Cloud edge model cooperation-based advanced geological prediction method and system

Abstract

The application belongs to the technical field of advanced geological prediction, and specifically discloses an advanced geological prediction method and system based on cloud edge model cooperation, which comprises the following steps: obtaining historical drilling parameters, seismic wave reflection spectra, tunnel face images and spectral data of a section to be excavated; obtaining a geological probability prediction spectrum by using a cloud end geological prediction model; regarding physical sensors on tunneling equipment as nodes to form a graph structure; obtaining real-time feature data of each node in the graph structure, and predicting a local geological state vector corresponding to each node by using an edge end geological prediction model; calculating a consistency index of each position based on the prediction results of the cloud end geological prediction model and the edge end geological prediction model, and triggering a three-level response mechanism according to the consistency index. The cloud end geological prediction model and the edge end geological prediction model realize cooperative evolution, which greatly improves the accuracy and advance nature of geological prediction.

Description

Technical Field

[0001] This invention relates to the field of advanced geological prediction technology, and in particular to an advanced geological prediction method and system based on cloud-edge model collaboration. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] During tunnel construction, tunnels often traverse complex geological structures such as water-rich fractured zones, karst formations, and areas with high ground stress, which can trigger major geological disasters such as sudden water inrushes and rock bursts. Therefore, accurate and rapid forecasting of geological conditions ahead is essential to ensuring construction safety.

[0004] Currently, advanced geological prediction technologies such as directional drilling, ground-penetrating radar, TSP advanced detection, real-time image analysis, and transient electromagnetic methods are widely used in tunnel construction. These technologies can predict the basic quality classification indicators of the rock mass in front of the tunnel face, as well as the presence of potential adverse geological conditions such as karst caves, fissures, internal structures, and water-rich areas.

[0005] However, the methods mentioned above can only be applied to single adverse geological conditions, such as identifying water-rich areas, identifying the lithology of the tunnel face, or identifying fractures ahead, and cannot achieve a comprehensive prediction of the geological conditions ahead.

[0006] While existing technologies have disclosed advanced geological exploration schemes based on multi-source data fusion, some schemes also disclose a hybrid architecture that uses edge computing nodes to process real-time drilling parameters and then runs machine learning algorithms on cloud servers; however, existing methods mostly use parallel feature stitching for data fusion, which cannot establish cross-modal dependencies (such as the coupling between seismic wave reflection and drilling rate anomalies); and, they only divide the edge and cloud ends at the functional level, with the edge model and cloud model being independent of each other, which cannot achieve collaborative optimization of geological prediction results. Summary of the Invention

[0007] To address the aforementioned issues, this invention proposes an advanced geological prediction method and system based on cloud-edge model collaboration. It utilizes drilling parameters, seismic wave reflection data, tunnel face photographs, and spectral data obtained from the engineering site, and employs a large cloud-based model to extract deep geological correlation patterns, resulting in a geological probability map. By collecting real-time tunneling parameter data, a lightweight edge-end model is used to correct cloud-based prediction biases in real time. Through the collaborative evolution of cloud-edge models, accurate advanced geological prediction is ultimately achieved.

[0008] In some implementations, the following technical solutions are adopted: Advanced geological prediction methods based on cloud-edge model collaboration include: Obtain historical drilling parameters, seismic wave reflection spectra, tunnel face images, and spectral data of the area to be excavated; A geological probability prediction map is obtained using a cloud-based geological prediction model. The geological probability prediction map includes the prediction results and corresponding confidence levels of rock mass integrity, water inrush risk, and fracture or weak interlayer at each location. The physical sensors on the tunneling equipment are regarded as nodes. Based on the principles of physical spatial proximity and functional coupling, the connection relationship between the nodes is constructed to form a graph structure. Real-time feature data of each node in the graph structure is obtained. Using the edge geological prediction model, the temporal latent feature representation of the node is predicted. The temporal latent feature representation of each node in the graph structure is inferred and calculated to obtain the local geological state vector corresponding to each node. The geological state vector includes the probability of rock mass integrity, the probability of water inrush risk, the probability of the occurrence of fractures or weak interlayers and their confidence information. The geological state vectors output by each node are spatially interpolated and topologically fused to form an edge geological probability distribution map, which is used to describe the real-time geological state of the tunnel face and the surrounding rock in front. Based on the prediction results of the cloud-based geological prediction model and the edge-based geological prediction model, the consistency index of each location is calculated. If the consistency index is within the set warning / protection range, the real-time node characteristics, seismic wave reflection spectrum of the excavated section, face image and spectral data are uploaded to the cloud-based geological prediction model for re-prediction, resulting in an updated geological probability prediction map.

[0009] As a further option, the cloud-based geological prediction model is a neural network model based on the Transformer architecture; The historical drilling parameters, seismic wave reflection spectra, tunnel face images, and spectral data of the area to be excavated are aligned according to the tunneling mileage and timestamp, and used as input data for the cloud-based geological prediction model. The input data is linearly mapped to a high-dimensional embedding space, and then the information of each modality is fused through a cross-modal attention mechanism to finally output a four-dimensional geological probability map.

[0010] As a further approach, three sets of vectors are generated for each modality of data: a query vector, a key vector, and a value vector; for the h-th attention head: ; in, This represents the query corresponding to the h-th attention head. AND key The rating; This represents the dimension (i.e., feature length) of each key vector. This represents the position offset term indicating the relative positional relationship between query position i and key position j in the tunnel spatial coordinate system or tunneling mileage coordinate system. This represents the graph bias derived from the sensor topological adjacency relationships. Indicates the modal confidence level of the key origin. This represents the time difference between the query and the key. Indicates the time decay coefficient; right Normalization is performed to obtain attention weights; based on the attention weights, the value vectors of each modality data are weighted and summed to obtain the data fusion representation.

[0011] As a further solution, the loss function of the cloud-based geological prediction model is specifically as follows: ; in, This represents the geological state regression error; Loss due to mask reconstruction For the latent variable distribution alignment constraint term, , , These are hyperparameters, used to balance the importance of the three losses in the total loss; The geological state regression error Specifically: ; in, Let i be the vector representing the true geological state at the i-th spatial location. This is the probability vector of geological state predicted by the model, where N is the spatial location or number of samples. The mask reconstruction loss Specifically: ; in, The original input features, The features to be reconstructed by the model, M is the set of mask locations; The latent variable distribution alignment constraint term Specifically: ; in, and Let be the mean and variance of the d-th dimension of the latent variable, respectively. This represents the total number of dimensions of the latent variable z.

[0012] As a further embodiment, the physical sensors on the tunneling equipment include: a cutterhead torque sensor, a cutterhead speed encoder, a cutter force sensor, a cutter vibration accelerometer, a high-definition camera at the tunnel face, a short-wave infrared spectral sensor, an earthquake source, a seismic detector, a ground-penetrating radar antenna, an ultrasonic sensor, a propulsion cylinder pressure sensor, a shield attitude sensor, and temperature, humidity, and gas sensors.

[0013] As a further solution, the edge geological prediction model is a spatiotemporal graph neural network model. The spatiotemporal graph neural network model uses data from multiple physical sensors on the tunneling equipment as nodes, and determines the connection relationship between nodes according to the principle of physical spatial proximity and the principle of functional coupling to construct a graph structure. The graph structure and the real-time feature data of each sensor node are used as input to the spatiotemporal graph neural network model. In each layer of graph convolution, each node aggregates information from its directly connected neighbor nodes through separable convolution operations. The aggregated information is then transformed by a linear transformation and a nonlinear activation function to generate a new feature representation of the node in the current layer that incorporates the local network state. After layer-by-layer operations, the node feature matrix after multi-layer processing is finally output. The feature vector of each node in the node feature matrix is ​​discriminated and mapped, and the mapping result is probabilistically processed to obtain the local geological state vector corresponding to each node.

[0014] As a further approach, based on the prediction results of the cloud-based geological prediction model and the edge-based geological prediction model, a consistency index is calculated for each location, specifically: For cloud-predicted probability vectors With edge prediction probability vector The consistency index is defined as: ; in: Represents Euclidean distance. As a normalization factor, make Cloud-based prediction probability vector Or edge prediction probability vector This includes the probability of rock mass integrity, the probability of water inrush risk, and the probability of fractures or weak interlayers.

[0015] As a further solution, when the consistency index is greater than the first predetermined threshold, the original prediction of the geological prediction model at the edge remains unchanged; when the consistency index is less than the second predetermined threshold, an alarm is issued and the equipment is linked to reduce the tunneling rate; when the consistency index is between the second predetermined threshold and the first predetermined threshold, an alarm is issued, and the real-time node characteristics, seismic wave reflection spectrum of the excavated section, face image and spectral data are uploaded to the cloud geological prediction model for re-prediction to obtain an updated geological probability prediction map.

[0016] As a further solution, the collaborative evolution process of cloud-based geological prediction models and edge-based geological prediction models is also included, specifically: The cloud-based geological prediction model utilizes a geological rule network extracted from its own cross-scenario data. This geological rule network uses geological semantic units as nodes and the coexistence states of different geological states as edges. The edge weights are determined based on the correlation strength between different geological elements in the cross-scenario geological data. The geological rule network is distributed to the edge geological prediction model, enabling the edge geological prediction model to have a wider range of geological knowledge that cannot be covered by local data. The edge-end geological prediction model provides highly refined value data to the cloud, including abnormal working condition data that triggers alarms and the corresponding geological prediction results, which are used to enrich the training sample library of the cloud-based geological prediction model.

[0017] As a further solution, the edge-end geological prediction model achieves cognitive synchronization with the cloud model through feature distribution alignment loss; the feature distribution alignment loss specifically refers to: ; in, This is the loss function for the edge-end geological prediction model, used to constrain the alignment of the feature distribution of the edge-end geological prediction model with the target distribution of the cloud-based rule map estimation. , These are the mean and covariance of the feature distribution at the edge, respectively. , These are the mean and covariance of the target distribution corresponding to the cloud-based rule graph, respectively.

[0018] In other embodiments, the following technical solutions are adopted: An advanced geological prediction system based on cloud-edge model collaboration includes: The data acquisition module is configured to acquire historical drilling parameters, seismic wave reflection spectra, tunnel face images, and spectral data of the area to be excavated. The cloud prediction module is configured to use a cloud-based geological prediction model to obtain a geological probability prediction map; the geological probability prediction map includes the rock mass integrity, water inrush risk, and fracture or weak interlayer prediction results and corresponding confidence levels for each location. The edge prediction module is configured to treat the physical sensors on the tunneling equipment as nodes, construct the connection relationship between the nodes based on the principles of physical spatial proximity and functional coupling, and form a graph structure; acquire the real-time feature data of each node in the graph structure, use the edge geological prediction model to predict the temporal latent feature representation of the node, and perform inference calculation on the temporal latent feature representation of each node in the graph structure to obtain the local geological state vector corresponding to each node. The geological state vector includes the probability of rock mass integrity, the probability of water inrush risk, the probability of the occurrence of fractures or weak interlayers, and their confidence information. The distribution map construction module is configured to use the geological state vectors output by each node to form an edge geological probability distribution map after spatial interpolation and topological fusion, which is used to describe the real-time geological state of the working face and the surrounding rock in front. The collaborative prediction module is configured to calculate the consistency index for each location based on the prediction results of the cloud-based geological prediction model and the edge-based geological prediction model. If the consistency index is within the set warning / protection range, the real-time node features, seismic wave reflection spectrum of the excavated section, face image and spectral data are uploaded to the cloud-based geological prediction model for re-prediction to obtain an updated geological probability prediction map.

[0019] In other embodiments, the following technical solutions are adopted: A terminal device includes a processor and a memory, the processor being used to implement instructions; the memory being used to store multiple instructions, the instructions being adapted to be loaded and executed by the processor to perform the aforementioned advanced geological prediction method based on cloud-edge model collaboration.

[0020] In other embodiments, the following technical solutions are adopted: A computer-readable storage medium storing a plurality of instructions adapted for loading and execution by a processor of a terminal device of the aforementioned cloud-edge model collaborative advanced geological prediction method.

[0021] Compared with the prior art, the beneficial effects of the present invention are: (1) The cloud-based geological prediction model of this invention is based on the cross-modal attention mechanism of Transformer, which deeply integrates drilling, seismic waves, images and spectral multimodal data, fully explores the deep correlation between multimodal data, breaks through the limitation of traditional single methods that can only detect a single geological target, and extends the effective prediction distance from within 20 meters to the range of 20-50 meters.

[0022] This invention deploys a lightweight graph neural network (GCN-Lite) in its edge-end geological prediction model, processing sensor data in real-time with a 200ms cycle. It achieves efficient computation through separable graph convolution and triggers a three-level response mechanism (maintain, warn, or slowdown) based on the consistency index, thus overcoming the industry pain points of traditional methods such as short prediction distance and high parsing latency. The millisecond-level risk perception capability of the edge-end geological prediction model drives dynamic control of the tunneling machine, significantly reducing unplanned downtime and equipment damage and repair costs caused by geological disasters such as sudden water inrush and rock bursts.

[0023] (2) The edge-end geological prediction model of the present invention responds promptly and analyzes the data obtained by the sensors quickly, but the prediction range is extremely short. The cloud-based geological prediction model has a slow calculation speed, but it can predict a relatively long section in front of the excavation and verify the prediction results at the edge end.

[0024] The geological prediction model sends compressed geological map data to the edge geological prediction model, and the edge geological prediction model uploads abnormal working condition data and parameter update volume between the edge and cloud models to the cloud geological prediction model. Through federated learning, multiple edge site samples are aggregated to achieve continuous iterative evolution of the cloud model sample library. The evolved knowledge is constrained by feature distribution alignment loss to ensure that it is both adapted to the local environment and does not deviate from the global cognition.

[0025] The invention enables the cloud-based geological prediction model and the edge-based geological prediction model to evolve in tandem, greatly improving the accuracy and predictability of geological forecasts.

[0026] Other features and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0027] Figure 1 This is a flowchart of the advanced geological prediction method based on cloud-edge model collaboration in an embodiment of the present invention. Detailed Implementation

[0028] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0029] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0030] Example 1 In one or more embodiments, a cloud-edge model-based advanced geological prediction method is disclosed, combining... Figure 1 Specifically, it includes the following process: S101: Obtain historical drilling parameters, seismic wave reflection spectra, tunnel face images, and spectral data of the area to be excavated.

[0031] In this embodiment, the historical drilling parameters of the section to be excavated include: sequential data such as drilling speed, thrust, mud pressure, and cutterhead torque recorded by the TBM construction control system, as well as information on surrounding exploration boreholes; seismic wave reflection spectra are obtained by the seismic source and detector array deployed in front of the tunnel boring machine, reflecting the strata reflection characteristics within tens of meters in front of the tunnel face; and tunnel face images are periodically acquired by an industrial camera installed at the front end of the TBM, with spectral data obtained by scanning the rock surface of the tunnel face using near-infrared or Raman spectral sensors.

[0032] The acquired data were preprocessed, including data smoothing and denoising, outlier filtering and removal, and data normalization. For image data, this included image enhancement and segmentation, and for spectral data, correction and standard normal variable transformation. The specific processing procedures are routine operations and will not be described in detail.

[0033] S102: Using a cloud-based geological prediction model, a geological probability prediction map is obtained; the geological probability prediction map includes the rock mass integrity, water inrush risk, and prediction results of fractures or weak interlayers at each location, as well as the corresponding confidence levels.

[0034] Specifically, the cloud-based geological prediction model is a neural network model based on the Transformer architecture; the historical drilling parameters, seismic wave reflection spectra, tunnel face images and spectral data of the area to be excavated are aligned according to the tunneling mileage and timestamp, and used as input data for the cloud-based geological prediction model. The core objective of cloud-based geological prediction models is to establish a model of latent geological variables within a 20-50 meter range ahead of the tunnel face by integrating multimodal information such as historical drilling parameters, seismic wave reflection spectra, tunnel face images, and spectral data. As a geological cognitive engine, cloud-based geological prediction models aim to uncover deep-seated relationships in historical disaster cases, including rock mass structure, seepage field, and joint and fracture networks.

[0035] The input data is linearly mapped to a high-dimensional embedding space, and then information from various modalities is fused through a cross-modal attention mechanism to finally output a four-dimensional geological probability map; the specific process is as follows: Let the drilling parameters be... The earthquake waveform is Image features are The spectrum is .

[0036] Each modal input is linearly mapped to a high-dimensional embedding space, in the form of: ; in, j Representing 4 different input data modes, d For drilling parameters, s This is an earthquake waveform. i For image features, spec For spectral data; Indicates the first j The original input features of each modality; This represents the linear mapping weight matrix for the corresponding modality, used to transform input features to a uniform embedding dimension; Represents the mapping bias vector; This represents a high-dimensional embedding vector obtained through linear transformation, used as the input token (smallest unit) of the Transformer geological cognition engine.

[0037] Subsequently, information is fused through a cross-modal attention mechanism, ultimately outputting a four-dimensional geological probability map. ;in, x Indicates the spatial position along the tunnel axis; y Indicates the lateral spatial location, used to describe the characteristics of the surrounding rock on the left and right sides of the working face; z Indicates vertical spatial location, used to describe the height distribution of the surrounding rock; t The time dimension represents the tunneling process, reflecting the dynamic changes in geological conditions as construction progresses. By introducing the time dimension, it is possible to achieve spatiotemporal joint prediction of the geological conditions of the area to be excavated in the future. The geological probability prediction map can reflect the rock mass integrity, water inrush risk, and prediction results of fractures or weak interlayers at each location, along with their corresponding confidence levels.

[0038] The core of the cross-modal attention mechanism lies in enabling the model to dynamically and selectively integrate data from different sources. First, the raw data of each modality (such as drilling parameters, seismic waveforms, etc.) are projected into a unified high-dimensional embedding space through independent linear transformations, transforming them into vectors with the same format, preparing for subsequent fusion. Next, the model generates three sets of vectors for each modality: a query vector, a key vector, and a value vector.

[0039] The fusion process begins with a "question-response" mechanism. For example, when the model processes a seismic wave signal (represented by its query vector), it treats this signal as a "question" and "asks" all modal key vectors (including drilling, images, etc.), obtaining a set of attention weights by calculating similarity. For the h-th attention head: ; in, This represents the query corresponding to the h-th attention head. AND key The rating; This represents the dimension (i.e., feature length) of each key vector. This represents the positional offset term used to characterize the relative positional relationship between query location i and key location j in the tunnel spatial coordinate system or tunneling mileage coordinate system, and is used to reflect the correlation differences between geological information at different spatial locations. This represents the graph bias derived from the topological adjacency relationships of physical sensors. Indicates the modal confidence level of the key origin. This represents the time difference between the query and the key. This represents the time decay coefficient.

[0040] This embodiment introduces parameters such as position bias, graph bias, modality reliability, and time difference into cross-modal attention calculation, so that the multimodal fusion process is simultaneously constrained by spatial continuity, equipment physical structure, data quality, and temporal consistency. This significantly improves the accuracy, stability, and overall robustness of geological risk prediction in noisy and modality-deficient scenarios.

[0041] right Normalization is performed to obtain attention weights; these weights quantify the importance of each other data type in interpreting the current seismic wave signal. Finally, the value vectors of all modes are weighted and summed according to these weights to generate a new feature representation that is based on the seismic wave perspective but incorporates global information.

[0042] For any modal data, it can be used as a query to obtain the corresponding fused representation. Then put all The tokens are concatenated and input into subsequent Transformer layers for deep interaction, or each modality is pooled and fused using a gated weighted approach to obtain the final spatiotemporal geological fusion representation. Finally, utilize The geological probability distribution is obtained by multilayer perceptron (MLP) prediction and used to generate a four-dimensional geological probability map. .

[0043] Through this mechanism, the model can automatically discover complex cross-modal correlations, such as when "high-frequency reflection peaks" (from seismic waves) and "sudden drops in drilling speed" (from drilling data) occur simultaneously, their joint probability points to "hidden fault zones." The reason for exploring these deep relationships is that major geological disasters (such as sudden water inrushes) are often the result of complex coupling of multiple factors, including rock mass structure, seepage field, and stress field. These nonlinear relationships far exceed the empirical summarization capabilities of human experts, and single-modal data cannot provide a unique solution. In practice, the model encounters massive amounts of historical case data during training. The underlying attention mechanism is responsible for discovering superficial, direct statistical correlations between different modal data (e.g., the co-occurrence relationship between fracture images and a certain type of seismic waveform). Deeper layers of the network then build upon this foundation, performing higher-order combinations and abstractions of these preliminary features to learn more complex coupling patterns (e.g., a specific "fracture-waveform-water pressure" combination pattern indicates an extremely high risk of rockburst). Through this data-driven approach, the model can ultimately uncover deep-seated disaster-causing patterns that are not explicitly defined but do exist, thereby achieving more accurate early warnings.

[0044] In this embodiment, the loss function of the cloud-based geological prediction model is specifically as follows: ; in, , , These are hyperparameters used to balance the importance of the three losses in the total loss.

[0045] Geological state regression error is used to measure the difference between the model's predicted probabilities of rock mass integrity, water inrush risk, and fractures or weak interlayers, and the actual geological annotation results. ; in, Let i be the vector representing the true geological state at the i-th spatial location. This is the probability vector of geological state predicted by the model, where N is the spatial location or number of samples.

[0046] The masked reconstruction loss is achieved by randomly masking some drilling parameters, seismic wave reflection characteristics, image features, or spectral features, and constraining the model's reconstruction error of the masked information. This enhances the model's robustness under conditions of missing or noisy multimodal data. Masked Reconstruction Loss Specifically: ; in, The original input features, M represents the features reconstructed by the model, and M is the set of mask locations.

[0047] As a latent variable distribution alignment constraint, KL divergence is used to measure the consistency between the prior and posterior distributions; it is used to constrain the distribution pattern of geological latent variables within the model so that it conforms to the preset prior distribution, thereby avoiding cognitive drift of the model under different construction scenarios and improving the generalization ability across engineering scenarios.

[0048] Assumption For model latent variables, For the posterior distribution, It is the prior distribution; KL divergence formula: ; Under Gaussian assumptions: ; in: , Let be the mean and variance of the d-th dimension of the latent variable, respectively. This represents the total number of dimensions of the latent variable z.

[0049] During training, a batch of samples is randomly selected, and a portion of the input data is masked (e.g., the entire drilling parameter sequence is set to zero, or certain blocks in an image are randomly masked) to simulate data gaps or incomplete data collection in real-world engineering projects. The model's task is then to predict the original values ​​of the masked portion based on the remaining unmasked data. For example, if drilling data is masked, the model must learn to observe other clues such as seismic wave reflection spectra and tunnel face images to infer that "a sudden drop in drilling speed likely occurred here." The loss function used to measure prediction accuracy (i.e., masked reconstruction loss) guides the model parameter updates. Through massive amounts of such practice, the model is forced to deeply understand the causal and correlational relationships between various types of data, thereby greatly enhancing its reasoning and generalization abilities in the face of incomplete information, making it more robust in practical applications.

[0050] S103: Treat the physical sensors on the tunneling equipment as nodes, and construct the connection relationship between the nodes based on the principles of physical spatial proximity and functional coupling to form a graph structure; acquire the real-time feature data of each node in the graph structure, and use the edge-end geological prediction model to predict the temporal latent feature representation of the node; perform inference calculation on the temporal latent feature representation of each node in the graph structure to obtain the local geological state vector corresponding to each node. The geological state vector includes the probability of rock mass integrity, the probability of water inrush risk, the probability of the occurrence of fractures or weak interlayers, and their confidence information.

[0051] In this embodiment, the physical sensors on the tunneling equipment include: a cutterhead torque sensor, a cutterhead speed encoder, a cutter force sensor, a cutter vibration accelerometer, a high-definition camera at the tunnel face, a short-wave infrared spectral sensor, a seismic source, a seismic detector, a ground-penetrating radar (GPR) antenna, an ultrasonic sensor, a propulsion cylinder pressure sensor, a shield attitude sensor (IMU), and temperature, humidity, and gas sensors. Specifically, torque, speed, and force sensors, as well as a high-definition camera and spectral sensor, are arranged in the cutterhead area (cutterhead spindle, cutter holder, and cutterhead back) to collect information on rock fracture characteristics, cutting resistance, and mineral composition. Seismic sources and detectors are arranged in the front and middle sections of the shield to obtain seismic wave reflection spectra within a 20–50 meter range ahead. Radar antennas are arranged on the top and sides of the shield to detect fractures and water-bearing areas ahead. Pressure and attitude sensors are arranged in the propulsion system cylinders, the shield main beam, and the joints of each section to reflect the weakness of the surrounding rock and possible deformation. Temperature, humidity, and gas sensors are arranged in the slurry pressure chamber and the slag removal system to monitor the construction environment.

[0052] The aforementioned sensors constitute a spatially distributed multi-node system, and a graph structure is constructed based on physical proximity and functional coupling relationships. ;in For a set of sensor nodes, For temporal edge sets, The graph is a feature matrix. In this embodiment, each physical sensor deployed on the tunneling machine is considered an independent node in the graph. The connections (edges) between nodes are established based on two main principles: first, the principle of physical spatial proximity, meaning that sensors installed on the same component of the equipment or adjacent mechanical structures are automatically connected, such as vibration sensors and temperature sensors on the same cutterhead; second, the principle of functional coupling, based on the physical logic of equipment operation, sensors with linkage or causal relationships are connected, for example, there is a strong correlation between the torque output of the main drive motor and the pressure of the propulsion hydraulic system, and an edge is established between their sensor nodes. In this way, the entire graph structure organically organizes the discrete sensor points on the tunneling machine into a network that reflects their physical layout and functional relationships.

[0053] In this embodiment, the edge-end geological prediction model is a lightweight spatiotemporal graph neural network model. The model uses separable graph convolution to achieve lightweight computation. ; , For the first , The input node feature matrix of a +1 layer graph convolutional network. Let be the adjacency matrix of the graph, representing the connection relationships between sensor nodes. The degree matrix is ​​a diagonal matrix. Yes Normalization operation, For the first Layer-learnable weight matrix, This is the activation function.

[0054] The input to separable graph convolution mainly consists of two parts: one is the graph structure describing the connection relationship of the sensor network, namely the adjacency matrix, which is a matrix that uses 0 and 1 to indicate whether there is a connection between nodes; the other is the real-time feature data of each sensor node, such as the amplitude values ​​of all vibration sensors, the readings of temperature sensors, etc. at a certain moment, which together form a feature matrix.

[0055] The processing involves a hierarchical feature extraction and fusion process: First, the adjacency matrix is ​​normalized to eliminate biases caused by excessive sensor connections. Then, in each graph convolutional layer, each node aggregates information from its directly connected neighbors through an efficient separable convolution operation. This process is analogous to each sensor "listening" to readings from its surrounding sensors and combining this information with its own. The aggregated information undergoes a linear transformation and a non-linear activation function to generate a new feature representation for the node in the current layer, incorporating the local network state. This operation is performed layer by layer, allowing higher-level node features to encompass a wider range of neighbor information.

[0056] The final output is a node feature matrix after multi-layer processing. The node feature matrix contains a high-dimensional feature representation of each sensor node after fusing its own real-time observations, neighbor node information, and time-series evolution features. It is used to characterize the correlation between sensor data and local surrounding rock geological conditions. At this point, the feature vector of each sensor is no longer just its original reading, but an enhanced "state representation" containing its surrounding environmental context information. By discriminative mapping of the feature vector of each node in the node feature matrix and probabilistic processing of the mapping results, the geological state vector corresponding to the node is obtained, including the probability of rock mass integrity, the probability of water inrush risk, and the probability of fractures or weak interlayers.

[0057] S104: The geological state vectors output by each node are spatially interpolated and topologically fused to form an edge geological probability distribution map, which is used to describe the real-time geological state of the tunnel face and the surrounding rock in front.

[0058] S105: Based on the prediction results of the cloud-based geological prediction model and the edge-based geological prediction model, calculate the consistency index for each location, and then... Triggering a three-level response mechanism.

[0059] In this embodiment, the consistency index is calculated as follows: For cloud-predicted probability vectors With edge prediction probability vector The consistency index is defined as: ; in: Represents Euclidean distance. This typically includes: probability of rock mass integrity, probability of water inrush risk, and probability of weak interlayers. As a normalization factor, make , can be set Or it can be obtained from historical sample statistics.

[0060] The system performs inference at a frequency of every 200ms, based on the consistency index. Triggering a three-level response mechanism: like >0.9, the original prediction of the edge geological prediction model remains unchanged; like When the time comes, an early warning is issued, and the real-time node characteristics, seismic wave reflection spectrum of the excavated section, face image and spectral data are uploaded to the cloud geological prediction model for re-prediction, resulting in an updated geological probability prediction spectrum.

[0061] like It issues an alarm and coordinates with equipment to reduce the tunneling rate, enhancing proactive response capabilities.

[0062] In this embodiment, the cloud-based geological prediction model and the edge-based geological prediction model achieve co-evolution through a two-way knowledge distillation and feature alignment mechanism.

[0063] The cloud-based geological prediction model utilizes a geological rule network extracted from its cross-scenario data. This network comprises nodes, edges, and weights. Nodes are geological semantic units, i.e., geological states or elements with clear engineering and geological semantics, such as surrounding rock type and integrity. Edges represent the correlation or evolutionary dependence of different geological elements under spatial, temporal, or engineering conditions, that is, the coexistence of different geological states. Edge weights are determined based on the correlation strength between different geological elements in the cross-scenario geological data, reflecting the degree of influence between geological states.

[0064] The geological rule network is essentially a compressed, massive knowledge base that significantly expands the geological cognitive boundaries of edge models. Edge models rely solely on single-point real-time data, limiting their perspective to the local conditions of the current tunnel and making them unable to identify rare geological patterns. In contrast, cloud-based models integrate massive historical cases, and their distributed maps contain rich information such as the probability distribution of rock mass types, the probability of water inrush risk, the probability of fractures or weak interlayers, and their confidence levels. This is equivalent to providing edge models with a "warning map" marked with various potential risks, enabling them to remain vigilant even about geological anomalies they haven't personally experienced, based on cloud-based knowledge, thus overcoming their inherent cognitive limitations.

[0065] The geological rule network is distributed to the edge geological prediction model, enabling the edge geological prediction model to have a wider range of geological knowledge that cannot be covered by local data; the data transmission adopts the ZFP lossless compression algorithm, and supports daily model differential updates. In the case of network outage, the local disaster plan library is automatically called to ensure the steady-state operation of the system.

[0066] The edge-end achieves cognitive synchronization with the cloud model through feature distribution alignment loss, specifically by constraining statistical features; the specific loss function is as follows: ; in, This is the loss function for the edge-end geological prediction model, used to constrain the alignment of the feature distribution of the edge-end geological prediction model with the target distribution of the cloud-based rule map estimation. , These are the mean and covariance of the feature distribution at the edge, respectively. , These are the mean and covariance of the target distribution corresponding to the cloud-based rule graph, respectively.

[0067] This mechanism aims to prevent "feature drift" in the edge model during its continuous learning of local data, where its internal data representation gradually deviates from the authoritative model in the cloud. In implementation, the system calculates the mean and covariance matrices of the feature outputs of both the edge and cloud models separately, and then forces the minimization of the difference between them using a loss function. This process acts like a calibrator for the edge model's learning, ensuring that its "mindset" remains consistent with the overall cloud model when adapting to the local environment.

[0068] Meanwhile, the edge-end geological prediction model provides highly refined valuable data to the cloud, including abnormal working condition data that triggers alarms and the corresponding geological prediction results, which are used to enrich the training sample library of the cloud-based geological prediction model.

[0069] The edge computing infrastructure provides the cloud with highly refined, valuable data rather than raw data streams. This data undergoes rigorous screening and compression to form a "knowledge increment." The cloud utilizes these high-value samples from various field sites to securely aggregate model update differences from each edge computing endpoint using a federated learning framework. This data significantly enriches the cloud's training sample library, especially rare geological disaster cases, enabling the large model to continuously evolve. Ultimately, this forms a virtuous cycle of "field practice, knowledge refinement, global sharing, and capability enhancement," driving the entire system to achieve a paradigm shift from passive perception to proactive prevention and control.

[0070] The method in this embodiment supports long-distance accurate prediction and millisecond-level early warning response. It absorbs new disaster samples monthly through federated learning to achieve periodic evolution, effectively promoting the transformation of advanced geological forecasting from "passive perception" to "proactive prevention and control", and providing a new intelligent and structured prediction and defense paradigm for deep tunnel engineering.

[0071] Example 2 In one or more embodiments, a cloud-edge model-based advanced geological prediction system is disclosed, specifically including: The data acquisition module is configured to acquire historical drilling parameters, seismic wave reflection spectra, tunnel face images, and spectral data of the area to be excavated. The cloud prediction module is configured to use a cloud-based geological prediction model to obtain a geological probability prediction map; the geological probability prediction map includes the rock mass integrity, water inrush risk, and fracture or weak interlayer prediction results and corresponding confidence levels for each location. The edge prediction module is configured to treat the physical sensors on the tunneling equipment as nodes, construct the connection relationship between the nodes based on the principles of physical spatial proximity and functional coupling, and form a graph structure; acquire the real-time feature data of each node in the graph structure, use the edge geological prediction model to predict the temporal latent feature representation of the node, and perform inference calculation on the temporal latent feature representation of each node in the graph structure to obtain the local geological state vector corresponding to each node. The geological state vector includes the probability of rock mass integrity, the probability of water inrush risk, the probability of the occurrence of fractures or weak interlayers, and their confidence information. The distribution map construction module is configured to use the geological state vectors output by each node to form an edge geological probability distribution map after spatial interpolation and topological fusion, which is used to describe the real-time geological state of the working face and the surrounding rock in front. The collaborative prediction module is configured to calculate the consistency index for each location based on the prediction results of the cloud-based geological prediction model and the edge-based geological prediction model. If the consistency index is within the set warning / protection range, the real-time node features, seismic wave reflection spectrum of the excavated section, face image and spectral data are uploaded to the cloud-based geological prediction model for re-prediction to obtain an updated geological probability prediction map.

[0072] The specific implementation methods of the above modules are the same as those in Example 1, and will not be described in detail again.

[0073] Example 3 In one or more embodiments, a terminal device is disclosed, comprising a processor and a memory, wherein the processor is used to implement instructions; and the memory is used to store multiple instructions adapted to be loaded by the processor and executed by the processor to perform the advanced geological prediction method based on cloud-edge model collaboration described in Embodiment 1.

[0074] It should be understood that in this embodiment, the processor can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0075] Memory may include read-only memory and random access memory, and provides instructions and data to the processor. A portion of memory may also include non-volatile random access memory. For example, memory may also store information about the device type.

[0076] In the implementation process, each step of the above method can be completed by the integrated logic circuits in the processor hardware or by software instructions.

[0077] Example 4 In one or more embodiments, a computer-readable storage medium is disclosed, wherein a plurality of instructions are stored, the instructions being adapted to be loaded by a processor of a terminal device and executed by the cloud-edge model collaborative advanced geological prediction method described in Embodiment 1.

[0078] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A cloud-edge model-based advanced geological prediction method, characterized in that, include: Obtain historical drilling parameters, seismic wave reflection spectra, tunnel face images, and spectral data of the area to be excavated; A geological probability prediction map is obtained using a cloud-based geological prediction model. The geological probability prediction map includes the prediction results and corresponding confidence levels of rock mass integrity, water inrush risk, and fracture or weak interlayer at each location. The physical sensors on the tunneling equipment are regarded as nodes. Based on the principles of physical spatial proximity and functional coupling, the connection relationship between the nodes is constructed to form a graph structure. Real-time feature data of each node in the graph structure is obtained. Using the edge geological prediction model, the temporal latent feature representation of the node is predicted. The temporal latent feature representation of each node in the graph structure is inferred and calculated to obtain the local geological state vector corresponding to each node. The geological state vector includes the probability of rock mass integrity, the probability of water inrush risk, the probability of the occurrence of fractures or weak interlayers and their confidence information. The geological state vectors output by each node are spatially interpolated and topologically fused to form an edge geological probability distribution map, which is used to describe the real-time geological state of the tunnel face and the surrounding rock in front. Based on the prediction results of the cloud-based geological prediction model and the edge-based geological prediction model, the consistency index of each location is calculated. If the consistency index is within the set warning / protection range, the real-time node characteristics, seismic wave reflection spectrum of the excavated section, face image and spectral data are uploaded to the cloud-based geological prediction model for re-prediction to obtain the updated geological probability prediction map. The edge-end geological prediction model is a spatiotemporal graph neural network model. The spatiotemporal graph neural network model uses multiple physical sensor data on the tunneling equipment as nodes, and determines the connection relationship between nodes according to the principle of physical spatial proximity and the principle of functional coupling to construct a graph structure. The graph structure and the real-time feature data of each sensor node are used as input to the spatiotemporal graph neural network model. In each layer of graph convolution, each node aggregates information from its directly connected neighbor nodes through separable convolution operations. The aggregated information is then transformed by a linear transformation and a nonlinear activation function to generate a new feature representation of the node in the current layer that incorporates the local network state. After layer-by-layer operations, the node feature matrix after multi-layer processing is finally output. The feature vector of each node in the node feature matrix is ​​discriminated and mapped, and the mapping result is probabilistically processed to obtain the local geological state vector corresponding to each node. It also includes the co-evolution process of cloud-based geological prediction models and edge-based geological prediction models, specifically: The cloud-based geological prediction model utilizes a geological rule network extracted from its own cross-scenario data. This geological rule network uses geological semantic units as nodes and the coexistence states of different geological states as edges. The edge weights are determined based on the correlation strength between different geological elements in the cross-scenario geological data. The geological rule network is distributed to the edge geological prediction model, enabling the edge geological prediction model to have a wider range of geological knowledge that cannot be covered by local data. The edge-end geological prediction model provides highly refined value data to the cloud, including abnormal working condition data that triggers alarms and the corresponding geological prediction results, which are used to enrich the training sample library of the cloud-based geological prediction model. The edge-end geological prediction model achieves cognitive synchronization with the cloud model through feature distribution alignment loss; the feature distribution alignment loss is specifically: ； in, This is the loss function for the edge-end geological prediction model, used to constrain the alignment of the feature distribution of the edge-end geological prediction model with the target distribution estimated by the cloud-based rule map. , These are the mean and covariance of the feature distribution at the edge, respectively. , These are the mean and covariance of the target distribution corresponding to the cloud-based rule graph, respectively.

2. The advanced geological prediction method based on cloud-edge model collaboration as described in claim 1, characterized in that, The cloud-based geological prediction model is a neural network model based on the Transformer architecture; The historical drilling parameters, seismic wave reflection spectra, tunnel face images, and spectral data of the area to be excavated are aligned according to the tunneling mileage and timestamp, and used as input data for the cloud-based geological prediction model. The input data is linearly mapped to a high-dimensional embedding space, and then the information of each modality is fused through a cross-modal attention mechanism to finally output a four-dimensional geological probability map.

3. The advanced geological prediction method based on cloud-edge model collaboration as described in claim 2, characterized in that, Each modality of data generates three sets of vectors: a query vector, a key vector, and a value vector; for the h-th attention head: ； in, This represents the query corresponding to the h-th attention head. AND key The rating; This represents the dimension of each key vector; This represents the position offset term indicating the relative positional relationship between query position i and key position j in the tunnel spatial coordinate system or tunneling mileage coordinate system. This represents the graph bias derived from the sensor topological adjacency relationships. Indicates the modal confidence level of the key origin. This represents the time difference between the query and the key. Indicates the time decay coefficient; right Normalization is performed to obtain attention weights; based on the attention weights, the value vectors of each modality data are weighted and summed to obtain the data fusion representation.

4. The advanced geological prediction method based on cloud-edge model collaboration as described in claim 1, characterized in that, The loss function of the cloud-based geological prediction model is as follows: ； in, This represents the geological state regression error; Loss due to mask reconstruction For the latent variable distribution alignment constraint term, , , These are hyperparameters, used to balance the importance of the three losses in the total loss; The geological state regression error Specifically: ； in, Let i be the vector representing the true geological state at the i-th spatial location. This is the probability vector of geological state predicted by the model, where N is the spatial location or number of samples. The mask reconstruction loss Specifically: ； in, The original input features, The features to be reconstructed by the model, M is the set of mask locations; The latent variable distribution alignment constraint term Specifically: ； in, and Let be the mean and variance of the d-th dimension of the latent variable, respectively. This represents the total number of dimensions of the hidden variables.

5. The advanced geological prediction method based on cloud-edge model collaboration as described in claim 1, characterized in that, The physical sensors on the tunneling equipment include: cutterhead torque sensor, cutterhead speed encoder, cutter force sensor, cutter vibration accelerometer, face high-definition camera, shortwave infrared spectral sensor, earthquake source, seismic detector, ground-penetrating radar antenna, ultrasonic sensor, propulsion cylinder pressure sensor, shield attitude sensor, and temperature, humidity and gas sensors.

6. The advanced geological prediction method based on cloud-edge model collaboration as described in claim 1, characterized in that, Based on the prediction results of the cloud-based geological prediction model and the edge-based geological prediction model, the consistency index for each location is calculated as follows: For cloud-predicted probability vectors With edge prediction probability vector The consistency index is defined as: ； in: Represents Euclidean distance. As a normalization factor, make Cloud-based prediction probability vector Or edge prediction probability vector This includes the probability of rock mass integrity, the probability of water inrush risk, and the probability of fractures or weak interlayers.

7. The advanced geological prediction method based on cloud-edge model collaboration as described in claim 6, characterized in that, When the consistency index is greater than the first predetermined threshold, the original prediction of the edge geological prediction model remains unchanged. When the consistency index is less than the second predetermined threshold, an alarm is issued and the equipment is linked to reduce the tunneling rate. When the consistency index is between the second predetermined threshold and the first predetermined threshold, an alarm is issued and the real-time node characteristics, seismic wave reflection spectrum of the excavated section, face image and spectral data are uploaded to the cloud geological prediction model for re-prediction to obtain an updated geological probability prediction map.

8. A system for implementing the cloud-edge model-based advanced geological prediction method as described in claim 1, characterized in that, include: The data acquisition module is configured to acquire historical drilling parameters, seismic wave reflection spectra, tunnel face images, and spectral data of the area to be excavated. The cloud prediction module is configured to use a cloud-based geological prediction model to obtain a geological probability prediction map; the geological probability prediction map includes the rock mass integrity, water inrush risk, and fracture or weak interlayer prediction results and corresponding confidence levels for each location. The edge prediction module is configured to treat the physical sensors on the tunneling equipment as nodes, construct the connection relationship between the nodes based on the principles of physical spatial proximity and functional coupling, and form a graph structure; acquire the real-time feature data of each node in the graph structure, use the edge geological prediction model to predict the temporal latent feature representation of the node, and perform inference calculation on the temporal latent feature representation of each node in the graph structure to obtain the local geological state vector corresponding to each node. The geological state vector includes the probability of rock mass integrity, the probability of water inrush risk, the probability of the occurrence of fractures or weak interlayers, and their confidence information. The distribution map construction module is configured to use the geological state vectors output by each node to form an edge geological probability distribution map after spatial interpolation and topological fusion, which is used to describe the real-time geological state of the working face and the surrounding rock in front. The collaborative prediction module is configured to calculate the consistency index for each location based on the prediction results of the cloud-based geological prediction model and the edge-based geological prediction model. If the consistency index is within the set warning / protection range, the real-time node features, seismic wave reflection spectrum of the excavated section, face image and spectral data are uploaded to the cloud-based geological prediction model for re-prediction to obtain an updated geological probability prediction map.

9. A terminal device comprising a processor and a memory, the processor for implementing instructions; the memory for storing multiple instructions, characterized in that, The instructions are adapted to be loaded by a processor and executed as described in any one of claims 1-7, which is a cloud-edge model-based advanced geological prediction method.

10. A computer-readable storage medium storing a plurality of instructions, characterized in that, The instructions are adapted to be loaded and executed by the processor of the terminal device, and to execute the cloud-edge model collaborative advanced geological prediction method according to any one of claims 1-7.

Images