A deep-buried underground cavern multi-source state fusion and GAN anomaly precursor mining method and system

Abstract

The application provides a deep-buried underground cavern multi-source state fusion and GAN abnormal precursor mining method and system, and is applied to the technical field of underground engineering safety monitoring. First, six types of underground cavern heterogeneous monitoring data are collected, a UCSM four-dimensional state tensor is constructed through three-dimensional gridding, spatial interpolation and time series normalization, and standardized time series samples are divided; relying on 3D-ConvGAN double branch network, a cavern normal working condition benchmark is learned in an adversarial manner. After model training is completed, the intermediate features in the discriminator are extracted, the VAI of the volume element is calculated by L2 distance, the global MAI is obtained by weighted aggregation, the spatial heat and time series health curve is generated. After the MAI is overlimited, the gradient attribution is used to quantify the risk contribution of each feature, the rock burst and water inrush coupling precursor is induced by combining the clustering algorithm, a mode library is established, the historical abnormal samples are stored to complete two-level induction, finally, the disaster early warning and precise risk tracing results are output by mapping the three-dimensional space, and multi-source data fusion and weak precursor intelligent identification are realized.

Description

Technical Field

[0001] This invention relates to the field of underground engineering safety monitoring technology, and in particular to a method and system for multi-source state fusion and GAN anomaly precursor mining of deeply buried underground caverns. Background Technology

[0002] With the continuous advancement of the national deep-earth strategy and major underground infrastructure projects, deep-buried cavern projects such as deep-buried transportation tunnels, underground powerhouses, and deep-earth laboratories are equipped with a variety of monitoring systems, including those for surrounding rock, microseismic monitoring, hydrology, tunneling equipment, and video monitoring. These systems rely on sensors, the Internet of Things, and video terminals to achieve full-domain perception and accumulate petabyte-level massive monitoring data. However, the existing monitoring system has multiple technical bottlenecks that hinder the implementation of intelligent safety management and control for these projects.

[0003] First, there are serious barriers to multi-source heterogeneous data. Monitoring data belong to different subsystems and cover various types of information such as stress, microseismic data, gas data, tunneling conditions, personnel coordinates, and video streams. The dimensions, data structures, and sampling frequencies vary greatly, and each manufacturer uses proprietary data protocols, forming data silos that are difficult to unify and integrate.

[0004] Second, the utilization of data mining is low. Massive amounts of monitoring data are only used for storage and tracing, relying on single indicator thresholds to trigger alarms. This is a passive prevention and control mode, which cannot uncover the inherent evolutionary patterns of the data and lacks the ability to predict disasters.

[0005] Third, it is difficult to identify weakly coupled precursors of major disasters. Disasters such as rock bursts, water inrushes, and large deformations in soft rock do not have the characteristic of a single parameter exceeding the limit significantly. Instead, they are mostly manifested by synchronous slight deviations in cross-dimensional indicators such as tunneling torque, micro-vibration b-value, and surrounding rock stress. Existing BIM and single monitoring platforms cannot integrate multi-source data, making it difficult to construct dynamic normal working condition benchmarks and capture hidden early precursors of multi-parameter coupling, resulting in delayed disaster early warning.

[0006] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0007] Other features and advantages of this application will become apparent from the following detailed description, or may be learned in part by practice of the invention.

[0008] According to one aspect of this application, a method for multi-source state fusion and GAN anomaly precursor mining in deeply buried underground caverns is provided, including: collecting multi-source heterogeneous monitoring data in six categories: surrounding rock stability, environmental ventilation, hydrogeology, construction equipment, video personnel, and static geology; executing a matrix construction process; unifying three-dimensional voxel meshing and four-dimensional tensor M(t) spatiotemporal reference; generating a UCSM unified cavern state four-dimensional raster tensor; constructing six feature channels based on Kriging / inverse distance weighted spatial interpolation, temporal normalization alignment algorithm, and static geological fixed channel mapping; calculating various sensor monitoring indicators with a single voxel (x,y,z) as the smallest statistical unit; constructing six-channel feature tensors for surrounding rock, environment, hydrology, equipment, video, and geology respectively; and simultaneously completing temporal windowing isolation of normal / disaster working condition samples; using 3D-ConvGAN to generate discriminative dual branches to extract spatiotemporally coupled evolution features, and using 3D convolution... A loop generator replicates normal dynamic sequences, and a 3D-Conv-LSTM discriminator outputs the probability of normal operating conditions. The adversarial game converges and solidifies the normal operating condition discrimination benchmark. A voxel L2 distance and a weighted double-layer anomaly quantification of the entire cavern space are configured. Multi-dimensional pattern deviation is modeled through the discriminator's intermediate convolutional feature space. Combined with spatial weight allocation and 3D heatmap mapping, the time-series curve of the entire cavern anomaly index (MAI) is reconstructed. Feature gradient attribution extracts the single-channel feature contribution of a single voxel. DBSCAN / K-Means clustering is used to mine coupled precursor combinations, forming a standardized library of precursor patterns for rockburst and water inrush large deformation disasters. When the MAI exceeds a threshold, an automatic early warning is triggered. Historical anomaly samples are stored based on gradient attribution feature vectors. Two-level induction is performed: feature contribution ranking and multi-source coupled pattern clustering. The results are mapped to the cavern's 3D spatial coordinates, outputting disaster early warning and refined risk root cause analysis information.

[0009] Another aspect of this application is a system for multi-source state fusion and GAN anomaly precursor mining of deeply buried underground caverns, used to execute executable instructions to perform the above-mentioned method for multi-source state fusion and GAN anomaly precursor mining of deeply buried underground caverns.

[0010] According to another aspect of this application, an electronic device includes: a first processor; and a memory for storing executable instructions of the first processor; wherein the first processor is configured to execute the above-described method for multi-source state fusion and GAN anomaly precursor mining of deep underground caverns by executing the executable instructions.

[0011] According to another aspect of this application, a computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a second processor, implements the above-described method for multi-source state fusion and GAN anomaly precursor mining of deeply buried underground caverns.

[0012] The beneficial technical effects of this application are as follows: It pioneers the UCSM unified cavern state matrix model, solving the problem of multi-source heterogeneous data fusion. For monitoring data of six categories with huge differences in sampling frequency, dimension, and structure, including surrounding rock, hydrology, equipment, microseismic, environmental, and geological data, spatiotemporal alignment is achieved through three-dimensional voxel meshing, spatial interpolation, and temporal normalization. The data is then uniformly encapsulated into a standardized four-dimensional high-dimensional tensor, realizing integrated mathematical modeling of all parameters from local units to the entire cavern, thus solving the industry shortcomings of underground engineering monitoring systems, such as the lack of a unified computing data foundation and data silos.

[0013] A 3D-ConvGAN baseline model for normal operating conditions is constructed to achieve proactive early warning of disasters, breaking away from the traditional passive threshold alarm mode. By leveraging adversarial learning to uncover the spatiotemporal coupling evolution patterns of multiple parameters, a standard normal feature benchmark is obtained through training. The discriminator calculates a two-layer anomaly index of VAI voxels and MAI global domain through intermediate features, quantifying the deviation degree of each spatial unit at the millisecond level. This can capture weak precursors with synchronous, minute coupling of multiple indicators, resulting in a significantly earlier warning time than single-parameter over-limit alarms.

[0014] A comprehensive intelligent assessment mechanism encompassing quantification, attribution, and clustering has been established to facilitate risk tracing and the digital accumulation of precursor knowledge. Leveraging gradient attribution algorithms, the risk contribution of each monitoring indicator is analyzed to accurately locate abnormal areas and identify dominant risk factors. Combined with DBSCAN and K-Means clustering, multi-parameter coupled precursors such as rockbursts and water inrushes are automatically summarized, and a standardized precursor pattern library is dynamically updated. This transforms fragmented monitoring data into reusable digital knowledge, addressing the loss of human experience and shifting cavern safety management from post-event tracing and experience-based judgment to pre-event intelligent prevention and control.

[0015] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0016] Figure 1 This document illustrates a flowchart of a method for multi-source state fusion and GAN anomaly precursor mining of deeply buried underground caverns, provided in an embodiment of this application. Figure 2 A schematic diagram of the overall modular architecture of a multi-source state fusion and GAN anomaly mining system for deeply buried underground caverns; Figure 3 A time-series comparison diagram of the entire process of rockburst disaster development; Figure 4 A schematic diagram of a layered deployment architecture for a multi-source monitoring and analysis system for deeply buried underground caverns; Figure 5 The degree of convergence difference in accuracy of a GAN adversarial training model; Figure 6 This is a comparison chart of the performance of weak precursor identification and early warning. Detailed Implementation

[0017] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0018] In one implementation, Figure 1 The diagram illustrates a flowchart of a method for multi-source state fusion and GAN anomaly precursor excavation in deeply buried underground caverns according to an embodiment of this application.

[0019] S101 collects multi-source heterogeneous monitoring data in six categories: surrounding rock stability, environmental ventilation, hydrogeology, construction equipment, video personnel, and static geology. It executes a matrix construction process, unifies three-dimensional voxel meshing and four-dimensional tensor M(t) spatiotemporal reference, and generates a UCSM unified cavern state four-dimensional raster tensor.

[0020] In one implementation, this data acquisition process is adaptable to various underground excavation projects, such as deeply buried highway and railway tunnels, underground powerhouses of pumped storage hydropower stations, deeply buried chambers in high-stress mines, and deep test tunnels. It comprehensively collects six categories of raw monitoring information, with each category corresponding to dedicated sensor acquisition hardware and time-series recording rules. The specific collection targets are detailed below: For monitoring data on the stability dimension of the surrounding rock, the acquisition hardware includes a deep multi-point displacement monitoring instrument, anchor bolt and cable stress sensors, steel arch surface strain gauges, and a microseismic real-time monitoring system. Recorded data includes the time-series values ​​of displacement at various measuring points in the deep surrounding rock, real-time tensile stress of the anchor bolts, circumferential strain of the steel arch, energy released by microseismic events, and a continuous time-series sequence of microseismic b-values. When a microseismic event occurs, the three-dimensional coordinates of the seismic source are recorded simultaneously, providing coordinate basis for subsequent spatial mapping. For example, the long-term monitoring range of anchor bolt stress at a monitoring section of a 1200-meter-deep hydraulic tunnel is 12.5 MPa to 28.3 MPa, the daily average b-value of microseismic events is stable between 1.90 and 2.20, and the maximum daily increase in local surrounding rock deformation is 0.32 mm.

[0021] For monitoring data related to environmental ventilation, data acquisition hardware is deployed at various ventilation sections of the tunnel and at monitoring sections behind the tunnel face. This includes temperature and humidity sensors, dust concentration detectors, and sensors for CO, H2S, and radon toxic gases. The hardware continuously collects real-time values ​​of temperature, relative humidity, suspended dust concentration, and the concentrations of various harmful gases at each monitoring section, with a uniform sampling period of 1 minute. For example, under normal ventilation conditions, the tunnel temperature is maintained between 16.2℃ and 21.5℃, and the CO safety threshold is 25ppm. Real-time values ​​exceeding these thresholds are recorded and their timelines are fully documented.

[0022] For monitoring data in the hydrogeological dimension, the data acquisition hardware includes a pore water pressure gauge around the tunnel and a segmented flow rate metering device. This device continuously records the pore water pressure, instantaneous flow rate, and cumulative flow rate at each monitoring section, with a sampling cycle of 1 hour. For example, the stable pore water pressure monitoring value around the fault fracture zone is 0.42 MPa, while the pore water pressure in intact rock mass without faults is 0.08 MPa.

[0023] For monitoring data on the operating conditions of construction equipment, the system connects to the TBM tunneling main control system and the drill-and-blast trolley electrical control terminal to collect real-time data on the cutterhead thrust, cutterhead torque, and three-dimensional coordinates of the equipment's tunneling space. The TBM outputs a set of operating condition time-series data every 50 milliseconds, and the drill-and-blast trolley outputs a set of operating parameters every 2 seconds. For example, during normal tunneling, the stable range of the TBM cutterhead thrust is 12500kN~14800kN, and the cutterhead torque is 2150kN. m, the device coordinates synchronously record the spatial points (xc, yc, zc).

[0024] For video monitoring data on personnel, the high-definition video stream from the entire cavern is integrated with an AI-powered structured analysis program. This program identifies in real-time rockfalls, support damage, and personnel encroachment violations within the footage, outputting three types of structured time-series indicators: personnel density, rockfall risk index, and equipment malfunction index. The video analysis output cycle is 30 seconds. For example, the personnel density threshold for a single monitoring section is 0.12 people / m²; exceeding this threshold automatically retains the corresponding time-period structured feature data.

[0025] For static geological data, digital results from the pre-excavation geological survey are imported, including the spatial distribution coordinates of fault fracture zones in the tunnel, rock mass integrity coefficient Kv, and fixed parameters such as uniaxial compressive strength (UCS). These parameters do not change with the excavation process and are fixedly bound to the corresponding tunnel space area throughout the entire process. For example, Kv=0.85 for intact granite rock mass and Kv=0.22 for fault fracture zone sections, corresponding to uniaxial compressive strengths of 86MPa and 18MPa, respectively.

[0026] The collected data is divided into two categories: dynamic time-series data and static geological parameters. The five categories of data—surrounding rock, environment, hydrology, equipment, and video—are all dynamic time-series data, independently archived and stored according to their respective hardware sampling cycles. Faults, rock mass integrity, and rock strength are static parameters, entered into a spatial coordinate matching library, and permanently bound to their corresponding spatial ranges. A unified adaptation to Modbus, OPCUA, and proprietary data transmission protocols of various tunneling equipment manufacturers is implemented to eliminate data silos caused by incompatibility between different monitoring subsystems. For the 1200-meter-deep hydraulic tunnel, four types of acquisition terminals are simultaneously deployed along the tunneling route: surrounding rock stress sensor array, tunnel gas monitoring section, TBM tunneling main control data interface, and four high-definition cameras at the tunnel face. All monitoring time-series and static geological parameters are collected according to the above six-category classification rules, fully covering all data types within the entire tunnel monitoring range.

[0027] Based on the tunnel design and excavation outline, the 50m excavation section ahead of the tunnel face, the high-stress concentration tunnel section, and the fault fracture zone section were designated as key areas for intensive monitoring. The remaining conventional tunnel sections were designated as standard monitoring areas, and all areas were uniformly included in the gridded subdivision. A unified three-dimensional scale of length L, width W, and height H was used to uniformly divide the designated monitoring areas into grids. After division, each independent three-dimensional spatial unit was defined as a basic voxel (x, y, z). All monitoring index calculations and spatial interpolation assignments used a single voxel as the smallest statistical carrier. For example, the tunnel face monitoring section has a transverse width of 20m, a longitudinal excavation extension length of 50m, and a vertical tunnel height of 12m. A uniform 2m grid division scale was selected; the transverse section could be divided into 10 columns, the longitudinal section into 25 segments, and the vertical section into 6 layers, forming 10×25×6 sets of independent three-dimensional voxels, each set of voxels corresponding to unique spatial coordinates (x, y, z). Each voxel after subdivision is assigned a unique spatial coordinate identifier. All subsequent sensor interpolation, feature channel numerical mapping, and anomaly index calculation are based on this identifier to ensure that the spatial data correspond one-to-one.

[0028] Two spatial extrapolation algorithms, Kriging interpolation and inverse distance weighted interpolation, are used to extrapolate the monitoring values ​​of discrete point sensors to all volume elements within the region, filling the data gaps in areas without sensor deployment. For example, if only three pore water pressure gauges are deployed longitudinally along the sidewall of a cavern, with a longitudinal spacing of 15m between the three gauges, the interpolation algorithm is used to extrapolate the pore water pressure values ​​corresponding to all 2m scale volume elements within the interval based on the measured water pressure values ​​at the three points, completely filling the monitoring data of the volume elements without measuring points between the two sections of pore water pressure gauges.

[0029] The original time-series data sampling period spans a wide range, including 50-millisecond TBM equipment data, 1-minute gas monitoring data, 1-hour pore water pressure data, and daily rock mass leveling measurement data. The physical dimensions of each indicator are completely different. A dimensionless normalization conversion is uniformly performed to map all dynamic time-series values ​​to the 0~1 standard value range, eliminating differences in units and amplitudes.

[0030] Static geological indicators such as fault distribution, rock mass integrity coefficient, and uniaxial compressive strength of rock do not change with time t. After gridding, they are directly and permanently bound to the corresponding spatial elements and are only reassigned when the cavern geological exploration data is updated. They do not participate in the temporal iterative update of each time step, realizing the hierarchical and independent management of dynamic monitoring data and static geological parameters. Interpolation is used to solve the problem of spatial data loss caused by sparse sensor deployment, normalize and unify the dimensions of multiple types of monitoring data, and store and manage dynamic and static parameters separately.

[0031] Using the decomposed 3D voxel (x, y, z) as the spatial basis, independent feature channel sets are constructed sequentially according to six categories: surrounding rock stability, environmental ventilation, hydrogeology, construction equipment, video personnel, and static geology. These six channels are merged to form an N-dimensional feature vector F. Each voxel corresponds to a complete N-dimensional feature vector at any time t. For example, channels 1 to K1 are assigned all features of the surrounding rock stability category; channels K1+1 to K2 are assigned features of the environmental ventilation category; channels K2+1 to K3 are assigned hydrogeological features; channels K3+1 to K4 are assigned features of the construction equipment category; channels K4+1 to K5 are assigned features of the video personnel category; and channels K5+1 to K6 are assigned permanent static geological features.

[0032] A unified time-series update step size is set to capture the dynamic monitoring data after normalization, and the values ​​of all voxels and all channels in the entire domain at the same moment are integrated into a snapshot of the cavern's holographic state at that moment. For example, if the time-series update step size is set to 10 minutes, the system will automatically read a round of six-channel monitoring values ​​of all voxels in the entire domain every 10 minutes to generate a complete state dataset for a single moment.

[0033] By integrating the three-dimensional spatial dimensions L, W, H, and N, the feature channel dimension, and the time dimension t, a unified four-dimensional raster tensor M(t) for the cavern is generated. The mathematical expression of the tensor is as follows: In the formula, L represents the number of cells in the length direction of the hole grid, W represents the number of cells in the width direction, H represents the number of cells in the height direction, N represents the total number of feature channels, and M(t) is the complete holographic state matrix of the hole chamber at time t.

[0034] The tensor input-output docking rules are as follows: the input content of the whole process is a set of gridded volumetric spatial coordinates and six types of monitoring datasets after interpolation and normalization; the output is a standardized four-dimensional tensor M(t), which can directly divide the time window to generate time series samples Xseq, which can be used as the standard training input of the 3D-ConvGAN model to complete the integrated mathematical modeling of multi-source heterogeneous monitoring data.

[0035] S102, relying on Kriging / inverse distance weighted spatial interpolation, temporal normalization alignment algorithm and static geological fixed channel mapping, constructs six types of feature channels. It uses a single volume element (x,y,z) as the smallest statistical unit to calculate the monitoring indicators of various sensors. It constructs feature tensors for six channels: surrounding rock, environment, hydrology, equipment, video and geology, and simultaneously completes temporal window isolation of normal / disaster working condition samples.

[0036] In one implementation, based on six heterogeneous monitoring data sources—surrounding rock, environmental ventilation, hydrogeology, construction equipment, video personnel, and static geology—it follows four technical requirements: L×W×H three-dimensional global voxel meshing, multi-source monitoring data interpolation, cross-frequency time-series dimension normalization, and fixed mapping of static geological parameter channels. It also adheres to the execution principles of minimum statistical calculation of a single voxel (x,y,z), independent modeling of the six types of monitoring features by channel, and partitioning and isolating time-series working condition samples. The process involves three-dimensional meshing, spatial interpolation, dimensionless time-series normalization, static geological channel binding, multi-channel four-dimensional tensor encapsulation, and time-series sample windowing to collaboratively construct a unified cavern state matrix four-dimensional tensor. All data input sources for the unified cavern state four-dimensional tensor are divided into six categories: surrounding rock stability monitoring data, environmental ventilation monitoring data, hydrogeological monitoring data, construction equipment operation data, structured video personnel data, and static geological fixed parameters. All calculation steps uniformly implement four types of hard technical constraints and three types of standardized execution criteria. The entire calculation is completed by six continuous processing steps.

[0037] The L×W×H global three-dimensional voxel meshing constraints are as follows: all monitoring areas of the tunnel are uniformly divided using a unified length, width, and height scale, and the resulting independent spatial units serve as the calculation carriers for all subsequent monitoring indicators. For example, a monitoring section at the face of a 1200-meter-deep hydraulic tunnel is selected. The section has a lateral width of 20 meters, a longitudinal excavation monitoring length of 50 meters, and a vertical tunnel height of 12 meters. A uniform 2-meter meshing standard is used throughout the section. After division, 10 columns of spatial units are formed laterally, 25 segments of spatial units are formed longitudinally, and 6 layers of spatial units are formed vertically. All monitoring data in the entire region are subsequently matched to spatial units of this scale for calculation.

[0038] The interpolation constraints for multi-source monitoring data are as follows: for monitoring values ​​from various sensors located only at discrete points, a spatial interpolation algorithm is used to extrapolate to all spatial units across the entire region, filling data gaps in sensor-free areas. For example, if three pore water pressure acquisition devices are deployed longitudinally on the tunnel wall at 15-meter intervals, only three points of water pressure monitoring values ​​can be obtained. Through interpolation calculations, the pore water pressure values ​​corresponding to all two-meter-scale spatial units between two devices are extrapolated, achieving complete coverage of data across the entire monitoring section.

[0039] The cross-frequency time series normalization constraint is as follows: for various dynamic monitoring time series data with significantly different sampling periods, a unified dimensionless conversion is performed to eliminate computational interference caused by different physical units and numerical ranges. For example, construction equipment monitoring data is collected every 50 milliseconds, tunnel environmental gas data is collected every minute, tunnel water pressure data is recorded every hour, and rock mass deformation measurement is collected on a daily basis. The original dimensions and numerical ranges of these four types of data are completely different. All of them are uniformly mapped to the zero-to-one interval to complete the standardization process and achieve time series synchronous matching calculation.

[0040] The static parameter channel has the following fixed mapping constraints: geological indicators that do not change with the excavation process, such as rock mass integrity, fault distribution, and rock strength, are assigned separate dedicated data channels and are bound to the corresponding spatial units throughout the process. They do not participate in each round of time-series update calculations. For example, the rock mass integrity value for a complete granite area is fixed at 0.85, and the rock mass integrity value for a fault fracture zone area is fixed at 0.22. These two types of geological indicators are permanently bound to all spatial units within their corresponding ranges, and the values ​​within this channel will not be changed during the dynamic data update operation every ten minutes.

[0041] The three types of standardized implementation guidelines and corresponding engineering examples are as follows. For the single-element minimum statistical accounting guideline, all interpolation operations, feature assignments, time-series matching, and anomaly calculations use a single length, width, and height spatial unit as the minimum statistical carrier, and monitoring indicators are not merged across units. For example, the anchor stress, microseismic indicators, and environmental values ​​within each two-meter-scale independent spatial unit are calculated independently, and data from adjacent units are not mixed, ensuring the accuracy of spatial anomaly location.

[0042] Six types of monitoring data are each assigned a dedicated set of feature channels. The calculation, updating, and storage of each type of data do not interfere with each other, and fusion is only completed in the tensor encapsulation stage. For example, surrounding rock related indicators are assigned to channels 1 to K1, environmental ventilation indicators to channels K1+1 to K2, hydrological indicators to channels K2+1 to K3, equipment operating conditions to channels K3+1 to K4, video analysis indicators to channels K4+1 to K5, and static geological parameters to channels K5+1 to K6. The six sets of channels are independently interpolated and normalized before being fused and encapsulated.

[0043] The time-series working condition sample partitioning and isolation criteria are as follows: After completing the time-series slicing, samples are archived and stored separately according to two types of working conditions: stable construction and disaster precursors. The two types of samples participate in network learning independently. For example, a single time-series window duration of 120 minutes is set, and continuous monitoring time series are divided into independent samples according to the window. The system automatically identifies stable tunneling periods without rock bursts or water inrush risks to generate normal samples, and identifies periods of continuous decrease in microseismic b-values ​​and water pressure pulse fluctuations to generate precursor samples. The two types of samples are retrieved separately for adversarial training of the 3D convolutional recurrent generative adversarial network.

[0044] First, a three-dimensional grid (length × width × height) is created to generate basic voxel units in the cavern monitoring area. Then, using Kriging interpolation and inverse distance weighted interpolation algorithms, monitoring indicators from various sparse sensors are assigned to corresponding (x, y, z) voxels across the entire domain. For multi-source time-series data spanning milliseconds to daily levels and with varying sampling frequencies, a unified dimensionless normalization and alignment process is performed. Geological model parameters that do not change with time t are permanently bound to the corresponding static channels of the voxels, completing the hierarchical construction of six categories of independent feature channels. The entire feature channel construction operation consists of four consecutive execution steps. All operations are carried out based on a uniform grid spatial reference (length × width × height), with each step accompanied by a corresponding engineering implementation example. The calculation results of each step are sequentially passed down, collectively completing the hierarchical construction of independent channels for the six types of monitoring data.

[0045] The first step is to uniformly divide the entire tunnel into three-dimensional grids, defining the entire monitoring range of the tunnel, with a focus on high-risk monitoring areas such as the tunnel face, high-stress sections, and fault fracture zones. A uniform length and width scale is used to uniformly divide the monitoring space, generating standardized independent spatial units. All sensor monitoring values ​​and geological parameters will subsequently be carried out using these spatial units. For example, in a 1200-meter-deep hydraulic tunnel face monitoring section with a 20-meter lateral width, a 50-meter longitudinal extension, and a 12-meter vertical tunnel height, a uniform 2-meter dimension is used to complete the full-area grid division. All the resulting spatial units will serve as the minimum statistical carrier for subsequent interpolation, normalization, and channel assignment.

[0046] The second step involves spatial interpolation of various sparse monitoring indicators. Two spatial extrapolation algorithms are selected: Kriging interpolation and inverse distance weighted interpolation. Monitoring values ​​from sensors with only a limited number of locations can be extrapolated outwards across the entire area using these algorithms, completely filling all spatial units within the grid and eliminating data gaps caused by sparse sensor deployment. For example, if three pore water pressure acquisition devices are placed longitudinally on the sidewall of a cavern with only 15-meter intervals, only three discrete points can be obtained. By using interpolation algorithms to extrapolate the water pressure values ​​for all two-meter-scale spatial units between the two devices, the data for the entire monitoring section without measurement points can be completely supplemented.

[0047] The third step involves completing the dimensionless normalization and alignment of cross-frequency time-series data. For five types of dynamic time-series monitoring data—equipment, environment, hydrology, surrounding rock, and video—a unified dimensionless conversion process is performed to eliminate differences in physical units, numerical ranges, and acquisition cycles. All time-series values ​​are mapped to the same standard value range, enabling synchronous matching and calculation of different time-series data. For example, tunneling equipment monitoring data is collected every 50 milliseconds, tunnel gas monitoring data every minute, pore water pressure every hour, and rock mass deformation with a daily acquisition cycle. These four types of data have completely different dimensions and numerical ranges. After unified conversion, they all fall within the zero-to-one range and can be synchronously participated in subsequent time-series slicing and network input calculations.

[0048] The fourth step involves completing the fixed binding of dedicated channels for static geological parameters and the layered construction of six types of channels. Geological indicators that do not change with excavation time, such as rock mass integrity, fault distribution, and uniaxial compressive strength of rock, are separately divided into dedicated feature channels and permanently bound to all spatial units within the corresponding range. Dynamic time-series data updates will not alter the values ​​within these channels. Independent channel sets are sequentially divided according to six categories: surrounding rock, environment, hydrology, construction equipment, video personnel, and static geology. After interpolation and normalization processing, each of the six types of channels achieves layered isolation. For example, the rock mass integrity value is fixed at 0.22 in the fault fracture zone area and 0.85 in the intact granite area. These two types of geological parameters are permanently bound to all spatial units within the corresponding range. The system's dynamic data update operation every ten minutes will not modify the stored content of this channel. The surrounding rock related indicators are classified into channels 1 to K1, environmental indicators into channels K1+1 to K2, hydrological indicators into channels K2+1 to K3, equipment operating conditions into channels K3+1 to K4, video analysis indicators into channels K4+1 to K5, and static geological parameters into channels K5+1 to K6. The six groups of channels are operated independently without interference.

[0049] Using a three-dimensional (x,y,z) volume element as the smallest calculation unit for the entire process, an N-channel four-dimensional holographic tensor M(t) is encapsulated according to the categories of surrounding rock, environment, hydrology, equipment, video, and geology. Simultaneously, a time series sample set is generated by dividing the continuous four-dimensional tensor time series into fixed time windows. This completes the block isolation of samples for normal and stable construction conditions and abnormal conditions with disaster precursors, providing standardized time series training samples Xseq for the input of the 3D-ConvGAN model. The entire operation involves pre-processing steps such as 3D mesh generation, interpolation normalization, and static channel binding. It consists of two main consecutive execution steps, with the output content directly serving as the standard input material for a 3D convolutional recurrent generative adversarial network. Specifically, the first step involves fusing six-category feature channels to construct a four-dimensional holographic tensor. The uniformly divided spatial units serve as the smallest carrier for all indicator calculations. Continuous feature channels are independently divided according to the six major monitoring data categories. After all channels are integrated, a unified four-dimensional holographic tensor is formed. The tensor simultaneously carries spatial dimensions (length, width, and height), multiple feature channel dimensions, and temporal dimensions, completely recording all monitoring information of the entire cavern at a single moment. The system generates a complete snapshot of the entire tensor every ten minutes, serving as the original data for the temporal slice. For example, the first to K1 interval channels uniformly record all monitoring indicators related to surrounding rock stability; the K1+1 to K2 interval channels record environmental ventilation monitoring indicators; the K2+1 to K3 interval channels record hydrogeological monitoring indicators; the K3+1 to K4 interval channels record construction equipment operating condition indicators; the K4+1 to K5 interval channels record structured indicators for video personnel; and the K5+1 to K6 interval channels permanently store static geological fixed parameters. The integrated four-dimensional holographic tensor is uniformly and standardizedly expressed as M(t), with tensor dimensions covering four categories: spatial length, spatial width, spatial height, and all features. Each spatial unit within the tensor corresponds to a complete set of six types of feature values, achieving one-to-one binding of spatial location with multi-source monitoring data, eliminating the problem of data fragmentation between different monitoring systems, and providing a standardized and unified data carrier for time-series slicing and network input.

[0050] The second step involves completing the time-series window segmentation and the classification and storage of working condition samples. For the continuously generated four-dimensional tensor snapshot data stream, a fixed duration is defined as a single time-series extraction interval, and the data is segmented to generate independent time-series samples. Based on the construction and operation status, samples are divided into two categories: stable, disaster-free samples and disaster precursor anomaly samples, which are archived separately. These two types of samples are independently fed into the network for adversarial iterative training. For example, a uniform 120-minute time-series window is set as the extraction duration for a single set of samples. The system automatically identifies tunneling periods without microseismic fluctuations, pore water pressure pulses, and stable surrounding rock deformation rates to generate stable working condition samples; it also identifies periods where microseismic b-values ​​continuously decrease and local pore water pressure exhibits periodic pulse fluctuations to generate precursor anomaly samples. These two types of samples are stored in separate databases and retrieved separately for model training.

[0051] The input and output specifications of the model are as follows: Input settings include two types of time-series samples stored separately and uniformly named Xseq, which are directly fed into a 3D convolutional recurrent generative adversarial network (GAN) as shared input material for both the generator and discriminator. The network consists of two independent computational branches: generation and discrimination. The generation branch uses a multi-layered 3D convolutional recurrent hierarchical cascaded structure, receiving random noise input to replicate the normal spatiotemporal evolution sequence of the cavern. The discrimination branch uses a 3D convolutional combined with a long short-term memory hierarchical cascaded architecture, simultaneously receiving real time-series samples and generating simulated samples, outputting normal operating probability values ​​in the range of 0 to 1. During the training phase, stable and abnormal samples are alternately retrieved to conduct adversarial games, continuously reducing the data distribution difference between real and simulated time-series samples until the model converges. After convergence, all parameters of the discrimination branch are locked to form a dynamic operating condition judgment benchmark for deeply buried caverns, used for subsequent real-time monitoring data anomaly identification.

[0052] S103 employs a 3D-ConvGAN dual-branch generator to extract spatiotemporal coupled evolutionary features. It replicates normal dynamic sequences through a 3D convolutional recurrent generator and outputs normal operating condition probabilities through a 3D-Conv-LSTM discriminator, thereby converging and solidifying the normal operating condition discrimination benchmark in adversarial game theory.

[0053] In one implementation, combining the characteristics of four-dimensional temporal tensor samples (Xseq) from deeply buried caverns with the 3D-ConvGAN model architecture specifications, random noise input mapping rules and spatiotemporal sequence feature extraction logic are introduced to perform feature analysis on real normal working condition time-series samples, generating multi-source coupled spatiotemporal feature adaptation information. The entire sample feature analysis process matches the inherent spatial-temporal structure of the four-dimensional temporal tensor samples (Xseq) from deeply buried caverns, fully conforms to the 3D-ConvGAN standard model architecture specifications, and follows the random noise independent input mapping rules and spatiotemporal hierarchical feature extraction logic throughout the process. This step is a preprocessing stage before network modeling, only completing the decomposition of multi-parameter coupling relationships within the sample, without incorporating network layer operations such as 3D convolution or LSTM loops.

[0054] The entire process comprises two independent input data streams. The two types of data are not numerically spliced ​​or fused. Specifically, the standard normal working condition time series sample Xseq is segmented and spliced ​​from a unified four-dimensional holographic tensor M(t) within a fixed time window. The dimensions include five dimensions: cavern space length L, space width W, space height H, multiple feature channels N, and continuous time series. All samples undergo pre-processing window isolation, selecting only construction time series with no rock bursts, no water inrushes, and stable surrounding rock deformation, while eliminating abnormal segments such as continuously decreasing microseismic b-values ​​and pore water pressure pulse fluctuations. For example, the cavern space uniformly adopts a 2m voxel partitioning standard. A single time series window contains multiple sets of M(t) holographic snapshots updated every 10 minutes. The six monitoring indicators are sequentially divided into independent feature channels K1 to K6, with static geological channel values ​​remaining fixed throughout.

[0055] For random noise tensors, the noise tensor dimension is exactly the same as Xseq. It is used only as independent input material for the 3D convolutional loop generator and is not merged with real sample features for feature parsing. For example, the noise tensor value range adopts conventional uniformly distributed values ​​with no fixed spatiotemporal evolution trend, and is only used to generate the dynamic sequence of the simulated cavern.

[0056] Using a single complete Xseq time series as the processing object, the synchronous evolution correlations of spatial, temporal, and cross-channel types are decomposed hierarchically to extract multi-parameter coupling adaptation information: Step 1, Single volume element spatial feature analysis: For each (x,y,z) spatial unit, the synchronous monitoring values ​​of six channels under the same holographic snapshot are extracted to establish spatial multi-parameter correspondence. For example, when the same spatial volume element collects three dynamic indicators—surrounding rock stress, pore water pressure, and TBM thrust—at the same time, the values ​​remain stable in a long-term range without synchronous sudden increases or decreases, forming a stable spatial coupling feature group. Step 2, Continuous temporal trend analysis: The long-term change law of multi-channel values ​​of the same spatial unit is extracted along the time axis to identify long-term synchronous fluctuation characteristics. For example, during a stable tunneling period, the temporal changes of three indicators—target body microseismic b-value, tunneling equipment thrust, and tunnel temperature and humidity—are synchronous and gradual, without abrupt coupling phenomena. Step 3, Cross-category dynamic channel coupling extraction: Only the five types of dynamic temporal parameters—surrounding rock, environment, hydrology, equipment, and video—are analyzed for correlation; static geological channel fixed values ​​are not included in the coupling calculation. For example, the static rock mass integrity coefficient of the fault fracture zone is fixed and does not change with time; the dynamic indicators of surrounding rock stress, water inflow, and tunneling torque fluctuate synchronously, forming multi-source coupled characteristic information.

[0057] Combining the needs of dynamic construction evolution analysis of underground caverns with the convergence specifications of adversarial training, a dual-branch allocation modeling logic is used for spatiotemporal feature adaptation information. This involves dividing the model into a 3D convolutional loop generator replication link and a 3D-Conv-LSTM discriminator identification link, configuring a sequence authenticity comparison and correlation mining mechanism, and generating branch-specific feature allocation information. Utilizing the multi-source coupled spatiotemporal feature adaptation information output from the time-series sample feature analysis process, and based on the needs of dynamic evolution analysis throughout the entire cycle of underground cavern excavation, support, and disturbance, and matching the GAN adversarial iterative convergence constraints, standardized coupled features are split into two independent computational links. A difference comparison and mining mechanism between real and simulated sequences is simultaneously configured to generate branch-specific feature allocation information that matches the computational logic of each adapter generator and discriminator. The entire link architecture is perfectly matched to the four-dimensional time-series sample Xseq dimension. For example, the window duration for a single time series sample is 120 minutes, the spatial partitioning scale is two meters, and the sample dimension is uniformly set as length × width × height × feature channel × time series step. All input and output data in the dual branches maintain this dimensional specification, and there is no dimension scaling or clipping operation.

[0058] The complete architecture and operational rules of the 3D convolutional loop generator replication link are as follows: the entire generation link is built by sequentially connecting multiple layers of 3D convolutional units and temporal loop units. The 3D convolutional units are responsible for extracting the coupling distribution patterns of multiple features in the entire spatial domain of the cavern, while the temporal loop units are responsible for capturing the synchronous evolution trends of surrounding rock, hydrology, and equipment parameters over continuous time. The two types of units are stacked alternately to form a complete generation network. The only input to the link is an independent random noise tensor. The length, width, height, feature channels, and time step number of the noise tensor are completely aligned with the real time-series sample Xseq, without mixing in any measured cavern monitoring data. The link finally outputs a simulated cavern dynamic sequence Xfake. The numerical range of each dimension and the channel division rules of Xfake are consistent with the real sample Xreal. For example, the real sample feature channels are divided into six groups from K1 to K6. Noise and simulation sequences are set with the same six groups of channels. The static geological channel values ​​in the simulation sequence are fixed to match the constant parameters of rock mass integrity and fault distribution. The dynamic channel replicates the smooth fluctuation characteristics of TBM thrust, micro-seismic b-value, and pore water pressure under stable tunneling, without pulse-like abrupt changes in values.

[0059] The complete architecture and operational rules of the 3D-Conv-LSTM discriminator identification link are as follows: the entire discriminator link consists of a 3D convolutional layer and a long short-term memory (LSM) temporal layer connected sequentially. The 3D convolutional layer extracts the multi-channel feature distribution of voxels in the entire space within a single temporal snapshot, while the LSM layer captures the feature evolution trend within a continuous temporal sequence. The two layers work together to distinguish between true and false sequences. The link synchronously receives two types of input data: the first type is real normal working condition samples (Xreal) isolated by temporal windows, and the second type is the simulated sequence (Xfake) output by the generator. The link outputs a normal working condition probability scalar with a value ranging from 0 to 1. The closer the value is to 1, the closer the input sequence is to the standard disaster-free evolution baseline. For example, when inputting a real sample of stable tunneling, the discriminator output probability is stable in the range of 0.92 to 0.98; when inputting an abnormal temporal sequence with continuously decreasing microseismic b-values ​​and pore water pressure pulse fluctuations, the output probability drops to below 0.35; when inputting a distorted simulated sequence generated in the non-convergence stage, the output probability remains in the range of 0.4 to 0.6.

[0060] After the discriminator extracts features from two types of input sequences, it calculates the difference between real and simulated features in the same space, sequence, and channel, generating a difference feedback signal. This feedback signal is then passed back to all layers of the 3D convolutional recurrent generator, updating the weight parameters within each convolutional and recurrent unit layer by layer. This process of repeated noise input, simulated sequence generation, feature difference calculation, and weight update completes one round of iterative game. Multiple iterations continue until the difference in feature distribution between the real and simulated sequences stabilizes and reaches the convergence criterion. At this point, parameter updates are stopped, and the weights of all discriminator layers are locked. For example, after a single iteration, if the average difference of features across the entire domain decreases by 0.8 to 1.2 units, and the difference fluctuation is less than 0.02 for twenty consecutive iterations, the model is considered converged, and the bi-branch parameter iteration operation is terminated.

[0061] In deeply buried caverns, disasters such as rock bursts and water inrushes do not present a single indicator exceeding the warning threshold; the risks are mostly manifested as synchronous micro-hourly time-series deviations of multiple monitoring parameters. The generation branch relies on random noise to replicate the standard spatiotemporal evolution baseline without disasters, fully reproducing the long-term coordinated changes in surrounding rock stress, water inflow, construction equipment, and environmental parameters during stable tunneling. The discrimination branch relies on convolution and temporal hierarchical mining of multi-dimensional feature distribution differences to quantify the degree of deviation between the real-time monitoring sequence and the normal baseline. The two links work together to complete adversarial learning, and the final solidified discriminator can directly interface with real-time four-dimensional tensor data streams to identify cross-system weakly coupled precursors that traditional single-point threshold alarms cannot capture, supporting advanced disaster assessment.

[0062] Based on the GAN adversarial iterative training standard, normal working condition probability output rules, and spatiotemporal evolution coupled learning specifications, game-like iterative computation is performed on branch feature allocation information to complete the differential fitting between the real tunnel dynamic sequence and the fake simulation sequence. After the model converges, the discriminator is solidified to form a benchmark model for the dynamic construction conditions of deeply buried underground caverns. The entire set of dual-branch network iterative computation strictly follows three types of established execution specifications: GAN standard adversarial iterative process, normal working condition probability output constraints in the 0 to 1 interval, and multi-dimensional spatiotemporal coupled feature synchronous learning requirements. The computational material is the branch-specific feature information allocated to the two modeling links in the early stage. The core objective of the computation is to continuously compress the feature distribution difference between the real cavern time series and the simulation sequence until the preset convergence judgment criteria are reached, the discriminator parameters are locked, and a benchmark model for judging the dynamic construction conditions of deeply buried caverns is generated. The entire training process follows the dual-branch network construction process and outputs a solidified discriminator network, providing a judgment basis for subsequent voxel and global anomaly quantification stages.

[0063] A single complete adversarial iteration consists of four consecutive computational steps, with all weights of the generator and discriminator updated synchronously in each round, as follows: Step 1: A uniform random noise tensor is input into the 3D convolutional recurrent generator. The noise's width, height, feature channels, and temporal step number are completely consistent with the temporal sample Xseq, without mixing in actual monitoring data. Step 2: The generator outputs a simulated cavern dynamic sequence Xfake based on the noise. The simulated sequence's six-channel classification and static geological parameter assignment rules are consistent with the real sample Xreal. Step 3: The real temporal sample and the generated simulated sequence are simultaneously fed into the 3D-Conv-LSTM discriminator to simultaneously extract the spatial and temporal coupling features of the two types of sequences. Step 4: The output probabilities of the two types of sequences are compared, the difference is calculated, and backpropagated to all layers of the generator and discriminator, updating the network weights layer by layer to complete one iteration.

[0064] For example, a single iteration uses a 120-minute time window sample, a spatial partitioning scale of two meters, and a single monitoring area is divided into 10 columns, 25 segments, and six groups of spatial units. The channels are divided into six categories of monitoring indicators according to K1 to K6. The output probability of the real sample in this round is 0.94, the output probability of the simulated sequence is 0.51, the probability difference is 0.43, and the difference signal updates the network parameters layer by layer.

[0065] The discriminator outputs only a single scalar value, with a fixed range of 0 to 1. The magnitude of the value represents the degree to which the input time series conforms to the normal construction baseline. The closer the value is to 1, the more synchronous the fluctuations of multiple parameters and the more the spatial distribution conforms to a stable state without disasters. A continuously low value indicates the presence of precursory characteristics of multi-parameter coupling. For example, after inputting a standard time series for stable tunneling, the output stabilizes in the range of 0.91 to 0.98. For precursory time series inputs with continuously decreasing microseismic b-values ​​and pore water pressure pulse fluctuations, the output drops to 0.28 to 0.42. For simulation sample inputs during the non-convergence phase of training, the output is concentrated in the range of 0.4 to 0.6.

[0066] In each iteration, the difference in the global feature distribution between the real and simulated sequences is recorded. If the difference consistently falls below a preset threshold for multiple iterations, convergence is determined, all weight updates are terminated, and all convolutional and long short-term memory layer parameters of the discriminator are permanently locked. For example, with a convergence threshold of 0.02, if the average global feature difference remains between 0.012 and 0.019 for twenty consecutive iterations, the convergence condition is met. The accompanying convergence curve shows that the global feature difference is close to 40 in the first iteration, and decreases to 0.35 in the twentieth iteration, reaching the convergence standard. If the difference is higher than 0.02 for five consecutive iterations, iterative computation continues.

[0067] The discriminator, with its parameters locked, serves as a unified benchmark for determining operating conditions across the entire domain and can be directly connected to the real-time generated four-dimensional time-series samples (Xseq). Traditional monitoring relies solely on a single indicator with a fixed threshold for alarms, only able to identify obvious disasters caused by significant parameter exceedances, but unable to capture minute synchronous deviations across systems. This model, however, relies on full-dimensional spatiotemporal coupled learning to identify hidden precursors through multi-dimensional feature distribution deviations. For example, after real-time time-series data from a 1200-meter-deep hydraulic tunnel is input into the discriminator, the output probability is 0.36. The system identifies precursors such as small oscillations in TBM thrust, slow decay of the micro-seismic b-value, and slight uplift of surrounding rock stress coupled with rockburst. It can push risk information 24 to 48 hours in advance, allowing the construction end to adjust the tunneling speed and strengthen support in a timely manner to avoid disasters.

[0068] S104 is configured with L2 distance of voxels and weighted double-layer anomaly quantization of the entire cavern space. It models the deviation of multi-dimensional modes through the convolutional feature space of the discriminator, and combines spatial weight allocation and three-dimensional heat map mapping to reconstruct the time series curve of the anomaly index MAI of the entire cavern.

[0069] In one implementation, combining the real-time four-dimensional temporal sample input of deeply buried caverns with the intermediate convolutional feature output characteristics of the 3D-ConvGAN discriminator, and relying on the voxel deviation evaluation standard, a dual computation mechanism is introduced: L2 distance calculation in the single voxel feature space and weighted summation of the global voxel space. This establishes a two-level hierarchical quantification system of VAI voxel anomaly index and MAI full cavern anomaly index. The entire anomaly quantification process is built on a 3D-ConvGAN discriminator network that has completed adversarial training and whose parameters are permanently locked. All computational materials are intermediate convolutional feature data extracted after real-time four-dimensional temporal sample input to the discriminator. The overall computational design employs a two-layer progressive computational logic. The first layer measures the degree of local deviation for a single three-dimensional spatial unit, while the second layer integrates all units across the entire domain to obtain the overall risk level. These correspond to two types of quantitative indicators: the VAI (Volume Element Anomaly Index) and the MAI (Mass All-Cavity Anomaly Index). The entire computation is carried out based on two established computational rules: L2 distance in the volume element feature space and spatial weighting. The computational benchmark is completely consistent with the previous three-dimensional mesh partitioning and six-channel four-dimensional tensor construction process, sharing the same L×W×H spatial scale and K1 to K6 feature channel division standard. There are no issues with misalignment of data scale or feature dimensions.

[0070] After inputting the real-time four-dimensional time-series samples of the cavern, the activation information output from the intermediate convolutional layer of the solidified discriminator is directly read and split to generate real-time working condition feature maps. Then, standard normal holographic snapshots are selected from a large number of disaster-free stable time-series samples stored during the training phase and input into the same discriminator network to extract corresponding standard baseline feature maps. The two types of feature maps strictly reuse the previous mesh division and channel configuration standards, and the number of spatial units in length, width, and height, as well as the arrangement order of the six types of feature channels, are completely consistent to ensure the consistency of the baseline for subsequent coordinate-by-coordinate matching calculations. For example, a key monitoring section at the face of a 1200-meter-deep hydraulic tunnel was selected. The grid was uniformly divided using a two-meter unit side length, with twenty-five units along the tunnel's length, ten units in the horizontal width, and six layers in the vertical height, forming a complete 10×25×6 spatial unit array. The feature channels were fixed into six segments: K1 to K1 for surrounding rock, K1+1 to K2 for environmental, K2+1 to K3 for hydrological, K3+1 to K4 for equipment, K4+1 to K5 for video, and K5+1 to K6 for static geological. Both real-time feature maps and baseline features adopted this spatial and channel configuration. Each pixel in the map corresponds to a unique (x, y, z) three-dimensional volumetric coordinate, preventing matching deviations such as inconsistent spatial unit numbers or disordered channel order.

[0071] Based on the real-time feature map and the baseline feature map that have been matched one by one, and using a single (x,y,z) voxel as the smallest calculation unit, the real-time full-channel feature vector and the baseline full-channel feature vector under the coordinate are extracted. The difference between the two types of vectors is calculated by L2 Euclidean distance, and the result is the VAI voxel anomaly index of the voxel. In the formula This represents the full-channel feature vector of a real-time sample at the (x,y,z) voxel position. The vector represents the feature vector of all channels under the same coordinates as the standard reference sample. The subscript 2 indicates that the deviation of the feature space is measured by L2 distance. For example, in a fault fracture zone, the coordinates of a certain volume element are column 5, segment 12, and layer 3. The real-time surrounding rock stress and pore water pressure values ​​of the two dynamic channels at this location are consistently slightly higher than the reference sample. After substituting into the L2 distance formula and calculating the differences of all six channels, the VAI calculation result of this volume element is 0.68. In the intact, unstructured granite segment, the coordinates of column 3, segment 8, and layer 4 show that the values ​​of the six channels are highly consistent with the reference sample, and the corresponding VAI calculation value is only 0.07. The magnitude of the value can be used to intuitively distinguish the risk intensity of a single spatial unit.

[0072] After completing the VAI numerical calculations for all independent volume elements, differentiated spatial weighting coefficients are configured for different regions of the cavern. Higher weights are assigned to the tunnel face, fault fracture zones, and areas of high ground stress concentration, while lower weights are assigned to areas of intact conventional rock mass. The VAI of each volume element is multiplied by its corresponding spatial weight, and then summed across the entire domain to obtain the MAI (Maintenance Index) for the entire tunnel. The summary expression is as follows: For example, all volume elements within the 50-meter excavation range in front of the tunnel face are uniformly assigned a weight coefficient of 1.8, volume elements in ordinary intact tunnel sections have a weight of 0.6, and volume elements in fault fracture zones have a weight of 1.5. A total of 1500 sets of three-dimensional spatial elements were monitored in this section. The VAI values ​​of all volume elements were multiplied sequentially by their corresponding weights and then summed to obtain a total MAI value of 0.32 for the entire monitoring section. This value indicates that the overall risk of the entire monitoring area is currently in a low range.

[0073] After all VAI and MAI calculations are completed, a 3D anomaly heatmap is generated based on unified volumetric coordinates. Visualization levels are defined according to VAI values, and MAI values ​​are continuously recorded at ten-minute time-series update intervals. These data are then stitched together to generate a continuous cavern health time-series curve, simultaneously outputting two types of analytical materials: spatial positioning visualization results and global time-series quantification results. The entire quantification process is preceded by 3D-ConvGAN adversarial training convergence and discriminator parameter solidification. The output VAI and MAI data directly flow into the gradient attribution feature contribution analysis and disaster precursor clustering post-processing stages. The entire process shares the same spatiotemporal benchmark with the earlier multi-source tensor construction and network training stages, ensuring a unified data standard and no matching bias throughout the entire technical chain. For example, the system updates a snapshot of the M(t) four-dimensional holographic tensor every ten minutes, and simultaneously refreshes the corresponding volume element VAI heatmap and global MAI record points; the heatmap color scale division standard is that VAI values ​​less than 0.2 render green stable areas, 0.2 to 0.5 render yellow warning areas, and values ​​greater than 0.5 render red high-risk areas, which can intuitively lock the three-dimensional spatial units where rockburst and water inrush risks are concentrated.

[0074] Based on the real-time monitoring feature parsing execution specification, real-time operating condition feature maps are first extracted respectively. Compared with standard reference feature map Then, the (x, y, z) 3D voxel coordinates within the two sets of feature maps are aligned one by one. Single voxel feature deviation is calculated through coordinate matching, forming a coherent processing chain from feature extraction to single voxel deviation calculation. This process relies on the intermediate convolutional output of the 3D-Conv-LSTM discriminator with fixed parameters. The preceding steps have already extracted real-time working condition feature maps. Standard reference feature map The two feature maps share a unified L×W×H three-dimensional mesh division scale and a six-category feature channel arrangement rule from K1 to K6. All spatial units are assigned unique (x,y,z) three-dimensional coordinates. The same coordinate position in the two feature maps corresponds one-to-one. There is no misalignment of the number of spatial units or the order of feature channels. All calculations use a single (x,y,z) volume element as the smallest independent calculation carrier and do not merge feature values ​​across units to perform difference calculations.

[0075] The algorithm iterates through all three-dimensional spatial units of the feature map and matches the corresponding feature vectors in the real-time feature map and the reference feature map one by one according to the three-dimensional coordinates of length, width, and height. At the same (x,y,z) coordinate position, it simultaneously retrieves all feature values ​​of six channels: surrounding rock, environment, hydrology, construction equipment, video personnel, and static geology. The two sets of feature vectors with the same coordinates and channels are extracted separately and used as two sets of input vectors for the L2 distance algorithm. The feature vectors of different spatial units are isolated from each other and complete the calculation independently without interfering with each other.

[0076] The formula for calculating the single-element anomaly index is as follows: , within the formula, This represents the complete six-channel feature vector of the real-time feature map in (x,y,z) coordinates. The six-channel complete feature vector represents the standard normal baseline feature map under the same coordinate. The subscript 2 indicates that the multi-dimensional feature difference is calculated by using the second-order Euclidean distance. The operation logic is to sum the squares of the differences of each dimension of the six features under the same spatial unit and then take the square root. This comprehensively measures the overall deviation of multiple parameters at this position. A single value can characterize the local comprehensive anomaly level without the need to compare the differences of a single channel.

[0077] A spatial unit was selected within the fault fracture zone of a hydraulic tunnel at a depth of 1200 meters, corresponding to columns 5, segments 12, and layers 3. The values ​​of the surrounding rock stress channel and pore hydrological channel within the real-time feature vector of this unit showed a slight increase compared to the baseline map. The values ​​of the equipment and video channels exhibited slight fluctuations, while the static geological channel values ​​remained constant. Substituting the two sets of vectors from the six channels into the L2 distance formula for comprehensive calculation, the final VAI calculation result for this unit was 0.68. This value indicates a high degree of deviation in multi-parameter coupling within this spatial unit, classifying it as a high-risk area. A complete granite structural spatial unit was selected within the same monitoring section, corresponding to columns 3, segments 8, and layers 4. The real-time values ​​of the six channels of this unit showed almost no deviation from the baseline map. Substituting these values ​​into the formula for comprehensive calculation, the VAI value was only 0.07, indicating that all monitoring parameters of this spatial unit conformed to the standard normal evolution baseline, with no abnormal deviations. The magnitude of VAI values ​​output by different voxels can intuitively distinguish the risk intensity of various local areas within the cavern. The closer the value is to zero, the more stable the monitoring status is, while the larger the value, the higher the risk of multi-parameter synchronous coupling.

[0078] After all spatial units have completed the VAI numerical solution, each set of (x,y,z) coordinates is matched with the corresponding VAI values ​​and synchronously transmitted to the weighted aggregation process of the full-domain MAI full-cavity anomaly index, serving as the basic material for full-domain risk aggregation calculation. At the same time, all VAI values ​​are bound to three-dimensional coordinates for three-dimensional anomaly heatmap rendering, providing quantitative indicators for accurate local risk positioning in the local underground cavities. The entire set of calculations and the previous mesh partitioning, feature channel construction, and GAN model training processes share the same spatiotemporal reference, and the data standards are completely unified.

[0079] Traditional monitoring relies solely on a single sensor and fixed thresholds to determine risk, and can only identify obvious disasters caused by a large exceedance of a single parameter. It cannot capture the coupling precursors formed by the synchronous and minute deviations of the six types of parameters. This process integrates the feature differences of all six channels through L2 distance, and uniformly quantifies the multi-dimensional synchronous deviation amplitude. It can identify weak coupling risks that traditional threshold alarms cannot capture, and provide quantitative identification for the minute precursors of rockbursts and water inrushes in fault and high-stress sections, supporting refined local risk assessment.

[0080] Based on the technical requirements for full-area risk visualization of deeply buried caverns, spatial weighting is performed on all VAI results. Weighted calculations are used to define the 3D anomaly heatmap based solely on volumetric spatial coordinates for rendering and mapping. Simultaneously, the entire cavern's MAI (Maintenance, Indication, and Quality) continuous health time-series curve is reconstructed, outputting both spatial thermal visualization and full-domain time-series quantification results. Risk levels are categorized based on the geological structure of the deeply buried cavern and the intensity of construction disturbance. Specific weighting coefficients are assigned to different areas, with higher weights allocated to areas with high disturbance and high structural risk, while reducing the weight of stable and intact rock mass areas to prevent localized minor fluctuations from inflating the overall risk assessment result. The regional division and assignment rules are as follows: continuously excavating sections at the tunnel face, high-stress concentration tunnel sections, and fault fracture zones are considered high-risk areas; conventional tunnel sections with intact granite without structural features are considered low-risk and stable areas. For example, all spatial units within the continuous excavation range of 50 meters in front of the tunnel face are uniformly configured with a weight of 1.8, the weight of the fault fracture zone covering unit is set to 1.5, and the weight of a single volume element in a conventional complete tunnel section is only 0.6. A total of 1,500 independent three-dimensional spatial units were divided into key monitoring sections of the tunnel face. The VAI was multiplied and accumulated with the corresponding weight by traversing all units, and the final MAI calculation result of the section was 0.32. This value indicates that the overall coupling deviation of various monitoring parameters in the entire monitoring section is low, and the entire area is in a low-hazard gestation range.

[0081] All spatial units are synchronously stored with two sets of associated data: three-dimensional coordinates and corresponding VAI indices. The heatmap uses only the (x,y,z) volume element coordinates as the sole spatial mapping benchmark, without employing other point mapping rules. Based on a fixed VAI value range, three levels of visual color are used to intuitively distinguish the local risk level of a single spatial unit. The unified color-grading standard is as follows: VAI values ​​less than 0.2 are filled with green, indicating that the coupling evolution of the six monitoring parameters in this spatial unit completely conforms to the normal construction baseline, with no abnormal deviations. VAI values ​​between 0.2 and 0.5 are filled with yellow, indicating that multiple monitoring parameters in this unit have simultaneously shown slight deviations, classifying it as a warning area. VAI values ​​greater than 0.5 are filled with red, indicating that the multi-channel characteristics of the surrounding rock, hydrology, etc., in this unit show significant coupling deviations, making it a key high-risk point for rockburst and water inrush. For example, the VAI calculation result of the volume element in column 5, segment 12, and layer 3 of the fault fracture zone is 0.68, and this three-dimensional location is rendered in red on the heat map; the VAI of the volume element in column 3, segment 8, and layer 4 of the complete granite section is only 0.07, and the corresponding area is filled in green, which can directly locate the concentrated and contiguous areas of disaster precursors from the spatial distribution map.

[0082] The system sets a 10-minute interval as the standard update interval for the four-dimensional holographic tensor M(t). Each time a global tensor snapshot is generated, a full VAI calculation, weighted MAI summation, and MAI time-series record are simultaneously completed. Values ​​are then continuously stitched together in chronological order to generate a continuous time-series curve of the cavern's overall health, tracking the dynamic evolution of overall risk as tunneling progresses. For example, every 10 minutes, the system completes a full set of global volume element feature extraction, VAI solving, and weighted MAI calculations, saving the corresponding MAI time-series nodes. After continuously collecting 120 hours of complete time-series data, a complete cavern health change curve can be generated. If all MAI values ​​within the curve remain stably within the 0 to 0.4 range for a long period, it indicates stable multi-parameter co-evolution throughout the tunneling process. If multiple consecutive time-series nodes on the curve show MAI exceeding 0.5, it indicates the simultaneous appearance of weak coupling precursors in multiple units within the cavern, and the system automatically initiates subsequent gradient attribution and precursor clustering operations.

[0083] The 3D anomaly heatmap is responsible for the precise spatial location of local risks, and can directly pinpoint the coordinates of single or contiguous high-risk 3D voxels. The MAI continuous time series curve is used to macroscopically assess the long-term risk evolution trend of the entire cavern. Both types of visualization results are output simultaneously as input materials for gradient attribution and disaster precursor clustering processes. The entire set of weighted calculations, visualization rendering processes, and the previous 3D mesh subdivision, 4D tensor construction, and VAI single voxel calculations share a unified 2-meter voxel scale, a 10-minute time series update step size, and a six-category feature channel division standard. There are no spatial scale, time series step size, or feature dimension matching deviations throughout the entire chain.

[0084] Traditional safety monitoring relies solely on fixed thresholds from single-point sensors to trigger alarms, only identifying obvious hazards with significant exceedances of single parameters. This differentiated weighting mechanism distinguishes the probability of disaster occurrence in different areas, allowing even minor anomalies in high-risk sections to raise the overall MAI (Matter of Indices) index, achieving a comprehensive quantitative assessment of weak cross-system coupling precursors. The system outputs dual analytical data—heat maps and time-series curves—for use by on-site tunneling and support teams. This enables precise location of areas requiring intensified support and slower tunneling, while also allowing for long-term tracking of the overall risk evolution of the tunnel, providing both quantitative and visual analytical support for proactive prevention and control of rockbursts and water inrushes.

[0085] S105, feature gradient attribution extracts the single-channel feature contribution of a single volume element, and DBSCAN / K-Means clustering is used to mine coupled precursor combinations to form a standardized precursor pattern library for rockburst and water inrush large deformation disasters.

[0086] In one implementation, real-time four-dimensional holographic tensor complete data is used as the input basis for gradient calculation. The gradient value of the global MAI index with respect to each three-dimensional voxel and each type of feature channel is calculated. The absolute value of the gradient represents the contribution of that channel and spatial unit to the overall anomaly of the cavern. A higher gradient value corresponds to the dominant factor inducing the anomaly in the monitoring indicator, while a gradient value close to zero indicates that the feature has no significant driving effect on the current anomaly. The calculation formula is as follows: The input material is a complete four-dimensional holographic tensor at the current moment. The tensor contains a unified length, width and height spatial dimension, six types of hierarchical feature channels, and all monitored values ​​of a single time-series snapshot. The output content is a one-to-one corresponding volume element channel contribution vector. The vector records the gradient contribution values ​​of each channel of surrounding rock, environment, hydrology, equipment, video, and static geology under each group of spatial units.

[0087] For example, after gradient calculation of the real-time tensor of a 1200-meter-deep hydraulic tunnel, the volume element of the third layer of the twelfth segment in the fifth column shows a gradient contribution value of 0.91 for the surrounding rock stress channel, 0.76 for the microseismic b-value channel, and 0.69 for the pore water pressure channel. The gradient values ​​of the remaining environmental, video, and geological channels are all below 0.1, indicating that the anomaly in this entire area is dominated by the synchronous shift of three types of indicators: surrounding rock stress, microseismic activity, and pore water pressure. In the intact granite section, the gradient values ​​of all channels are below 0.05, showing no characteristic contribution. Traditional monitoring can only identify the overall level of anomalies and cannot distinguish which indicator or spatial area among multiple parameters causes the risk. Gradient attribution completes the quantification of the contribution of a single volume and single channel, enabling precise anomaly tracing, distinguishing between dominant risk indicators and irrelevant stationary indicators, and providing standardized feature vector materials for subsequent clustering and summarization.

[0088] After each round of MAI exceeding the limit triggering gradient attribution calculation, the duration of the current time series window, the gradient contribution vectors of all voxel channels, the corresponding global MAI values, and the 3D spatial coordinate information are fully retained to construct a historical anomaly sample material library. All stored materials are standardized with six channels, a uniform two-meter voxel scale, and a uniform 120-minute time series window standard to ensure a consistent data benchmark for clustering operations. For example, during continuous system operation, multiple sets of anomaly sample records are archived separately. The first type of sample records the high gradient contribution combination of TBM thrust, microseismic activity, and surrounding rock stress, corresponding to the time series of the tunnel face section; the second type of sample records the high gradient contribution combination of fault pore water pressure and rock mass acoustic emission, corresponding to the time series of the tectonic fracture zone. The two types of samples are retained completely separately and without confusion.

[0089] All archived historical anomaly gradient contribution vectors were retrieved, and feature grouping was performed using both density clustering and mean clustering algorithms. Time-series samples with highly similar gradient value trends and multi-channel contribution matching patterns were grouped into the same coupled precursor group. Scattered, irregular, and isolated anomaly samples were removed, and multi-parameter synchronous deviation combination patterns with recurring patterns were extracted. The input material is the full set of historical anomaly gradient contribution vectors, with a unified six-channel feature dimension specification. The output content is a multi-channel coupled precursor classification group, with fixed high-contribution feature matching rules extracted for each group.

[0090] After simultaneous grouping and calculation using two clustering algorithms, the first cluster group uniformly exhibited three high-contribution features: slight oscillation of TBM thrust, continuous decrease of microseismic b-value, and local uplift of surrounding rock stress. The gradients of other environmental and video channels remained low for a long time. The second cluster group uniformly exhibited two high-contribution features: pulse fluctuation of pore water pressure in fault areas and increase of acoustic frequency of rock mass. The other equipment indicators showed no significant gradient changes. The two groups corresponded to two recurring coupling anomaly patterns.

[0091] Small fluctuations from a single sensor do not have disaster reference value. Only when multiple indicators show slight shifts simultaneously can they be considered precursors to disasters. Clustering algorithms automatically filter out recurring multi-parameter coupling combinations from massive historical time series, filter out meaningless single random numerical fluctuations, remove invalid abnormal records, and extract standardized precursor combinations with engineering early warning value.

[0092] Engineering disaster labels are matched for various coupling precursor groups that have been clustered. Combining the gestation characteristics of three typical disasters in deep-buried caverns—rock bursts, water inrushes, and large deformations in soft rock—multi-channel coupling rules that match the corresponding risk types are stored in the database and archived to form a standardized precursor pattern library that can be retrieved in real time and matched automatically. Each pattern in the database is simultaneously labeled with its corresponding high-contribution features, typical spatial regions, and temporal fluctuation features.

[0093] For example, the first type of clustering coupling features matches rockburst disasters, and the rockburst classification entries are entered into the model library. The typical occurrence area is marked as the 50-meter tunneling section in front of the tunnel face. The core coupling features are small fluctuations in TBM thrust, a continuous decrease in microseismic b-value, and nonlinear uplift of local surrounding rock stress. The second type of clustering coupling features matches water inrush disasters, and the water inrush classification entries are entered into the model library. The typical occurrence area is marked as the fault fracture zone. The core coupling features are periodic pulses of pore water pressure and a continuous increase in rock mass acoustic emission count. Both types of models are permanently stored, and subsequent real-time temporal gradient vectors can be directly and automatically compared and matched with entries in the library.

[0094] The model library has completed the digital accumulation of disaster precursor experience from long-term on-site construction, avoiding the loss of experience due to personnel turnover. The gradient contribution vector generated by real-time monitoring can be automatically compared with the standard coupling model in the library. After a successful match, the corresponding disaster early warning is directly pushed. The system also outputs multi-dimensional judgment results of risk three-dimensional spatial coordinates, dominant anomaly characteristics, and precursor type, realizing the upgrade from simple anomaly value identification to accurate disaster type determination.

[0095] S106: When MAI exceeds the threshold, an automatic early warning is triggered. Based on the gradient attribution feature vector to store historical abnormal samples, two-level induction is carried out in sequence: feature contribution ranking and multi-source coupling pattern clustering. The results are mapped to the three-dimensional spatial coordinates of the cavern, and information on disaster early warning and refined risk root cause analysis is output.

[0096] In one implementation, a gradient attribution solution program is used to analyze the feature contribution of samples exceeding the anomaly index limits across all caverns, establishing a complete workflow for cavern anomaly tracing and disaster precursor clustering analysis. The four components—MAI exceeding limit judgment rules, gradient feature extraction logic, historical sample storage specifications, and clustering classification criteria—are integrated and processed simultaneously to establish a one-to-one correspondence between anomaly time-series samples and single-channel feature contribution vectors, generating a standardized anomaly induction and analysis workflow. The entire workflow maintains consistent input and output data specifications with the initial VAI and MAI quantization stages, using a two-meter voxel scale, a ten-minute time-series update interval, and a K1 to K6 six-category feature channel division benchmark throughout, eliminating data specification mismatch issues. The four core execution standards are as follows: 1. MAI Exceedance Judgment Rule: A global MAI risk threshold is pre-set. Only time-series samples whose MAI values ​​exceed the threshold in the current period will trigger the entire gradient tracing operation. 2. Gradient Feature Extraction Logic: Gradient solving uses the global MAI index as the calculation target and performs partial derivative operations on all volumes and channels of the four-dimensional tensor. 3. Historical Sample Storage Standard: For each group of exceeding-limit samples, the complete set of data including the time-series window duration, global MAI values, all three-dimensional coordinates, and six-channel gradients is retained. 4. Clustering Classification Criteria: DBSCAN and K-Means algorithms are used simultaneously for group verification to remove invalid isolated samples formed by single-point random fluctuations.

[0097] A unified risk threshold for MAI (Mean Income Index) is defined as the boundary for source tracing. Every ten minutes, the system completes a round of weighted summation calculation of the current MAI across all voxels. Samples meeting the threshold are stored in the historical data library, while samples failing to meet the threshold terminate the current analysis process. Example: The system sets the MAI risk threshold to 0.4. A single-round weighted calculation of the MAI for a 1200-meter hydraulic tunnel monitoring section yields 0.51, exceeding the threshold. The entire time series is then archived and saved. If the single-round MAI calculation result is 0.28, subsequent gradient and clustering operations are not performed.

[0098] The gradient calculation uses the global MAI exponent as the derivative target and employs the gradient calculation formula: In the formula, MAI represents the total cavern anomaly index. The original feature of the nth channel of the volume element (x,y,z) at time t. To assess the risk contribution of each channel, the calculation only uses the absolute value of the gradient to measure the strength of the contribution. Unified computational constraints: the gradient calculation material must be a complete four-dimensional holographic tensor of the out-of-limit samples. The tensor must be a standardized 10-column × 25-segment × 6-layer two-meter grid with six channels. After calculation, the coordinates are bound to the gradient vectors one-to-one, and the vector format is standardized, allowing it to be directly used as clustering input material.

[0099] All archived gradient vectors are retrieved and two types of clustering algorithms are run sequentially. First, DBSCAN is used to filter out scattered single-point anomalies, and then K-Means is used to normalize similar coupling patterns. The grouping results of the two algorithms are cross-compared, and only time series that are synchronously classified are determined as valid precursor samples, while isolated fluctuation records are directly discarded. The entire process uses unified time series, grid, and channel calculation standards to eliminate the problem of inconsistent calculation benchmarks for samples from different time periods. Gradient vectors can be directly fed into the subsequent multi-source coupled precursor clustering process without secondary transformation, reducing the workload of data preprocessing and avoiding calculation errors caused by differences in specifications.

[0100] The system integrates and processes the MAI (Mean Interference of Interest) exceedance judgment rules, gradient feature extraction logic, historical sample storage specifications, and clustering classification criteria to establish a one-to-one correspondence between anomalous time-series samples and single-channel feature contribution vectors, generating a standardized anomaly induction and analysis process. Once the system calculates that the global MAI value exceeds a preset threshold, it automatically runs a gradient attribution program. Using this exceedance signal as the computation start condition, the system distinguishes the magnitude of risk contribution based on the gradient values ​​calculated by each voxel and channel, extracting all feature information that plays a major role in the anomaly. The computation input consists of two parts: first, the complete four-dimensional holographic tensor corresponding to this exceedance; second, all parameters of the trained, fixed-parameter 3D-Conv-LSTM discriminator. The computation output is a gradient vector, with each vector bound to the corresponding three-dimensional voxel coordinates, and the vector recording the gradient values ​​of each of the six monitoring channels of that unit.

[0101] Gradients are calculated for each individual element and each monitoring channel category. Only the absolute value of the gradient is used to determine the contribution; a higher value indicates that the monitoring indicator is the primary cause of the anomaly. Supporting calculation formula: In the formula, MAI represents the overall cavern anomaly index. Let (x, y, z) be the nth type of monitoring feature of the volume element at time t. The calculation result represents the risk contribution degree of this indicator at this location.

[0102] In a volumetric element within a fault zone of a 1200-meter-deep hydraulic tunnel, the gradient values ​​for the surrounding rock, microseismic, and hydrological channels are all above 0.65, while the gradient values ​​for the environmental, video, and geological channels are all below 0.1. This indicates that the anomaly was caused by synchronous changes in these three indicators. Within the intact granite region, all channel gradient values ​​are less than 0.05, indicating that the monitoring status in this area is consistent with the normal baseline and does not contribute to any risk.

[0103] Conventional monitoring can only detect anomalies in the overall cavern, but it cannot pinpoint the location of the risk or identify which monitoring data triggered the risk. Gradient calculation can quantify the risk contribution of each monitoring indicator, generate a uniformly formatted feature vector, and directly use it for subsequent clustering operations.

[0104] Based on a standardized anomaly induction process, high-contribution feature information is extracted. Using the MAI over-limit trigger signal as the initiation source, the gradient contribution value of voxel channels as the division criterion, and the DBSCAN / K-Means clustering algorithm as the processing method, the system outputs multiple results item by item, including single-feature contribution ranking, multi-source coupled precursor clustering, 3D spatial coordinate mapping, disaster early warning, and root cause analysis. After the gradient attribution operation is completed, the system collects all voxel channel gradient vectors bound to 3D coordinates. All vectors adhere to the standard of two-meter voxels, six feature channels, and a 120-minute time series window. Grouping operations are performed using DBSCAN and K-Means clustering algorithms, and four types of results are output simultaneously: single-feature contribution ranking, multi-source coupled precursor grouping, 3D spatial coordinate mapping results, and disaster early warning and root cause analysis information. The entire clustering process inherits the gradient attribution output materials, maintaining consistency with the data benchmark and the preceding four-dimensional tensor construction, GAN training, and two-level anomaly quantification processes.

[0105] Clustering is performed in two sequential steps, with two algorithms each performing their respective functions: filtering invalid data and extracting stable precursor patterns. The first step uses DBSCAN density clustering, relying on the numerical distribution density of gradient vectors to filter out scattered and isolated anomalous samples caused by single equipment failures or instantaneous environmental fluctuations, retaining only valid gradient records with continuous temporal variation characteristics. The second step uses K-Means mean clustering on the filtered valid samples, using the similarity of multi-channel gradient values ​​as the grouping criterion, grouping time series with similar synchronous shift patterns of indicators into the same group, and extracting recurring multi-parameter coupled anomaly combinations from the field.

[0106] After two-step clustering, the archived gradient vectors of a 1200-meter-deep hydraulic tunnel were stably divided into two typical precursor groups. The first group of samples were all distributed in the 50-meter tunneling section in front of the tunnel face, where the gradients of TBM thrust, microseismic b-value, and surrounding rock stress channels were consistently higher than 0.65, while the gradients of environmental and video channels were lower than 0.1. The second group of samples was concentrated in the fault fracture zone area, where the gradients of pore water pressure and rock mass acoustic emission channels simultaneously rose to above 0.6, while the equipment operating conditions showed no significant gradient changes.

[0107] For each set of coupled precursors, matching three typical hazards—rockburst, water inrush, and large deformation of soft rock—the typical spatial regions and core contribution monitoring indicators corresponding to each set are recorded simultaneously. The three-dimensional coordinates of all high-risk volume elements within the set are imported into the global anomaly heatmap, and the high-risk points are marked by overlaying using the original VAI numerical color scale rules, intuitively presenting the spatial range of risk concentration.

[0108] After completing group matching and spatial mapping, the system automatically pushes early warnings, with the warning information labeled with the disaster type, three-dimensional risk coordinates, and the dominant monitoring indicators that triggered the anomalies. Traditional monitoring can only display fluctuations of a single parameter and cannot identify hidden precursors formed by multiple indicators simultaneously. The clustering process automatically summarizes standardized coupling patterns from massive time series, transforming scattered monitoring data into directly applicable construction judgment criteria. On-site teams can adjust tunneling speed and increase support density in corresponding areas based on the warnings, thus addressing potential disaster hazards in advance.

[0109] like Figure 2 As shown, Figure 2 This diagram illustrates the overall modular architecture of a multi-source state fusion and GAN anomaly detection system for deeply buried underground caverns. The left side displays the main title box, while the right side features four rectangular functional modules arranged sequentially. Labeled straight arrows illustrate the overall functional divisions and data flow logic between modules. The main title box on the left represents the multi-source state fusion and GAN anomaly detection method and system for deeply buried underground caverns, serving as the overall framework for the entire solution. It branches outwards into three functional branches, each corresponding to one of the four core functional modules.

[0110] The first branch constructs the labeled text matrix, with an arrow pointing right to the topmost matrix construction module. The bottom of this module outputs a downward arrow indicating the construction of the labeled matrix, connecting to the deep learning module below. This module is responsible for the unified processing of multi-source monitoring data, generating a standardized four-dimensional time-series volumetric tensor to provide uniformly sized input data for subsequent deep learning operations.

[0111] The second branch splits into two labeled lines: deep learning and anomaly index quantization. The arrow labeled deep learning points to the right towards the deep learning module, which has a built-in 3D-ConvGAN network. It uses the tensors output by the matrix construction module to complete adversarial learning under normal working conditions and outputs the solidified discriminator's intermediate convolutional features. The arrow labeled anomaly index quantization points to the right towards the anomaly index quantization module, which receives the convolutional features output by the deep learning module and completes two-level quantization operations for the VAI voxel anomaly index and the MAI full-cavity anomaly index. The bottom of the anomaly index quantization module outputs a downward arrow labeled anomaly index quantization, which connects to the mining and summarization module.

[0112] The third branch is for text mining and summarization. The arrow points to the bottom right to the mining and summarization module. This module receives the global anomaly data output by the anomaly index quantification module, and sequentially performs gradient attribution feature contribution calculation, DBSCAN and K-Means coupled precursor clustering to complete the disaster precursor pattern summarization and early warning output.

[0113] The overall data flow sequence is as follows: the matrix construction module outputs data to the deep learning module, the deep learning module outputs feature data to the anomaly index quantification module, and the anomaly index quantification module outputs quantification results to the mining and summarization module. The four modules are connected in sequence, fully covering the entire process of multi-source data fusion, model training, anomaly quantification, and precursor mining.

[0114] like Figure 3 As shown, Figure 3 This diagram illustrates the timeline comparison of the entire process of rockburst disaster development. The timeline runs horizontally to the right and is divided into five key time nodes. The nodes are connected sequentially by dashed arrows. Vertically, the diagram is divided into three information display areas, with vertical dashed lines distinguishing the warning interval of this solution from the identification interval of existing technologies.

[0115] The timeline is divided horizontally into four functional description boxes, corresponding to the technical methods used at different stages: The first box marks the normal working baseline of the deeply buried underground cavern, enabling dynamic data matrix construction and deep learning training. It includes three types of monitoring indicator illustrations: local surrounding rock stress, TBM thrust fluctuation, and microseismic values, used to collect basic multi-source monitoring data. The second box marks the extraction, discrimination, quantification, and anomaly mining and summarization of weak precursors using GAN technology. This is further subdivided into three core processing steps: matrix learning, feature extraction, and anomaly summarization. The trained GAN model identifies early, weakly coupled anomalies. The third box marks existing technology identification, corresponding to the alarm mode of traditional single-point monitoring equipment. The fourth box marks the occurrence of rockburst disasters, representing the final eruption point of the disaster.

[0116] The time sequence and monitoring data changes at each time point are as follows: On day 1, the monitoring index values ​​are: local surrounding rock stress 29MPa, TBM thrust fluctuation range ±2.5%, microseismic b value stable at 2.1, and the overall working condition is in line with the normal baseline. The system only continuously collects data and constructs a dynamic matrix.

[0117] On the third day, the monitoring indicators showed that the local surrounding rock stress increased slightly to 29.2 MPa, the TBM thrust fluctuation expanded to ±3%, and the microseismic b value decreased slightly to 2.05. The parameters showed only slight shifts. Traditional monitoring did not respond, and this scheme continued to complete matrix learning and feature extraction.

[0118] On the 5th day, the monitoring indicators showed that the local surrounding rock stress increased by 3% compared with the normal baseline, the TBM thrust fluctuation increased to ±5%, the microseismic b value decreased to 1.8, and multiple parameters showed simultaneous coupled shifts. These were identified and marked as anomalies by the GAN discriminator, completing the early weak precursor capture and initiating subsequent gradient attribution and precursor clustering operations.

[0119] On the 8th or 9th day, the vertical dotted line marks the dividing point for the current technology to start identification. At this time, all monitoring indicators exceed the limits: the local surrounding rock stress reaches 35MPa, and the TBM thrust fluctuation range is ±8%. Only then will the multi-indicator fusion system trigger an alarm and can only identify obvious risks of significant exceedance.

[0120] On the 9th day, after multiple monitoring indicators continued to exceed the limits, a rockburst disaster occurred directly.

[0121] like Figure 4 As shown, Figure 4 This is a schematic diagram of the layered deployment architecture of a multi-source monitoring and analysis system for deeply buried underground caverns. The system is divided into four layers from top to bottom: the cavern acquisition edge layer, the transmission layer, the cloud layer, and the user terminal. Data flows unidirectionally between layers via vertical arrows. Each layer is divided into internal functional units by dashed boxes, fully demonstrating the entire data flow from on-site acquisition to terminal early warning.

[0122] The first layer: the cavern acquisition edge layer, responsible for real-time data processing. This layer comprises three parts: a monitoring terminal cluster, edge computing nodes, and a cavern ring network switch, interconnected by dashed arrows. The monitoring terminal cluster includes six types of acquisition devices: surrounding rock area sensors, gas sensors, hydrogeological sensors, equipment condition controllers, personnel positioning coordinate acquisition units, and underground cavern video streams, responsible for acquiring multi-source raw monitoring data on-site. The edge computing nodes are equipped with three components: a protocol adaptation engine, a local computing unit, and a 1-hour cache module, performing unified conversion of multiple data protocols, short-term local caching, and basic preprocessing. The cavern ring network switch, equipped with a data aggregation and forwarding unit, consolidates all monitoring data from the edge nodes and outputs it to the next transmission layer.

[0123] The second layer is the transport layer, responsible for secure data transfer. This layer comprises a firewall and intrusion detection system, a data transmission gateway, and a cloud access switch. The firewall and intrusion detection system incorporate AES encryption and unauthorized access blocking mechanisms to encrypt transmitted data and intercept abnormal access requests. The data transmission gateway uses the OPC protocol for data exchange and is configured with dual-link backup to ensure uninterrupted data transmission. The cloud access switch forwards data to the cloud, pushing the encrypted, complete monitoring data to the cloud layer.

[0124] The third layer, the cloud layer, is for intelligent analysis and early warning. This layer consists of three main units: a distributed storage cluster, a core computing engine, and an application service layer. Dashed arrows connect these units to transmit computational data. The distributed storage cluster includes three types of storage media: time-series databases, relational databases, and object storage, storing time-series monitoring data, device files, and model tensor files in layers. The core computing engine is equipped with three core computing engines: a matrix generation engine, a GAN training engine, and an anomaly calculation engine. These engines sequentially complete the construction of four-dimensional time-series tensors, 3D-ConvGAN adversarial training, and VAI and MAI two-level anomaly index quantization. The application service layer integrates anomaly early warning, visual monitoring, and pattern library construction. It relies on gradient attribution and clustering algorithms to summarize precursors and stores a standardized rockburst and water inrush disaster precursor pattern library.

[0125] The fourth layer is the user terminal, which receives visualized monitoring images, early warnings of disasters, and risk assessment results from the cloud layer, allowing on-site construction and safety management personnel to view 3D heat maps, time-series risk curves, and disaster response prompts.

[0126] Raw data is collected by various monitoring devices on site, uniformly organized and cached at the edge layer, and securely uploaded to the cloud through an encrypted backup transmission link. The cloud completes data storage, GAN model training, anomaly quantification, and precursor pattern mining. Finally, the visualized monitoring and early warning information is sent to the user terminal, forming a complete closed-loop link of collection-transmission-intelligent analysis-terminal early warning.

[0127] like Figure 5 As shown, Figure 5 The graph represents the convergence difference in accuracy of the GAN adversarial training model, and is a line graph used to illustrate the convergence effect of the 3D-ConvGAN network adversarial training iterations. The horizontal axis represents the training iteration stage, from left to right: the initial model, and epochs 1 to 20; the vertical axis represents the convergence difference with the optimal standard normal operating condition, in percentage, with a value range of 0 to 40; the legend is a blue solid line, meaning the convergence difference (%) with the optimal standard normal operating condition, and each node of the line is labeled with the corresponding difference value.

[0128] The initial model, without iterative training, showed a convergence discrepancy of 39.88%, indicating a low match between the model and standard normal operating conditions. During training epochs 1 to 10, the discrepancy decreased rapidly: 38.52% in epoch 1, 32.28% in epoch 5, and 20.97% in epoch 10, demonstrating how adversarial training quickly corrected the model's feature extraction bias. From epochs 11 to 16, the rate of decrease slowed: 11.02% in epoch 13 and 4.47% in epoch 16, as the model gradually aligned with the distribution of characteristics characteristic of normal operating conditions in caverns. From epochs 17 to 20, the discrepancy continued to decrease slightly: 1.52% in epoch 18 and 0.35% in epoch 20.

[0129] As the number of iterations increased, the GAN generator and discriminator networks underwent continuous adversarial optimization, and the difference between the model's output features and the standard normal operating conditions of the cavern continued to narrow. By the 20th iteration, the convergence difference was only 0.35%, meeting the preset convergence criteria. At this point, all weight parameters of the discriminator network could be fixed for subsequent extraction of convolutional features from real-time tensors, and for performing VAI and MAI anomaly index quantization calculations. The curve generally showed a trend of rapid decline in the early stages and slow convergence in the later stages, intuitively reflecting the entire convergence process of complete adversarial training, verifying that the network can stably learn the normal baseline distribution of multi-source monitoring data from deeply buried caverns.

[0130] like Figure 6 As shown, Figure 6 The performance comparison chart for weak precursor identification and early warning is presented in the form of a three-dimensional grouped bar chart. Four identification schemes are set horizontally as comparison objects, and each scheme is matched with three bars of different colors, representing three evaluation indicators respectively. The legend is as follows: the blue bar is the accuracy rate of weak precursor identification, the orange bar is the false alarm rate, and the yellow bar is the false alarm rate. The vertical axis is the percentage value.

[0131] The corresponding indicator values ​​for the four schemes are as follows. For the traditional threshold method, the accuracy rate for weak precursor identification is 35.20%, the false negative rate is 42.10%, and the false positive rate is 28.70%. Relying solely on a fixed threshold for a single monitoring indicator cannot identify early disaster precursors with simultaneous weak shifts in multiple parameters, resulting in a large number of false negatives and the lowest identification accuracy.

[0132] For the CS-LSTM model, the accuracy rate for weak precursor identification is 78.60%, the false negative rate is 19.30%, and the false positive rate is 15.40%. However, since the model only focuses on single time-series data and does not integrate multi-source monitoring information, it has a significant weakness in identifying weak precursors across parameter coupling.

[0133] For the multi-indicator fusion system, the accuracy rate of weak precursor identification is 91.53%, the false negative rate is 2%, and the false positive rate is 8.90%. It achieves the fusion of multiple monitoring indicators, but lacks a dynamic normal operating condition baseline constructed by deep learning, and there are still some misjudgments.

[0134] The dynamic matrix + GAN scheme of this invention achieves a weak precursor identification accuracy of 98.80%, a false negative rate of 1.20%, and a false positive rate of 3.50%. By relying on multi-source four-dimensional dynamic matrix fusion and 3D-ConvGAN adversarial learning to construct a dynamic normal baseline, it can accurately capture the hidden coupling of weak precursors of rockbursts and water inrushes. Among the four schemes, it has the highest identification accuracy, with the lowest false negative and false positive rates.

[0135] A horizontal comparison of the four scheme indicators shows that the accuracy of weak precursor identification of the present invention is significantly higher than that of the traditional threshold method, CS-LSTM model and conventional multi-indicator fusion system. At the same time, the false alarm rate and false alarm rate are greatly reduced, which intuitively verifies that the technical route of dynamic matrix fusion combined with GAN model has better early warning performance in the early hidden disaster precursor identification scenario of deep buried caverns.

[0136] In one embodiment, this application also provides a multi-source state fusion and GAN anomaly precursor mining system for deeply buried underground caverns, comprising: The matrix construction module is used to collect six categories of heterogeneous monitoring data: surrounding rock, environment, hydrology, equipment, video, and static geology. It completes the L×W×H three-dimensional voxel meshing, builds six independent feature channels with the help of kriging, inverse distance interpolation and temporal normalization, generates a unified spatiotemporal reference UCSM four-dimensional tensor M(t), and windows and isolates the temporal data to generate GAN standardized training samples. The working condition deep learning module is used to build a 3D-ConvGAN dual-branch network for generation and discrimination. The 3D convolutional recurrent generator replicates the dynamic sequence of normal caverns, and the 3D-Conv-LSTM discriminator outputs the normal working condition probability. After multiple rounds of adversarial game iteration and convergence, the dynamic normal working condition discrimination benchmark of deep buried caverns is solidified. The anomaly index quantization module is used to extract the intermediate convolution features of the discriminator, solve the single-volume VAI anomaly index through the feature L2 distance, and combine spatial weights to weight and summarize the global MAI full-cavity index to generate a three-dimensional anomaly heat map and a continuous health time series curve, quantifying the spatial and overall anomaly degree. The precursor clustering module is used to perform gradient attribution calculations on abnormal samples to determine the risk contribution of each element and feature channel. It uses DBSCAN and K-Means clustering algorithms to mine multi-parameter coupling patterns and build a standardized disaster precursor pattern library for rockbursts and water inrushes. The early warning and judgment module is used to monitor the MAI index. When the value exceeds the preset threshold, the source tracing process is initiated, the gradient feature historical abnormal samples are archived, the feature contribution ranking and coupling pattern clustering are completed in sequence, the judgment results are mapped to the three-dimensional coordinates of the cavern, and the results of advanced disaster early warning and refined risk root cause analysis are output. The engineering application module is used to accurately capture minute anomalies across systems in projects such as deep-buried tunnels and underground power plants by relying on the multi-source data fusion and GAN weak precursor identification system. It replaces the traditional single threshold alarm and provides digital technical support for dynamic management and control of underground engineering construction safety and disaster prevention and control.

[0137] The computer-readable storage medium provided in the above embodiments of this application and the multi-source state fusion and GAN anomaly precursor mining method for deep buried underground caverns provided in the embodiments of this application are based on the same inventive concept and have the same beneficial effects as the methods adopted, run or implemented by the application stored therein.

[0138] The various embodiments in this application are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments for evaluating the method, system, electronic device, and readable storage medium for multi-source state fusion and GAN anomaly precursor mining of deeply buried underground caverns are basically similar to the embodiments of the above-described method for multi-source state fusion and GAN anomaly precursor mining of deeply buried underground caverns, so the description is relatively simple. Relevant parts can be referred to in the description of the embodiments of the above-described method for multi-source state fusion and GAN anomaly precursor mining of deeply buried underground caverns.

Claims

1. A method for multi-source state fusion and GAN anomaly precursor mining in deeply buried underground caverns, characterized in that, include: Collect six categories of multi-source heterogeneous monitoring data: surrounding rock stability, environmental ventilation, hydrogeology, construction equipment, video personnel, and static geology. Execute the matrix construction process, unify the three-dimensional voxel meshing and the four-dimensional tensor M(t) spatiotemporal reference, and generate the UCSM unified cavern state four-dimensional raster tensor. Based on Kriging / inverse distance weighted spatial interpolation, temporal normalization alignment algorithm and static geological fixed channel mapping, six types of feature channels are constructed. The monitoring indicators of various sensors are calculated with a single volume element (x,y,z) as the smallest statistical unit. Feature tensors of six channels, namely surrounding rock, environment, hydrology, equipment, video and geology, are constructed respectively, and temporal window isolation of normal / disaster working condition samples is completed simultaneously. A 3D-ConvGAN dual-branch generator is used to extract spatiotemporal coupled evolution features. A 3D convolutional recurrent generator replicates the normal dynamic sequence, and a 3D-Conv-LSTM discriminator outputs the probability of normal working conditions. The adversarial game converges and solidifies the normal working condition discrimination benchmark. The system configures L2 distance of voxels and weighted double-layer anomaly quantization of the entire cavern space. It models the degree of deviation of multi-dimensional modes through the convolutional feature space of the discriminator, and combines spatial weight allocation and three-dimensional heat map mapping to reconstruct the time series curve of the anomaly index MAI of the entire cavern. Feature gradient attribution is used to extract the single-channel feature contribution of a single volume element, and DBSCAN / K-Means clustering is used to mine coupled precursor combinations to form a standardized database of precursor patterns for rockburst, water inrush and large deformation disasters. When the MAI exceeds the threshold, an automatic early warning is triggered. Based on the gradient attribution feature vector to store historical abnormal samples, two-level induction is carried out in sequence: feature contribution ranking and multi-source coupling pattern clustering. The results are mapped to the three-dimensional spatial coordinates of the cavern, and information on disaster early warning and refined risk root cause analysis is output.

2. The method as described in claim 1, characterized in that, Based on Kriging / inverse distance weighted spatial interpolation, temporal normalization alignment algorithm, and static geological fixed channel mapping, six types of feature channels are constructed. Using a single volume element (x,y,z) as the smallest statistical unit, various sensor monitoring indicators are calculated. Six feature tensors for surrounding rock, environment, hydrology, equipment, video, and geology are constructed respectively. Simultaneously, temporal windowing isolation of normal / disaster condition samples is completed, including: Based on six heterogeneous monitoring data sources—surrounding rock, environmental ventilation, hydrogeology, construction equipment, video personnel, and static geology—this approach follows four technical requirements: L×W×H three-dimensional global voxel meshing, multi-source monitoring data interpolation, cross-frequency time-series dimension normalization, and fixed mapping of static geological parameter channels. It adheres to the execution principles of minimum statistical calculation of a single voxel (x,y,z), independent modeling of the six types of monitoring features by channel, and partitioning and isolating time-series working condition samples. The process involves three-dimensional meshing, spatial interpolation, dimensionless time-series normalization, static geological channel binding, multi-channel four-dimensional tensor encapsulation, and time-series sample windowing collaboratively constructing a unified cavern state matrix four-dimensional tensor. First, the cavern monitoring area is divided into a three-dimensional grid of length × width × height to generate basic voxel units. Then, based on Kriging interpolation and inverse distance weighted interpolation algorithms, the monitoring indicators of various sparse sensors are assigned to the corresponding (x,y,z) voxels across the entire domain. A unified dimensionless normalization and alignment process is performed on multi-source time series data spanning millisecond to daily levels and with different sampling frequencies. Geological model parameters that do not change with time t are permanently bound to the corresponding voxel's dedicated static channel, and the six major categories of independent feature channels are layered and constructed. Using a three-dimensional (x,y,z) volume element as the smallest calculation unit for the entire process, an N-channel four-dimensional holographic tensor M(t) is encapsulated according to the categories of surrounding rock, environment, hydrology, equipment, video, and geology. Simultaneously, a time series sample set is generated by dividing the continuous four-dimensional tensor time series into fixed time windows. This completes the block isolation of samples for normal and stable construction conditions and abnormal conditions with disaster precursors, providing standardized time series training samples Xseq for the input of the 3D-ConvGAN model.

3. The method as described in claim 1, characterized in that, A 3D-ConvGAN dual-branch generator is used to extract spatiotemporally coupled evolutionary features. A 3D convolutional recurrent generator replicates normal dynamic sequences, and a 3D-Conv-LSTM discriminator outputs the probability of normal operating conditions. This approach, through adversarial game theory, converges and solidifies the normal operating condition discrimination benchmark, including: Combining the characteristics of Xseq data from four-dimensional temporal tensors of deep-buried caverns with the 3D-ConvGAN model architecture specifications, random noise input mapping rules and spatiotemporal sequence feature extraction logic are introduced to perform feature analysis on real normal working condition time series samples and generate multi-source coupled spatiotemporal feature adaptation information. Combining the needs of dynamic construction evolution analysis of underground caverns and the convergence specifications of adversarial training, the spatiotemporal feature adaptation information is modeled by a dual-branch allocation logic, the 3D convolutional loop generator replication link and the 3D-Conv-LSTM discriminator identification link are divided, the sequence authenticity comparison association mining mechanism is configured, and branch-specific feature allocation information is generated. Based on the GAN adversarial iterative training standard, normal working condition probability output rules, and spatiotemporal evolution coupled learning specifications, game-like iterative calculations are carried out on the branch feature allocation information to complete the differential fitting of the real tunnel dynamic sequence and the fake simulation sequence. After the model converges, the discriminator is solidified to form a benchmark model of dynamic construction conditions for deep underground caverns.

4. The method as described in claim 1, characterized in that, The system employs L2 distance configuration and weighted two-layer anomaly quantization across the entire cavern space. It models multi-dimensional pattern deviation using convolutional feature space within the discriminator, and combines this with spatial weight allocation and 3D heatmap mapping to reconstruct the MAI time-series curve of the entire cavern anomaly index, including: Combining the real-time four-dimensional time series sample input of deep buried caverns with the intermediate convolutional feature output characteristics of the 3D-ConvGAN discriminator, and relying on the voxel deviation evaluation standard, a dual operation mechanism of single voxel feature space L2 distance calculation and global voxel space weighted summation is introduced to build a two-level hierarchical quantification system of VAI voxel anomaly index and MAI full cavern anomaly index. Based on the real-time monitoring feature parsing execution specification, real-time operating condition feature maps are first extracted respectively. Compared with standard reference feature map Then, the coordinates of the three-dimensional voxels (x, y, z) in the two sets of feature maps are aligned one by one. The deviation of a single voxel is solved by coordinate matching, forming a coherent processing link from feature extraction to single voxel deviation calculation. Based on the technical requirements for full-area risk visualization of deeply buried caverns, spatial weighting is performed on all VAI results. Weighted calculations are used to limit the rendering and mapping of the three-dimensional anomaly heat map to volumetric spatial coordinates, and the reconstruction of the continuous health time-series curve of the entire cavern MAI is completed simultaneously, outputting two types of results: spatial thermal visualization and full-domain time-series quantization.

5. The method as described in claim 4, characterized in that, When the MAI exceeds the threshold, an automatic early warning is triggered. Based on the gradient attribution feature vector storing historical abnormal samples, a two-level induction process is performed: feature contribution ranking and multi-source coupling pattern clustering. The results are mapped to the three-dimensional spatial coordinates of the cavern, outputting disaster early warning and refined risk root cause analysis information, including: A gradient attribution solution program was used to analyze the feature contribution of samples with out-of-limit anomaly indices in all caverns, and a complete process for cavern anomaly tracing and disaster precursor cluster analysis was established. The rules for MAI exceeding the limit, gradient feature extraction logic, historical sample storage specifications, and clustering classification criteria are integrated and processed to establish a one-to-one correspondence between abnormal time series samples and single-channel feature contribution vectors, and to generate a standardized abnormal induction and analysis process. Based on the standardized anomaly induction process, high-contribution feature information is extracted. The MAI over-limit trigger signal is used as the starting source, the gradient contribution value of the volume element channel is used as the division basis, and the DBSCAN / K-Means clustering algorithm is used as the processing method. This enables the output of multiple results, including single feature contribution ranking, multi-source coupled precursor clustering, three-dimensional spatial coordinate mapping, disaster early warning and root cause analysis.

6. A multi-source state fusion and GAN anomaly precursor discovery system for deeply buried underground caverns, characterized in that, The system is used to execute executable instructions to perform the multi-source state fusion and GAN anomaly precursor mining method for deep buried underground caverns as described in any one of claims 1 to 5.

7. An electronic device, characterized in that, include: First processor; The processor also includes a memory for storing executable instructions of the first processor; wherein the first processor is configured to execute the method for multi-source state fusion and GAN anomaly precursor mining of deep underground caverns according to any one of claims 1 to 5 by executing the executable instructions.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the second processor, it implements the method for multi-source state fusion and GAN anomaly precursor mining of deep underground caverns as described in any one of claims 1 to 5.

Images