A machine learning-based tunnel water gushing rapid prediction method and system

Abstract

The application provides a tunnel water gushing rapid prediction method and system based on machine learning, relates to the field of tunnel engineering, and aims at the problem that the existing technology has single and limited data sources, single prediction model and insufficient prediction accuracy, and cannot adapt to dynamic change conditions for real-time prediction.The method comprises the following steps: acquiring multi-source data, performing space-time alignment on the multi-source data, and generating a unified data tensor; inputting the unified data tensor into a dynamic feature distillation network model for feature extraction to obtain unified semantic space features; and inputting the unified semantic space features into a three-layer dynamic integrated model for parallel prediction to obtain a final prediction result.The application improves the prediction accuracy and efficiency through multi-source data fusion, adaptive feature extraction and real-time integrated learning, can timely warn potential water gushing risks, provides technical support for tunnel construction safety, and effectively reduces the incidence of tunnel construction safety accidents.

Description

Technical Field

[0001] This invention belongs to the field of tunnel engineering, and in particular relates to a method and system for rapid prediction of tunnel water inrush based on machine learning. 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] Water inrush is a common and dangerous geological hazard during tunnel construction, which can not only affect the construction progress but also lead to engineering accidents and casualties. With the continuous expansion of tunnel engineering scale and the increase in tunnel construction under complex geological conditions, how to accurately predict the risk of tunnel water inrush has become a key technical problem that urgently needs to be solved in engineering construction.

[0004] Currently, methods for predicting tunnel water inflow mainly include empirical formula methods, numerical simulation methods, and data-driven methods. Among these, empirical formula methods rely on calculation formulas summarized from previous engineering experience for prediction; numerical simulation methods simulate groundwater flow processes by establishing mathematical models. In recent years, with the development of artificial intelligence technology, data-driven prediction methods have gradually gained attention.

[0005] Furthermore, to improve prediction accuracy, researchers have also attempted to combine multiple methods. For example, they have employed semi-data, semi-empirical models, statistically analyzing multiple real-world engineering cases and adjusting relevant parameters using existing empirical formulas. Other methods include prediction based on key indicators and decoupling indicators. Still others involve training geological data with various machine learning models and selecting the best-performing model for geological risk prediction.

[0006] However, existing methods for predicting tunnel water inrush still have the following shortcomings: existing technologies rely on a single data source, often focusing only on a single or limited geological parameter or local monitoring data; the prediction models are singular, prone to errors, unable to adapt to predictions of complex data relationships, and are mostly static models, making it difficult to respond to changes in the construction environment; traditional prediction methods involve complex calculation processes and are time-consuming, making it difficult to meet the needs of construction sites for rapid assessment of water inrush risks and unable to provide timely and effective early warnings of water inrush risks for tunnel construction. Summary of the Invention

[0007] To overcome the shortcomings of the existing technology, this invention provides a method and system for rapid prediction of tunnel water inrush based on machine learning. By comprehensively collecting and integrating multi-source data such as geological data, hydrological data, meteorological data, engineering process data and monitoring data, the comprehensiveness and accuracy of information are improved. A tunnel water inrush risk prediction model is constructed using a dynamic feature distillation model and a three-layer dynamic integration model for effective prediction.

[0008] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions: The first aspect of this invention provides a method for rapid prediction of tunnel water inrush based on machine learning, comprising: Acquire multi-source data, which should include at least geological data, hydrological data, meteorological data, engineering process data, and monitoring data; Spatiotemporal alignment of multi-source data is performed to generate a unified data tensor; A unified data tensor is input into a dynamic feature distillation network model for feature extraction, resulting in unified semantic space features. The unified semantic space features are input into a three-layer dynamic ensemble model for prediction and optimization to obtain the final prediction result. The three-layer dynamic ensemble model consists of a model prediction layer, a feature fusion optimization layer, and an output layer connected in sequence. The model prediction layer makes predictions to obtain independent prediction values ​​for each model. Based on the independent prediction values ​​of each model, the feature fusion optimization layer extracts and fuses features to obtain intermediate prediction features after fusion optimization. The output layer performs dynamic error weighting to obtain the final prediction result.

[0009] As one implementation method, spatiotemporal alignment of multi-source data is performed, and the specific process is as follows: A sliding time window resampling method is used to align the time of multi-source data. A spatial inverse distance weighted interpolation algorithm based on tunnel cross-section coordinates is used to spatially align multi-source data; Controlled imputation is performed on spatially aligned outlier data, and features of different dimensions are normalized to obtain a unified data tensor.

[0010] As one implementation method, a unified data tensor is input into a dynamic feature distillation network model for feature extraction. This model includes an original feature encoder layer, a feature distillation layer, and a residual enhancement layer. The specific process is as follows: A unified data tensor is input into the original feature encoder layer to extract primary features from multi-source data. The initial features of multi-source data are input into the feature distillation layer. The initial features of multi-source data are weighted and fused through a multi-source attention mechanism and a channel selection mechanism to obtain a unified feature representation after weighted fusion. The weighted and fused unified feature representation is input into the residual enhancement layer to obtain unified semantic space features.

[0011] As one implementation method, the dynamic feature distillation network is optimized through a dual-supervision mechanism, as shown in the following formula:

[0012] in, This is a hyperparameter used to control the weight of distillation loss in the total loss; To predict the loss of the main task, it represents the standard deviation between the predicted value and the actual inflow. The distillation consistency loss represents the difference between the fused features and the high-quality features of the teacher model.

[0013] As one implementation method, prediction is performed through a model prediction layer to obtain independent prediction values ​​for each model. Specifically, unified semantic space features are input into the model prediction layer, and predictions are performed through parallel LSTM, CNN, XGBoost, and Transformer models to generate independent prediction values ​​for the corresponding models.

[0014] As one implementation method, based on the independent predictions of each model, feature extraction and fusion are performed through a feature fusion optimization layer to obtain the fused and optimized intermediate prediction features. The specific process is as follows: Based on the independent predictions of each model, a prediction vector is established; Sparse prediction features are extracted from the prediction vector using a sparse coding mechanism. The sparse prediction features are input into the nonlinear mapping layer to obtain the fused and optimized intermediate prediction features.

[0015] As one implementation method, dynamic error weighting is performed on the output layer to obtain the final prediction result. The specific process is as follows: Calculate the average error of the corresponding model based on the independent predictions of each model; Based on the average error of the corresponding model, the weights of the corresponding model are calculated according to the inverse error principle; Based on the weights of the corresponding models, the independent predictions of each model are weighted and fused to obtain the final prediction result.

[0016] A second aspect of the present invention provides a machine learning-based rapid prediction system for tunnel water inrush, comprising: The data acquisition module is used to acquire multi-source data, which includes at least geological data, hydrological data, meteorological data, engineering process data, and monitoring data. The data processing module is used to perform spatiotemporal alignment on multi-source data and generate a unified data tensor. The feature extraction module is used to input the unified data tensor into the dynamic feature distillation network model for feature extraction, and obtain unified semantic space features. The prediction module is used to input unified semantic space features into a three-layer dynamic ensemble model for prediction and optimization to obtain the final prediction result. The three-layer dynamic ensemble model includes a model prediction layer, a feature fusion optimization layer, and an output layer. The model prediction layer performs prediction to obtain the independent prediction value of each model. Based on the independent prediction value of each model, the feature fusion optimization layer performs feature extraction and fusion to obtain the fused and optimized intermediate prediction features. The output layer performs dynamic error weighting to obtain the final prediction result.

[0017] A third aspect of the present invention provides a computer device including a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to perform the steps of the method described in the first aspect of the present invention.

[0018] A fourth aspect of the present invention is to provide a computer-readable storage medium having a program stored thereon that, when executed by a processor, performs the steps described in the first aspect of the present invention.

[0019] The above one or more technical solutions have the following beneficial effects: In this embodiment, by comprehensively collecting and fusing multimodal data such as geological data, hydrological data, meteorological data, engineering process data, and monitoring data, compared with existing methods that rely on only a single or limited data source for prediction, the comprehensiveness and accuracy of information are improved, providing richer feature inputs for machine learning models and ensuring the reliability of prediction results.

[0020] In this embodiment, multiple machine learning models such as LSTM, CNN, XGBoost, and Transformer are used for tunnel water inrush risk prediction. This allows the system to select the most suitable model for different types of data and problems, improving the overall accuracy and robustness of predictions compared to traditional single-model prediction methods and effectively reducing the bias that may be caused by a single model. Through an online model weight update mechanism, the system can automatically adjust model parameters based on real-time monitored dynamic data from the construction site, solving the problem that traditional static models cannot adapt to dynamic changes in the construction environment and achieving continuous optimization of the prediction model. By analyzing multimodal data in real time, the system can quickly identify potential water inrush risks and promptly issue alerts to the construction team, providing effective data support for construction management and improving construction safety and efficiency.

[0021] Advantages of additional aspects 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

[0022] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0023] Figure 1 This is a flowchart of a rapid prediction method for tunnel water inrush based on machine learning in this embodiment. Figure 2 This is a schematic diagram of feature extraction using the dynamic feature distillation network model in this embodiment. Detailed Implementation

[0024] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0025] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0026] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0027] Example 1 This embodiment discloses a method for rapid prediction of tunnel water inrush based on machine learning.

[0028] To more clearly illustrate this embodiment, a rapid prediction process for tunnel water inflow based on machine learning can be specifically described as follows: A machine learning-based method for rapid prediction of tunnel water inrush includes: S1. Acquire multi-source data, which includes at least geological data, hydrological data, meteorological data, engineering process data, and monitoring data; S2. Perform spatiotemporal alignment on multi-source data to generate a unified data tensor; S3. Input the unified data tensor into the dynamic feature distillation network model to extract features and obtain unified semantic space features; S4. Input the unified semantic space features into the three-layer dynamic ensemble model for prediction and optimization to obtain the final prediction result; The three-layer dynamic ensemble model includes a model prediction layer, a feature fusion optimization layer, and an output layer. The model prediction layer performs predictions to obtain independent prediction values ​​for each model. Based on the independent prediction values ​​of each model, the feature fusion optimization layer extracts and fuses features to obtain intermediate prediction features after fusion optimization. The output layer performs dynamic error weighting to obtain the final prediction result.

[0029] like Figure 1As shown, in step S1, multi-source data is acquired, which includes at least geological data, hydrological data, meteorological data, engineering process data, monitoring data, and historical water inrush records.

[0030] In this embodiment, to achieve comprehensive data collection, various types of intelligent sensors and acquisition devices are deployed at the tunnel construction site, including but not limited to ground-penetrating radar (GPR), water pressure gauges, water level gauges, piezometers, displacement gauges, stress gauges, temperature and humidity sensors, video monitoring devices, image acquisition units, tunneling parameter recorders, and blasting control terminals, forming a high-density, multi-type data sensing network.

[0031] The multi-source data collected by the aforementioned devices specifically includes: Geological data, including lithology distribution, surrounding rock grade, joint and fracture density, etc., are extracted by ground-penetrating radar, advanced geological prediction system and image recognition system; Hydrological data includes groundwater level, water pressure, water temperature, and permeability coefficient, which are collected in real time using water level gauges, water pressure gauges, and piezometers. Meteorological data: including rainfall, temperature, humidity and air pressure in the tunnel area, provided by the meteorological module and temperature and humidity sensors; Engineering process data includes construction progress, blasting parameters, shotcrete support structure and tunneling parameters, obtained from the tunneling control terminal, blasting control system and construction log; Monitoring data includes changes in surrounding rock stress, displacement, and water pressure, which are collected using stress gauges, displacement gauges, and distributed fiber optic monitoring equipment. Historical water inrush records: These include the time, intensity, location, and handling methods of water inrushes that occurred in previous tunnel sections. They are compiled from construction unit logs, manual inspection records, and historical monitoring databases.

[0032] The aforementioned multi-source data are uniformly uploaded to the edge computing device after collection for preprocessing and format conversion. Then, they proceed to the spatiotemporal alignment process in step S2, ultimately forming a unified data input tensor for dynamic feature extraction and predictive analysis.

[0033] like Figure 1 As shown, in step S2, the multi-source data is spatiotemporally aligned to generate a unified data tensor.

[0034] In actual tunnel engineering, data sources are diverse, dimensions are inconsistent, and sampling frequencies vary, resulting in significant spatiotemporal heterogeneity. Therefore, it is necessary to fuse and spatiotemporally align multi-source heterogeneous data, mapping perceptual data from different dimensions onto a unified spatiotemporal framework to ensure the timeliness and completeness of data input.

[0035] To achieve data alignment at different frequencies and times, a strategy combining sliding time window resampling and spatial interpolation algorithms is adopted. The specific process is as follows: S2-1. Use the sliding time window resampling method to align the time of multi-source data.

[0036] Set the time window length to Δt, and slide forward for τ duration each time. Define the resampling function for each type of data, as follows: ; in, For the standardized first Class data in The mean of the time window at any given moment; The original time series data (such as water pressure, rainfall, etc.); This is the width of the sliding window.

[0037] S2-2. A spatial inverse distance weighted interpolation algorithm based on tunnel cross-section coordinates is used to spatially align multi-source data to obtain a unified data tensor.

[0038] Spatial alignment is performed using an inverse spatial distance weighted interpolation (IDW) algorithm based on tunnel cross-section coordinates, as shown in the formula: ; in, Interpolation for the predicted points; The data values ​​are for known points; The Euclidean distance between the points; For powers, generally take .

[0039] Specifically, a sliding window mechanism is employed to divide multi-source data into fixed-length window segments (e.g., the past n hours) in chronological order, and all features are aligned to a unified time axis using linear interpolation, spline interpolation, or nearest neighbor methods. Simultaneously, the mapping and alignment of various spatial variables are achieved using tunnel construction progress data, monitoring point geographic coordinates, and sensor deployment maps. Missing or outlier data are imputed (e.g., using mean imputation, forward imputation, etc.), and features of different dimensions are normalized (e.g., Z-score standardization, min-max scaling, etc.), ultimately constructing a unified format multidimensional tensor. This serves as the input for subsequent models. The tensor contains a three-dimensional structure of time, space, and variable categories, comprehensively representing the evolution of the tunnel construction environment at the current moment and over a past period.

[0040] After the above steps, a consistent data tensor with aligned dimensions is constructed through spatiotemporal unified processing, providing a consistent input for subsequent deep learning modeling.

[0041] like Figure 1 , Figure 2 As shown, in step S3, the unified data tensor is input into the dynamic feature distillation network model for feature extraction to obtain unified semantic space features.

[0042] To fully extract key features from multi-source data and improve the model's generalization ability, this method introduces an innovative Multi-source Feature Distillation Network (MFDN), which integrates redundant feature compression, cross-source feature transfer, and task-related feature enhancement into a unified framework.

[0043] The Dynamic Feature Distillation Network (MFDN) model includes an original feature encoder layer, a feature distillation layer, and a residual enhancement layer.

[0044] The original feature encoder layer encodes each type of data source separately to extract primary features; the feature distillation layer extracts task-related shared features through self-attention and channel selection mechanisms; and the residual enhancer strengthens temporal variations and spatial gradient features to improve expressive power.

[0045] The specific process is as follows: (1) Input the unified data tensor into the original feature encoder layer to extract the primary features of the multi-source data respectively.

[0046] Unified data tensor Primary features of multi-source data were obtained through various encoders. The primary features of multi-source data are specifically: geological features. Hydrological characteristics Meteorological characteristics and engineering features .

[0047] (2) Input the primary features of multi-source data into the feature distillation layer, and perform weighted fusion of the initial features of multi-source data through multi-source attention mechanism and channel selection mechanism to obtain a unified feature representation after weighted fusion.

[0048] In this embodiment, a multi-source attention fusion distillation method is introduced, and the formula is: ; ; in, These are local features extracted from each data source; For linear or nonlinear converters (such as FC layers); Weights are fused to represent the features of the attention mechanism; This represents the unified feature representation after weighted fusion.

[0049] Fusion weights Dynamic control feature contribution, all The unified feature representation obtained after attention weighting is obtained through weighted fusion. This result is used in subsequent prediction layers and also for distillation alignment during the training process. This mechanism enables dynamic weighting of multi-source information fusion while filtering out noise information irrelevant to the prediction target.

[0050] (3) Input the weighted fusion unified feature representation into the residual enhancement layer, and obtain the unified semantic space features through residual blocks and skip connections.

[0051] To improve the stability and deep expressive power of features, a unified feature representation after weighted fusion is proposed. The input is fed into the residual enhancement layer, which includes residual blocks and skip connections, to form the final predicted output features. This refers to unified semantic space features. This module can employ multi-layer convolutions or gating mechanisms with residual connections, represented as: ; Here, ResidualBlock represents one or more nonlinear subnetworks with skip connections (such as the Conv1D-ReLU-BN structure), which can improve the stability of feature representation and mitigate the gradient vanishing problem.

[0052] (4) Optimize the dynamic feature distillation model using the feature loss function.

[0053] Training is performed using a dual-supervision mechanism, with the prediction loss for the main task plus the distillation consistency loss, as shown in the formula: ; ; ; in, This is a hyperparameter used to control the weight of distillation loss in the total loss; To predict the loss of the main task, it represents the standard deviation between the predicted value and the actual inflow. The distillation consistency loss represents the difference between the fused features and the high-quality features of the teacher model.

[0054] Final prediction target The accuracy of the output from the subsequent three-layer ensemble model is highly dependent on the quality of the input feature representation. The unified feature representation after weighted fusion... Feature distillation is the core of semantic understanding in a model and directly impacts prediction performance. If feature representation is distorted, even a robust ensemble model structure will struggle to achieve satisfactory prediction results. Therefore, feature distillation loss is introduced. High-quality features output by the teacher model The learning objective is to guide students to optimize the feature representation ability of their models, thereby improving the overall semantic modeling effect. This is achieved by jointly optimizing the total loss function. To achieve prediction accuracy (by Control) and characterization stability (by The dual improvement of control ultimately significantly enhances the effectiveness, stability, and generalization ability of tunnel water inrush prediction.

[0055] Specifically, to achieve real-time and high-quality water inrush prediction, a modeling approach based on dynamic feature distillation is adopted. A teacher model with strong expressive power (such as a deep Transformer or a stacked Bi-LSTM network) is constructed. The input to this model is the fused multi-source feature sequence, and the output is a high-dimensional feature representation tensor. Simultaneously, a lightweight student model (such as a CNN-LSTM hybrid network) is constructed as a real-time prediction model for actual deployment on edge devices, with the output being... By employing a feature distillation mechanism, the Euclidean distance between the feature representations of the two elements is minimized, as shown in the formula: .

[0056] To ensure that the student model can both maintain the teacher's network feature learning ability and possess good predictive ability, a total loss function is constructed: .

[0057] like Figure 1 As shown, in step S4, the unified semantic space features are input into the three-layer dynamic ensemble model for parallel prediction and optimization to obtain the final prediction result.

[0058] Considering the dynamic changes in geological conditions during tunnel construction, a single model is insufficient to adapt to all situations. In order to fully integrate the predictive capabilities of multiple models, this invention proposes a three-layer dynamic ensemble model for multi-level and multi-strategy integrated prediction of unified semantic space features extracted by Dynamic Feature Distillation Network (MFDN). The ensemble is performed in the model prediction layer, feature fusion optimization layer, and output layer, respectively. The three layers are interconnected and progressively function to form a prediction structure with strong generalization ability and dynamic adaptability.

[0059] (1) First layer: Model prediction layer (heterogeneous integration).

[0060] Four basic predictors are used: LSTM model, CNN model, XGBoost model and Transformer model.

[0061] LSTM models are used to model time series trends; CNN models extract local spatiotemporal patterns; XGBoost models capture nonlinear relationships; and Transformers model long dependencies.

[0062] Unified semantic space features are input into the model's prediction layer module, where parallel LSTM, CNN, XGBoost, and Transformer models generate independent predictions for each model. The output formula for each model is: ; in, For the first Individual models (such as LSTM, Transformer, etc.); For the first Independent predictions from each model.

[0063] All predictive sub-models are based on unified semantic space features. Predictions are performed in parallel, and multiple candidate predicted values ​​are output separately. … The prediction results of multiple sub-models are combined into a prediction vector, using the following formula: ; Where M is the number of candidate predicted values.

[0064] The above steps enable model diversity and obtain multi-faceted predictive understanding, providing information sources for subsequent fusion.

[0065] (2) Second layer: Feature fusion optimization layer.

[0066] This layer primarily addresses issues such as information redundancy, volatility differences, and low consistency that may exist in the outputs of multiple models. Through sparse coding mechanisms and nonlinear mapping structures, the outputs of multiple models are compressed, optimized, and refined to extract key prediction patterns.

[0067] 1) Sparse prediction features are extracted from the prediction vector through a sparse coding mechanism.

[0068] Using eigentransformation matrices with sparsity requirements (in The original prediction vector is compressed into a sparse representation, which yields the sparse prediction features, as shown in the formula: ; To ensure sparsity, an L1 norm constraint is introduced: ; in: It is a sparse prediction feature; Represents the prediction vector; Adjust the parameter selection range to control the sparsity threshold; It is the feature transformation matrix with sparsity requirements.

[0069] 2) Input the sparse prediction features into the nonlinear mapping layer to enhance the feature representation capability, and obtain the fused and optimized intermediate prediction features, as shown in the formula: ; in, This is the mapping weight matrix; For bias terms; This represents the intermediate features after fusion and optimization.

[0070] After the above steps, the optimized intermediate prediction features are fused. This process compresses the output of multiple models, removes redundancy and noise, and extracts stable and representative prediction features, providing a stable foundation for the weighting of the next output layer.

[0071] (3) Third layer: Output layer, used for dynamic error weighting.

[0072] This layer is responsible for dynamically and reliably modeling the fused prediction features from the previous layer and calculating the final tunnel water inflow prediction result. A sliding window error feedback mechanism is adopted to quantify the credibility of each model based on its prediction error within the recent time window, and dynamically adjust the contribution weight of each model.

[0073] 1) Calculate the model error weights.

[0074] With a window length of Within the time frame, calculate the first The average error of each model is given by the formula: ; in, Indicates the first Each prediction model at time step The predicted tunnel water inflow, i.e., the model The predicted value at a certain point in the past; This indicates the time step. The actual observed value or the true tunnel water inflow.

[0075] Furthermore, the weights of each model are defined according to the inverse error principle, using the following formula: ; 2) Output the final prediction Based on the output of the second layer The weight adjustment process guides the combination of the model predictions from the first layer according to their weights to arrive at the final predicted value. The formula is as follows: ; The final prediction result is the final predicted value. This is the final estimate of the water inflow.

[0076] After the above steps, the influence of each model in the prediction is dynamically adjusted to improve the adaptability and robustness of the model under different environments and time periods. The result is the final estimate of the inflow.

[0077] In this embodiment, a three-layer dynamic ensemble model is used. The first layer provides parallel predictions from multiple model perspectives, constructing a prediction "candidate set." The second layer optimizes these candidate predictions through sparse mapping, extracting stable key prediction features. The third layer weights the candidate predictions based on error feedback, ultimately obtaining a highly reliable prediction result. This three-layer prediction structure, consisting of "multi-perspective → feature purification → dynamic weighting," not only effectively integrates multi-source information but also possesses strong generalization ability and robustness, making it particularly suitable for the constantly changing data distribution characteristics of tunnel construction environments.

[0078] To achieve a balance between real-time performance and scalability in the prediction system, this embodiment proposes a hybrid deployment strategy of "edge-cloud collaboration".

[0079] Based on an edge-cloud collaborative computing architecture, real-time prediction and continuous optimization of the model are achieved. Specifically: (1) Edge task A lightweight Dynamic Feature Distillation Network (MFDN) model is deployed at the tunnel site and integrated with a sliding window prediction mechanism on an edge computing device (such as a control box or on-site industrial computer) to achieve real-time processing and rapid prediction of multi-source sensor data at the site.

[0080] The edge model employs a simplified CNN-LSTM hybrid structure, balancing feature extraction and time series modeling capabilities. It supports prediction response within 10 seconds, making it suitable for real-time demanding conditions in tunnel construction environments. The system also has offline operation capabilities, ensuring independent completion of water inflow prediction tasks even without a network connection.

[0081] (2) Cloud Tasks A full version of the MFDN model and multi-model integrated prediction module are deployed on the cloud server to perform more complex data analysis tasks. The cloud system receives raw data and prediction results uploaded from the edge devices daily, and retrains and updates the model's parameters to achieve continuous model optimization and improved trend analysis capabilities. Simultaneously, the cloud is responsible for managing and versioning each model version, ensuring stable and traceable model evolution.

[0082] (3) Edge-cloud collaboration mechanism To improve deployment efficiency and meet real-time requirements on-site, an edge-cloud collaborative mechanism is adopted for operation and management. At the edge, a compressed and distilled student model is deployed, and a sliding window strategy is used to continuously analyze the sensor input data stream, achieving second-level water inflow prediction. This model supports operation in offline environments, ensuring system stability and continuity. Simultaneously, a complete training architecture and integration system are deployed in the cloud, centrally performing the following tasks: integrating prediction data and raw inputs from multiple edge nodes, periodically retraining the model and optimizing parameters, and generating the latest version of the model weights.

[0083] To achieve efficient communication and hot updates between the edge and cloud, a message middleware (such as Kafka or MQTT) is built between the edge and cloud. This middleware is used for the edge to upload raw data and prediction results to the cloud. The cloud sends new model versions and control information in real time. The system manages model versions using identification codes to ensure that the models are traceable and rollbackable. This mechanism not only ensures the timeliness of model updates but also supports parallel deployment and collaborative optimization of multiple edge nodes.

[0084] Specifically, the edge and cloud communicate bidirectionally through lightweight message middleware such as Kafka or MQTT, building a highly reliable asynchronous message transmission mechanism. The cloud can push model update instructions and new version parameters to edge devices based on training results, enabling hot model updates; simultaneously, edge prediction results can be synchronized back to the cloud in real time for global situation analysis and early warning strategy adjustments. This collaborative mechanism ensures the system's adaptive and evolvable capabilities, significantly improving the reliability and intelligence of the prediction system in complex engineering environments.

[0085] Through the above steps, we can make full use of multi-source spatiotemporal data, dynamically mine key features, integrate the predictive advantages of multiple models, and achieve the accuracy, real-time performance, and scalability of predictions through edge-cloud collaboration.

[0086] This embodiment enables high-frequency monitoring, real-time prediction, and dynamic adjustment of tunnel water inrush risk. It features flexible deployment, high accuracy, and fast response, making it suitable for practical applications in tunnel engineering under complex geological conditions.

[0087] Example 2 The purpose of this embodiment is to provide a rapid prediction system for tunnel water inrush based on machine learning, including: The data acquisition module is used to acquire multi-source data, which includes at least geological data, hydrological data, meteorological data, engineering process data, and monitoring data. The data processing module is used to perform spatiotemporal alignment on multi-source data and generate a unified data tensor. The feature extraction module is used to input the unified data tensor into the dynamic feature distillation network model for feature extraction, and obtain unified semantic space features. The prediction module is used to input unified semantic space features into a three-layer dynamic ensemble model for parallel prediction and optimization to obtain the final prediction result.

[0088] This embodiment adopts a modular structure design, and the system consists of five core functional modules: a sensing and acquisition module, an edge computing module, a cloud-based intelligent analysis module, a data communication and collaboration module, and an early warning and visualization module. These modules work collaboratively through a standardized communication protocol to form a closed-loop intelligent prediction system, which can be widely applied to water inrush early warning tasks in tunnel excavation construction.

[0089] The sensing and acquisition module is deployed at the tunnel construction site to acquire multi-source data on construction and the geological environment in real time. This module integrates various sensor devices, including piezometers, water level gauges, temperature and humidity sensors, anemometers, ground-penetrating radar, video monitoring equipment, and interfaces with construction equipment. All acquisition devices are connected to edge computing nodes via a unified data bus, enabling full-cycle monitoring of multi-dimensional information such as geology, hydrology, and construction activities. The sensing and acquisition module features breakpoint resume, clock synchronization, and anomaly data identification capabilities to ensure the continuity, accuracy, and temporal consistency of the acquired data.

[0090] The edge computing module serves as the frontline decision-making unit of this system, deployed in locations such as the tunnel control box and front-end equipment rooms. It possesses low-power, highly integrated industrial edge computing capabilities. This module incorporates a student prediction model optimized by compressed distillation, employing a CNN-LSTM structure to process real-time input time-series data using a sliding window, enabling rapid prediction of future water inflow trends. The prediction results are updated every 10 seconds, providing sub-second response time. The edge computing module also features a local caching mechanism, allowing it to operate independently for at least 12 hours without network connectivity, ensuring continuous prediction support even under extreme conditions at the tunnel construction site.

[0091] The cloud-based intelligent analysis module is deployed in a remote data center or on a dedicated enterprise cloud server, undertaking the overall high-performance computing and modeling tasks of the system. This module integrates a complete MFDN dynamic feature modeling network and a multi-model ensemble prediction structure, responsible for daily aggregation of raw monitoring data and prediction results uploaded from the edge, and performing offline retraining, parameter fine-tuning, and trend analysis. The cloud system has model scheduling services and a version control system, capable of generating differentiated model parameter packages and dynamically transmitting them back to the edge for hot updates based on actual geological evolution. Furthermore, the cloud module provides a global prediction view of geological zones, offering project managers a basis for medium- and long-term risk assessment.

[0092] The data communication and collaboration module employs a highly reliable, low-latency communication protocol to achieve data interoperability between the edge and the cloud. This module integrates a Kafka / MQTT message queue mechanism to enable asynchronous, batch, and high-frequency data transmission, and supports TLS encryption and multi-device access. Edge devices upload current monitoring status, predicted values, and model running status through this module; the cloud synchronizes the latest model weights, strategy parameters, and anomaly warning commands back to the edge. This module also handles edge-cloud collaboration safeguards such as device heartbeat detection, automatic failover in case of anomalies, and data retransmission mechanisms, ensuring the overall stability and fault tolerance of the system.

[0093] The warning and visualization module serves as the system's human-computer interaction terminal, deployed in the construction management center or a portable dispatch terminal. This module visualizes current and historical water inflow predictions, hydrological change curves, model reliability heatmaps, and geological profile information, and can generate printable warning reports for project management records. When a predicted value exceeds a set threshold, the module automatically triggers a three-tiered response mechanism: Level 1 is a local audible and visual alarm; Level 2 is a push notification to the safety manager via SMS and WeChat; and Level 3 is a synchronized update to the project risk management system. The warning module supports personalized configuration, allowing construction units to set risk level classifications and trigger threshold strategies based on different working conditions.

[0094] Based on a machine learning-based rapid prediction system for tunnel water inrush, the method steps in Embodiment 1 are implemented.

[0095] The tunnel water inrush rapid prediction system provided in this embodiment achieves a fully intelligent closed loop from data acquisition to prediction and response through the coordinated operation of five modules: sensing, calculation, analysis, collaboration, and response. The system boasts advantages such as fast prediction and response, strong update capability, flexible deployment, and high adaptability, demonstrating strong engineering practicality and expansion potential in complex and ever-changing tunnel construction scenarios.

[0096] Example 3 The purpose of this embodiment is to provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the above-described method.

[0097] Example 4 The purpose of this embodiment is to provide a computer-readable storage medium.

[0098] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the steps of the above method.

[0099] Example 5 The purpose of this embodiment is to provide a computer program product containing instructions that, when run on a computer, cause the computer to perform the methods and functions involved in any of the above embodiments. The steps and methods involved in the apparatus of the above embodiments correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0100] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0101] 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 method for rapid prediction of tunnel water inflow based on machine learning, characterized in that, include: Acquire multi-source data, which includes at least geological data, hydrological data, meteorological data, engineering process data, and monitoring data; The multi-source data is spatiotemporally aligned to generate a unified data tensor; The unified data tensor is input into the dynamic feature distillation network model for feature extraction to obtain unified semantic space features. The dynamic feature distillation network model includes an original feature encoder layer, a feature distillation layer, and a residual enhancement layer. The specific process is as follows: The unified data tensor is input into the original feature encoder layer to extract primary features from multi-source data. The initial features of the multi-source data are input into the feature distillation layer. The initial features of the multi-source data are weighted and fused through a multi-source attention mechanism and a channel selection mechanism to obtain a unified feature representation after weighted fusion. The weighted and fused unified feature representation is input into the residual enhancement layer to obtain unified semantic space features; The unified semantic space features are input into a three-layer dynamic ensemble model for prediction and optimization to obtain the final prediction result. The three-layer dynamic ensemble model includes a model prediction layer, a feature fusion optimization layer, and an output layer connected in sequence. Prediction is performed through the model prediction layer to obtain the independent prediction value of each model. Based on the independent predictions of each model, feature extraction and fusion are performed through a feature fusion optimization layer to obtain the fused and optimized intermediate prediction features. The specific process is as follows: Based on the independent predictions of each model, a prediction vector is established; Sparse prediction features are extracted from the prediction vector using a sparse coding mechanism. The sparse prediction features are input into the nonlinear mapping layer to obtain the fused and optimized intermediate prediction features; The final prediction result is obtained by dynamically weighting the error in the output layer.

2. The method for rapid prediction of tunnel water inrush based on machine learning as described in claim 1, characterized in that, The spatiotemporal alignment of the multi-source data is performed as follows: The sliding time window resampling method is used to align the time of the multi-source data. A spatial inverse distance weighted interpolation algorithm based on tunnel cross-section coordinates is used to spatially align the multi-source data; Controlled imputation is performed on spatially aligned outlier data, and features of different dimensions are normalized to obtain a unified data tensor.

3. The method for rapid prediction of tunnel water inrush based on machine learning as described in claim 1, characterized in that, The dynamic feature distillation network is optimized using a dual-supervision mechanism, as shown in the following formula: in, This is a hyperparameter used to control the weight of distillation loss in the total loss; To predict the loss of the main task, it represents the standard deviation between the predicted value and the actual inflow. The distillation consistency loss represents the difference between the fused features and the high-quality features of the teacher model.

4. The method for rapid prediction of tunnel water inrush based on machine learning as described in claim 1, characterized in that, The prediction is performed through the model prediction layer to obtain the independent prediction value of each model. Specifically, the unified semantic space features are input into the model prediction layer, and prediction is performed through parallel LSTM model, CNN model, XGBoost model and Transformer model respectively to generate the independent prediction value of the corresponding model.

5. The method for rapid prediction of tunnel water inrush based on machine learning as described in claim 1, characterized in that, The final prediction result is obtained by dynamically weighting the error in the output layer. The specific process is as follows: Calculate the average error of the corresponding model based on the independent predictions of each model; Based on the average error of the corresponding model, the weights of the corresponding model are calculated according to the inverse error principle; Based on the weights of the corresponding models, the independent predictions of each model are weighted and fused to obtain the final prediction result.

6. A rapid prediction system for tunnel water inrush based on machine learning, characterized in that, Implementing a machine learning-based method for rapid prediction of tunnel water inflow as described in any one of claims 1-5, comprising: The data acquisition module is used to acquire multi-source data, which includes at least geological data, hydrological data, meteorological data, engineering process data, and monitoring data. The data processing module is used to perform spatiotemporal alignment on the multi-source data and generate a unified data tensor; The feature extraction module is used to input the unified data tensor into the dynamic feature distillation network model for feature extraction to obtain unified semantic space features; The prediction module is used to input the unified semantic space features into a three-layer dynamic ensemble model for prediction, and obtain the final prediction result. The three-layer dynamic ensemble model includes a model prediction layer, a feature fusion optimization layer and an output layer connected in sequence. The model prediction layer performs prediction to obtain the independent prediction value of each model. Based on the independent prediction value of each model, the feature fusion optimization layer performs feature extraction and fusion to obtain the fused and optimized intermediate prediction features. The output layer performs dynamic error weighting to obtain the final prediction result.

7. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method described in any one of claims 1-5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it performs the steps of the method described in any one of claims 1-5.

Images