Rock intelligent grading and support decision method, device and medium based on multi-source heterogeneous data fusion

Abstract

The application discloses a kind of based on multi-source heterogeneous data fusion surrounding rock intelligent grading and support decision method, equipment and medium, it is related to tunnel engineering technical field, comprising: obtaining geological sketch image in construction, monitoring sensor time series data, microseismic event data and drilling equipment time series data;Subsequently, all data are unified to the same space-time reference frame;Respectively, feature extraction is carried out on each registered data, to obtain corresponding image, monitoring, microseismic and drilling feature vector;These feature vectors are fused, and the dynamic stability index of surrounding rock is calculated based on the fusion result;Finally, the index, current excavation process information and adjacent section support effect feedback information are used as decision state, input into pre-trained support decision model, and the model outputs quantitative support parameters for the current construction section.The application realizes the effective cooperation and real-time fusion of multi-source data, and provides dynamic, accurate surrounding rock stability evaluation and adaptive support decision.

Description

Technical Field

[0001] This invention relates to the field of tunnel engineering technology, and in particular to a method, equipment and medium for intelligent classification and support decision-making of surrounding rock based on multi-source heterogeneous data fusion. Background Technology

[0002] In tunnel construction, various types of data are typically collected, including geological sketches, stress and displacement monitoring, microseismic signals, and drilling parameters. These data often differ in format, temporal scale, and spatial scale, making direct integration difficult. This results in isolated information silos, hindering their full synergy and value. Furthermore, dynamic rock grading relies primarily on phased geological surveys and manual judgment, making it difficult to integrate the massive amounts of real-time data generated during construction. Consequently, rapid, accurate, and dynamic updates of grading results are impossible. Moreover, the determination of support schemes usually depends on static grading results and standard lookup tables, lacking a decision-making mechanism that can adaptively adjust to changes in the actual rock condition and output quantitative support parameter recommendations. This leads to limitations in the real-time performance, accuracy, and economy of support design. Summary of the Invention

[0003] This invention provides a method, equipment, and medium for intelligent classification and support decision-making of surrounding rock based on multi-source heterogeneous data fusion. The technical problem it aims to solve is the lack of coordination of multi-source information, the lag in condition assessment, and the rigidity of support decision-making in tunnel construction.

[0004] To address the aforementioned technical problems, in a first aspect, embodiments of the present invention provide a method for intelligent rock mass classification and support decision-making based on multi-source heterogeneous data fusion, comprising: Acquire multi-source heterogeneous data during the tunnel construction process, including geological sketch images, time-series data from monitoring sensors, microseismic event data, and time-series data from drilling equipment. Spatiotemporal registration is performed on the multi-source heterogeneous data to unify the geological sketch images, the time-series data of the monitoring sensors, the microseismic event data, and the time-series data of the drilling equipment under the same spatiotemporal reference. Feature extraction is performed on the spatiotemporally registered geological sketch image, the time-series data of the monitoring sensor, the microseismic event data, and the time-series data of the drilling equipment to obtain the corresponding image feature vector, monitoring feature vector, microseismic feature vector, and drilling feature vector. The image feature vector, the monitoring feature vector, the microseismic feature vector, and the drilling feature vector are fused together, and the dynamic stability index of the surrounding rock is calculated based on the fused feature vector. The dynamic stability index of the surrounding rock, the current excavation process information, and the support effect feedback information of adjacent supported sections are used as the decision state and input into the pre-trained support decision model. The support decision model then outputs quantitative support parameters for the current construction section.

[0005] The spatiotemporal registration of multi-source heterogeneous data includes: Establish a three-dimensional spatial coordinate system associated with the tunnel design axis, and construct a timeline with the construction progress time as the main axis; For a geological sketch image, based on the corresponding shooting location and shooting time, the geological sketch image is mapped to the corresponding position in the three-dimensional spatial coordinate system and associated with the corresponding time on the timeline; For the time-series data of the monitoring sensors and the drilling equipment, based on the spatial location of the collected sensors and the data timestamp, each data point is mapped to the corresponding position in the three-dimensional spatial coordinate system and associated with the corresponding time on the timeline; For microseismic event data, based on the spatial location results of the seismic source and the triggering time, each microseismic event is mapped to the corresponding position in the three-dimensional spatial coordinate system and associated with the corresponding time on the timeline; For the drilling equipment time-series data, based on the real-time spatial position of the drill bit and the data timestamp recorded during the acquisition and obtained through the integrated navigation system, each data point is mapped to the corresponding position in the three-dimensional spatial coordinate system and associated with the corresponding time on the timeline.

[0006] Optionally, the extraction of the image feature vector includes: The geological sketch image is input into a pre-trained semantic segmentation network for geological structural surfaces, which is trained based on a pre-defined geological annotation image library for tunnel engineering. The geological structural surface semantic segmentation network outputs a pixel-level binary map of structural surface distribution and a structural surface type label map. Geometric feature calculations are performed on the binary map of the structural surface distribution to extract topological network feature vectors of structural surface trace length, spacing, and intersection relationships; The topological network feature vector is concatenated with the proportion features of various structural surfaces statistically obtained from the structural surface type label map to form the image feature vector.

[0007] Optionally, the extraction of the monitoring feature vector includes: The time-series data of the monitoring sensors are grouped according to sensor type and spatial location to form multiple parallel time series. For each time series, a one-dimensional convolutional neural network is used to extract local mutation pattern features, and a gated recurrent unit network is used to extract long-term evolution trend features. Early fusion of the local mutation mode features and the long-term evolution trend features from different types of sensors in the same spatial region generates a local stability symptom vector for the same spatial region. The local stability symptom vectors of all monitored areas are aggregated and encoded into the global monitoring feature vector through a fully connected layer.

[0008] Optionally, the step of fusing the image feature vector, the monitoring feature vector, the microseismic feature vector, and the drilling feature vector, and calculating the surrounding rock dynamic stability index based on the fused feature vector, includes: The image feature vector, monitoring feature vector, microseismic feature vector, and drilling feature vector are input into a preset gated multimodal fusion network; The gated multimodal fusion network sets a learnable gate weight vector for each type of feature vector, and generates a gate value between 0 and 1 through the sigmoid function; The gate value is used to perform element-wise weighted modulation on each feature vector; The modulated feature vectors are concatenated and passed through a regression subnetwork to output a continuous scalar value as the dynamic stability index of the surrounding rock.

[0009] Optionally, the support decision-making model is trained based on a hierarchical reinforcement learning framework, including: The high-level strategy model takes the decision state as input and outputs an abstract support strategy type. The underlying strategy model receives the support strategy type and decision state output by the higher-level strategy model, and outputs the specific quantitative support parameters. The high-level strategy model and the low-level strategy model are trained collaboratively using an asynchronous advantage executor-commentator algorithm. The reward signal of the high-level strategy model is generated based on the long-term cost and safety benefit assessment results generated after the quantitative support parameters output by the low-level strategy model are executed.

[0010] Optionally, the acquisition of the support effect feedback information includes: Deploy a distributed fiber optic sensing system in the vicinity of the supported section to obtain continuous spatial distribution data of the strain of the support structure and the deformation of the surrounding rock. Based on the continuous spatial distribution data, the stress utilization rate of the support structure, the development depth of the plastic zone of the surrounding rock, and the deformation compatibility coefficient are calculated. The stress utilization rate of the support structure, the development depth of the plastic zone of the surrounding rock, and the deformation coordination coefficient are compared with the design expected values ​​to generate a multi-dimensional support effectiveness evaluation vector, which serves as feedback information on the support effect.

[0011] Optionally, the reward function of the hierarchical reinforcement learning framework is constructed and calculated through the following steps: A safety assessment item is set, which is calculated based on the monitored surrounding rock deformation data. When the surrounding rock deformation data does not exceed the preset safety threshold, the safety assessment item is zero. When the surrounding rock deformation data exceeds the safety threshold, the safety assessment item is a negative value, and the absolute value of the safety assessment item is positively correlated with the degree to which the surrounding rock deformation data exceeds the safety threshold. An economic evaluation item is defined, and the calculation of the economic evaluation item includes: calculating the direct material cost based on the support parameters output by the action space, and estimating the future maintenance cost through a long-term maintenance cost prediction model based on the support effectiveness evaluation vector in the support effect feedback information; the economic evaluation item is the negative value of the sum of the direct material cost and the future maintenance cost. A construction continuity assessment item is set up. The construction continuity assessment item is used to penalize drastic fluctuations in support parameters. The calculation method is as follows: calculate the Euclidean distance between the support parameter vector output in the current decision cycle and the support parameter vector output in the previous decision cycle, and multiply the Euclidean distance by a preset negative penalty coefficient. The safety assessment item, the economic assessment item, and the construction continuity assessment item are weighted and summed to obtain the reward value of the reward function at each decision step.

[0012] Optionally, the method employs a hybrid triggering mechanism combining event-driven and period-driven approaches, including: Set a fixed minimum decision-making period as the base period; Within the basic period, the microseismic feature vector and the monitoring feature vector are monitored in real time. If a cluster of microseismic events with energy exceeding a preset threshold or an accelerating trend in deformation rate is identified, an emergency decision-making process is immediately triggered. If no emergency occurs, a routine decision-making process will be triggered at the end of the basic cycle. The emergency decision-making process and the regular decision-making process call upon support decision-making models with different risk preferences.

[0013] Optionally, the method further includes a decision simulation and verification step based on digital twin technology, including: Based on current geological sketch images and monitoring data, update the mechanical parameters of the digital twin of the tunnel surrounding rock; In the digital twin, a rapid finite element numerical simulation is performed on the quantitative support parameter scheme output by the support decision model; The simulated predicted deformation and stress distribution are matched with a pre-set historical case library for similarity analysis, and a safety margin assessment and risk warning are generated based on the matching results.

[0014] Optionally, the method further includes interpretability analysis of the gated multimodal fusion network, including: Record and analyze the gated value sequence of the gated multimodal fusion network in historical decision-making; Using association rule mining algorithms, we analyze strong association rules between different gate value patterns and specific geological conditions, construction conditions and final surrounding rock stability results; The discovered strong association rules are transformed into readable expert knowledge entries, which can be used to verify the model logic or assist in decision-making understanding under new operating conditions.

[0015] Optionally, the method further includes a step of periodically performing incremental updates to the model, including: Set a model update cycle. At the end of each update cycle, collect all multi-source heterogeneous data obtained in the actual construction during the update cycle, the actual support parameters used, and the actual deformation and stress response data of the surrounding rock monitored subsequently, to form an incremental training sample set. Using the incremental training sample set, the original model is collaboratively fine-tuned. The original model includes a geological structure surface semantic segmentation network for feature extraction, a one-dimensional convolutional neural network and a gated recurrent unit network, a gated multimodal fusion network for feature fusion, and the support decision model. The collaborative fine-tuning employs an incremental learning algorithm, which updates the connection weight parameters while retaining the knowledge of the original model and with the goal of minimizing the prediction error of newly added samples. The finely tuned and updated connection weight parameters are deployed to the online system to replace the original model for subsequent real-time decision-making.

[0016] Secondly, embodiments of the present invention also provide a computer device, which includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-described method.

[0017] Thirdly, embodiments of the present invention also provide a computer-readable storage medium storing a computer program that, when executed by a processor, can implement the above-described method.

[0018] This invention effectively solves the problems of uncoordinated multi-source information, delayed condition assessment, and rigid support decisions in tunnel construction through a systematic data alignment, feature fusion, and intelligent decision-making process. First, by constructing a unified spatiotemporal benchmark, geological, monitoring, microseismic, and drilling data from different sources, formats, and acquisition frequencies are transformed into standardized information that can be correlated and analyzed, breaking down information barriers and enabling in-depth mining and complementarity of data value. Second, by utilizing multimodal feature fusion and dynamic calculation, a continuous, real-time surrounding rock stability index is generated, replacing the traditional discrete grading method that relies on staged manual judgment, significantly improving the sensitivity and update frequency of condition perception. Finally, based on this dynamic index and combined with construction progress and historical effect feedback, a pre-trained intelligent decision-making model outputs quantitative support parameters, enabling the support scheme to respond to changes in the surrounding rock in real time and adaptively optimize, thereby improving the economy and scientific nature of decision-making while ensuring safety, achieving a fundamental shift from static experience-based lookup to dynamic intelligent optimization. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 A flowchart illustrating a method for intelligent classification and support decision-making of surrounding rock based on multi-source heterogeneous data fusion, provided in an embodiment of the present invention; Figure 2 This is a schematic block diagram of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0023] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0024] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0025] As used in this specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if [described condition or event] is detected" may be interpreted, depending on the context, as "once determined," "in response to determination," "once [described condition or event] is detected," or "in response to detection of [described condition or event]."

[0026] Please see Figure 1 This invention provides a method for intelligent classification and support decision-making of surrounding rock based on multi-source heterogeneous data fusion. The method includes the following steps: S1, acquire multi-source heterogeneous data during tunnel construction, including geological sketch images, time-series data from monitoring sensors, microseismic event data, and time-series data from drilling equipment.

[0027] In practice, acquiring multi-source heterogeneous data during tunnel construction involves simultaneous collection from multiple independent subsystems. High-resolution geological sketches of the tunnel face and tunnel walls are obtained by deploying explosion-proof digital cameras or 3D laser scanners at the tunnel face, collecting data once or multiple times per excavation cycle. The image format is RGB or point cloud data.

[0028] Furthermore, by pre-installing fiber optic grating sensor arrays or electronic sensors in the form of grids or key sections inside the surrounding rock and on the surface of the support structure, the monitoring sensor time-series data of various physical quantities such as internal rock stress, surface displacement, convergence deformation, and anchor bolt axial force are collected synchronously at a sampling frequency of 1 to several times per minute. The data is transmitted in real time in the form of a time-stamped numerical stream.

[0029] Furthermore, a multi-channel microseismic monitoring array deployed within the rock mass surrounding the tunnel is used to collect vibration waveform signals generated by rock mass fracturing. After being processed by a real-time processing system, a structured list of microseismic events is generated, containing attributes such as the three-dimensional coordinates of the seismic source, the time of occurrence, the released energy, and the dominant frequency.

[0030] Furthermore, by using embedded sensors installed on the rock drilling rig or TBM drill bit drive unit, drilling equipment timing data such as drilling speed, feed pressure, rotational torque, and impact frequency, which are directly related to the drilling process, are collected in real time and uploaded in real time via industrial bus.

[0031] S2, perform spatiotemporal registration on the multi-source heterogeneous data, and unify the geological sketch image, the time series data of the monitoring sensor, the microseismic event data, and the time series data of the drilling equipment under the same spatiotemporal reference.

[0032] In practice, the first step is to establish a unified spatiotemporal reference. This reference consists of a three-dimensional spatial coordinate system fixed to the tunnel design axis and a time axis based on the absolute time of the project. The spatial coordinate system adopts the Frenet frame system established along the design centerline, where the position of any point is uniquely determined by the mileage S along the axis, the lateral offset L, and the vertical offset H. The time axis begins at the start of the project, and all times are based on this origin.

[0033] Furthermore, for geological sketch images, by reading the shooting station number and time in their metadata, or by using image-tunnel BIM model registration technology, the corresponding spatial section position (S, L fixed) and shooting time point t are determined, and they are mapped to a two-dimensional data plane in the spatiotemporal reference.

[0034] Furthermore, for the time-series data of the monitoring sensors, their spatial coordinates (S, L, H) are determined based on the predefined installation location of each sensor. Each timestamped reading collected by the sensor is directly mapped to a sequence of data points with fixed spatial coordinates that change over time in a spatiotemporal reference.

[0035] Furthermore, for microseismic event data, spatial coordinates (X, Y, Z) are obtained based on the source location results, and (S, L, H) are obtained through coordinate transformation. The trigger time t is then aligned with the time axis, thereby mapping each event to a discrete point (S, L, H, t) in a spatiotemporal reference.

[0036] Furthermore, for the timing data of the drilling equipment, a high-precision clock and a combined navigation and positioning system are integrated for each piece of equipment. For each drilling parameter (such as drilling speed and thrust) collected, the system records its acquisition time t and simultaneously obtains the spatial coordinates of the drill bit's working point at that time. Through coordinate transformation, these coordinates are mapped to the Frenet system to obtain the real-time position (S(t), L(t), H(t)). Thus, each drilling parameter data point is converted into a dynamic data point (S(t), L(t), H(t), t) whose spatial position changes continuously over time in a spatiotemporal reference.

[0037] Through the above steps, all data from different sources, in different formats, and with different collection frequencies are converted and registered under the same spatiotemporal reference framework with clear physical meaning, providing a precise spatiotemporal alignment basis for subsequent feature extraction and fusion.

[0038] For example, in some preferred embodiments, the spatiotemporal registration of multi-source heterogeneous data includes: establishing a three-dimensional spatial coordinate system associated with the tunnel design axis and constructing a timeline with the construction progress time as the main axis; for geological sketch images, mapping the geological sketch images to the corresponding positions in the three-dimensional spatial coordinate system according to the corresponding shooting positions and shooting times, and associating them with the corresponding times on the timeline; for monitoring sensor time-series data and drilling equipment time-series data, mapping each data point to the corresponding position in the three-dimensional spatial coordinate system according to the spatial position of the acquisition sensors and data timestamps, and associating them with the corresponding times on the timeline; for microseismic event data, mapping each microseismic event to the corresponding position in the three-dimensional spatial coordinate system according to the spatial positioning results of the seismic source and the triggering time, and associating it with the corresponding times on the timeline; for drilling equipment time-series data, mapping each data point to the corresponding position in the three-dimensional spatial coordinate system according to the real-time spatial position of the drill bit recorded during acquisition and obtained through the integrated navigation system and data timestamps, and associating it with the corresponding times on the timeline.

[0039] In practice, spatiotemporal registration of multi-source heterogeneous data is a process of sequentially registering different types of data to the same spatiotemporal reference according to a unified architecture. First, a Frenet three-dimensional spatial coordinate system is established with the tunnel design centerline as the reference line. This coordinate system defines the spatial position with the longitudinal mileage S, lateral offset L, and vertical offset H along the tunnel axis. At the same time, a continuously increasing timeline is constructed with the absolute time of project commencement as the origin.

[0040] Furthermore, for geological sketch images, the system reads the shooting station number and the time stamp accurate to the second from its metadata. The station number is converted into (S, L, H) coordinates in the Frenet coordinate system using design data, thereby determining the location of the two-dimensional spatial section represented by the image; at the same time, the shooting timestamp is aligned to the absolute time on the timeline, completing the static positioning of the image in the spatiotemporal reference.

[0041] Furthermore, for monitoring the time-series data of the sensors, each sensor has its fixed (S, L, H) coordinates determined during installation. During data acquisition, each monitored value (such as stress or displacement readings) is accompanied by a timestamp of the acquisition time. Based on the sensor's fixed spatial coordinates and the timestamp of the reading, the system directly maps each data point to a sequence of points with fixed spatial coordinates and continuously distributed on the timeline in a spatiotemporal reference.

[0042] Furthermore, dynamic spatiotemporal mapping is achieved for the timing data of drilling equipment. Each piece of equipment is equipped with a combined navigation system that outputs the three-dimensional absolute coordinates of the drill bit's working point in real time, while a high-precision clock provides a unified timestamp. The system converts the absolute coordinates into Frenet coordinates (S(t), L(t), H(t)) in real time using a preset coordinate transformation model. Each drilling parameter data point (such as drilling speed and torque) is thus associated with a dynamic spatial location (S(t), L(t), H(t)) and a precise time t, thereby forming a spatial trajectory and parameter sequence that varies with time within a spatiotemporal reference.

[0043] Furthermore, for microseismic event data, the system receives the three-dimensional coordinates and trigger time output by its source location module. These coordinates are then transformed to the Frenet coordinate system to obtain (S, L, H), and the trigger time is aligned to the timeline. Each microseismic event is thus mapped to a discrete event point in the spatiotemporal reference, possessing a definite spatial location and occurrence time.

[0044] Through the above methods, all four types of data, with different sources, formats, and acquisition modes, are converted into standardized spatiotemporal data with a unified (S, L, H, t) as the basic index, thus realizing the foundation for accurate alignment and fusion of cross-modal information.

[0045] The spatiotemporal registration method provided in this embodiment establishes a natural coordinate system that closely matches the physical morphology of the tunnel by introducing the Frenet frame system. This allows the data to be uniformly represented as intuitive physical quantities of distance along the tunnel axis, lateral and vertical offsets, regardless of the data source, greatly facilitating the spatial correlation analysis and visualization of the data.

[0046] S3, respectively extract features from the spatiotemporally registered geological sketch image, the time-series data of the monitoring sensor, the microseismic event data, and the time-series data of the drilling equipment to obtain the corresponding image feature vector, monitoring feature vector, microseismic feature vector, and drilling feature vector.

[0047] In practice, feature extraction is performed on the spatiotemporally registered data. For example, for a geological sketch image, a deep convolutional neural network feature extractor pre-trained on a large rock mass image dataset is input. The network outputs a dense vector of fixed length as an image feature vector, which encodes visual features such as texture, structure, and fissure distribution in the image.

[0048] Furthermore, for the time-series data of the monitoring sensors, a sliding time window is used to divide it into segments. For each segment, a one-dimensional convolutional layer is used to capture local mutation patterns, and a bidirectional gated recurrent unit layer is used to capture long-term trends that depend on previous and subsequent data. The output features of the two networks are fused and pooled to obtain a monitoring feature vector that characterizes the deformation dynamics of that time period.

[0049] Furthermore, for microseismic event data, within a dynamically updated spatiotemporal cube, the cumulative energy of events per unit time, event frequency, b-value of energy-frequency relationship, and spatial clustering index of events (such as kernel density) are statistically analyzed, and these statistics are combined into a microseismic feature vector.

[0050] Furthermore, for the time-series data of the drilling equipment, the statistical characteristics such as the mean, variance, peak value and area under the change curve of the parameters within each drilling cycle are calculated, and combined with the correlation coefficient between the drilling rate and the propulsion pressure, a drilling feature vector is formed.

[0051] In some preferred embodiments, the extraction of the image feature vector includes: inputting the geological sketch image into a pre-trained geological structure surface semantic segmentation network, which is trained based on a preset tunnel engineering geological annotation image library; outputting a pixel-level binary map of structure surface distribution and a structure surface type label map through the geological structure surface semantic segmentation network; performing geometric feature calculation on the binary map of structure surface distribution to extract topological network feature vectors of structure surface trace length, spacing, and intersection relationship; and concatenating the topological network feature vectors with the proportion features of various structure surfaces statistically obtained from the structure surface type label map to jointly constitute the image feature vector.

[0052] In practice, the pre-trained geological structural surface semantic segmentation network adopts a U-Net++ architecture with an attention gating mechanism. The network encoder uses ResNet-34 pre-trained on ImageNet as the backbone to extract multi-scale features. The decoder path adopts a nested dense skip connection structure and introduces an attention gating module, which can automatically focus on the region related to the structural surface in the feature map based on deep semantic features, suppressing background interference. The tunnel engineering geological annotation image library used for training contains tens of thousands of images collected under different geological conditions and finely annotated by senior geological engineers. The annotation content is pixel-level structural surface contours and types (such as joints, faults, bedding, unloading fractures). This tunnel engineering geological annotation image library covers a variety of typical surrounding rock grades, structural surface development characteristics, and lighting conditions to ensure the generalization ability of the network model. The network's training objective function is a composite loss function combining cross-entropy loss and Dice coefficient to optimize the segmentation accuracy of structural surface boundaries.

[0053] Furthermore, after inputting the geological sketch image into the network, the network propagates forward, ultimately generating a (C+1)-dimensional probability vector for each pixel in the output layer, where C is the number of structural surface types. By performing an Argmax operation, pixel-level semantic segmentation results are obtained, i.e., a structural surface type label map. Simultaneously, by merging all non-background category pixel regions, a binary map of structural surface distribution is obtained.

[0054] Furthermore, geometric features are calculated from the binary image of the structural surface distribution. First, the Zhang-Suen parallel thinning algorithm is used to skeletonize the binary image, extracting a single-pixel-wide network of structural surface traces. Next, connective component analysis and polygon approximation algorithms are used to decompose the skeleton into a series of continuous line segments (i.e., trace segments). For each trace segment, its length, midpoint coordinates, orientation (angle with the hole axis), and dip angle are calculated. Then, based on the statistics of all trace segments, the average trace length, trace density (total trace length per unit area), and the dominant trace orientation determined using a rose diagram are calculated. To analyze topological relationships, a graph model is constructed with trace midpoints as nodes and trace intersections as edges. From this graph model, network connectivity indices, average clustering coefficients, and the number and size distribution of structural surface clusters identified using a community detection algorithm are extracted. These graph theory features are combined into a topological network feature vector.

[0055] Furthermore, from the structural surface type label map, the number of pixels belonging to each type of structural surface (such as type I joints, type II joints, faults, etc.) is counted, and the percentage of each type to the total number of pixels of all structural surfaces is calculated, resulting in a C-dimensional percentage feature vector. This vector represents the relative development degree of each type of structural surface in the currently exposed surface.

[0056] Furthermore, the calculated topological network feature vector (containing geometric and topological information) and the structural surface type proportion feature vector (containing rock mass composition and tectonic information) are concatenated along the feature dimensions to form a comprehensive image feature vector. This image feature vector provides a comprehensive and quantitative description of the geological structure of the tunnel face from multiple dimensions such as morphology, distribution, connectivity, and type composition.

[0057] This embodiment achieves fully automated, high-precision, and multi-dimensional quantitative analysis of geological sketch image information by integrating advanced semantic segmentation networks with refined geometric topology analysis. Employing a U-Net++ network structure with an attention mechanism, it effectively handles the challenges of varying contrast and scale between structural surfaces and background in geological images, significantly improving the accuracy and boundary clarity of structural surface identification under complex conditions. Based on skeletonization and graph theory-based feature extraction methods, visual fracture patterns are transformed into rigorously computable mathematical descriptions, such as trace density, dominant orientation, and network connectivity index. These indicators are crucial parameters in rock mass quality assessment but have traditionally been difficult to obtain quickly and accurately. Furthermore, statistical analysis of the proportion of different types of structural surfaces provides quantitative information on the rock mass tectonic background. The final image feature vector, formed by stitching together the data, represents a deep digital representation of the core geological factor of "rock mass structure."

[0058] In some preferred embodiments, the extraction of the monitoring feature vector includes: grouping the time-series data of the monitoring sensors according to sensor type and spatial location to form multiple parallel time series; for each time series, using a one-dimensional convolutional neural network to extract local mutation pattern features and using a gated recurrent unit network to extract long-term evolution trend features; performing early fusion of the local mutation pattern features and the long-term evolution trend features from different types of sensors in the same spatial region to generate a local stability symptom vector for the same spatial region; aggregating the local stability symptom vectors of all monitoring regions and encoding them into a global monitoring feature vector through a fully connected layer.

[0059] In practice, when grouping the time-series data of the monitoring sensors according to sensor type and spatial location, a hierarchical grouping strategy is adopted. First, based on the type of physical quantity, it is divided into major categories such as displacement group (convergence gauges, multi-point displacement gauges), stress-strain group (anchor stress gauges, concrete strain gauges), and pressure group (earth pressure cells). Under each major category, secondary grouping is performed according to the mileage of the sensor installation section, grouping similar sensors on the same section or adjacent sections (e.g., within 5 meters) together to form a "physical quantity-spatial cluster".

[0060] Furthermore, feature extraction is performed in parallel for each time series. Local mutation pattern features are extracted using a small CNN containing three one-dimensional convolutional layers and max-pooling layers. Kernel sizes of 3, 5, and 7 are used to capture local fluctuations and spikes at different time scales. Long-term evolutionary trend features are extracted using a two-layer bidirectional GRU network, whose hidden states can fuse information from preceding and following time steps to capture monotonically increasing, decreasing, or periodic trends in the sequence. The final output features of the CNN and GRU are flattened and then processed through a fully connected layer to generate local feature vectors and trend feature vectors, respectively.

[0061] Furthermore, early fusion occurs within "physical quantity-spatial clusters." For sensors of different physical quantities within the same spatial cluster (e.g., convergence meters and anchor stress meters on the same cross-section), their extracted local feature vectors and trend feature vectors are concatenated to form a longer feature vector. This concatenated vector is then fused and dimensionality-reduced through a small shared fully connected network, outputting a local stability symptom vector that comprehensively reflects the coordinated response of multiple physical quantities within the spatial cluster. This vector characterizes the "overall health status" of the local region.

[0062] Furthermore, when aggregating the local symptom vectors of all monitored areas, all local symptom vectors are arranged into a sequence according to their spatial location along the tunnel longitudinal direction. This sequence is input into a one-dimensional convolutional layer or Transformer encoder, which can capture the spatial correlation between the states of different monitored areas (e.g., whether an anomaly at the front section is propagating backward).

[0063] Furthermore, through a global average pooling layer and a fully connected layer, the spatially structured sequence information is encoded into a fixed-dimensional, global monitoring feature vector. This vector encapsulates the spatial distribution pattern and overall evolution trend of the surrounding rock deformation and stress field throughout the entire monitoring segment.

[0064] This embodiment achieves a deep perception of the dynamic response of surrounding rock, from microscopic anomalies to macroscopic trends, and from local signs to the global situation, through multi-level and multi-scale feature extraction and fusion of monitoring data. The hierarchical grouping strategy respects the physical meaning and spatial attributes of the monitoring data, providing a reasonable structure for subsequent feature fusion. Using a multi-scale one-dimensional CNN to extract local mutations can sensitively capture short-term, high-frequency anomaly signals that indicate local rock mass fracturing or stress redistribution; these signals are often precursors to instability. Using a bidirectional GRU to extract long-term trends allows for a grasp of the macroscopic direction of whether surrounding rock deformation tends to converge or continue to develop, providing a temporal basis for stability assessment. Early fusion of multiple physical quantities within spatial clusters enables cross-validation and complementarity of displacement and stress information, making the judgment of local states more robust and avoiding misjudgments caused by single sensor failure or interference. Furthermore, utilizing network layers capable of capturing spatial correlations (such as Transformers) to aggregate global information allows the system to understand whether anomalies are isolated or have formed regularly propagating disease zones, thereby making a more accurate assessment of overall stability.

[0065] S4, the image feature vector, the monitoring feature vector, the microseismic feature vector and the drilling feature vector are fused, and the dynamic stability index of the surrounding rock is calculated based on the fused feature vector.

[0066] In practice, the four feature vectors are input into a feature fusion module. This module first projects each vector onto the same feature dimension using a fully connected layer. Then, it employs a multi-head attention-based fusion mechanism to calculate the correlation weights between the feature vectors, weighting and aggregating them to generate a unified fused feature vector. This fused feature vector is then fed into a multilayer perceptron regressor, which outputs a continuous, dimensionless real value, namely the Dynamic Stability Index (DSI). Its value range can be limited to 0 to 1 using the sigmoid function; a higher value indicates better stability.

[0067] In some preferred embodiments, the step of fusing the image feature vector, the monitoring feature vector, the microseismic feature vector, and the drilling feature vector, and calculating the surrounding rock dynamic stability index based on the fused feature vector, includes: inputting the image feature vector, the monitoring feature vector, the microseismic feature vector, and the drilling feature vector into a preset gated multimodal fusion network; setting a learnable gate weight vector for each type of feature vector by the gated multimodal fusion network, and generating a gate value between 0 and 1 using the sigmoid function; using the gate value to perform element-wise weighted modulation on each feature vector; concatenating the modulated feature vectors, and outputting a continuous scalar value as the surrounding rock dynamic stability index through a regression subnetwork.

[0068] In practice, the pre-defined gated multimodal fusion network is a neural network module with a specialized structure. The network receives four input tensors: image feature vector V_img (dimension d_i), monitoring feature vector V_mon (dimension d_m), microseismic feature vector V_ms (dimension d_s), and drilling feature vector V_dr (dimension d_d).

[0069] Furthermore, the learnable gating weight vectors set for each type of feature vector are four independent parameter tensors with the same dimension as the corresponding input vector, denoted as G_img, G_mon, G_ms, and G_dr. These parameters are initialized and trained along with the network. The gating weight vectors are initialized with small random values ​​in the early stages of training so that the model can adaptively learn the importance weights of different data sources and feature dimensions under different engineering scenarios during training.

[0070] Furthermore, the process of generating gate values ​​is as follows: for each gate weight vector G, the Sigmoid activation function is applied element-wise, i.e., Gate = σ(G). Since the range of the Sigmoid function is (0, 1), each scalar element in each generated gate vector Gate is an independent gate value, the magnitude of which is adaptively determined by the training process based on the importance of that feature dimension.

[0071] Furthermore, the element-wise weighted modulation operation using the gate value is an element-wise multiplication (Hadamard product). Specifically: V'_img = V_img⊙σ(G_img), V'_mon = V_mon⊙σ(G_mon), V'_ms = V_ms⊙σ(G_ms), V'_dr = V_dr⊙σ(G_dr). This operation is equivalent to applying an adaptive soft switch or attenuator to each dimension of the original feature vector. A gate value close to 1 allows the feature in that dimension to pass through almost completely; a gate value close to 0 suppresses the feature in that dimension to near zero, which is equivalent to performing a dynamic, fine-grained feature selection at the feature level.

[0072] Furthermore, the four modulated feature vectors V'_img, V'_mon, V'_ms, and V'_dr are concatenated along the feature dimension to form a fused feature vector of length (d_i+d_m+d_s+d_d).

[0073] Furthermore, this fused feature vector is then fed into a regression subnetwork. This subnetwork consists of two fully connected layers, with a ReLU activation function in between, and a final output layer using a linear activation function, outputting a single, continuous real-valued scalar, namely the Dynamic Stability Index (DSI). This index is constrained through training to reflect the overall stability of the surrounding rock.

[0074] This embodiment introduces a learnable, element-level gating mechanism to provide a dynamic and adaptive feature selection and weighting scheme for multi-source information fusion. Traditional fusion methods, such as concatenation or average pooling, treat all feature dimensions equally, failing to handle variations in signal-to-noise ratio and contribution differences between different data sources under varying construction stages and geological conditions. The core innovation of this method lies in generating an independent weight (gating value) for each feature dimension using the Sigmoid gating function. This weight is not fixed but learned from the data through network training. This allows the model to automatically determine which features are critical and reliable based on the current context (implicit in the input features), thus assigning them high weights (gating values ​​close to 1); while suppressing redundant, noisy, or irrelevant features (gating values ​​close to 0). For example, when the surrounding rock is intact and monitoring data is stable, the model may rely more on geological image features; while when rock fracturing is active and microseismic events are frequent, the model will automatically increase the weight of the microseismic feature vector. This fine-grained, dynamic feature modulation capability endows the fusion process with high intelligence and context awareness, effectively enhancing the discriminative power of the fused features. The dynamic stability index of the surrounding rock calculated based on this high-quality fused features can therefore more accurately and robustly reflect the true state of the surrounding rock.

[0075] S5, using the surrounding rock dynamic stability index, current excavation process information, and support effect feedback information of adjacent supported sections as decision states, inputs them into the pre-trained support decision model, and the support decision model outputs quantitative support parameters for the current construction section.

[0076] In practice, a decision state vector is constructed by combining the surrounding rock dynamic stability index, the process information vector encoding the current excavation cycle number and process (such as upper section excavation, lower bench excavation), and the support effect feedback vector calculated from monitoring data of adjacent completed support sections. This decision state vector is input into a pre-trained deep neural network model for support decision-making. The model undergoes end-to-end training, directly receiving the state vector and, through its internal multi-layer nonlinear transformations, ultimately outputting a vector containing multiple specific support parameter values, such as quantified suggested values ​​for anchor length, spacing, row spacing, shotcrete thickness, and steel arch frame type and spacing, thus completing the mapping from state perception to parameter decision-making.

[0077] This invention addresses the core issues of insufficient utilization of multi-source information and delayed decision-making in tunnel construction by constructing a complete closed-loop process for automatic data perception, alignment, parsing, fusion, and decision-making. As seen in the implementation, by constructing a unified spatiotemporal benchmark, heterogeneous image, time-series, and event list data are transformed into comparable and correlated spatiotemporal data objects, creating prerequisites for information fusion. The deep feature extraction method designed for each type of data characteristic can automatically uncover deep feature patterns that surpass manual interpretation, such as quantifying the complexity of structural surface networks from images and capturing invisible deformation precursor trends from time-series data. The fusion method employing an attention mechanism can dynamically evaluate the confidence and contribution of different data sources in the current scenario, achieving intelligent weighting and generating a dynamic stability index that more comprehensively reflects the true state of the surrounding rock. Finally, decisions are made based on this dynamic index, accurate construction progress, and real support effect feedback, enabling support schemes to break free from dependence on static geological classification and achieve true real-time response and adaptive adjustment. The entire process is data-driven, reducing the delays and subjectivity of human intervention and improving the real-time, objectivity, and accuracy of hierarchical and decision-making processes.

[0078] In some preferred embodiments, the training of the support decision model is based on a hierarchical reinforcement learning framework, including: a high-level strategy model taking the decision state as input and outputting an abstract support strategy type; a low-level strategy model receiving the support strategy type output by the high-level strategy model and the decision state, and outputting specific quantitative support parameters; wherein the high-level strategy model and the low-level strategy model are co-trained through an asynchronous advantage executor-critic algorithm, and the reward signal of the high-level strategy model is generated based on the long-term cost and safety benefit assessment results generated after the quantitative support parameters output by the low-level strategy model are executed.

[0079] In practice, the support decision-making model based on the hierarchical reinforcement learning framework is trained using a simulator-learner architecture. Both the high-level strategy model (Manager) and the low-level strategy model (Worker) are multilayer perceptron neural networks. The input to the high-level strategy is a decision state vector, and the output is a Softmax probability distribution leading to K abstract support strategy types (such as "active reinforcement," "moderate yielding," and "monitoring and early warning"), with the strategy type selected for the current period determined through sampling. The input to the low-level strategy is the concatenation of the strategy type output from the high-level strategy (converted into a dense vector through an embedding layer after one-hot encoding) and the decision state vector, and the output is a multidimensional continuous action vector directly corresponding to quantitative support parameters such as anchor spacing, length, and prestress.

[0080] Furthermore, the training environment is a parameterized tunnel construction numerical simulator that receives support parameter actions and calculates the surrounding rock deformation response for the next stage based on a simplified rock mechanics model. Co-training of the high-level and low-level models employs a variant of the Asynchronous Advantageous Actor-Critic (A3C) algorithm. Multiple parallel threads each hold a copy of the model and explore independently within the simulator. At each time step, the high-level policy selects a policy type based on the current simulation state, and the low-level policy outputs specific parameters based on that type and state. The simulator executes these parameters and advances to the next state, simultaneously calculating an immediate reward based on the reward function.

[0081] Furthermore, the key design lies in the calculation of the reward signal for the high-level policy model. This reward is not an immediate reward at each step, but rather an evaluation based on the cumulative cost generated in the simulator over a fixed step length (H steps) in the future, according to the action sequence output by the low-level policy. Specifically, after the high-level policy makes a decision at time t, the low-level policy continuously executes the parameters it generates over the period from t to t+H. At time t+H, the total cost (including material cost, over-limit penalty, etc.) and safety benefits (such as whether deformation is controlled) within this period are calculated, and this long-term cumulative result is discounted and used as the reward obtained by the high-level policy at time t. The reward of the low-level policy is updated using the standard A3C algorithm based on the immediate reward and advantage function at each step.

[0082] In this way, the higher-level policy learns how to select a macro-policy that guides the lower-level policy to generate a long-term superior action sequence, while the lower-level policy learns how to optimize the specific parameters at each step given the macro-policy. The two models share global parameters through asynchronous gradient updates, achieving collaborative optimization.

[0083] This embodiment decomposes the complex tunnel support sequence decision-making problem into two solvable levels: macro-strategy selection and micro-parameter optimization, through a hierarchical reinforcement learning framework and the design of long-term reward signals. It guides the model to learn decision logic with a long-term perspective. Traditional methods or single-layer reinforcement learning are prone to getting trapped in local optima, pursuing single-step rewards while ignoring long-term consequences. In this method, the high-level strategy is responsible for setting the direction, and its reward comes directly from the long-term cumulative effect of the lower-level action sequences. This forces the high-level strategy to be forward-looking, evaluating the long-term benefits of different macro-strategies. For example, choosing an active reinforcement strategy may have higher initial costs, but if it can avoid large deformations and emergency repairs later, it is superior in the long run. Under the guidance of the high-level strategy, the lower-level strategies focus on fine-grained operations, learning how to achieve optimal parameter settings for each step within that direction. The use of the A3C asynchronous framework for training improves sampling efficiency and exploration capabilities, helping to find better global strategies. This hierarchical decision-making architecture not only reduces the learning difficulty and improves training stability, but more importantly, it enables the final learned decision model to inherently balance short-term costs and long-term security, and to output support solutions that meet immediate control requirements while also having good economic efficiency and sustainability. This achieves a qualitative leap from imitating static rules to learning dynamic optimization strategies.

[0084] In some preferred embodiments, the acquisition of the support effect feedback information includes: deploying a distributed optical fiber sensing system in the vicinity of the supported section to acquire continuous spatial distribution data of the strain of the support structure and the deformation of the surrounding rock; calculating the stress utilization rate of the support structure, the development depth of the plastic zone of the surrounding rock, and the deformation coordination coefficient based on the continuous spatial distribution data; comparing the stress utilization rate of the support structure, the development depth of the plastic zone of the surrounding rock, and the deformation coordination coefficient with the design expected values ​​to generate a multi-dimensional support effectiveness evaluation vector as the support effect feedback information.

[0085] In practice, to obtain feedback information on the support effect, a distributed fiber optic sensing system based on Brillouin optical time-domain reflectometry technology is systematically deployed in adjacent supported sections. Specifically, sensing fibers are tightly arranged in a zigzag or spiral pattern on the surface of the shotcrete layer, or pre-embedded within the shotcrete layer, or tied to the main reinforcement of the steel arch frame, forming a continuous sensing network covering the surface of the support structure. This system achieves centimeter-level spatial resolution and micro-strain level measurement accuracy.

[0086] Furthermore, the system continuously acquires strain data at every point along the fiber optic path at a frequency of minutes, forming continuous spatial distribution data of strain along the longitudinal and circumferential directions of the tunnel. By differentially processing the strain data of adjacent measuring points, curvature distribution data can be further calculated, indirectly reflecting the deformation.

[0087] Furthermore, based on this high-density spatial distribution data, three core indicators were quantitatively calculated. The stress utilization rate of the support structure was calculated: a representative monitoring section was selected, and the average strain of the concrete surface measured by optical fiber was multiplied by the measured elastic modulus of the shotcrete to convert it into an average stress σ_actual. This stress was then compared with the design compressive strength f_c of the shotcrete or a certain allowable stress threshold σ_allow to calculate the stress utilization rate η = σ_actual / σ_allow. This value reflects the degree to which the bearing potential of the support material is utilized; an excessively high η indicates that the strength limit may have been reached, while an excessively low η indicates that the support potential is not being fully utilized.

[0088] Furthermore, the depth of the plastic zone development in the surrounding rock is calculated by analyzing the attenuation gradient of strain measured by optical fiber along the tunnel radial direction (by comparing the strain difference between the surface and the optical fiber at slightly deeper burial depths). When the strain gradient undergoes a sudden change or the strain value exceeds the strain threshold corresponding to the elastic limit of the rock mass, that point can be considered to have entered the plastic state. By identifying the elastoplastic interfaces on multiple radial paths, the approximate depth and shape of the plastic zone can be estimated.

[0089] Further, the deformation compatibility coefficient is calculated: At the same monitoring section, the circumferential strain distribution curves of the support structure (concrete layer) and the circumferential strain distribution curve of the surrounding rock adjacent to the support, estimated through back analysis, are extracted respectively. The correlation coefficient ρ between the two curves is calculated, or their root mean square error (RMSE) is calculated. The deformation compatibility coefficient ξ can be defined as ξ = ρ or ξ = 1 / (1 + RMSE). This coefficient reflects whether the support and the surrounding rock deform in unison and share the load; a higher ξ indicates better compatibility.

[0090] Furthermore, the calculated scalars η (stress utilization rate), plastic zone depth D_p, and compatibility coefficient ξ are compared with the expected target values ​​or theoretical optimal values ​​from the design stage to generate a multi-dimensional support effectiveness evaluation vector [Δη, ΔD_p, Δξ]. This vector serves as quantitative and objective feedback information on support effectiveness, inputting it into the decision-making model to evaluate the actual effectiveness of the previous stage's decisions and guide subsequent optimization.

[0091] This embodiment establishes a high-precision, full-field, real-time quantitative evaluation system for support effectiveness by applying distributed optical fiber sensing and multi-dimensional performance index calculation. This provides a key technical approach to solving the problem of lacking real-world feedback in decision-making models. Traditional monitoring points are sparse, making it difficult to comprehensively grasp the working state of the support structure and its interaction with the surrounding rock. This method utilizes distributed optical fibers to achieve holographic monitoring of the strain field of the support structure, enabling the detection of local weak points or areas of abnormal stress concentration—something point sensors cannot do. The stress utilization rate index, calculated based on full-field data, can accurately assess the safety reserve of the support, avoiding the risk of overall failure due to local overload. The estimation of the plastic zone development depth directly quantifies the support's control effect on the relaxation range of the surrounding rock, and is one of the core indicators for evaluating the effectiveness of the support scheme. The introduction of the deformation coordination coefficient evaluates, from the perspective of mechanical interaction, whether the support and the surrounding rock have formed a good collaborative bearing system, which is crucial for optimizing support stiffness and leveraging the self-supporting capacity of the surrounding rock. The performance evaluation vector composed of these three indicators provides the intelligent decision-making model with an unprecedented, high-information-density, quantifiable closed-loop feedback signal.

[0092] In some preferred embodiments, the reward function of the hierarchical reinforcement learning framework is constructed and calculated through the following steps: setting a safety assessment term, which is calculated based on the monitored surrounding rock deformation data; when the surrounding rock deformation data does not exceed a preset safety threshold, the safety assessment term is zero; when the surrounding rock deformation data exceeds the safety threshold, the safety assessment term is a negative value, and the absolute value of the safety assessment term is positively correlated with the degree to which the surrounding rock deformation data exceeds the safety threshold; setting an economic assessment term, the calculation of which includes: calculating the direct material cost based on the support parameters output by the action space, and based on the support effect feedback information... The support effectiveness evaluation vector is used to estimate future maintenance costs through a long-term maintenance cost prediction model. The economic evaluation term is the negative value of the sum of the direct material cost and the future maintenance cost. A construction continuity evaluation term is set to penalize drastic fluctuations in support parameters. The calculation method is as follows: calculate the Euclidean distance between the support parameter vector output in the current decision cycle and the support parameter vector output in the previous decision cycle, and multiply the Euclidean distance by a preset negative penalty coefficient. The safety evaluation term, the economic evaluation term, and the construction continuity evaluation term are weighted and summed to obtain the reward value of the reward function at each decision step.

[0093] In practice, the reward function R of the hierarchical reinforcement learning framework is constructed by calculating each term separately and then performing linear weighted fusion. The design of the safety evaluation term R_safe aims to treat deformation control as a hard constraint. Let the currently monitored critical deformation value be u (such as arch settlement), and its preset safety threshold be u_critical. The calculation logic of R_safe is as follows: if u≤u_critical, then R_safe=0, indicating that the deformation is within the safe range, and this term does not penalize it. If u>u_critical, then R_safe=-α*exp(β*(u-u_critical)), where α and β are positive coefficients. This is an exponentially increasing penalty term, meaning that when the deformation exceeds the limit, the penalty value increases sharply with the increase of the excess, thus strongly compelling the agent to control the deformation within the threshold.

[0094] Furthermore, the design of the economic evaluation item R_econ comprehensively considers both immediate and long-term costs. The direct material cost C_mat is calculated: based on the support parameter vector output by the underlying strategy model (such as the number of anchor bolts and the thickness of the sprayed layer), combined with a material unit price database, the direct material cost of the current support scheme is calculated. The future maintenance cost C_maint is estimated: the support effectiveness evaluation vector [Δη, ΔD_p, Δξ] is input into a pre-trained maintenance cost prediction model. This maintenance cost prediction model is a gradient boosting decision tree regressor, trained using historical engineering data, including but not limited to support effectiveness indicators, corresponding support parameters, basic rock mass quality indicators, and actual maintenance records and costs within a certain period. The model fits the mapping relationship between support effectiveness and future maintenance costs through supervised learning. It learns the complex nonlinear relationship between support effectiveness indicators and the probability of possible maintenance and estimated costs within a future period (such as the next kilometer of tunnel). C_maint is the output value of the maintenance cost prediction model. Finally, R_econ = -(γ1*C_mat + γ2*C_maint), where γ1 and γ2 are cost weighting coefficients, and the negative sign indicates that cost is a negative factor that needs to be minimized.

[0095] Furthermore, the construction continuity assessment term R_cont aims to smooth the decision output and avoid drastic fluctuations in the scheme. Let the support parameter vector output in the current decision step be a_t, and the output in the previous decision step be a_t-1. Calculate the Euclidean distance between them: dist = ||a_t - a_t-1||_2. Then, R_cont = -λ*dist, where λ is a positive penalty coefficient. This term imposes a linear negative penalty on any change in the parameter vector; the greater the change, the heavier the penalty, thus encouraging the model to output relatively stable and continuous support schemes between adjacent decision steps, which is beneficial to on-site construction organization.

[0096] Ultimately, the total reward for each step is: R = ω_s * R_safe + ω_e * R_econ + ω_c * R_cont, where ω_s, ω_e, and ω_c are pre-set weighting coefficients used to balance the importance of safety, economy, and continuity. Typically, ω_s is set relatively large to ensure the primary importance of safety objectives.

[0097] This embodiment provides a clear, comprehensive, and quantifiable optimization objective guide for the reinforcement learning agent by designing a structured multi-objective reward function. The safety evaluation term employs an exponential penalty, creating a safety cliff where the reward drops sharply once deformation approaches or exceeds a critical value. This non-linear design, compared to linear penalties, forces the agent to strictly control deformation within a safe range, effectively ensuring the bottom-line safety of the project. The innovation of the economic evaluation term lies in the introduction of a long-term maintenance cost prediction model based on support effectiveness feedback. This expands the agent's optimization perspective from the single current construction cost to the entire lifecycle cost, including future operation and maintenance expenses. The agent will therefore learn to select solutions that achieve the best balance between initial construction cost and long-term durability; for example, moderately increasing initial support strength in exchange for a lower probability and cost of later repairs. The introduction of a construction continuity assessment term addresses the potential for leaps in solutions that can result from pure optimization. It penalizes drastic changes in decision output, prompting the agent to learn and output smooth and stable support parameter sequences. This not only meets the continuity requirements of on-site construction processes, reducing management costs and error risks associated with frequent changes, but also makes the model's decision output more consistent with the inertia of engineering practice. By adjusting the weights of the three sub-items, the model can flexibly adapt to the specific needs of different projects regarding risk tolerance, investment strategies, and construction pace.

[0098] In some preferred embodiments, the method employs a hybrid triggering mechanism combining event-driven and period-driven approaches, including: setting a fixed minimum decision cycle as a base cycle; within the base cycle, real-time monitoring of the microseismic feature vector and the monitored feature vector; if a cluster of microseismic events with energy exceeding a preset threshold or an accelerating trend in deformation rate is identified, an emergency decision-making process is immediately triggered; if no emergency event occurs, a regular decision-making process is triggered at the end of the base cycle; wherein the emergency decision-making process and the regular decision-making process invoke the support decision-making model with different risk preferences.

[0099] In practice, the implementation of the hybrid triggering mechanism relies on a background real-time monitoring and scheduling engine. The base cycle T_base is set to a time length that matches the construction cycle, such as the time required for a complete excavation-muck removal-initial support cycle, or a fixed advance length (e.g., 2 meters).

[0100] Furthermore, within this basic cycle, the monitoring engine executes two real-time analysis tasks in parallel. The first is microseismic event cluster analysis: the system continuously receives microseismic event streams, maintaining a sliding time window (e.g., the most recent 30 minutes) and a spatial window (e.g., 50 meters before and after the excavation face). Within the window, spatiotemporal clustering analysis is performed on the microseismic events (e.g., based on the DBSCAN algorithm). If a cluster containing at least N events (e.g., N=5) is identified, and the average energy of the events within the cluster exceeds a preset threshold E_th, it is determined to be an abnormal microseismic event cluster. The second is deformation acceleration trend analysis: the system calculates the time series of deformation rates of key monitoring points (e.g., arch crown, arch waist) in real time over a recent period (e.g., the most recent 1 hour), and uses the Mann-Kendall trend test or linear regression analysis to analyze its slope. If the test results show a statistically significant acceleration trend (p-value less than 0.05), and the acceleration slope is greater than the preset threshold S_th, it is determined to be a deformation acceleration trend.

[0101] Furthermore, the emergency decision-making process is preemptively triggered. Once the monitoring engine identifies any of the aforementioned emergency events, it immediately sends a high-priority interrupt signal to the main decision-making process. The main process suspends any current waiting or computational tasks, immediately captures a snapshot of all data at the latest instant, and invokes a dedicated "emergency decision-making pipeline." This pipeline may employ a simplified but faster feature extraction and fusion model to output a set of support parameter recommendations with minimal latency (e.g., seconds) and push them to the field via audible and visual alarms.

[0102] Furthermore, the regular decision-making process is automatically triggered by a timer at the end of each base cycle T_base. It invokes a standard, more sophisticated but slightly slower decision-making model to perform periodic state assessments and scheme optimizations.

[0103] To enable model invocation based on different risk appetites, two copies of the support decision-making model are trained and coexist during system deployment. They share the same network structure but develop different strategies in different training environments. During training, the emergency model has its safety weight ω_s in its reward function set to be extremely high, while its economic weight ω_e is relatively low, resulting in a conservative "safety-first" strategy. The regular model achieves a more balanced weight setting between safety and economics. The scheduling engine routes requests to the corresponding model based on the trigger source (emergency or scheduled).

[0104] This embodiment, through a hybrid triggering mechanism driven by both events and cycles, and a differentiated model strategy, enables the intelligent decision-making system to possess scenario-adaptive response capabilities, achieving a balance between high efficiency in normal situations and safety in emergency situations. The basic cycle ensures regular and systematic review of support design, aligning with the planned nature of engineering management. Meanwhile, the real-time micro-seismic and deformation trend monitoring embedded within the cycle acts like a tireless "sentinel," capturing early signs of dynamic instability in the surrounding rock 24 / 7. Once the sentinel alarms, the system immediately switches from cruise mode to emergency mode, activating a rapid decision-making channel. This significantly shortens the lag time from anomaly identification to response decision, buying valuable time for timely engineering measures to curb the development of the danger. Calling the emergency model with a higher risk aversion ensures that the recommended solution in emergency situations naturally tends to strengthen protection and reserve a greater safety margin—an intelligent risk avoidance behavior consistent with safety engineering principles. In non-emergency, routine situations, calling the balanced model allows for full economic optimization while ensuring safety. This mechanism makes the system both intelligent and robust, capable of handling routine incremental optimizations and effectively responding to sudden, nonlinear risk events, significantly improving the proactive safety management capabilities and overall decision-making efficiency of tunnel construction in the face of uncertain geological conditions.

[0105] In some preferred embodiments, the method further includes a decision simulation and verification step based on digital twin technology, including: updating the mechanical parameters of the digital twin of the tunnel surrounding rock according to the current geological sketch image and monitoring data; performing a rapid finite element numerical simulation on the quantitative support parameter scheme output by the support decision model in the digital twin; matching the simulated predicted deformation and stress distribution with a preset historical case library, and generating a safety margin assessment and risk warning based on the matching results.

[0106] In practice, the decision simulation and verification step based on digital twin technology runs as an independent, optional advanced verification module. The digital twin is a high-fidelity three-dimensional finite element numerical model that is updated synchronously with the physical tunnel, and its geometry is consistent with the physical tunnel design BIM model.

[0107] Furthermore, the process of updating mechanical parameters is dynamic and localized. Based on the latest geological sketch images identifying lithological distribution and structural features, the system differentially updates rock mass mechanical parameters in the corresponding areas of the digital twin. For example, for areas identified as fault influence zones, parameters such as elastic modulus and cohesion are reduced by a percentage. Simultaneously, using the latest spatiotemporally registered multi-point displacement monitoring data, an inversion algorithm based on sensitivity analysis is employed to fine-tune the initial geostress field parameters in the model, making the initial mechanical state of the twin more closely resemble reality.

[0108] Furthermore, during rapid finite element numerical simulation, the system, within a digital twin environment, "virtually constructs" the support structure in a programmed manner according to the quantitative parameters output by the support decision model. For example, rod elements (anchors) are generated at specified locations and assigned calculated mechanical properties, while shell elements (sprayed layers) are generated on the tunnel wall and assigned thickness properties. Subsequently, the construction process of one or more subsequent excavation steps is simulated, performing rapid elastoplastic finite element calculations. To balance speed and accuracy, this simulation may employ a reduced-order model or a pre-calculated response surface method.

[0109] Furthermore, the simulated prediction results, such as displacement values ​​of key points around the tunnel, maximum stress values ​​of the support structure, and the range of the plastic zone, are matched with a historical case database for similarity analysis. The historical case database is a structured database where each case entry contains its key geological features, support parameters, and a vector of surrounding rock response indicators that were ultimately observed or simulated with high precision. The matching algorithm employs a k-nearest neighbor search based on Euclidean or Mahalanobis distance to find the k historical cases in the database that are most similar to the current predicted response vector.

[0110] Furthermore, based on the matching results, multi-dimensional outputs are generated. Safety margin assessment: By comparing the currently predicted displacement and stress values ​​with the critical values ​​of similar cases that resulted in failure, the safety factor or safety margin percentage of the current solution is calculated. Risk warning: The most similar cases that ultimately resulted in significant deformation or required engineering intervention are listed, and their geological and construction characteristics at the time are extracted, indicating the risks that the current solution may face regarding "what problems occurred under what conditions." These assessments and warnings, along with the model's original decision recommendations, are presented for engineers to make a final comprehensive judgment.

[0111] This embodiment integrates digital twin simulation and case analogy reasoning to provide a second opinion verification layer based on physical mechanisms and historical experience for data-driven intelligent decision-making, which significantly improves the interpretability and credibility of decision recommendations.

[0112] In some preferred embodiments, the method further includes interpretability analysis of the gated multimodal fusion network, including: recording and analyzing the gated value sequence of the gated multimodal fusion network in historical decisions; using an association rule mining algorithm to analyze strong association rules between different gated value patterns and specific geological conditions, construction conditions and final surrounding rock stability results; and converting the mined strong association rules into readable expert knowledge entries for verifying model logic or assisting in decision understanding under new conditions.

[0113] In practice, interpretability analysis of the gated multimodal fusion network is performed as an offline or background analysis service. During each online decision-making process, the system records the four feature vectors (V_img, V_mon, V_ms, V_dr) and their corresponding gate value vectors (Gate_img, Gate_mon, Gate_ms, Gate_dr) calculated using the Sigmoid function into a dedicated log database. Each gate value vector is a set of numbers between 0 and 1, representing the degree of attention given to each feature dimension during that decision-making process.

[0114] The analysis process is initiated periodically. First, the accumulated gating value sequence data is preprocessed. For each decision record, in addition to the gating value, a context label for that decision is also associated. These labels are obtained through automatic analysis of the raw data at that time or in combination with construction logs, and include: discretized geological labels (such as "Class III surrounding rock", "containing argillaceous interlayers", "dense joints"), construction condition labels (such as "bench excavation", "passing through fractured zone"), and finally, stability result labels based on subsequent monitoring (such as "stable", "slight deformation", "support adjustment required").

[0115] Furthermore, association rule mining algorithms (such as the FP-Growth algorithm) are used to mine strong association rules from these labeled transactional data. The algorithm treats each decision as a "transaction," where "items" include various labels and discretized high-gated feature identifiers (e.g., "image feature_directional variance_high weight", "monitoring feature_vault settlement rate_high weight"). The algorithm identifies frequently occurring itemsets and generates rules in the form of "preceding item => subsequent item," where the subsequent item is typically a stability outcome label.

[0116] For example, a rule might be discovered: {Geology = Dense Joints, Working Condition = Post-Blasting, Image Features_Structural Surface Density_High Weight, Microseismic Features_b-Value_Low Weight} => {Result = Need for Enhanced Support}. This rule can be interpreted as: In densely jointed rock strata, immediately after blasting, if the model pays particular attention to the structural surface density features reflected in the image, and also pays particular attention to the decreasing b-value feature in the microseismic data, which represents a high proportion of high-energy events, then the final result will often be a situation where enhanced support is needed.

[0117] Furthermore, the system transforms these machine-mined strong association rules into readable knowledge entries described in natural language through a templated translation engine, and stores them in a "model decision knowledge base." This knowledge base can be used for: 1. Model auditing, checking for absurd rules that violate common sense in engineering; 2. Decision assistance, where, when the model makes a decision under new conditions, it can retrieve and display the key rules that triggered the decision, thereby explaining to engineers "why the model thinks this way."

[0118] This embodiment successfully endows complex deep fusion models with interpretability and transparency by systematically analyzing the key intermediate variable of gating weights within the model and associating it with external engineering semantics.

[0119] In some preferred embodiments, the method further includes a step of periodically incrementally updating the model, comprising: setting a model update cycle; at the end of each update cycle, collecting all multi-source heterogeneous data obtained during actual construction, the actual support parameters used, and the subsequently monitored actual deformation and stress response data of the surrounding rock, to form an incremental training sample set; using the incremental training sample set to collaboratively fine-tune the original model, wherein the original model includes a geological structure surface semantic segmentation network for feature extraction, a one-dimensional convolutional neural network and a gated recurrent unit network, a gated multimodal fusion network for feature fusion, and the support decision model; the collaborative fine-tuning adopts an incremental learning algorithm, which updates the connection weight parameters with the goal of minimizing the prediction error of the new samples while retaining the knowledge of the original model; and deploying the finely tuned and updated connection weight parameters to the online system to replace the original model for subsequent real-time decision-making.

[0120] In practice, the incremental updates to the model are scheduled and executed by a separate model management service. The update cycle is set based on the project progress and data accumulation rate, for example, triggered after every 50 excavation cycles or when a sufficient number of new samples have been accumulated.

[0121] Furthermore, upon triggering an update cycle, the model management service initiates a data collection pipeline. This pipeline extracts all raw data generated within the current cycle from the engineering database: including multi-source heterogeneous data snapshots acquired at each decision-making stage, the final support parameters actually recorded on-site (which may differ from model-recommended values, reflecting human intervention or on-site constraints), and the actual deformation and stress response data of the surrounding rock collected in a new monitoring cycle after the implementation of these support parameters. This data is strictly aligned chronologically and spatially to ensure that the "decision input - actual action - environmental feedback" constitute a complete sample triplet suitable for supervised learning, thus forming the incremental training sample set.

[0122] Furthermore, the original model is fine-tuned using this sample set. The original model here is a set of models, including: a semantic segmentation network for geological structure surfaces responsible for geological image feature extraction; a one-dimensional CNN and GRU network responsible for monitoring / drilling time-series feature extraction; a gated multimodal fusion network responsible for feature fusion; and a support decision model responsible for the final decision (including its internal high-level and low-level policy networks). The fine-tuning process is not retraining, but rather based on the existing parameters of these models.

[0123] Furthermore, collaborative fine-tuning employs the Elastic Weight Consolidation (EWC) algorithm, an incremental learning method. For each neural network requiring fine-tuning, the EWC algorithm calculates the importance of its existing parameters to its performance on the old dataset (the diagonal approximation of the Fisher information matrix). In the objective function of fine-tuning, in addition to minimizing the loss function L_new(θ) on the new sample set (such as the mean squared error of support parameter prediction and the error of stability exponential regression), a constraint term is added that penalizes drastic changes to important parameters. Mathematically, the objective function is L_new(θ) + Σ_i(λ / 2)*F_i*(θ_i - θ_i_old)^2, where θ is the current model parameter vector to be trained (updated); θ_i is the i-th parameter (scalar) in θ, i.e., the i-th connection weight or bias term in the current model; θ_i_old is the old model parameter corresponding to θ_i; F_i is an approximation of the Fisher information of parameter i, reflecting the importance of the parameter to the old task; λ is a hyperparameter that controls the balance between new and old knowledge; and the squared term (θ_i - θ_i_old)^2 penalizes the deviation between the current and old parameters. By optimizing this joint objective, the model is constrained to retain the memory of old knowledge while learning new knowledge.

[0124] Furthermore, the connection weights of all the aforementioned networks are updated by optimizing this joint objective function through backpropagation and gradient descent. This process enables the model to adapt to new data and learn project-specific patterns while retaining general knowledge learned from a large amount of historical data, effectively mitigating the problem of catastrophic forgetting.

[0125] Furthermore, after the update is complete, a new set of model parameter files is generated. After a simplified validation test, the model management service seamlessly hot-deploys this new set of parameters to the online inference system, replacing the old parameters currently in use. Thereafter, all new real-time decision requests will be handled by this evolved model, which has incorporated the latest engineering best practices.

[0126] In this embodiment of the invention, tunnel engineering exhibits strong site specificity, making it impossible for a general pre-trained model to perfectly adapt to all projects. This method continuously collects closed-loop data from the "plan-execution-result" process during project implementation and uses this data for periodic incremental fine-tuning of the model, essentially allowing the model to learn through practical experience in a real engineering environment. Employing advanced incremental learning algorithms such as elastic weight consolidation is crucial. This ensures that while absorbing new knowledge and adapting to the specific geological behavior of the project (such as encountering special expansive rocks) or construction characteristics, the model does not forget general, fundamental engineering principles (such as basic support principles for weak surrounding rock), thus maintaining the stability of its core capabilities.

[0127] Please see Figure 2 , Figure 2 This is a schematic block diagram of a computer device provided in an embodiment of this application. The computer device 500 can be a terminal or a server, wherein the server can be a standalone server or a server cluster composed of multiple servers.

[0128] The computer device 500 includes a processor 502, a memory, and a network interface 505 connected via a system bus 501. The memory may include a non-volatile storage medium 503 and internal memory 504.

[0129] The non-volatile storage medium 503 can store an operating system 5031 and a computer program 5032. When the computer program 5032 is executed, it enables the processor 502 to execute a method for intelligent rock grading and support decision-making based on multi-source heterogeneous data fusion.

[0130] The processor 502 provides computing and control capabilities to support the operation of the entire computer device 500.

[0131] The internal memory 504 provides an environment for the operation of the computer program 5032 in the non-volatile storage medium 503. When the computer program 5032 is executed by the processor 502, the processor 502 can execute a method for intelligent classification and support decision-making of surrounding rock based on multi-source heterogeneous data fusion.

[0132] The network interface 505 is used for network communication with other devices. Those skilled in the art will understand that the above structure is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device 500 to which the present application is applied. A specific computer device 500 may include more or fewer components than shown in the figures, or combine certain components, or have different component arrangements.

[0133] The processor 502 is used to run a computer program 5032 stored in a memory to implement the steps of a method for intelligent classification and support decision-making of surrounding rock based on multi-source heterogeneous data fusion provided in any of the above method embodiments.

[0134] It should be understood that in the embodiments of this application, the processor 502 may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0135] It will be understood by those skilled in the art that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program may be stored in a storage medium, which is a computer-readable storage medium. The computer program is executed by at least one processor in the computer system to implement the process steps of the embodiments of the above methods.

[0136] Therefore, the present invention also provides a storage medium. This storage medium can be a computer-readable storage medium. The storage medium stores a computer program. When executed by a processor, the computer program causes the processor to perform the steps of the intelligent rock grading and support decision-making method based on multi-source heterogeneous data fusion provided in any of the above method embodiments.

[0137] The storage medium is a physical, non-transient storage medium, such as a USB flash drive, external hard drive, read-only memory (ROM), magnetic disk, or optical disk, or any other physical storage medium capable of storing program code. The computer-readable storage medium can be non-volatile or volatile.

[0138] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0139] In the several embodiments provided by this invention, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For example, the division of each unit is merely a logical functional division, and there may be other division methods in actual implementation. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0140] The steps in the method of this invention can be adjusted, merged, or reduced in order according to actual needs. The units in the device of this invention can be merged, divided, or reduced according to actual needs. Furthermore, the functional units in the various embodiments of this invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0141] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a terminal, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.

[0142] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0143] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Since these modifications and variations fall within the scope of the claims and their equivalents, this invention also intends to include these modifications and variations.

[0144] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for intelligent classification and support decision-making of surrounding rock based on multi-source heterogeneous data fusion, characterized in that, include: Acquire multi-source heterogeneous data during the tunnel construction process, including geological sketch images, time-series data from monitoring sensors, microseismic event data, and time-series data from drilling equipment. Spatiotemporal registration is performed on the multi-source heterogeneous data to unify the geological sketch images, the time-series data of the monitoring sensors, the microseismic event data, and the time-series data of the drilling equipment under the same spatiotemporal reference. Feature extraction is performed on the spatiotemporally registered geological sketch image, the time-series data of the monitoring sensor, the microseismic event data, and the time-series data of the drilling equipment to obtain the corresponding image feature vector, monitoring feature vector, microseismic feature vector, and drilling feature vector. The image feature vector, the monitoring feature vector, the microseismic feature vector, and the drilling feature vector are fused together, and the dynamic stability index of the surrounding rock is calculated based on the fused feature vector. The dynamic stability index of the surrounding rock, the current excavation process information, and the support effect feedback information of adjacent supported sections are used as the decision state and input into the pre-trained support decision model. The support decision model then outputs quantitative support parameters for the current construction section.

2. The method according to claim 1, characterized in that, The spatiotemporal registration of multi-source heterogeneous data includes: Establish a three-dimensional spatial coordinate system associated with the tunnel design axis, and construct a timeline with the construction progress time as the main axis; For a geological sketch image, based on the corresponding shooting location and shooting time, the geological sketch image is mapped to the corresponding position in the three-dimensional spatial coordinate system and associated with the corresponding time on the timeline; For the time-series data of the monitoring sensors and the drilling equipment, based on the spatial location of the collected sensors and the data timestamp, each data point is mapped to the corresponding position in the three-dimensional spatial coordinate system and associated with the corresponding time on the timeline; For microseismic event data, based on the spatial location results of the seismic source and the triggering time, each microseismic event is mapped to the corresponding position in the three-dimensional spatial coordinate system and associated with the corresponding time on the timeline; For the drilling equipment time-series data, based on the real-time spatial position of the drill bit and the data timestamp recorded during the acquisition and obtained through the integrated navigation system, each data point is mapped to the corresponding position in the three-dimensional spatial coordinate system and associated with the corresponding time on the timeline.

3. The method according to claim 1, characterized in that, The extraction of the image feature vector includes: The geological sketch image is input into a pre-trained semantic segmentation network for geological structural surfaces, which is trained based on a pre-defined geological annotation image library for tunnel engineering. The geological structural surface semantic segmentation network outputs a pixel-level binary map of structural surface distribution and a structural surface type label map. Geometric feature calculations are performed on the binary map of the structural surface distribution to extract topological network feature vectors of structural surface trace length, spacing, and intersection relationships; The topological network feature vector is concatenated with the proportion features of various structural surfaces statistically obtained from the structural surface type label map to form the image feature vector.

4. The method according to claim 1, characterized in that, The extraction of the monitoring feature vector includes: The time-series data of the monitoring sensors are grouped according to sensor type and spatial location to form multiple parallel time series. For each time series, a one-dimensional convolutional neural network is used to extract local mutation pattern features, and a gated recurrent unit network is used to extract long-term evolution trend features. Early fusion of the local mutation mode features and the long-term evolution trend features from different types of sensors in the same spatial region generates a local stability symptom vector for the same spatial region. The local stability symptom vectors of all monitored areas are aggregated and encoded into the global monitoring feature vector through a fully connected layer.

5. The method according to claim 1, characterized in that, The process of fusing the image feature vector, the monitoring feature vector, the microseismic feature vector, and the drilling feature vector, and calculating the surrounding rock dynamic stability index based on the fused feature vector, includes: The image feature vector, monitoring feature vector, microseismic feature vector, and drilling feature vector are input into a preset gated multimodal fusion network; The gated multimodal fusion network sets a learnable gate weight vector for each type of feature vector, and generates a gate value between 0 and 1 through the sigmoid function; The gate value is used to perform element-wise weighted modulation on each feature vector; The modulated feature vectors are concatenated and passed through a regression subnetwork to output a continuous scalar value as the dynamic stability index of the surrounding rock.

6. The method according to claim 1, characterized in that, The support decision-making model is trained based on a hierarchical reinforcement learning framework, including: The high-level strategy model takes the decision state as input and outputs an abstract support strategy type. The underlying strategy model receives the support strategy type and decision state output by the higher-level strategy model, and outputs the specific quantitative support parameters. The high-level strategy model and the low-level strategy model are trained collaboratively using an asynchronous advantage executor-commentator algorithm. The reward signal of the high-level strategy model is generated based on the long-term cost and safety benefit assessment results generated after the quantitative support parameters output by the low-level strategy model are executed.

7. The method according to claim 1, characterized in that, The acquisition of the support effect feedback information includes: Deploy a distributed fiber optic sensing system in the vicinity of the supported section to obtain continuous spatial distribution data of the strain of the support structure and the deformation of the surrounding rock. Based on the continuous spatial distribution data, the stress utilization rate of the support structure, the development depth of the plastic zone of the surrounding rock, and the deformation compatibility coefficient are calculated. The stress utilization rate of the support structure, the development depth of the plastic zone of the surrounding rock, and the deformation coordination coefficient are compared with the design expected values ​​to generate a multi-dimensional support effectiveness evaluation vector, which serves as feedback information on the support effect.

8. The method according to claim 6, characterized in that, The reward function of the hierarchical reinforcement learning framework is constructed and calculated through the following steps: A safety assessment item is defined, which is calculated based on the surrounding rock deformation data obtained from monitoring. When the surrounding rock deformation data does not exceed the preset safety threshold, the safety assessment item is zero; When the surrounding rock deformation data exceeds the safety threshold, the safety assessment item is a negative value, and the absolute value of the safety assessment item is positively correlated with the degree to which the surrounding rock deformation data exceeds the safety threshold. An economic evaluation item is defined, and the calculation of the economic evaluation item includes: calculating the direct material cost based on the support parameters output by the action space, and estimating the future maintenance cost through a long-term maintenance cost prediction model based on the support effectiveness evaluation vector in the support effect feedback information; the economic evaluation item is the negative value of the sum of the direct material cost and the future maintenance cost. A construction continuity assessment item is set up. The construction continuity assessment item is used to penalize drastic fluctuations in support parameters. The calculation method is as follows: calculate the Euclidean distance between the support parameter vector output in the current decision cycle and the support parameter vector output in the previous decision cycle, and multiply the Euclidean distance by a preset negative penalty coefficient. The safety assessment item, the economic assessment item, and the construction continuity assessment item are weighted and summed to obtain the reward value of the reward function at each decision step.

9. A computer device, characterized in that, The computer device includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method as described in any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, The storage medium stores a computer program that, when executed by a processor, can implement the method as described in any one of claims 1-8.

Images