Tunnel rock burst adaptive early warning method and system based on multi-source dynamic data fusion

Abstract

The application relates to the technical field of construction risk management, and discloses a tunnel rock burst self-adaptive early warning method and system based on multi-source dynamic data fusion, which aims to solve the problem of poor accuracy of existing rock burst early warning methods, and mainly comprises the following steps: collecting microseismic data, construction data and geological data; performing space-time alignment, redundancy elimination and complementary enhancement processing on the data to obtain a standardized fusion data set including basic parameters and fusion derivative parameters; extracting 12-dimensional dynamic features from the data set, determining the current scene based on real-time construction and geological parameters, combining a pre-defined scene risk coefficient to generate a scene-based dynamic threshold corresponding to the features; and an early warning model adjusts the feature weight and network structure according to the scene, and comprehensively judges the input features in combination with the dynamic threshold, and finally outputs an early warning result containing a risk level and a confidence degree. The application realizes high-accuracy and high-real-time early warning of rock burst risks, and is particularly suitable for high ground stress environments.

Description

Technical Field

[0001] This invention relates to the field of construction risk management technology, specifically to a tunnel rockburst adaptive early warning method and system based on multi-source dynamic data fusion. Background Technology

[0002] Drill-and-blast method, due to its high cost-effectiveness and strong geological adaptability, has become the mainstream technology for deep-buried tunnel construction. However, with the increase in tunnel depth, rockburst disasters caused by high ground stress environments are becoming increasingly frequent. Rockbursts manifest as sudden brittle failure of the surrounding rock, explosive ejection, and are accompanied by strong vibrations and sounds, seriously threatening the safety of construction personnel and equipment. Therefore, it is necessary to conduct real-time and accurate early warning of rockburst disasters. Existing rockburst early warning methods mainly include the following:

[0003] The first method is an early warning system based on human experience. This method relies on on-site technicians conducting regular inspections, visually observing phenomena such as cracks and spalling in the surrounding rock at the tunnel face and walls, or using tools to listen to sounds within the rock mass to assess risk. This method relies entirely on personal experience, lacks objective quantitative standards, and has low accuracy and high randomness in its early warnings. Furthermore, information acquisition is intermittent and not real-time; there is a significant time lag between the occurrence of a risk and its detection by humans, making it unable to meet the rapid response requirements due to the suddenness and destructive power of rockbursts.

[0004] The second type is an automated early warning method based on a single sensor and a fixed threshold. This method typically deploys a network of microseismic monitoring sensors to collect microseismic signals generated by rock fractures in real time, and presets fixed physical quantity thresholds (e.g., setting an alarm when the microseismic event rate exceeds 10 times per hour). This technology relies solely on microseismic information and fails to integrate construction activity (such as blasting) data, resulting in an inability to distinguish between rockburst incubation signals and blasting operation interference signals. This leads to a high false alarm rate and frequent unnecessary construction interruptions. Furthermore, its early warning threshold is static and globally uniform, unable to be dynamically adjusted based on the specific geological conditions of a section (such as surrounding rock grade and ground stress level) and the construction stage (such as the intense blasting period and the support period). In tunnels with complex geology or variable construction conditions, this can easily lead to missed alarms in high-risk sections and excessive warnings in low-risk sections.

[0005] The third type is a static machine learning early warning model trained on historical data. This technology uses algorithms such as Long Short-Term Memory (LSTM) networks to train on previously collected microseismic and partial geological data, learning historical patterns before rock bursts, and then predicting risks based on real-time data. The drawback of this technology is that once deployed, the number of network layers and decision-making logic are fixed, making it a "static" system. When tunnel construction moves from one geological-construction scenario (e.g., low stress, weak blasting) to a completely different scenario (e.g., high stress, strong blasting), the static model cannot adjust itself online to adapt to the data characteristics and risk patterns in the new environment. This leads to a significant decrease in early warning accuracy as construction progresses, requiring frequent manual retraining and parameter tuning in later maintenance, resulting in insufficient intelligence.

[0006] In summary, existing technologies generally suffer from two major drawbacks: insufficient fusion of multi-source dynamic data and a lack of adaptability of early warning models and thresholds to dynamic changes in construction and geological scenarios. These shortcomings make it difficult to achieve stable and accurate rockburst risk early warning in drill-and-blast tunnel construction where the entire process is dynamically changing. Summary of the Invention

[0007] This invention aims to address the problem that existing rockburst early warning methods suffer from poor accuracy due to insufficient fusion of multi-source dynamic data and the lack of adaptability of early warning models and thresholds to dynamic changes in construction and geological scenarios. It proposes an adaptive early warning method and system for tunnel rockbursts based on multi-source dynamic data fusion.

[0008] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows:

[0009] In a first aspect, the present invention provides a tunnel rockburst adaptive early warning method based on multi-source dynamic data fusion, the method comprising:

[0010] Collect microseismic data, construction data, and geological data, and dynamically adjust the sampling interval according to the scenario. The geological data includes the quasi-static background value of the maximum principal stress, the dynamic fluctuation value of the maximum principal stress, the uniaxial compressive strength of the rock, and the joint density.

[0011] The microseismic data, construction data, and geological data are subjected to spatiotemporal alignment, redundancy removal, and complementary enhancement processing to obtain a standardized fusion dataset including basic parameters and fusion-derived parameters.

[0012] 12-dimensional dynamic features are extracted from the standardized fusion dataset to form a feature vector;

[0013] Based on the basic parameters in the standardized fusion dataset, the current construction disturbance intensity and geological risk level are determined. The scene to which the current construction disturbance intensity and geological risk level belong are determined according to the current construction disturbance intensity and geological risk level. Based on the scene risk coefficient of the scene, a scene-based dynamic threshold corresponding to the 12-dimensional dynamic features is generated.

[0014] The feature weights and network structure of the early warning model are adaptively adjusted according to the current scenario. The feature vector is then input into the adjusted early warning model, and a comprehensive judgment is made in combination with the scenario-based dynamic threshold. The early warning result, which includes risk level and confidence level, is then output.

[0015] Furthermore, the complementary enhancement process specifically includes:

[0016] Multiple complementary data sets are constructed, including: construction-microseismic complementary pairs, stress-microseismic complementary pairs, and geological-stress complementary pairs.

[0017] For each complementary data pair, a weighted fusion is performed to generate fusion-derived parameters, calculated using the following formula:

[0018] ;

[0019] in, This refers to the fusion value, i.e., the fusion-derived parameter. and This refers to a pair of correlation parameters selected from microseismic data, construction data, and geological data after spatiotemporal alignment and redundancy removal. and These are the weighting coefficients;

[0020] After complementary enhancement processing, the basic parameters in the standardized fusion dataset include: microseismic event rate, maximum microseismic energy, average microseismic energy, microseismic source coordinates, charge quantity, blasting time, tunneling speed, quasi-static background value of maximum principal stress, dynamic fluctuation value of maximum principal stress, joint density, and uniaxial compressive strength of rock; the fusion-derived parameters include: fusion value of construction-microseismic complementary pair, fusion value of stress-microseismic complementary pair, and fusion value of geology-stress complementary pair.

[0021] Furthermore, the construction-microseismic complementary alignment, For the amount of explosives, This represents the microseismic event rate; when the explosive charge is greater than 50 kg, When the charge weight is less than or equal to 50 kg, ;

[0022] In the stress-microseismic complementary pair This represents the maximum dynamic fluctuation value of the principal stress. The maximum energy of the microseismic event; when the dynamic fluctuation value of the maximum principal stress relative to its background value increases by more than 5 MPa, ,otherwise, ;

[0023] In the geological-stress complementary pair For the uniaxial compressive strength of rock, This is the quasi-static background value of the maximum principal stress; when the quasi-static background value of the maximum principal stress is greater than 20 MPa... When the maximum principal stress quasi-static background value is less than or equal to 20 MPa, .

[0024] Furthermore, the scenarios include construction scenarios and geological scenarios;

[0025] The construction scenarios include three categories: strong blasting scenarios with a charge of more than 50kg, weak blasting scenarios with a charge of less than or equal to 50kg and a tunneling speed of more than 0mm / min, and support period scenarios with a charge of 0kg and a tunneling speed of 0mm / min.

[0026] The geological scenarios are categorized into four types: high-stress sparse joint scenarios with a background value greater than 20 MPa and a joint density less than 2 joints / meter; high-stress dense joint scenarios with a background value greater than 20 MPa and a joint density greater than or equal to 2 joints / meter; medium-low stress sparse joint scenarios with a background value less than or equal to 20 MPa and a joint density less than 2 joints / meter; and medium-low stress dense joint scenarios with a background value less than or equal to 20 MPa and a joint density greater than or equal to 2 joints / meter.

[0027] The three types of construction scenarios are combined with the four types of geological scenarios to form 12 scenarios. Each scenario has a unique scenario risk coefficient, which ranges from 0.1 to 0.6.

[0028] Furthermore, the 12-dimensional dynamic features include microseismic features, coupling features, and scene features;

[0029] The microseismic features include microseismic event rate, microseismic cumulative energy, average microseismic magnitude, microseismic event cluster density, source-excavation face distance, and dominant frequency of microseismic waveform; the calculation method for the microseismic event cluster density includes: dividing the number of microseismic events within an adaptive sliding window by the volume of the corresponding smallest geological unit;

[0030] The coupling characteristics include the correlation between charge amount and microseismic energy, the correlation between ground stress increment and microseismic event rate, the stress-strength matching coefficient, and the correlation between tunneling speed and microseismic frequency; the calculation method of the stress-strength matching coefficient includes: dividing the uniaxial compressive strength of rock by the dynamic fluctuation value of the maximum principal stress, and then standardizing it;

[0031] The scene features include construction scene markers and geological scene markers;

[0032] The method for adjusting the adaptive sliding window includes:

[0033] When the microseismic event rate increases by more than 30% for two consecutive windows, the window length is shortened from 60 minutes to 15 minutes; when the microseismic event rate fluctuates by less than 10% for three consecutive windows, the window length is restored to 60 minutes.

[0034] Furthermore, the scenario-based dynamic threshold is calculated and generated based on the following formula:

[0035] ;

[0036] in, For scenario-based dynamic thresholds, As the feature threshold baseline, This represents the scenario risk coefficient for the current situation.

[0037] The feature threshold baseline is calibrated and updated according to a preset period. Its calculation priority is as follows: the average feature value of the most recent 3 risk-free periods is preferred as the real-time baseline; when the real-time baseline is missing, the historical baseline calculated from the historical risk-free data of similar tunnels is used; when both the real-time baseline and the historical risk-free data of similar tunnels are missing, the preset default baseline is used.

[0038] Furthermore, the feature weights and network structure of the early warning model are adaptively adjusted according to the current scenario, specifically including:

[0039] The random forest algorithm is used to calculate the feature contribution of 12-dimensional dynamic features in real time, and the weight of each feature is dynamically adjusted according to the feature contribution. Features with a feature contribution of ≥25% are defined as core features and their weight is set to 0.3; features with a feature contribution of less than 5% are defined as weak features and their weight is set to 0.01; the remaining features are ordinary features and their initial weights are linearly distributed between 0.1 and 0.2. Then, all initial weights are normalized so that the sum of the final weights is 1, which is used as the input weights of the early warning model.

[0040] When the scenario risk coefficient is greater than 0.4, a 6-layer long short-term memory network is used; when the scenario risk coefficient is less than or equal to 0.4, a 3-layer gated recurrent unit network is used.

[0041] Furthermore, the comprehensive determination specifically includes:

[0042] The number of feature values ​​in a 12-dimensional dynamic feature that exceed their corresponding scene-specific dynamic threshold is denoted as . ;

[0043] Obtain the rockburst risk probability output by the adjusted early warning model. ;

[0044] Calculate confidence level : ,in, This refers to the number of feature values ​​in the core features that exceed their corresponding scene-specific dynamic threshold. The total number of core features;

[0045] The final risk level is determined according to preset mapping rules, which include:

[0046] High-risk level: Meets the requirements , and ;

[0047] Medium risk level: Meets the requirements , and And it does not meet the criteria for a high-risk level;

[0048] Low risk level: Meets the requirements , and And it does not meet the criteria for determining high or medium risk levels;

[0049] Risk-free level: Meets the requirements , and .

[0050] Furthermore, the method also includes: processing steps for the early warning results and coordinated response steps;

[0051] The steps for processing the early warning results include:

[0052] The warning results are associated and stored with the corresponding feature vectors, scenario-based dynamic thresholds, and triggering instructions.

[0053] The warning results are marked with a timestamp, tunnel mileage coordinates, and final handling feedback information;

[0054] The tagged early warning results are used as key case data for iterative optimization of the early warning model and thresholds;

[0055] The coordinated response steps include:

[0056] Trigger on-site audible and visual alarm devices to issue graded warnings;

[0057] The equipment control interface outputs stop or speed reduction commands to the tunneling machine and lock commands to the blasting authority management terminal.

[0058] Evacuation instructions are sent to the wristbands of people in the risk area via personnel positioning units.

[0059] Secondly, the present invention provides a tunnel rockburst adaptive early warning system based on multi-source dynamic data fusion, used to implement the tunnel rockburst adaptive early warning method based on multi-source dynamic data fusion as described in the first aspect, the system comprising:

[0060] The multi-source dynamic data acquisition module is used to acquire microseismic data, construction data, and geological data, and dynamically adjust the sampling interval according to the scenario. The geological data includes the quasi-static background value of the maximum principal stress, the dynamic fluctuation value of the maximum principal stress, the uniaxial compressive strength of the rock, and the joint density.

[0061] The data fusion processing module is used to perform spatiotemporal alignment, redundancy removal, and complementary enhancement processing on the microseismic data, construction data, and geological data to obtain a standardized fusion dataset including basic parameters and fusion-derived parameters.

[0062] The dynamic feature mining module is used to extract 12-dimensional dynamic features from the standardized fusion dataset to form a feature vector.

[0063] The scenario-based threshold adjustment module is used to determine the current construction disturbance intensity and geological risk level based on the basic parameters in the standardized fusion dataset, determine the scenario to which the current construction disturbance intensity and geological risk level belong, and generate a scenario-based dynamic threshold corresponding to the 12-dimensional dynamic features based on the scenario risk coefficient of the scenario.

[0064] The adaptive early warning model module is used to adaptively adjust the feature weights and network structure of the early warning model according to the current scenario. The feature vector is input into the adjusted early warning model, and a comprehensive judgment is made in combination with the scenario-based dynamic threshold. The output is an early warning result that includes risk level and confidence level.

[0065] The beneficial effects of this invention are as follows: The tunnel rockburst adaptive early warning method and system based on multi-source dynamic data fusion provided by this invention solves the problem of confusion between disturbances and rockburst precursors caused by a single data source by deeply integrating microseismic, construction dynamics and geological data, and significantly reduces the false alarm rate; by introducing scene judgment and dynamic threshold generation mechanisms, as well as the adaptive adjustment capability of model parameters and structure, it ensures that the early warning system can accurately adapt to the dynamically changing construction and geological conditions throughout the entire cycle, thereby achieving high accuracy and high real-time early warning of rockburst risk in complex tunnel engineering, effectively ensuring construction safety and efficiency. Attached Figure Description

[0066] Figure 1 A flowchart illustrating the adaptive early warning method for tunnel rockburst based on multi-source dynamic data fusion provided in this embodiment;

[0067] Figure 2 This is a schematic diagram of the structure of a tunnel rockburst adaptive early warning system based on multi-source dynamic data fusion, provided as an example. Detailed Implementation

[0068] To overcome the shortcomings of existing rockburst early warning technologies, such as high false alarm rates, poor scenario adaptability, and a sharp drop in early warning accuracy due to single data fusion dimensions and static, fixed models and thresholds, this invention proposes a technical solution. In this invention, a comprehensive multi-source dynamic data foundation is first constructed by dynamically collecting three types of data according to scenarios: microseismic data, construction data (such as charge quantity and tunneling speed), and geological data (such as ground stress, rock mass strength, and joints). Then, through a three-step fusion process of spatiotemporal alignment, redundancy removal, and complementary enhancement, a standardized fusion dataset is generated that simultaneously contains original basic parameters and weighted fusion-derived parameters, thereby achieving the separation and enhancement of construction disturbance signals and actual rock mass fracture signals at the data level. Next, 12-dimensional dynamic features covering microseismic, coupling, and scenario dimensions are extracted from this dataset. Based on these features and basic parameters, the current "construction-geology" combined scenario is determined in real time, and the risk coefficient bound to this scenario is invoked to dynamically generate a matching 12-dimensional feature threshold, achieving refined and adaptive early warning benchmarks. Finally, the early warning model adjusts its internal feature weights and network structure in sync with the current scenario (e.g., a more complex LSTM network is used in high-risk scenarios), and then combines the dynamic threshold to make a comprehensive judgment on the input feature vector, thereby outputting an accurate early warning result that includes both risk level and confidence level.

[0069] The technical solutions in this embodiment will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0070] Figure 1 A flowchart illustrating an adaptive early warning method for tunnel rockburst based on multi-source dynamic data fusion is shown. Please refer to [link / reference]. Figure 1 The method includes the following steps:

[0071] Step 1: Collect microseismic data, construction data, and geological data, and dynamically adjust the sampling interval according to the scenario. The geological data includes the quasi-static background value of the maximum principal stress, the dynamic fluctuation value of the maximum principal stress, the uniaxial compressive strength of the rock, and the joint density.

[0072] This step dynamically collects three types of data according to the scenario:

[0073] 1. Microseismic Data Acquisition: A set of three-component microseismic sensors is staggered at intervals of 10-20 meters along the tunnel's extension direction, on both sidewalls and the tunnel arch, forming a microseismic network. The sampling frequency is 10-20kHz, and the positioning accuracy is required to be ≤5 meters. The sensor installation depth is determined based on the integrity of the surrounding rock to ensure good coupling with the rock mass, and data is transmitted via optical fiber to avoid electromagnetic interference.

[0074] 2. Construction data acquisition: The tunneling speed (unit: mm / min) is collected in real time by connecting to the programmable logic controller (PLC) system of the tunneling machine; at the same time, the blasting monitoring terminal is connected to collect the charge amount of a single blast (unit: kg / blast) and the blasting time accurate to the second.

[0075] 3. Geological Data Acquisition: A dynamic, encrypted data acquisition technology based on scenario risk coefficients is employed for advanced drilling. During each drilling operation (drilling depth 30-50 meters), the following three types of parameters are simultaneously acquired and linked to corresponding mileage coordinates:

[0076] Joint density: Statistically measured in real time during core drilling, counted per 1-meter core segment, and the final average value for the entire borehole is taken (unit: joints / m).

[0077] Maximum principal stress: A two-layer acquisition scheme was adopted, consisting of "background value determination using the stress relief method + long-term monitoring of fluctuation values ​​using strain gauges". First, the quasi-static background value of the maximum principal stress in the pre-drilled section was obtained using the hollow inclusion stress relief method, serving as the regional stress benchmark. Second, strain gauges were retained in the borehole for long-term monitoring, collecting real-time strain data every 2 hours and converting it into the dynamic fluctuation value of the maximum principal stress, with an accuracy requirement of ±0.5 MPa.

[0078] Uniaxial compressive strength of rock: Standard specimens were prepared from intact rock cores obtained from the borehole and measured by laboratory uniaxial compression tests, with an accuracy of ±1 MPa. If laboratory conditions are not available, a rebound hammer can be used to perform rapid testing on the rock cores (error ±5 MPa) as temporary supplementary data.

[0079] The dynamic adjustment logic for the sampling interval is as follows:

[0080] Dynamically adjust the sampling frequency of microseismic and construction data based on the real-time construction scenario:

[0081] High-explosive scenarios (charge > 50kg): microseismic data sampling interval is 1 minute / time, and construction data sampling interval is 5 minutes / time.

[0082] Other scenarios (weak blasting, conventional tunneling, support period): microseismic data sampling interval is 5 minutes / time, and construction data sampling interval is 10 minutes / time.

[0083] The mileage of tunnels constructed using advanced drilling is dynamically adjusted based on the scenario's risk factor, as specified below:

[0084] ① Low-risk scenarios (scenario risk coefficient < 0.2): 150~200m. For situations without rockburst precursors, this reduces the frequency of advance drilling and improves construction efficiency.

[0085] ② Medium-risk scenario (scenariosity coefficient of 0.2~0.4): 100~150m. For sporadic low-energy microseismic conditions, it is necessary to balance detection accuracy and cost, and monitor changes in local risk.

[0086] ③ High-risk scenarios (scenarios with a risk coefficient of 0.4~0.6): 30~100m. For high ground stress (>20MPa) conditions, microseismic energy increases in a stepwise manner, requiring more frequent detection to identify the disaster-causing geological bodies in advance;

[0087] For the maximum principal stress, when the fluctuation value of the long-term monitoring is greater than 5 MPa compared with the background value, it is judged as a stress mutation and triggers encrypted sampling (the sampling frequency of the maximum principal stress is increased from 2 hours / time to 30 minutes / time).

[0088] All collected data must be stored with a timestamp accurate to the second and tunnel mileage coordinates.

[0089] Step 2: Perform spatiotemporal alignment, redundancy removal, and complementary enhancement processing on the microseismic data, construction data, and geological data to obtain a standardized fusion dataset including basic parameters and fusion-derived parameters.

[0090] This step is used to perform a three-step processing flow of spatiotemporal alignment, redundancy removal, and complementary enhancement on the raw multi-source data collected in step 1, in order to generate a high-quality, standardized fusion dataset.

[0091] First, spatiotemporal alignment is performed to unify the spatial and temporal references of multi-source data and eliminate misalignments caused by differences in sensor location and acquisition frequency. Specifically, this includes:

[0092] Spatial alignment: Using the tunnel mileage coordinate system as a unified benchmark, starting from mileage 0 km and increasing along the tunneling direction with an accuracy ≤0.1m. Microseismic source coordinates, construction data (bound to the tunnel face or blasting area mileage), and geological data (bound to the mileage interval covered by the boreholes) are all mapped to the smallest geological unit with an accuracy of 5m. For geological parameters in the boundary area of ​​the advance boreholes, a weighted average method is used for smooth transition, ensuring the continuity of spatial information, ultimately ensuring that the spatial error between all data and the corresponding smallest geological unit is ≤5m.

[0093] Time alignment: Using the second-level timestamp of the microseismic data as the core time reference, strategies such as linear interpolation are used to complete or adapt the tunneling speed, charge quantity, and maximum principal stress dynamic fluctuation value in the construction data and geological data to the sampling time series consistent with the microseismic data, ensuring coordination in the time dimension with a time error ≤10 seconds.

[0094] Then, redundancy removal is performed to clean the data based on spatiotemporal alignment, removing interfering data that is irrelevant to rockburst risk or of low quality. Specifically, this includes:

[0095] Redundancy removal of microseismic data: Data identified as blasting disturbances are directly removed based on the following criteria: the event occurs shortly after the blast, the epicenter is located in the blasting zone, the event rate increases sharply immediately after the blast and then decreases rapidly, and the energy decays rapidly in a cliff-like manner; sensor noise is removed based on low energy; and microseismic events with completely identical characteristics are removed.

[0096] Redundant construction data elimination: During the support period and other stages without construction disturbance, continuously collected steady-state zero-value data, such as tunneling speed of 0 for 1 consecutive hour, are compressed and recorded to eliminate intermediate redundant sampling points.

[0097] Redundancy removal in geological data: Data that is obviously contradictory, such as extremely high joint density but abnormally high rock strength, is marked and corrected by weighted substitution using data from adjacent boreholes or by combining observations from the working face.

[0098] Fault data removal: Continuous abnormal signals are identified as sensor faults, triggering maintenance alarms and activating backup data sources.

[0099] The data removed must meet preset verification standards for data integrity, relevance, and redundancy. Among these:

[0100] Data integrity verification: The data removed must cover the entire construction process, with a data retention rate of ≥95% for high-risk scenarios and ≥70% for other scenarios;

[0101] Correlation verification: The retained data must be strongly correlated with rockburst risk indicators (stress mutation, microseismic energy anomaly) (Pearson correlation coefficient ≥ 0.6).

[0102] Redundancy verification: The overall data redundancy rate is ≤30%, and the amount of data after removal is 70%~95% of the original data amount, avoiding excessive removal of valid data.

[0103] All rejected data must be stored separately and labeled with the reason for rejection (such as "blasting disturbance" or "duplicate sampling"), and the traceability path must be preserved; the filtered valid data should be stored in categories of "spatiotemporal unit-parameter type-time stamp" for subsequent complementary enhancement and fusion.

[0104] Finally, complementary enhancement is performed to construct three sets of key complementary data pairs based on the cleaned and effective data. These pairs are then dynamically weighted and fused to generate derived parameters that strengthen risk associations. Specifically, this includes:

[0105] Multiple complementary data sets are constructed, including: construction-microseismic complementary pairs, stress-microseismic complementary pairs, and geological-stress complementary pairs.

[0106] For each complementary data pair, a weighted fusion is performed to generate fusion-derived parameters, calculated using the following formula:

[0107] ;

[0108] in, This refers to the fusion value, i.e., the fusion-derived parameter. and This refers to a pair of correlation parameters selected from microseismic data, construction data, and geological data after spatiotemporal alignment and redundancy removal. and These are the weighting coefficients.

[0109] Among them, the construction-microseismic complementary alignment, For the amount of explosives, This represents the microseismic event rate; when the explosive charge is greater than 50 kg, When the charge weight is less than or equal to 50 kg, Through conversion, the fusion value's dimensions are unified to J;

[0110] In the stress-microseismic complementary pair This represents the maximum dynamic fluctuation value of the principal stress. The maximum energy of the microseismic event; when the dynamic fluctuation value of the maximum principal stress relative to its background value increases by more than 5 MPa, ,otherwise, Through conversion, the fusion value's dimensions are unified to J;

[0111] In the geological-stress complementary pair For the uniaxial compressive strength of rock, This is the quasi-static background value of the maximum principal stress; when the quasi-static background value of the maximum principal stress is greater than 20 MPa... When the maximum principal stress quasi-static background value is less than or equal to 20 MPa, .

[0112] Please refer to Table 1 for the complementary data and fusion logic in this embodiment.

[0113] Table 1 Complementary Data and Fusion Logic Table

[0114]

[0115] After complementary enhancement processing, the final output is a standardized fusion dataset containing basic parameters and fusion-derived parameters, providing a unified input for subsequent feature mining. The basic parameters include: microseismic event rate, maximum microseismic energy, average microseismic energy, microseismic source coordinates, charge quantity, blasting time, tunneling speed, quasi-static background value of maximum principal stress, dynamic fluctuation value of maximum principal stress, joint density, and uniaxial compressive strength of rock. The fusion-derived parameters include: fusion values ​​of construction-microseismic complementary pairs, stress-microseismic complementary pairs, and geology-stress complementary pairs.

[0116] Step 3: Extract 12-dimensional dynamic features from the standardized fusion dataset to form a feature vector.

[0117] In this embodiment, the 12-dimensional dynamic features include three categories:

[0118] The first category is microseismic characteristics, comprising six dimensions, reflecting the spatiotemporal and intensity information of rock mass fracturing. These include: microseismic event rate (the number of microseismic events recorded per unit time); cumulative microseismic energy (the sum of the energies of all microseismic events within the statistical period); average microseismic magnitude; microseismic event cluster density (characterizing the spatial concentration of fracturing events); source-excavation face distance (the real-time distance from the source to the excavation face); and the dominant frequency of the microseismic waveform obtained through spectral analysis, used to distinguish fracturing types.

[0119] The second category is coupling characteristics, comprising four dimensions, revealing the interaction relationships between different physical processes. These include: the correlation between charge amount and microseismic energy; the correlation between ground stress increment and microseismic event rate; the stress-strength matching coefficient, calculated from the dynamic fluctuation value of the uniaxial compressive strength of rock and the maximum principal stress; and the correlation between tunneling speed and microseismic frequency.

[0120] The third category is scene features, which are two-dimensional and represent the current working condition. This includes: construction scene identifiers coded based on the charge quantity and tunneling status, and geological scene identifiers coded based on the ground stress level and joint density.

[0121] Please refer to Table 2 for the 12-dimensional dynamic features in this embodiment.

[0122] Table 2 12-dimensional dynamic feature table

[0123]

[0124]

[0125]

[0126] When extracting the aforementioned microseismic features, an adaptive sliding time window mechanism is employed. This means the window length is not fixed but is adjusted in real-time based on the dynamic changes in microseismic activity. Specifically, the adjustment rule is as follows: when the microseismic event rate increases by more than 30% within two consecutive time windows, the sliding window length is automatically shortened from 60 minutes to 15 minutes to improve sensitivity in detecting rapid changes in risk; conversely, when the microseismic event rate remains stable within three consecutive time windows with fluctuations less than 10%, the system restores the sliding window length to 60 minutes to ensure the stability of feature statistics and reduce computational load.

[0127] Finally, the feature values ​​of the 12 dimensions calculated in real time are arranged and combined in a preset order to form a standardized feature vector. This feature vector fully integrates information on microseismic activity, multi-physics coupling relationships, and the current geological environment of the construction site, providing a unified, structured, and high-quality input for subsequent risk assessment models.

[0128] Step 4: Based on the basic parameters in the standardized fusion dataset, determine the current construction disturbance intensity and geological risk level, determine the scene to which the current construction disturbance intensity and geological risk level belong, and generate a scene-based dynamic threshold corresponding to the 12-dimensional dynamic features based on the scene risk coefficient of the scene.

[0129] In this embodiment, the scenarios include construction scenarios and geological scenarios;

[0130] The construction scenarios include three categories: strong blasting scenarios with a charge of more than 50kg, weak blasting scenarios with a charge of less than or equal to 50kg and a tunneling speed of more than 0mm / min, and support period scenarios with a charge of 0kg and a tunneling speed of 0mm / min.

[0131] The geological scenarios are categorized into four types: high-stress sparse joint scenarios with a background value greater than 20 MPa and a joint density less than 2 joints / meter; high-stress dense joint scenarios with a background value greater than 20 MPa and a joint density greater than or equal to 2 joints / meter; medium-low stress sparse joint scenarios with a background value less than or equal to 20 MPa and a joint density less than 2 joints / meter; and medium-low stress dense joint scenarios with a background value less than or equal to 20 MPa and a joint density greater than or equal to 2 joints / meter.

[0132] The three types of construction scenarios are combined with the four types of geological scenarios to form 12 scenarios. Each scenario has a unique scenario risk coefficient, which ranges from 0.1 to 0.6. Please refer to Table 3.

[0133] Table 3. Various Scenarios and Their Corresponding Scenarios Risk Coefficients

[0134]

[0135]

[0136] In practical applications, the scenario determination process is first executed, which includes the following two levels of rules:

[0137] Level 1: Determining the Construction Scenarios: Based on the real-time collected parameters of explosive charge and tunneling speed, the current working condition is classified into one of three scenarios. Specifically: when the explosive charge is greater than 50kg, it is classified as a strong blasting scenario; when the explosive charge is less than or equal to 50kg and the tunneling speed is greater than 0mm / min, it is classified as a weak blasting scenario; when the explosive charge is 0kg and the tunneling speed is 0mm / min, it is classified as a support phase scenario.

[0138] The second level involves determining the geological scenario: based on two geological parameters—the maximum principal stress quasi-static background value and the joint density—the geological conditions are classified into one of four geological scenarios. Specifically: when the background value is greater than 20 MPa and the joint density is less than 2 joints / meter, it is classified as a high-stress sparse joint scenario; when the background value is greater than 20 MPa and the joint density is greater than or equal to 2 joints / meter, it is classified as a high-stress dense joint scenario; when the background value is less than or equal to 20 MPa and the joint density is less than 2 joints / meter, it is classified as a medium-low stress sparse joint scenario; and when the background value is less than or equal to 20 MPa and the joint density is greater than or equal to 2 joints / meter, it is classified as a medium-low stress dense joint scenario.

[0139] Final scenario combination and risk coefficient matching: By combining the construction scenario categories determined above with the geological scenario categories, a unique category among the twelve preset construction geological scenarios can be identified. Each scenario category has a predefined unique scenario risk coefficient, which ranges from 0.1 to 0.6. A higher value indicates a higher inherent risk of rockburst in that scenario. For example, the scenario with strong blasting combined with high-stress sparse joints has the highest risk coefficient, while the scenario with support period combined with low-stress dense joints has the lowest risk coefficient.

[0140] Then, a scenario-based dynamic threshold is calculated based on the determined scenario. In this embodiment, the scenario-based dynamic threshold is generated based on the following formula:

[0141] ;

[0142] in, For scenario-based dynamic thresholds, As the feature threshold baseline, This represents the scenario risk coefficient for the current situation.

[0143] The feature threshold baseline characterizes the typical level of the feature under a risk-free state. In this embodiment, calibration and updates are performed according to a preset period, and the calculation priority is as follows: the average value of the feature over the most recent three risk-free periods is preferred as the real-time baseline; when the real-time baseline is missing, the historical baseline calculated from historical risk-free data of similar tunnels is used; when both the real-time baseline and historical risk-free data of similar tunnels are missing, the preset default baseline is used.

[0144] Ultimately, all generated scenario-based dynamic thresholds must pass triple verification: logical consistency, engineering experience range, and data relevance. After successful verification, combining the 12-dimensional dynamic features and 12 scenario categories, a total of 12 × 12 = 144 scenario-based dynamic thresholds are generated. These 144 scenario-based dynamic thresholds will be output according to feature category and scenario number, seamlessly integrated into the subsequent early warning model as the benchmark for risk level determination.

[0145] Step 5: Adaptively adjust the feature weights and network structure of the early warning model according to the current scenario, input the feature vector into the adjusted early warning model, and make a comprehensive judgment in combination with the scenario-based dynamic threshold, and output an early warning result including risk level and confidence level.

[0146] In this embodiment, the feature weights and network structure of the early warning model are adaptively adjusted according to the current scenario, specifically including:

[0147] The random forest algorithm is used to calculate the feature contribution of 12-dimensional dynamic features in real time, and the weight of each feature is dynamically adjusted according to the feature contribution. Features with a feature contribution of ≥25% are defined as core features and their weight is set to 0.3; features with a feature contribution of less than 5% are defined as weak features and their weight is set to 0.01; the remaining features are ordinary features and their initial weights are linearly distributed between 0.1 and 0.2. Then, all initial weights are normalized so that the sum of the final weights is 1, which is used as the input weights of the early warning model.

[0148] When the scenario risk coefficient is greater than 0.4, a 6-layer long short-term memory network is used; when the scenario risk coefficient is less than or equal to 0.4, a 3-layer gated recurrent unit network is used.

[0149] In practical applications, the early warning model parameters are first adaptively adjusted, including two parallel adjustment mechanisms: feature weight adaptation and model structure adaptation.

[0150] The adaptive feature weight mechanism is implemented using the random forest algorithm. This algorithm uses the reduction in impurity of decision tree nodes as the core indicator, and calculates the contribution of each of the 12 dynamic features to the current risk assessment in real time within a set time window; this contribution is called the feature contribution. Based on the calculated feature contribution, the weight allocation of each feature in the model input is dynamically adjusted. Specifically, features with a contribution of 25% or more are defined as core features, with a weight set to 0.3; features with a contribution of less than 5% are considered weak features, with a weight set to 0.01; the weights of the remaining features are linearly distributed between 0.1 and 0.2 according to their contribution; then, all initial weights are normalized so that the final sum of weights is 1, which is used as the input weights of the warning model. This mechanism ensures that the model can always focus on the most indicative risk signals in the current scenario, filtering out interference from irrelevant or weakly correlated information.

[0151] The adaptive model structure mechanism automatically switches the neural network architecture used by the early warning model based on the scenario risk coefficient output in step 4. Specifically, when the scenario risk coefficient is greater than 0.4, indicating a high-risk condition, a 6-layer Long Short-Term Memory (LSTM) network is automatically selected as the early warning model to leverage its powerful long-term dependency capture capability and accurately identify the progressive evolution of rockburst precursors. When the scenario risk coefficient is less than or equal to 0.4, indicating a medium-to-low-risk condition, the model automatically switches to a 3-layer gated recurrent unit (GRU) network, significantly improving computational efficiency while ensuring necessary accuracy, and ensuring that the early warning response time meets real-time requirements.

[0152] Then, a rockburst risk assessment and comprehensive determination are performed, specifically including:

[0153] The number of feature values ​​in a 12-dimensional dynamic feature that exceed their corresponding scene-specific dynamic threshold is denoted as . ;

[0154] Obtain the rockburst risk probability output by the adjusted early warning model. ;

[0155] Calculate confidence level : ,in, This refers to the number of feature values ​​in the core features that exceed their corresponding scene-specific dynamic threshold. The total number of core features;

[0156] The final risk level is determined according to preset mapping rules, which include:

[0157] High-risk level: Meets the requirements , and

[0158] Medium risk level: Meets the requirements , and And it does not meet the criteria for a high-risk level;

[0159] Low risk level: Meets the requirements , and And it does not meet the criteria for determining high or medium risk levels;

[0160] Risk-free level: Meets the requirements , and .

[0161] Specifically, firstly, the feature vector generated in step 3 is input into the adaptively adjusted early warning model, which then outputs a rockburst risk probability between 0 and 100%. Simultaneously, combining the scene-specific dynamic thresholds generated in step 4 that correspond one-to-one with the scene and the 12-dimensional features, the number of features in the current feature vector whose feature values ​​exceed their corresponding thresholds is counted, denoted as... .

[0162] Then, calculate the confidence level of this warning result. Confidence level Equal to the risk probability output by the early warning model Multiply by the number of core features that exceed their corresponding scene-specific dynamic threshold. accounting for the total number of core features The proportion.

[0163] Finally, based on the above... , and Three key indicators are used to comprehensively assess the risk level, progressing from high to low risk according to a pre-defined mapping rule. The risk level is divided into four levels, from level zero to level three, corresponding to no risk, low risk, medium risk, and high risk, respectively. The final output warning result includes the determined risk level, the calculated confidence level, and the approximate tunnel mileage location of the risk.

[0164] In this embodiment, the method further includes: a step of processing the early warning result and a step of coordinated response.

[0165] The processing steps for the early warning results include:

[0166] The warning results are associated and stored with the corresponding feature vectors, scenario-based dynamic thresholds, and triggering instructions.

[0167] The warning results are marked with a timestamp, tunnel mileage coordinates, and final handling feedback information;

[0168] The tagged warning results are used as key case data for iterative optimization of the warning model and threshold.

[0169] In practical applications, firstly, the system automatically associates and binds the early warning results (including risk level, confidence level, and risk location) with the complete feature vector used in its generation process, the corresponding scenario-based dynamic threshold set, and the automatically triggered linkage instructions (such as "evacuate" and "reduce speed") based on the risk level, and stores them in a dedicated database. Secondly, each early warning record is automatically stamped with a timestamp accurate to the second and tunnel mileage coordinates accurate to the meter. Simultaneously, a feedback interface is designed, allowing on-site commanders to input the final on-site verification results and handling measures (such as "verified as blasting disturbance, construction has resumed" or "confirmed rockburst precursor, reinforcement has been completed") via a remote management platform or mobile terminal. This feedback information is used as a marker associated with the early warning record. Early warning records that have been fully tagged, especially those subsequently confirmed as rockbursts or clearly false alarms, are automatically identified as key case data. During the weekly model iteration cycle, key case data is extracted first for retraining the early warning model, optimizing feature contribution calculation rules, and calibrating feature threshold baselines, thereby continuously improving early warning performance as the project progresses.

[0170] The coordinated response steps include:

[0171] Trigger on-site audible and visual alarm devices to issue graded warnings;

[0172] The equipment control interface outputs stop or speed reduction commands to the tunneling machine and lock commands to the blasting authority management terminal.

[0173] Evacuation instructions are sent to the wristbands of people in the risk area via personnel positioning units.

[0174] In practical applications, firstly, based on the determined risk level, instructions are sent via the industrial network to the audible and visual alarms deployed within the tunnel. For example, if a high-risk situation is identified, the alarm at a designated location behind the tunnel face will trigger a high-frequency (2Hz), high-loudness (85dB) flashing red warning; a medium-risk situation will trigger a medium-frequency (1Hz), medium-loudness (75dB) flashing yellow warning; and a low-risk situation will illuminate a solid green light. This provides on-site personnel with intuitive and tiered risk perception. Secondly, instructions are sent directly to the control systems of relevant equipment via hardwiring or safety protocols. When the risk level is high, an emergency stop instruction is immediately sent to the tunneling machine's PLC system to ensure that tunneling operations cease within 3 seconds; simultaneously, a lock instruction is sent to the blasting permission management terminal to prohibit blasting operations until the risk is cleared. For medium-risk situations, an instruction to reduce the tunneling speed is sent. In addition, alarms are sent to the UWB positioning wristbands worn by construction personnel via the personnel positioning and notification unit. The alarm combines vibration, on-screen text, and voice prompts, with content corresponding to the risk level; for example, in the case of high risk, a clear announcement of "High risk, evacuate immediately" is broadcast. Meanwhile, the precise extent of the risk area and the location of personnel within the area are displayed in real time on the remote management platform, providing decision support for evacuation command.

[0175] In summary, the tunnel rockburst adaptive early warning method based on multi-source dynamic data fusion provided in this embodiment constructs a standardized fusion dataset covering microseismic, construction, and geological information. This fundamentally solves the industry pain point of existing technologies, which suffer from confusion between blasting disturbances and actual rockburst precursors due to the single data dimension, significantly reducing the false alarm rate. By establishing a construction-geological scenario judgment system based on precise quantitative parameters and a dynamic threshold generation mechanism driven by scenario risk coefficients, accurate matching between the early warning benchmark and real-time working conditions is achieved, effectively overcoming the problems of poor adaptability of fixed thresholds in complex and variable tunnel environments, and the coexistence of missed and over-warnings. The early warning model, with dual adaptive adjustment of feature weights and model structure, ensures the stability and real-time performance of early warning accuracy under different risk scenarios. Finally, through automated early warning result processing and a hierarchical linkage closed loop, seamless connection from intelligent risk perception to rapid emergency response is achieved, providing a fully adaptive and highly reliable rockburst safety early warning guarantee for drill-and-blast tunnel construction.

[0176] Please see Figure 2 Based on the above technical solution, this embodiment also provides a tunnel rockburst adaptive early warning system based on multi-source dynamic data fusion, used to implement the tunnel rockburst adaptive early warning method based on multi-source dynamic data fusion as described in the embodiment. The system includes:

[0177] The multi-source dynamic data acquisition module is used to acquire microseismic data, construction data, and geological data, and dynamically adjust the sampling interval according to the scenario. The geological data includes the quasi-static background value of the maximum principal stress, the dynamic fluctuation value of the maximum principal stress, the uniaxial compressive strength of the rock, and the joint density.

[0178] The data fusion processing module is used to perform spatiotemporal alignment, redundancy removal, and complementary enhancement processing on the microseismic data, construction data, and geological data to obtain a standardized fusion dataset including basic parameters and fusion-derived parameters.

[0179] The dynamic feature mining module is used to extract 12-dimensional dynamic features from the standardized fusion dataset to form a feature vector.

[0180] The scenario-based threshold adjustment module is used to determine the current construction disturbance intensity and geological risk level based on the basic parameters in the standardized fusion dataset, determine the scenario to which the current construction disturbance intensity and geological risk level belong, and generate a scenario-based dynamic threshold corresponding to the 12-dimensional dynamic features based on the scenario risk coefficient of the scenario.

[0181] The adaptive early warning model module is used to adaptively adjust the feature weights and network structure of the early warning model according to the current scenario. The feature vector is input into the adjusted early warning model, and a comprehensive judgment is made in combination with the scenario-based dynamic threshold. The output is an early warning result that includes risk level and confidence level.

[0182] It is understood that the tunnel rockburst adaptive early warning system based on multi-source dynamic data fusion described in this embodiment is a system used to implement the tunnel rockburst adaptive early warning method based on multi-source dynamic data fusion described in the embodiment. As the system disclosed in the embodiment corresponds to the method disclosed in the embodiment, the description is relatively simple. For relevant parts, please refer to the description of the method. It will not be repeated here.

Claims

1. A tunnel rockburst adaptive early warning method based on multi-source dynamic data fusion, characterized in that, The method includes: Collect microseismic data, construction data, and geological data, and dynamically adjust the sampling interval according to the scenario. The geological data includes the quasi-static background value of the maximum principal stress, the dynamic fluctuation value of the maximum principal stress, the uniaxial compressive strength of the rock, and the joint density. The microseismic data, construction data, and geological data are subjected to spatiotemporal alignment, redundancy removal, and complementary enhancement processing to obtain a standardized fusion dataset including basic parameters and fusion-derived parameters. 12-dimensional dynamic features are extracted from the standardized fusion dataset to form a feature vector; Based on the basic parameters in the standardized fusion dataset, the current construction disturbance intensity and geological risk level are determined. The scene to which the current construction disturbance intensity and geological risk level belong are determined according to the current construction disturbance intensity and geological risk level. Based on the scene risk coefficient of the scene, a scene-based dynamic threshold corresponding to the 12-dimensional dynamic features is generated. The feature weights and network structure of the early warning model are adaptively adjusted according to the current scenario. The feature vector is then input into the adjusted early warning model, and a comprehensive judgment is made in combination with the scenario-based dynamic threshold. The early warning result, which includes risk level and confidence level, is then output.

2. The tunnel rockburst adaptive early warning method based on multi-source dynamic data fusion according to claim 1, characterized in that, The complementary enhancement process specifically includes: Multiple complementary data sets are constructed, including: construction-microseismic complementary pairs, stress-microseismic complementary pairs, and geological-stress complementary pairs. For each complementary data pair, a weighted fusion is performed to generate fusion-derived parameters, calculated using the following formula: ； in, This refers to the fusion value, i.e., the fusion-derived parameter. and This refers to a pair of correlation parameters selected from microseismic data, construction data, and geological data after spatiotemporal alignment and redundancy removal. and These are the weighting coefficients; After complementary enhancement processing, the basic parameters in the standardized fusion dataset include: microseismic event rate, maximum microseismic energy, average microseismic energy, microseismic source coordinates, charge quantity, blasting time, tunneling speed, quasi-static background value of maximum principal stress, dynamic fluctuation value of maximum principal stress, joint density, and uniaxial compressive strength of rock; the fusion-derived parameters include: fusion value of construction-microseismic complementary pair, fusion value of stress-microseismic complementary pair, and fusion value of geology-stress complementary pair.

3. The tunnel rockburst adaptive early warning method based on multi-source dynamic data fusion according to claim 2, characterized in that, The construction-microseismic complementary alignment For the amount of explosives, This represents the microseismic event rate; when the explosive charge is greater than 50 kg, When the charge weight is less than or equal to 50 kg, ; In the stress-microseismic complementary pair This represents the maximum dynamic fluctuation value of the principal stress. The maximum energy of the microseismic event; when the dynamic fluctuation value of the maximum principal stress relative to its background value increases by more than 5 MPa, ,otherwise, ; In the geological-stress complementary pair For the uniaxial compressive strength of rock, This is the quasi-static background value of the maximum principal stress; when the quasi-static background value of the maximum principal stress is greater than 20 MPa... When the maximum principal stress quasi-static background value is less than or equal to 20 MPa, .

4. The tunnel rockburst adaptive early warning method based on multi-source dynamic data fusion according to claim 2, characterized in that, The scenarios include construction scenarios and geological scenarios; The construction scenarios include three categories: strong blasting scenarios with a charge of more than 50kg, weak blasting scenarios with a charge of less than or equal to 50kg and a tunneling speed of more than 0mm / min, and support period scenarios with a charge of 0kg and a tunneling speed of 0mm / min. The geological scenarios are categorized into four types: high-stress sparse joint scenarios with a background value greater than 20 MPa and a joint density less than 2 joints / meter; high-stress dense joint scenarios with a background value greater than 20 MPa and a joint density greater than or equal to 2 joints / meter; medium-low stress sparse joint scenarios with a background value less than or equal to 20 MPa and a joint density less than 2 joints / meter; and medium-low stress dense joint scenarios with a background value less than or equal to 20 MPa and a joint density greater than or equal to 2 joints / meter. The three types of construction scenarios are combined with the four types of geological scenarios to form 12 scenarios. Each scenario has a unique scenario risk coefficient, which ranges from 0.1 to 0.

6.

5. The tunnel rockburst adaptive early warning method based on multi-source dynamic data fusion according to claim 1, characterized in that, The 12-dimensional dynamic features include microseismic features, coupling features, and scene features; The microseismic features include microseismic event rate, microseismic cumulative energy, average microseismic magnitude, microseismic event cluster density, source-excavation face distance, and dominant frequency of microseismic waveform; the calculation method for the microseismic event cluster density includes: dividing the number of microseismic events within an adaptive sliding window by the volume of the corresponding smallest geological unit; The coupling characteristics include the correlation between charge amount and microseismic energy, the correlation between ground stress increment and microseismic event rate, the stress-strength matching coefficient, and the correlation between tunneling speed and microseismic frequency; the calculation method of the stress-strength matching coefficient includes: dividing the uniaxial compressive strength of rock by the dynamic fluctuation value of the maximum principal stress, and then standardizing it; The scene features include construction scene markers and geological scene markers; The method for adjusting the adaptive sliding window includes: When the microseismic event rate increases by more than 30% for two consecutive windows, the window length is shortened from 60 minutes to 15 minutes; when the microseismic event rate fluctuates by less than 10% for three consecutive windows, the window length is restored to 60 minutes.

6. The tunnel rockburst adaptive early warning method based on multi-source dynamic data fusion according to claim 1, characterized in that, The scenario-based dynamic threshold is calculated and generated based on the following formula: ； in, For scenario-based dynamic thresholds, As the feature threshold baseline, This represents the scenario risk coefficient for the current situation. The feature threshold baseline is calibrated and updated according to a preset period. Its calculation priority is as follows: the average feature value of the most recent 3 risk-free periods is preferred as the real-time baseline; when the real-time baseline is missing, the historical baseline calculated from the historical risk-free data of similar tunnels is used; when both the real-time baseline and the historical risk-free data of similar tunnels are missing, the preset default baseline is used.

7. The tunnel rockburst adaptive early warning method based on multi-source dynamic data fusion according to claim 1, characterized in that, The feature weights and network structure of the early warning model are adaptively adjusted according to the current scenario, specifically including: The random forest algorithm is used to calculate the feature contribution of 12-dimensional dynamic features in real time, and the weight of each feature is dynamically adjusted according to the feature contribution. Features with a feature contribution of ≥25% are defined as core features and their weight is set to 0.3; features with a feature contribution of less than 5% are defined as weak features and their weight is set to 0.01; the remaining features are ordinary features, and their initial weights are linearly distributed between 0.1 and 0.

2. Then, all initial weights are normalized so that the sum of the final weights is 1, which is used as the input weights of the early warning model. When the scenario risk coefficient is greater than 0.4, a 6-layer long short-term memory network is used; when the scenario risk coefficient is less than or equal to 0.4, a 3-layer gated recurrent unit network is used.

8. The tunnel rockburst adaptive early warning method based on multi-source dynamic data fusion according to claim 7, characterized in that, The comprehensive determination specifically includes: The number of feature values ​​in a 12-dimensional dynamic feature that exceed their corresponding scene-specific dynamic threshold is denoted as . ; Obtain the rockburst risk probability output by the adjusted early warning model. ; Calculate confidence level : ,in, This refers to the number of feature values ​​in the core features that exceed their corresponding scene-specific dynamic threshold. The total number of core features; The final risk level is determined according to preset mapping rules, which include: High-risk level: Meets the requirements , and ; Medium risk level: Meets the requirements , and And it does not meet the criteria for a high-risk level; Low risk level: Meets the requirements , and And it does not meet the criteria for determining high or medium risk levels; Risk-free level: Meets the requirements , and .

9. The tunnel rockburst adaptive early warning method based on multi-source dynamic data fusion according to any one of claims 1 to 8, characterized in that, The method also includes: processing steps for the early warning results and coordinated response steps; The steps for processing the early warning results include: The warning results are associated and stored with the corresponding feature vectors, scenario-based dynamic thresholds, and triggering instructions. The warning results are marked with a timestamp, tunnel mileage coordinates, and final handling feedback information; The tagged early warning results are used as key case data for iterative optimization of the early warning model and thresholds; The coordinated response steps include: Trigger on-site audible and visual alarm devices to issue graded warnings; The equipment control interface outputs stop or speed reduction commands to the tunneling machine and lock commands to the blasting authority management terminal. Evacuation instructions are sent to the wristbands of people in the risk area via personnel positioning units.

10. A tunnel rockburst adaptive early warning system based on multi-source dynamic data fusion, characterized in that, For implementing the tunnel rockburst adaptive early warning method based on multi-source dynamic data fusion as described in any one of claims 1 to 9, the system comprises: The multi-source dynamic data acquisition module is used to acquire microseismic data, construction data, and geological data, and dynamically adjust the sampling interval according to the scenario. The geological data includes the quasi-static background value of the maximum principal stress, the dynamic fluctuation value of the maximum principal stress, the uniaxial compressive strength of the rock, and the joint density. The data fusion processing module is used to perform spatiotemporal alignment, redundancy removal, and complementary enhancement processing on the microseismic data, construction data, and geological data to obtain a standardized fusion dataset including basic parameters and fusion-derived parameters. The dynamic feature mining module is used to extract 12-dimensional dynamic features from the standardized fusion dataset to form a feature vector. The scenario-based threshold adjustment module is used to determine the current construction disturbance intensity and geological risk level based on the basic parameters in the standardized fusion dataset, determine the scenario to which the current construction disturbance intensity and geological risk level belong, and generate a scenario-based dynamic threshold corresponding to the 12-dimensional dynamic features based on the scenario risk coefficient of the scenario. The adaptive early warning model module is used to adaptively adjust the feature weights and network structure of the early warning model according to the current scenario. The feature vector is input into the adjusted early warning model, and a comprehensive judgment is made in combination with the scenario-based dynamic threshold. The output is an early warning result that includes risk level and confidence level.

Images