A tunnel crossing active fault zone disaster early warning system

Abstract

The application discloses a tunnel crossing active fault zone disaster early warning system and relates to the technical field of tunnel early warning.The system comprises a multi-modal sensing layer for acquiring microseismic wave, vibration, tilt and strain data along the tunnel fault zone.The multi-modal sensing layer comprises: a fiber distributed acoustic sensing unit DAS, wherein the optical fiber is arranged in a section of every 10 m along the inner side of the tunnel lining, and is used for monitoring the micro acoustic wave signals of the fracture surface in real time; a MEMS accelerometer array arranged at the four corners of each section, and used for detecting high-frequency vibration; a GNSS and INS combined tilt and strain sensor installed on the tunnel roof and side wall, and used for acquiring millimeter-level tilt and strain changes; and an acoustic cavity sensor arranged in the vicinity of the sliding surface of the fault zone, and used for capturing air acoustic precursor signals.The application realizes all-around capture of fracture microseismic, vibration and deformation functions through multi-modal fusion sensing, solves the problem of single sensing missing report and false report, improves the sensitivity by an order of magnitude, and reduces the false report rate.

Description

Technical Field

[0001] This invention relates to the field of tunnel early warning technology, specifically to a disaster early warning system for tunnels crossing active fault zones. Background Technology

[0002] Existing early warning systems for tunnel fault zones mostly rely on single sensors or centralized processing, which have the following main technical problems: the data source is singular, and it can only capture single signals in vibration or surface deformation, making it prone to missed and false alarms; communication and storage are centralized, resulting in high latency in remote processing and difficulty in preventing tampering, leading to poor security; the prediction models are simple, mostly based on static thresholds or linear regression, and cannot identify complex nonlinear fault precursors; emergency response is lagging, lacking automated multi-level decision-making and rapid response mechanisms.

[0003] Patent CN118167430B discloses a dynamic early warning system and method for rockburst in TBM tunnels based on multi-source information fusion. The patent realizes advanced classification, dynamic correction and real-time early warning of rockburst disasters in TBM tunnels, thereby avoiding rockburst disasters and improving the efficiency and safety of TBM tunneling.

[0004] The aforementioned patents provide advanced classification, dynamic correction, and real-time early warning for tunnel rockburst disasters, thereby avoiding rockburst disasters and improving the efficiency and safety of TBM tunneling. However, the reliance on a single sensor for disaster data collection within the tunnel presents a problem of limited information.

[0005] Therefore, this application proposes a disaster early warning system for tunnels crossing active fault zones that can achieve comprehensive capture of fault micro-seismic activity, vibration, and deformation. Summary of the Invention

[0006] The purpose of this invention is to provide a disaster early warning system for tunnels crossing active fault zones, in order to solve the technical problems mentioned in the background art that rely heavily on a single sensor or centralized processing.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a disaster early warning system for tunnels crossing active fault zones, the system comprising a multimodal sensing layer for acquiring microseismic waves, vibration, tilt, and strain data along the tunnel fault zone, the multimodal sensing layer comprising:

[0008] The fiber optic distributed acoustic wave sensing unit (DAS) has an optical fiber laid every 10m along the inner side of the tunnel lining for real-time monitoring of tiny acoustic wave signals on the fracture surface.

[0009] MEMS accelerometer arrays are placed at the four corners of each section to detect high-frequency vibrations;

[0010] A combined GNSS and INS tilt strain sensor is installed on the tunnel roof and sidewalls to acquire millimeter-level tilt and strain changes.

[0011] An acoustic cavity sensor is placed in the vicinity of the sliding surface of the fracture zone to capture precursor signals of airborne acoustic waves.

[0012] Preferably, the multimodal sensing layer is directly connected to an edge smart gateway via an industrial Ethernet network. The edge smart gateway includes an FPGA and an ARM Cortex-A series processor architecture, used to perform:

[0013] Noise filtering is performed using an adaptive wavelet threshold algorithm to remove environmental and construction noise.

[0014] Online extraction of multiple features including time domain, frequency domain, time-frequency domain, and entropy value;

[0015] The online transfer learning module adaptively updates the feature extraction and classification model based on the initial fracture simulation data.

[0016] Preferably, the system further includes a distributed blockchain storage module, built on Hyperledger Fabric, with the edge smart gateway serving as a blockchain network node. The distributed blockchain storage module is used for lightweight on-chain storage of feature vectors and preliminary warning results to ensure that the data is tamper-proof and fully traceable.

[0017] Preferably, the system further includes a digital twin simulation module, deployed on the central control server, comprising:

[0018] A three-dimensional tunnel-strata-fracture zone coupled model was established based on the finite element method.

[0019] The stress-displacement coupled simulation engine supports GPU acceleration and has a simulation latency of ≤1s, which is used to meet the needs of hourly rapid updates. The stress-displacement coupled simulation engine receives real-time feature input from the edge smart gateway to simulate the fracture evolution process.

[0020] Preferably, the system further includes a prediction and decision module, which integrates a graph neural network (GNN) combined with a long-term short-term network (LTS) and a network LSTM, to fuse simulation data with historical monitoring sequences and generate multi-scale disaster early warnings from level I to level IV.

[0021] Preferably, the system further includes an emergency linkage module and a remote diagnosis and maintenance module. The emergency linkage module is used to push the early warning results to the mobile supervision APP and the central control room visualization screen via 5G wireless and industrial Ethernet respectively, and automatically trigger the tunnel ventilation, lighting and escape indication system.

[0022] The emergency response module uses RESTful and MQTT dual-protocol communication, supports dual platforms on mobile devices, and generates automatic reports in standard PDF format, including warning level, simulation snapshot, and emergency response suggestions.

[0023] The remote diagnostics and maintenance module triggers the gateway's health status to be recorded on the blockchain via a smart contract, and combines this with an off-chain data analysis platform to proactively maintain and warn about the status of the gateway and sensors.

[0024] Preferably, the fiber optic distributed acoustic wave sensing unit uses φ9 / 125μm single-mode fiber, employs phase-sensitive OTDR technology, has a spatial resolution ≤0.5m, and a sampling frequency ≥10kHz.

[0025] The MEMS accelerometer array uses a triaxial ±4g range sensor, with each sensor having a bandwidth of 0.5Hz-2kHz, and is equipped with both hardware and software temperature compensation.

[0026] The combined GNSS and INS tilt strain sensor achieves a tunnel cross-section micro-deformation detection accuracy of ≤0.2mm through a four-point differential layout.

[0027] Preferably, the online transfer learning module deployed in the edge smart gateway is pre-trained based on large-scale source domain fracture simulation data and fine-tuned on newly collected data using self-supervised comparative learning; when the model performance decreases by more than 5%, a remote update is automatically triggered.

[0028] Preferably, the blockchain storage module supports dynamic node management, with a single-certificate block size ≤10KB, an on-chain latency ≤100ms, and implements hash-based notarization of large-scale feature vectors through an off-chain Merkle tree structure.

[0029] Preferably, in the prediction decision module, the graph neural network submodule is used to construct the spatial topological relationship between sensing points, and the long short-term memory network submodule is used to capture temporal features; the two are fused through an attention mechanism, and the overall model has a single prediction accuracy of ≥92%.

[0030] Compared with the prior art, the beneficial effects of the present invention are:

[0031] 1. This invention achieves comprehensive capture of fracture micro-vibrations, vibrations and deformations through multimodal fusion sensing, solving the problem of missed and false alarms by single sensing, improving sensitivity by orders of magnitude and reducing the false alarm rate;

[0032] 2. This invention uses edge smart gateways and blockchain to achieve local real-time classification decision-making and data traceability and tamper-proofing, solving the problems of high latency and security risks in centralized systems, and improving data real-time performance, data integrity and trust assurance.

[0033] 3. This invention achieves multi-scale risk prediction through stress-displacement simulation and multi-scale models, solving the problem that linear models are unable to capture complex nonlinear precursors, extending the early warning time window and improving prediction accuracy;

[0034] 4. This invention achieves multi-level early warning automatic triggering of emergency control and proactive maintenance work order generation through automatic linkage and maintenance closed loop, solving the problems of slow manual response and passive maintenance, reducing response latency and maintenance work order processing time, and improving system availability. Attached Figure Description

[0035] Figure 1 This is a schematic diagram of the disaster early warning logic flow of the present invention. Detailed Implementation

[0036] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0037] Example 1

[0038] Please see Figure 1 A disaster early warning system for tunnels crossing active fault zones, the system comprising a multimodal sensing layer for acquiring microseismic wave, vibration, tilt, and strain data along the tunnel fault zone, the multimodal sensing layer comprising:

[0039] The fiber optic distributed acoustic wave sensing unit (DAS) has an optical fiber laid every 10m along the inner side of the tunnel lining for real-time monitoring of tiny acoustic wave signals on the fracture surface.

[0040] MEMS accelerometer arrays are placed at the four corners of each section to detect high-frequency vibrations;

[0041] A combined GNSS and INS tilt strain sensor is installed on the tunnel roof and sidewalls to acquire millimeter-level tilt and strain changes.

[0042] Acoustic cavity sensors are placed in the vicinity of the sliding surface of the fracture zone to capture precursor signals of airborne sound waves;

[0043] The fiber-optic distributed acoustic wave sensing unit uses φ9 / 125μm single-mode fiber and phase-sensitive OTDR technology, with a spatial resolution ≤0.5m and a sampling frequency ≥10kHz.

[0044] The MEMS accelerometer array uses a triaxial ±4g range sensor, with each sensor having a bandwidth of 0.5Hz-2kHz, and is equipped with both hardware and software temperature compensation.

[0045] The combined GNSS and INS tilt strain sensor achieves a tunnel cross-section micro-deformation detection accuracy of ≤0.2mm through a four-point differential layout;

[0046] Furthermore, early-stage microseismic wave fusion monitoring of the fault zone:

[0047] In a mountainous highway tunnel fracture zone, a φ9 / 125μm single-mode optical fiber was laid every 10m along the inner side of the lining, and fixed to the lining sidewalls and roof. Triaxial ±4g MEMS accelerometers were installed at the four corners of each cross-section, and temperature sensors were set at the same locations for dual temperature compensation. A set of GNSS and INS combined tilt strain sensors was arranged at the tunnel roof and sidewalls, with a four-point differential arrangement, achieving a micro-deformation resolution of ≤0.2mm after calibration. Acoustic cavity sensors were suspended in the potential sliding area of ​​the fracture surface, with a frequency response range of 20Hz–20kHz.

[0048] An edge intelligent gateway is installed every 100m next to the lining and is directly connected to the aforementioned sensing unit via industrial Ethernet. The DAS unit uses phase-sensitive OTDR technology with a sampling frequency of 10kHz and a spatial resolution of 0.5m to acquire echo signals to the FPGA in real time. The FPGA runs an adaptive wavelet threshold algorithm to filter out man-made construction noise greater than 80Hz and extract features such as time-domain envelope, spectral peak, time-frequency plot, and time-series entropy. The MEMS accelerometer array data is transmitted to the gateway via the SPI bus, and the gateway also performs multiple feature extraction such as entropy and frequency band energy ratio.

[0049] The edge gateway is pre-installed with a contrastive learning model pre-trained based on a fracture simulation library to classify newly acquired features in real time. When the feature matching degree of microseismic events is ≥0.85 and the peak acceleration is >0.2g within 5 consecutive seconds, it is judged as a Level I warning (attention), and the feature vector and warning label are uploaded to the blockchain via Fabric lightweight blocks. After the central control server obtains the data uploaded to the blockchain by the edge nodes, it drives the three-dimensional digital twin model to run stress-displacement coupled simulation to simulate the possible opening of the fracture surface in the next hour. The simulation results and historical data are input into the GNN+LSTM prediction module. If the risk of Level II or above is predicted in the next hour, it is pushed to the central control screen and then to the supervision APP via 5G.

[0050] Example 2

[0051] Please see Figure 1 A disaster early warning system for tunnels crossing active fault zones, the system further includes a distributed blockchain storage module, built on Hyperledger Fabric, with the edge smart gateway serving as a blockchain network node. The distributed blockchain storage module is used for lightweight on-chain storage of feature vectors and preliminary early warning results to ensure that the data is tamper-proof and fully traceable.

[0052] Furthermore, in the implementation of multi-scale coordinated emergency response:

[0053] Based on Example 1, each edge smart gateway is registered as a Fabric network node and a dynamic node management contract is configured; the central control system connects with the local earthquake bureau's emergency platform through a RESTful interface and pre-configures emergency linkage rules.

[0054] The GNN submodule constructs a fracture point cloud map based on the spatial topology of the fractured fiber and the acceleration array; the LSTM submodule takes the entropy sequence of the past 24 hours as input and outputs the sliding trend for the next 24 hours; the two outputs are fused using an attention mechanism to generate a risk probability distribution of levels I-IV, defined as follows:

[0055] Level I (Safe): Risk <10%;

[0056] Level II (Note): 10% ≤ Risk < 30%;

[0057] Level III (Warning): 30% ≤ Risk < 60%;

[0058] Level IV (Emergency): Risk ≥ 60%;

[0059] When the warning level reaches Level III, the system automatically triggers the switching of tunnel lighting and ventilation to emergency mode via MQTT, and simultaneously generates a PDF report containing a simulation snapshot; when the warning level reaches Level IV, the escape instruction system is further activated, and an SMS and email alarm is sent to the earthquake bureau platform; the blockchain smart contract monitors the health data of the gateway on the chain, and if a node is offline for more than 30 minutes, it automatically alarms to prompt maintenance and generates a diagnostic work order on the off-chain platform.

[0060] Example 3

[0061] Please see Figure 1 A disaster early warning system for tunnels traversing active fault zones, wherein the multimodal sensing layer is directly connected to an edge intelligent gateway via industrial Ethernet, and the edge intelligent gateway includes an FPGA and an ARM Cortex-A series processor architecture for executing:

[0062] Noise filtering is performed using an adaptive wavelet threshold algorithm to remove environmental and construction noise.

[0063] Online extraction of multiple features including time domain, frequency domain, time-frequency domain, and entropy value;

[0064] The online transfer learning module adaptively updates the feature extraction and classification model based on the initial fracture simulation data.

[0065] The online transfer learning module deployed in the edge smart gateway is pre-trained based on large-scale source domain fracture simulation data and fine-tuned on newly collected data using self-supervised contrastive learning; when the model performance decreases by more than 5%, a remote update is automatically triggered.

[0066] Furthermore, each sensor data is encapsulated into UDP packets via industrial Ethernet and sent to the FPGA DMA buffer. The FPGA first performs phase demodulation and OTDR differential calculation on the DAS data and outputs a time-domain amplitude sequence. For accelerometer and acoustic cavity data, the FPGA samples the data and forwards it directly to the ARM side, where it is packaged and stored in a circular buffer queue in parallel.

[0067] The FPGA segment applies a fourth-order FIR bandpass filter to the DAS and acceleration data to remove signals outside the frequency band; the ARM side runs an adaptive wavelet thresholding algorithm module to further denoise each channel.

[0068] Divide the data into 1-second blocks each time:

[0069] 1) Wavelet decomposition up to level 5;

[0070] 2) Based on the background noise estimation, adaptively calculate the threshold λ = σ√(2lnN);

[0071] 3) Apply soft threshold shrinkage to high-frequency coefficients;

[0072] 4) The denoised signal is obtained by reconstruction.

[0073] After multiple features are extracted online, the ARM side outputs the feature vectors to the circular feature queue every 0.5 seconds for the prediction model to use.

[0074] Large-scale fracture simulation experimental data (N≈10) were used in the cloud. 6 A pre-trained contrastive learning encoder (with samples) and a lightweight fully connected network for binary classification based on feature vectors;

[0075] Self-monitoring fine-tuning:

[0076] 1. Randomly select 512 new samples from the local feature queue and construct positive and negative pairs together with the pre-training samples;

[0077] 2. Calculate the contrast loss (NT-Xent);

[0078] 3. Perform 5 gradient updates on ARM;

[0079] Edge nodes evaluate the accuracy of the local model on the validation set every hour; if the accuracy drops by more than 5% compared to the last evaluation, they communicate with the cloud model management service via gRPC to automatically pull the latest pre-trained weights and perform a seamless hot update.

[0080] Example 4

[0081] Please see Figure 1 A disaster early warning system for tunnels crossing active fault zones, the system further comprising a digital twin simulation module deployed on a central control server, including:

[0082] A three-dimensional tunnel-strata-fracture zone coupled model was established based on the finite element method.

[0083] The stress-displacement coupled simulation engine supports GPU acceleration and has a simulation latency of ≤1s, which is used to meet the needs of hourly rapid updates. The stress-displacement coupled simulation engine receives real-time feature input from the edge smart gateway to simulate the fracture evolution process.

[0084] The system also includes a prediction and decision module, which integrates a graph neural network (GNN) with a long-term short-term network (LSTM) combined architecture to fuse simulation data with historical monitoring sequences and generate multi-scale disaster early warnings from level I to IV.

[0085] Furthermore, the central control unit features digital twin simulation and multi-scale early warning decision-making processes:

[0086] Construction of a 3D tunnel-strata-fault zone coupled model: Based on the tunnel design blueprints and geological survey report, a 3D model was built in Ansys Mechanical.

[0087] Geometry: Tunnel lining, circumferential perforations, and fault zone discontinuities are constructed according to the measured strike and dip angle;

[0088] Mesh generation: The fault zone area uses locally refined tetrahedral meshes with a minimum element size of 0.1m. The lining and surrounding bottom layer meshes have a size of 0.5m, with a total of approximately 5 × 10n elements. 6 ;

[0089] Material properties: Lining - high-strength fiber reinforced concrete (elastic modulus 40GPa, Poisson's ratio 0.2), surrounding rock - layered sandstone-shale (E=10GPa-15GPa, v=0.25), fault zone filling - clay-fine sand (E=0.5GPa, v=0.3);

[0090] Contact definition: The fracture surface adopts surface-to-surface contact elements, with a friction coefficient μ=0.6, and opening and closing are allowed;

[0091] Stress-displacement coupled simulation engine:

[0092] Solution strategy: Employ implicit time-domain integral Newmark-β, β=0.25, γ=0.5;

[0093] Parallel computing: The model domain is divided into 4 subdomains, which are mapped to two GPUs respectively. Two MPIs are run on each GPU to accelerate the solution of sparse matrices.

[0094] Real-time feature input: The edge smart gateway uploads feature vectors (including 30 dimensions such as entropy, main frequency, and strain increment) every 30 seconds.

[0095] The Python service layer adjusts the load boundary conditions based on the latest features: applying equivalent normal stress Δσ and shear stress Δτ to the fracture surface;

[0096] Simulation delay: Single simulation step size Δt = 60s, total calculation time is about 1s, which meets the requirements for hour-level fast feedback;

[0097] Predictive decision-making module structure and process:

[0098] Get in real time:

[0099] Latest simulation results: Maximum open distance δ of the fracture surface max Stress concentration ratio C=σ local / σ far ;

[0100] Historical monitoring sequence: multimodal features every 30 seconds over the past 24 hours;

[0101] The two parts of data are concatenated into a feature sequence with a time window length of 48 (24h) × 1 (one sequence per hour in simulation) = 48 + 48 = 96;

[0102] GNN input graph construction:

[0103] Node: Represents each sensor substation (N≈5);

[0104] Edge weight: A weighted matrix calculated based on spatial distance and historical relevance;

[0105] Number of layers and parameters: 2-layer GraphSAGE, 64 hidden dimensions, and the aggregation function is mean;

[0106] Output: The spatial context vector Si∈R for each node 64 ;

[0107] LSTM input: concatenated GNN output S = concat i（Si） ∈R （N×64） With the time series feature matrix F∈R （96×30） ;

[0108] Structure: Frost-layer LSTM, 128 hidden units per layer, dropout=0.2;

[0109] Output: Time series context vector T∈R 128 ;

[0110] Concatenating Sflat (reduced to 128 dimensions via a fully connected layer) with T yields H∈R. 256 ;

[0111] Through a two-layer fully connected network, softmax obtains four levels of risk probability P;

[0112] The initial model was built in the cloud using labeled data (N≈10). 4 Training (covering four levels of risk samples);

[0113] After local deployment, fine-tuning is performed using newly labeled data during off-peak hours each day;

[0114] The current warning level is determined by the highest probability level L* in P.

[0115] If L* ≥ Level III, then:

[0116] The warning level, P vector, simulation snapshot and time series trend chart are packaged into JSON and pushed to the supervision APP and central control screen through dual channels of 5G and industrial Ethernet.

[0117] At the same time, the emergency response module is activated to switch the corresponding ventilation, lighting and escape instructions.

[0118] Example 5

[0119] Please see Figure 1 A disaster early warning system for tunnels crossing active fault zones, the system further includes an emergency linkage module and a remote diagnosis and maintenance module. The emergency linkage module is used to push the early warning results to the mobile supervision APP and the central control room visualization screen via 5G wireless and industrial Ethernet respectively, and automatically trigger the tunnel ventilation, lighting and escape indication system.

[0120] The emergency response module uses RESTful and MQTT dual-protocol communication, supports dual platforms on mobile devices, and generates automatic reports in standard PDF format, including warning level, simulation snapshot, and emergency response suggestions.

[0121] The remote diagnostics and maintenance module triggers the gateway's health status to be recorded on the blockchain via a blockchain smart contract, and combines it with an off-chain data analysis platform to proactively maintain and warn about the status of the gateway and sensors.

[0122] Furthermore, the edge gateway collects its own CPU / GPU utilization, memory usage, on-chain transaction latency, and sensor communication packet loss rate every 30 minutes; calculates status codes (0=normal, 1=minor anomaly, 2=major anomaly); uploads the status code and log digest (SHA-256) to the blockchain to trigger a HealthReport event; subscribes to the Fabric event stream and stores the HealthReport in a time-series database (InfluxDB); the Grafana dashboard displays the health curves of each node in real time and configures alarm rules: if any node has 3 consecutive status codes ≥1, maintenance personnel will be notified via Email and SMS.

[0123] When a single-node memory leak is detected (memory usage > 85% and lasts > 15 minutes) or on-chain transaction latency > 200ms:

[0124] Automatically generate and email work orders (including fault snapshots and log download links);

[0125] Highlight abnormal nodes on the "System Status" page of the central control screen;

[0126] An emergency drill is conducted monthly to simulate the switching between Level I to Level IV early warnings and maintenance alarms, ensuring the availability of the entire chain from linkage to maintenance; the average response time of the drill (target ≤30s) and the average processing time of maintenance work orders are statistically analyzed.

[0127] Working principle: Fiber optic DAS, MEMS accelerometers, GNN and INS tilt strain and acoustic cavity sensors are deployed along the fracture zone. The data is processed by FPGA+ARM gateway to complete bandpass filtering, adaptive wavelet denoising and multi-domain feature extraction to obtain highly sensitive and low-noise real-time monitoring information.

[0128] The edge smart gateway uses an online transfer learning module based on pre-trained and self-supervised contrastive learning to make local classification decisions on extracted features and stores important feature vectors and early warning results on the Hyperledger Fabric blockchain to ensure the real-time nature and traceability of the data.

[0129] The central control server constructs a three-dimensional finite element coupled model, and uses a GPU-accelerated stress-displacement simulation engine to simulate fracture evolution on an hourly basis. Then, it combines a GNN+LSTM multi-scale deep neural network to merge simulation and historical sequences and generate a graded early warning system of levels I-IV.

[0130] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A disaster early warning system for tunnels crossing active fault zones, characterized in that: The system includes a multimodal sensing layer for acquiring microseismic waves, vibration, tilt, and strain data along the tunnel fault zone. The multimodal sensing layer includes: The fiber optic distributed acoustic wave sensing unit (DAS) has an optical fiber laid every 10m along the inner side of the tunnel lining for real-time monitoring of tiny acoustic wave signals on the fracture surface. MEMS accelerometer arrays are placed at the four corners of each section to detect high-frequency vibrations; A combined GNSS and INS tilt strain sensor is installed on the tunnel roof and sidewalls to acquire millimeter-level tilt and strain changes. Acoustic cavity sensors are placed in the vicinity of the sliding surface of the fracture zone to capture precursor signals of airborne sound waves; The multimodal sensing layer is directly connected to the edge smart gateway via industrial Ethernet. The edge smart gateway includes an FPGA and ARM Cortex-A series processor architecture, used to execute: Noise filtering is performed using an adaptive wavelet threshold algorithm to remove environmental and construction noise. Online extraction of multiple features including time domain, frequency domain, time-frequency domain, and entropy value; The online transfer learning module adaptively updates the feature extraction and classification model based on the initial fracture simulation data; The system also includes a distributed blockchain storage module, built on Hyperledger Fabric. The edge smart gateway is the blockchain network node. The distributed blockchain storage module is used to perform lightweight on-chain storage of feature vectors and preliminary warning results to ensure that the data is tamper-proof and fully traceable. The system also includes a digital twin simulation module, deployed on the central control server, including: A three-dimensional tunnel-strata-fracture zone coupled model was established based on the finite element method. The stress-displacement coupled simulation engine supports GPU acceleration and has a simulation latency of ≤1s, which is used to meet the needs of hourly rapid updates. The stress-displacement coupled simulation engine receives real-time feature input from the edge smart gateway to simulate the fracture evolution process.

2. The disaster early warning system for tunnels crossing active fault zones according to claim 1, characterized in that: The system also includes a prediction and decision-making module, which integrates a graph neural network (GNN) with a long-term short-term and short-term network (LSTM) combined architecture to fuse simulation data with historical monitoring sequences and generate multi-scale disaster early warnings from level I to level IV.

3. The disaster early warning system for tunnels crossing active fault zones according to claim 1, characterized in that: The system also includes an emergency linkage module and a remote diagnosis and maintenance module. The emergency linkage module is used to push the early warning results to the mobile supervision APP and the central control room visualization screen via 5G wireless and industrial Ethernet respectively, and automatically trigger the tunnel ventilation, lighting and escape indication system. The emergency response module uses RESTful and MQTT dual-protocol communication, supports dual platforms on mobile devices, and generates automatic reports in standard PDF format, including warning level, simulation snapshot, and emergency response suggestions. The remote diagnostics and maintenance module triggers the gateway's health status to be recorded on the blockchain via a smart contract, and, in conjunction with an off-chain data analysis platform, proactively maintains and issues early warnings regarding the status of the gateway and sensors.

4. The disaster early warning system for tunnels crossing active fault zones according to claim 1, characterized in that: The fiber-optic distributed acoustic wave sensing unit uses φ9 / 125μm single-mode fiber and phase-sensitive OTDR technology, with a spatial resolution ≤0.5m and a sampling frequency ≥10kHz. The MEMS accelerometer array uses a triaxial ±4g range sensor, with each sensor having a bandwidth of 0.5Hz-2kHz, and is equipped with both hardware and software temperature compensation. The combined GNSS and INS tilt strain sensor achieves a tunnel cross-section micro-deformation detection accuracy of ≤0.2mm through a four-point differential layout.

5. A disaster early warning system for tunnels crossing active fault zones according to claim 1, characterized in that: The online transfer learning module deployed in the edge smart gateway is pre-trained based on large-scale source domain fracture simulation data and fine-tuned for newly collected data using self-supervised comparative learning; when the model performance decreases by more than 5%, a remote update is automatically triggered.

6. A disaster early warning system for tunnels crossing active fault zones according to claim 1, characterized in that: The blockchain storage module supports dynamic node management, with a single-certificate block size of ≤10KB and an on-chain latency of ≤100ms. It also implements hash-based evidence storage of large-scale feature vectors through an off-chain Merkle tree structure.

7. A disaster early warning system for tunnels crossing active fault zones according to claim 2, characterized in that: The graph neural network submodule in the prediction and decision module is used to construct the spatial topological relationship between sensing points, and the long short-term memory network submodule is used to capture temporal features; the two are fused through an attention mechanism, and the overall model has a single prediction accuracy of ≥92%.

Images