with the intensity of energy disturbance; that is, the correction rate is accelerated when the energy disturbance is large, and the correction magnitude is gradually reduced when the disturbance tends to stabilize. Subsequently, the correction results are evaluated by comparing the corrected model output with the latest monitoring data to verify the effectiveness of the correction. If the model output highly matches the actual support stress state, the correction command stream is solidified into a long-term memory path for the model to directly call in subsequent sudden change scenarios; if the deviation still exists, iterative correction continues through a counterfactual playback mechanism until the model's prediction results are consistent with the causal laws of historical events. Through the closed-loop correction process of the counterfactual replay chain and the correction instruction flow, the model not only realizes the causal reproduction of past abrupt events, but also has the ability to self-correct based on historical experience, thereby enabling it to proactively prevent support mismatch and risk accumulation in future construction scenarios.

[0115] Through the above implementation steps, based on the adaptive adjustment of feature weights, the construction of the counterfactual replay chain and the establishment of a causal correction mechanism were achieved. The entire process begins with the temporal inversion of energy trajectories, and through the reproduction analysis of model responses, the comparison and verification of causal consistency, and the dynamic generation of correction instruction streams, forms a highly interpretable self-correcting closed-loop learning system. By introducing the concept of counterfactual reasoning, the machine learning model can understand the causes of its own prediction bias in a physical sense, thereby achieving a fundamental shift from data-driven to causal-driven approaches.

[0116] A closed-loop support optimization and control unit is constructed based on the correction command stream. The correction command stream is used to dynamically correct the support parameters output in real time step by step. The correction results are synchronized to the non-stationary feature migration network through the parameter write-back mechanism, thereby realizing the closed-loop adaptive optimization between support parameters and dynamic energy distribution of surrounding rock to maintain the overall stability of tunnel construction process under low-frequency sudden change scenario.

[0117] The specific steps to achieve this process are as follows:

[0118] Based on the generated correction command stream, the real-time output support parameters are dynamically corrected step by step to ensure consistency between the support parameters and the transient response of the surrounding rock energy field. Specifically, firstly, the feature weight correction coefficients, time delay compensation, and support adjustment factors contained in the correction command stream are mapped one-to-one with the model's current prediction parameters to form a parameter correction matrix. Then, under the action of the parameter correction matrix, the key support parameters output by the model are adjusted in stages, including shotcrete thickness, anchor bolt spacing, steel arch stiffness, and anchorage depth. The adjustment range of each support parameter is dynamically controlled by the gain factor of the correction command stream. When the energy distribution gradient is large, the gain weighting coefficient is automatically increased to accelerate the correction response; when the energy disturbance tends to be stable, the gain coefficient gradually decreases to prevent over-adjustment from causing secondary disturbances. Through the above step-by-step dynamic correction method, the real-time correspondence between support parameters and changes in surrounding rock energy can be achieved, ensuring that the mechanical support system remains coordinated and consistent during energy abrupt changes in the construction process.

[0119] After initial dynamic correction, the adjusted support parameters are verified through energy response to assess the matching degree between the support correction effect and the surrounding rock energy field. Specifically, the corrected support parameters are input into the surrounding rock energy distribution calculation framework to recalculate the energy release rate, stress concentration factor, and deformation compatibility at the current moment. By comparing with the time-series evolution curve of the dynamic energy distribution factor, it is determined whether the support adjustment has effectively reduced the stress peak and abrupt change intensity in the energy concentration area. When the calculation results show that the energy release rate tends to stabilize and the stress concentration decreases significantly, it indicates that the support correction has achieved the expected effect; otherwise, the weight distribution of the parameter correction matrix needs to be further adjusted according to the correction command flow. To avoid misjudgment due to short-term fluctuations, the energy response data also needs to be smoothed through multiple windows to obtain a long-term stable trend. This energy response verification process not only achieves physical verification of the support correction results but also provides feedback data for subsequent parameter write-back, thereby ensuring the causal consistency and closed-loop continuity of the entire control process.

[0120] After energy response verification is completed, the correction results are synchronized to the non-stationary feature transfer network via a parameter write-back mechanism to achieve dynamic consistency between the model's internal parameters and the external construction state. Specifically, the corrected support parameters, energy response change rate, and corresponding dynamic energy distribution factor generated during the verification phase are first uniformly encoded into a feature feedback vector. Then, this feature feedback vector is input into the feature update layer of the non-stationary feature transfer network to adjust the model's feature weight distribution and transfer mapping function, enabling the model's internal parameter structure to adaptively reflect the real-time mechanical state of the construction site. Simultaneously, a time-delay feedback control strategy is introduced to match the time step of support parameter write-back with the energy disturbance cycle, avoiding oscillations and instability caused by frequent model updates. Through this parameter write-back mechanism, the model can obtain new feature calibration information after each low-frequency mutation event, thus maintaining higher timeliness and accuracy in subsequent learning and prediction processes. This process essentially constructs a two-way coupling mechanism between data-driven and energy-driven approaches, enabling the model to no longer rely solely on historical samples for learning but to absorb construction feedback in real time for structural optimization, forming an adaptive evolutionary system with learning, correction, and relearning capabilities.

[0121] After the parameter write-back is completed, a global stability evaluation and continuous iteration are performed on the closed-loop support optimization and control process to ensure long-term stability of the tunnel construction process under low-frequency abrupt change scenarios. Specifically, a stability criterion function is constructed by conducting multidimensional statistical analysis on the changes in support parameters, changes in surrounding rock energy distribution, and model prediction errors over multiple consecutive energy disturbance cycles. If the output value of this function remains within a stable range, it indicates that the support optimization closed loop has reached equilibrium, and the model can automatically complete parameter adaptive adjustment when new disturbances occur. If the output value exceeds the threshold range, the correction command flow generation process needs to be retried, and the parameter correction matrix and energy response verification process need to be updated. To prevent the accumulation of offsets caused by overlearning or delayed correction, force field inertial constraints are also introduced in the stability evaluation stage to keep the support adjustment of the model continuously controllable in a short period of time. Through this global stability evaluation and iteration process, the closed-loop control mechanism can achieve dynamic evolution from real-time correction to long-term self-stability, ensuring that the adjustment of support parameters always keeps in line with the changes in surrounding rock energy distribution, ultimately achieving continuous safety and structural reliability in the construction process.

[0122] Through the above implementation steps, a closed-loop adaptive control process for optimizing support parameters was achieved based on the modified command flow. This process uses causal correction as its core driver and parameter write-back as its self-learning basis, forming a complete feedback chain from model prediction to on-site feedback and then to model updating. This implementation method achieves, for the first time, dynamic closed-loop coupling between the machine learning model and the surrounding rock energy field, enabling support parameters to adaptively adjust with the non-stationary evolution of the geostress field, thereby maintaining the overall stability and construction safety of the tunnel structure under low-frequency abrupt changes.

[0123] This invention achieves dynamic characterization and adaptive model control of the energy field of surrounding rock in deep tunnels by introducing force field perturbation calculation and energy feature transfer mechanisms. By establishing a dynamic energy distribution factor based on multi-scale stress monitoring baselines and combining the force field perturbation calculation mechanism with the generation process of feature transition maps, the non-stationary characteristics of energy changes are fully captured. This method can identify the temporal variation trend of surrounding rock energy in real time during abrupt low-frequency stress release scenarios and transmit this variation characteristic to the model's feature weight allocation process, enabling the model to have self-adjustment and proactive adaptation capabilities. This significantly improves the accuracy of surrounding rock stability assessment and the ability to respond early to abrupt risks, avoiding structural mismatch problems caused by lag in support parameters.

[0124] This invention achieves self-feedback and self-correction between support parameter prediction and energy distribution state by constructing a counterfactual replay chain and a closed-loop support optimization and control mechanism. Through counterfactual replay and causal comparison of historical abrupt change trajectories, a correction command stream is generated and drives the stepwise dynamic correction of support parameters. Simultaneously, the correction results are written back to a non-stationary feature transfer network, enabling the model to self-learn and update features after each energy disturbance. This closed-loop control method not only maintains real-time consistency between support parameters and changes in the surrounding rock energy field but also achieves continuous self-stabilization and dynamic risk minimization during construction, significantly improving the safety and intelligence level of deep tunnel construction.

[0125] In summary, by utilizing the technical solution of this invention, a self-feedback and self-correction mechanism between support parameter prediction and energy distribution is achieved through the construction of a counterfactual replay chain and a closed-loop support optimization and control mechanism. By replaying historical abrupt change trajectories counterfactually and comparing causal relationships, a correction command stream is generated and drives the stepwise dynamic correction of support parameters. Simultaneously, the correction results are written back to the non-stationary feature transfer network, enabling the model to self-learn and update features after each energy disturbance. This closed-loop control method not only maintains real-time consistency between support parameters and changes in the surrounding rock energy field but also achieves continuous self-stabilization and dynamic risk minimization during construction, significantly improving the safety and intelligence level of deep tunnel construction.

[0126] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for optimizing tunnel excavation advance and support parameters based on machine learning, characterized in that, Includes the following steps: Step 1: Establish a multi-scale stress monitoring baseline in the deep tunnel construction environment, continuously collect low-frequency stress fluctuation data of the surrounding rock using a high-precision sensor array, and calculate the energy distribution of the surrounding rock based on the collected data to obtain the dynamic energy distribution factor. Step 2: Construct a force field disturbance calculation mechanism based on the dynamic energy distribution factor, perform time-series expansion on the dynamic energy distribution factor, extract the potential triggering time window of low-frequency stress change in the surrounding rock, and generate a characteristic transition map based on the energy change trajectory within the triggering time window; Step 3: Train a non-stationary feature transfer network based on the feature transition map, and guide the dynamic redistribution of feature weights within the non-stationary feature transfer network through the energy patterns in the feature transition map to obtain a model with adaptively adjusted feature weights; Step 4: Based on the model after adaptive adjustment of feature weights, construct a counterfactual replay chain, replay the historical low-frequency mutation trajectories in the feature transition map using counterfactual methods, compare and analyze the replay feature response results with historical collapse event samples, and generate a correction instruction stream with causal consistency. Step 5: Based on the correction command stream, construct a closed-loop support optimization and control unit, use the correction command stream to dynamically correct the support parameters output in real time, and synchronize the correction results to the non-stationary feature migration network through the parameter write-back mechanism to realize closed-loop adaptive optimization between support parameters and dynamic energy distribution of surrounding rock. Step 2 specifically includes: Based on the dynamic energy distribution factor, a force field disturbance response model for surrounding rock is established. The force field disturbance response matrix is ​​calculated by coupling the stress-strain energy density relationship with the energy release rate. The dynamic energy distribution factor is subjected to time-series expansion, and the time-frequency response surface of energy fluctuation is constructed using multi-scale time window analysis and Hilbert transform. Based on the time series unfolding results, the energy change rate is calculated and combined with the force field disturbance response matrix to determine the spatial concentration region of energy anomalies, thereby extracting the potential triggering time window; Energy change trajectory features are extracted within the potential triggering time window, and the feature transition spectrum is generated using principal component analysis and nonlinear embedding.

2. The method for optimizing tunnel excavation progress and support parameters based on machine learning according to claim 1, characterized in that, Step 1 specifically includes: Based on the tunnel burial depth, geological structure and distribution of surrounding rock joints, the surrounding rock space is divided into three characteristic sections: the main control stress zone, the secondary control stress zone and the local disturbance zone. High-precision stress sensor arrays are arranged in each section to form the multi-scale stress monitoring baseline. The high-precision stress sensor array is used to continuously collect low-frequency stress fluctuation data of the surrounding rock, and the sampling error is corrected by a baseline drift self-correction mechanism. Based on the corrected low-frequency fluctuation data, the energy distribution of the surrounding rock is calculated using mechanical energy field theory and wave energy spectrum analysis method, and a three-dimensional energy distribution field is established. The dynamic energy distribution factor is constructed based on the temporal and spatial gradient variation characteristics of the three-dimensional energy distribution field.

3. The method for optimizing tunnel excavation progress and support parameters based on machine learning according to claim 2, characterized in that, The high-precision stress sensor array includes a combination structure of triaxial stress sensors and distributed fiber optic strain sensors; and the baseline drift self-correction mechanism uses the average stress value in the short-period stable range to correct the sampled data.

4. The method for optimizing tunnel excavation progress and support parameters based on machine learning according to claim 1, characterized in that, Generating the feature transition map includes: Cluster analysis was performed on the energy change trajectories within the potential triggering time window, and they were divided into several energy transition clusters based on similarity. Calculate the center trajectory of each energy transition cluster; The central trajectories are connected sequentially in the time series to form a continuous feature transition map.

5. The method for optimizing tunnel excavation progress and support parameters based on machine learning according to claim 1, characterized in that, Step 3 specifically includes: Based on the aforementioned feature transition map, a feature mapping set is constructed, and the energy density change rate, energy gradient direction, stress response sensitivity, and energy release rate are decomposed in multiple dimensions. The weight mapping function of the energy mode is obtained through the eigenvalue decomposition of the covariance matrix. The weight mapping function is used to dynamically redistribute the features within the non-stationary feature transfer network, and the feature weight gradient is corrected according to the fluctuation trend of the energy pattern. The dynamically redistributed feature set is input into the non-stationary feature transfer network, and a feature alignment relationship is established between stable geological samples and low-frequency mutation samples through a transfer learning strategy.

6. The method for optimizing tunnel excavation progress and support parameters based on machine learning according to claim 5, characterized in that, When dynamically redistributing the internal features of the non-stationary feature transfer network, a force field perturbation regularization term is introduced to constrain the spatial continuity of the energy pattern.

7. The method for optimizing tunnel excavation progress and support parameters based on machine learning according to claim 1, characterized in that, Step 4 specifically includes: The characteristic transition map is subjected to time-series inversion, and the energy transition segment with the most drastic change in energy gradient is selected as the counterfactual replay sample to reconstruct the historical energy evolution path; A time series analysis is performed on the characteristic response results generated by the model during the counterfactual replay process. The phase difference and correlation coefficient between the model output and the energy trajectory are calculated to form a model response deviation distribution map. The model response deviation distribution map is compared with historical collapse event samples to establish a causal consistency matrix between model behavior and geological events; The correction instruction stream is generated based on the causal consistency matrix to synchronously correct the model feature weights and support parameter outputs.

8. The method for optimizing tunnel excavation progress and support parameters based on machine learning according to claim 7, characterized in that, When generating the correction command stream, the adjustment parameters of the correction command include the feature weight correction coefficient, the time delay compensation amount, and the support parameter adjustment factor; and the correction amplitude is adaptively adjusted through the dynamic gain function corresponding to the energy change rate.

9. The method for optimizing tunnel excavation progress and support parameters based on machine learning according to claim 1, characterized in that, Step 5 specifically includes: Based on the correction command flow, a parameter correction matrix is ​​established, and the feature weight correction coefficient, time delay compensation amount and support adjustment factor are mapped to the support parameters output by the model. The shotcrete thickness, anchor bolt spacing, steel arch stiffness and lock foot anchorage depth are dynamically corrected step by step according to the gain factor of the correction command flow. The corrected support parameters are verified by energy response. The correction results are input into the surrounding rock energy distribution calculation framework to calculate the energy release rate and stress concentration factor, so as to judge the correction effect and optimize the parameter correction matrix. The energy-verified support parameter correction values, energy response change rate, and dynamic energy distribution factor are encoded as feature feedback vectors and input into the non-stationary feature transfer network. The model parameters are dynamically synchronized through the parameter write-back mechanism. The stability of the closed-loop support optimization and control process is evaluated, and the construction process is kept stable in the low-frequency sudden change scenario through multidimensional statistical analysis and force field inertial constraints.

Images