End-side cloud integrated tunnel structure health monitoring and early warning system

Abstract

The invention relates to the technical field of tunnel engineering structure health monitoring and informatization management, in particular to an end-side cloud integrated tunnel structure health monitoring and early warning system which comprises an end-side sensing subsystem. An edge computing subsystem which is in communication connection with the end side sensing subsystem and is used for carrying out semantic analysis, feature extraction and quality scoring on received data based on a cross-layer data semantics and data consanguinity model, carrying out flow priority scheduling and congestion control on feature data according to a preset delay budget, and executing risk initial judgment and local linkage control; a cloud analysis and model governance subsystem; and a cross-layer data fusion and linkage control module. According to the method, the functions of cross-layer data semantic consistency management, risk state evolution reasoning, delay budget arrangement, federal online learning, uncertainty quantification and the like can be realized, so that the timeliness of tunnel structure disease identification, the accuracy of risk assessment and the stability of system operation are improved.

Description

Technical Field

[0001] This invention relates to the field of tunnel engineering structural health monitoring and information management technology, and in particular to an end-to-end cloud integrated tunnel structural health monitoring and early warning system. Background Technology

[0002] During long-term service, tunnel structures are susceptible to defects such as cracks, misalignments, leaks, lining deformation, and voids behind the lining due to geological conditions, construction quality, operational loads, and environmental factors. Failure to detect and address these defects in a timely manner can lead to structural performance degradation and even safety accidents. To ensure the safety of tunnel structures, structural health monitoring (SHM) systems have been widely adopted in recent years for long-term, continuous monitoring and assessment of tunnel conditions.

[0003] Currently, Chinese invention patent CN116055413B discloses a cloud-edge collaborative method for identifying anomalies in tunnel networks. This method addresses the difficulties in identifying network anomalies within tunnels under existing technologies, given the complexity of Ethernet devices within tunnels, the high sampling rate of network traffic, and the heavy pressure on transmission and servers. To achieve tunnel network anomaly identification and improve system response speed, this method involves edge computing nodes collecting and acquiring network traffic features corresponding to different services, and then extracting features from the high-dimensional traffic features using a deep autoencoder. Preliminary data processing is performed at the edge, and the feature information of the traffic data is preserved while reducing dimensionality, thereby alleviating the pressure on data transmission and sharing the task of network anomaly identification in the cloud. Furthermore, based on the characteristics of tunnel networks, improvements are made to the traditional density peak clustering algorithm in two key technologies: similarity measurement and adaptive parameter selection, achieving reasonable and effective tunnel network anomaly identification.

[0004] Existing tunnel structural health monitoring systems typically consist of multiple types of sensors, data acquisition units, transmission networks, data processing platforms, and visualization systems. With the increase in the number of sensors and sampling frequency, the monitoring data exhibits characteristics of multimodality, high frequency, and large scale, leading to the following prominent issues in engineering practice:

[0005] The data processing hierarchy is too simple and lacks coordination; some systems directly transmit all collected data to the central server or cloud for processing. Although this facilitates centralized management, it is highly dependent on network bandwidth and has a large data transmission delay. Once communication is interrupted or the delay increases, it will seriously affect the timeliness of early warning.

[0006] The edge computing capabilities are underutilized; some solutions have set up edge computing nodes at the ends of tunnels or along the tunnel for preliminary screening and compression of data, but most are limited to basic feature extraction or simple threshold judgment, lacking a collaborative mechanism with cloud-based deep analysis, and failing to fully utilize the real-time advantages of the edge.

[0007] There is a lack of integration and semantic unification of multi-source heterogeneous data. Tunnel monitoring involves multi-modal data such as stress, strain, displacement, temperature and humidity, images, and radar point clouds. Existing systems often only perform simple temporal or spatial matching when integrating data, lacking a unified data semantic model and traceability mechanism, which makes it difficult to guarantee the accuracy and traceability of the integration results.

[0008] The risk analysis and early warning mechanisms are limited; most existing systems rely on anomaly detection of single points or single types of data for early warning, lacking global risk evolution analysis across measurement points, components, and even tunnel sections, making it difficult to identify potential propagation paths and cascading risks of structural defects in a timely manner.

[0009] Insufficient model updates and adaptability: Existing monitoring systems based on machine learning or deep learning rarely perform incremental model updates in response to changes in on-site working conditions during operation, resulting in a decrease in model accuracy over time and an inability to adapt to factors such as environmental changes and load pattern changes.

[0010] The system lacks real-time and reliability metrics; in the data transmission and processing of multi-layer architectures, there is a lack of clear latency budgets and service level agreements (SLAs), as well as a lack of unified management of redundancy and self-healing mechanisms for communication links, edge nodes, and cloud platforms.

[0011] In summary, existing technologies still have shortcomings in multi-source monitoring data fusion, cross-layer collaborative analysis, risk linkage early warning, and model lifecycle governance under the edge-cloud collaborative architecture. There is an urgent need for a tunnel structure monitoring and intelligent analysis system that can achieve cross-layer data semantic consistency management, risk state evolution reasoning, online model collaborative updates, and has a clear real-time and reliability guarantee mechanism to improve the timeliness of tunnel structure defect detection, the accuracy of risk assessment, and the scientific nature of disposal decisions. Summary of the Invention

[0012] Based on the above description, the present invention provides an edge-cloud integrated tunnel structure health monitoring and early warning system, which realizes functions such as cross-layer data semantic consistency management, risk state evolution reasoning, delayed budget orchestration, federated online learning and uncertainty quantification, thereby improving the timeliness of tunnel structure defect identification, the accuracy of risk assessment and the stability of system operation.

[0013] On the one hand, the technical solution of the present invention to solve the above-mentioned technical problems is as follows: an end-edge-cloud integrated tunnel structure health monitoring and early warning system, comprising:

[0014] The end-side sensing subsystem is deployed at different monitoring sections and key components of the tunnel structure to collect multimodal raw data and control the sampling time accuracy to ±5 milliseconds through a local synchronization unit.

[0015] The edge computing subsystem is communicatively connected to the end-side sensing subsystem. It is used to perform semantic parsing, feature extraction, and quality scoring on received data based on cross-layer data semantics and data lineage models. It also performs traffic priority scheduling and congestion control on feature data according to a preset delay budget, and performs initial risk assessment and local linkage control.

[0016] The cloud-based analysis and model governance subsystem communicates with the edge computing subsystem and is used to perform causal reasoning and global risk assessment on feature data uploaded by multiple edge nodes based on the risk state evolution diagram, generate disposal strategies and distribute them to the edge side, and at the same time perform federated online learning, version canary release and consistency verification on the risk identification model on the edge side.

[0017] The cross-layer data fusion and linkage control module spans the edge, cloud, and terminal layers. It is used to realize multimodal data fusion, uncertainty quantification, and conflict evidence processing, and dynamically adjust the early warning thresholds and control strategies of each layer to meet real-time and reliability requirements.

[0018] Through the above technical solutions, it is possible to achieve full-chain coordination of tunnel monitoring from data acquisition, edge processing to cloud analysis, bringing significant technical effects. First, the end-side perception subsystem ensures the consistency of multi-source sensing data in the sampling link through a high-precision time synchronization mechanism, enabling multi-modal information such as stress, strain, images, temperature, and humidity to be compared and fused under a unified time scale, solving the analysis deviation problem caused by data asynchrony in the prior art. Second, the edge computing subsystem realizes the traceability of data sources, processing steps, and versions through a cross-layer data semantics and data lineage model, and combines a delay budget and a priority scheduling strategy to effectively alleviate network congestion, improve the real-time transmission of key features and the accuracy of initial risk judgment, so that rapid linkage control can still be achieved under communication-limited conditions. Further, the cloud analysis and model governance subsystem conducts causal relationship reasoning based on the risk state evolution diagram, can identify cascading risks across components and cross-sections, and realizes comprehensive evaluation at the global level. At the same time, a federated online learning and gray release mechanism is introduced to enable the model to adapt to changes in working conditions and be updated safely, solving the problems of model aging and accuracy decline in traditional monitoring systems. Finally, the cross-layer data fusion and linkage control module improves the credibility and robustness of the results while ensuring real-time performance through an improved fusion and uncertainty quantification strategy, can effectively reduce false alarms and missed alarms, and enhance the robustness of the system to complex environments and abnormal data. In summary, the present invention can significantly improve the real-time performance, accuracy, and reliability of tunnel structure monitoring, providing technical support for the safety management of the entire life cycle of tunnels.

[0019] Based on the above technical solutions, the present invention can also be improved as follows.

[0020] Further, the cross-layer data semantics and data lineage model includes a monitoring event object model, a data sampling benchmark, a quality scoring system, a time stamp, and version information, and calculates a comprehensive quality index through the following formula:

[0021] ;

[0022] Where, is the sampling accuracy score, is the time synchronization score, is the data integrity score, .

[0023] Further, the delay budget includes three parts: the end→edge link, the edge→cloud link, and the cloud→edge / end feedback link, and is calculated by the formula:

[0024] ;

[0025] Where, is the end-to-edge transmission delay, is the edge and cloud data processing delay, To control the delay in issuing commands, and .

[0026] Furthermore, the nodes of the risk state evolution diagram include: tunnel structure component nodes, monitoring event nodes, sensor nodes, and equipment status nodes; the edges include load-response relationships, environment-structure coupling relationships, and event causal relationships, with each edge having a corresponding weight. Timeliness parameters This is used to describe the intensity and delay characteristics of risk propagation; the cloud-based analysis and model governance subsystem constructs a causal reasoning model based on the evolutionary graph to analyze target risk events. In the collection of evidence The probability of occurrence of the following event is updated in real time according to the following formula:

[0027] ;

[0028] Among them, prior probability The likelihood function is obtained from historical monitoring data. Combining the current sensor state with dynamic adjustments to edge weights, new events uploaded by edge nodes trigger cascading updates of the probabilities of the corresponding node and its neighboring nodes; when continuous prediction results satisfy... And the risk growth rate At the same time, risk management strategies are generated in the cloud across components or tunnel sections and distributed to relevant edge nodes through the linkage control module to achieve early warning of risks and regional collaborative management.

[0029] Furthermore, the multimodal data fusion employs an improved Dempster-Shafer evidence theory, with the following fusion rules:

[0030] ;

[0031] in, For the conflict coefficient, in When the conflicting data source exceeds a preset threshold, a weight reduction process is applied.

[0032] Furthermore, the uncertainty quantification is based on the Monte Carlo Dropout method, which calculates the prediction mean by randomly deactivating neurons multiple times during the inference phase. With variance ,when When the preset threshold is exceeded, multi-source data review and manual intervention are triggered.

[0033] Furthermore, the federated online learning includes: edge nodes continuously monitoring the difference between local data distribution and historical baseline distribution during operation, and when the distribution offset index... satisfy At this time, a small-sample incremental learning process is triggered, using the current window data to update local model parameters and generate an updated model; the cloud analysis and model governance subsystem periodically receives updated model parameters and performance metadata from multiple edge nodes, based on the data volume ratio of each node. loss function of local model Perform global model parameter aggregation, with the optimization objective being:

[0034] ;

[0035] in, These are global model parameters. The number of participating nodes; after aggregation, the cloud uses a distillation optimization strategy to compress and enhance the generalization ability of the global model, and calculates the accuracy difference between the updated model and the original model on a unified test set. and inference delay difference ,when and At this point, the gray-scale release phase begins, where the new model is distributed to select edge nodes in batches for execution, and performance and consistency metrics are collected during runtime. If continuous... If the new model performs stably and does not trigger an abnormal rollback within a monitoring period, the model will be promoted to all edge nodes in batches. If the above conditions are not met, a rollback to the previous stable version will be performed, and the reasons for the model update failure and the differences in data characteristics will be recorded to optimize subsequent training and deployment strategies.

[0036] Secondly, the technical solution of the present invention to solve the above-mentioned technical problems is as follows: an edge-cloud integrated method for monitoring and early warning of tunnel structure health, comprising the following steps:

[0037] Multimodal sensors are deployed at different monitoring sections of the tunnel structure. Raw data on stress, strain, displacement, crack images, temperature and humidity, leakage and environmental conditions are collected through the end-side sensing subsystem and synchronized with high precision in time.

[0038] The raw data is transmitted to the edge computing subsystem, where semantic parsing and feature extraction are performed. Based on the data semantics and data lineage model, the data source, processing steps, model version and strategy adjustment information are recorded. The feature data is then prioritized and retransmitted in case of packet loss according to the delay budget.

[0039] At the edge, a pre-set model is used to make a preliminary risk assessment, and the feature data and assessment results are uploaded to the cloud analysis and model governance subsystem.

[0040] The cloud-based system performs causal reasoning and global risk assessment based on the risk state evolution diagram, generates cross-component or cross-tunnel segment handling strategies, and distributes them to the edge side for execution.

[0041] The cloud collects incremental update results of the model from multiple edge nodes, performs federated online learning and distillation optimization, and achieves secure model updates through canary release and consistency verification.

[0042] The cross-layer data fusion and linkage control module fuses multimodal data from the edge, cloud, and terminal to process conflict evidence and dynamically adjusts the early warning thresholds of each layer to achieve real-time closed-loop risk management.

[0043] Through the aforementioned technical solutions, the edge-cloud integrated tunnel structure monitoring data fusion and analysis method significantly improves the real-time performance and reliability of the monitoring system through the synergy of multi-source sensing, edge intelligence, and cloud governance. By deploying multimodal sensors on key tunnel sections and employing high-precision time synchronization technology, the method ensures that various data types, such as stress, strain, images, and environmental data, are aligned under a unified time scale, solving the accuracy problem caused by asynchronous multi-source data in traditional monitoring. In the edge computing stage, the method introduces cross-layer data semantics and data lineage models to record and track data sources and processing processes. Simultaneously, through delay budgeting and priority scheduling mechanisms, it ensures the priority transmission of key risk characteristics, improving the timeliness of initial risk assessment and local handling. The cloud portion conducts causal reasoning and global risk assessment based on the risk state evolution graph, enabling cascaded risk identification across components and tunnel sections, and issuing targeted handling strategies, significantly enhancing global decision-making capabilities. Furthermore, the method combines federated online learning and distillation optimization, allowing each edge node to incrementally update the model during data distribution drift, and completing aggregation and canary release in the cloud, ensuring the model maintains long-term adaptability and security. Finally, the cross-layer data fusion and linkage control module employs conflict evidence processing and uncertainty quantification techniques to effectively reduce the risks of false alarms and missed alarms caused by abnormal data, achieving closed-loop control of risk identification, early warning, and response. In summary, this method significantly improves monitoring accuracy and system stability while ensuring real-time performance, providing technical support for the safety management of tunnels throughout their entire lifecycle.

[0044] Furthermore, the delayed budget adopts a hierarchical constraint approach, including:

[0045] End-to-edge link delay This represents the time required for the edge sensing subsystem to transmit the raw data to the edge computing subsystem.

[0046] Edge-to-cloud link latency This represents the time required for the edge computing subsystem to transmit the processed data to the cloud analysis and model governance subsystem.

[0047] Cloud-to-edge / device feedback latency This indicates the time required for the cloud to distribute the processing strategy to the edge or endpoint.

[0048] The total delay budget Calculate according to the following formula:

[0049] ;

[0050] The constraints are as follows: When prioritizing feature data scheduling, the latency sensitivity coefficients of various features are used as a basis. Calculate scheduling weights The formula for its calculation is:

[0051] ;

[0052] in, The total number of feature types, A larger value indicates that the feature is more sensitive to latency, and the weight... The higher the value, the more transmission bandwidth will be allocated to ensure the real-time nature of risk closed-loop handling.

[0053] Furthermore, the cross-layer data fusion employs an improved Dempster-Shafer evidence theory for multimodal information fusion, with the following fusion rules:

[0054] ;

[0055] in, and These represent the basic probability assignment functions from different data sources. The conflict coefficient is calculated as follows:

[0056] ;

[0057] when When the (preset conflict threshold) is reached, the data sources involved in the conflict are downweighted, and the basic probability allocation function after downweighting is updated as follows:

[0058] ;

[0059] in, This is the conflict reduction factor, with a range of values. , This method, by reducing the weight of highly conflicting data sources, can reduce the interference of a single abnormal sensor on the fusion result, thereby improving the reliability and stability of the fusion decision.

[0060] Compared with the prior art, the technical solution of this application has the following beneficial technical effects:

[0061] 1. This invention achieves end-to-end management and high-precision fusion of multimodal data in tunnel structures through an integrated edge-cloud architecture. The edge sensing subsystem employs a ±5ms time synchronization mechanism to ensure that stress, strain, and image data are aligned under a unified time scale; the edge computing subsystem combines semantic parsing and delay budget scheduling to prioritize the transmission and processing of key risk features, significantly reducing the latency of risk identification; the cloud performs causal reasoning and global risk assessment based on the risk state evolution diagram, which can identify cascading risks across components or tunnel sections in advance and issue coordinated response strategies, thereby achieving rapid closed-loop risk control and improving the real-time performance and reliability of the system.

[0062] 2. This invention introduces a cross-layer data semantics and data lineage model, along with an improved DS evidence fusion method, enabling quantitative evaluation of data sampling accuracy, time synchronization, and completeness. Furthermore, it effectively reduces the interference of anomalous sensors on the fusion results through a conflict evidence weighting strategy. Simultaneously, it incorporates the Monte Carlo Dropout method to quantify uncertainty; when the prediction variance exceeds a threshold, multi-source verification and manual intervention can be triggered, thereby avoiding false positives and false negatives. Through these mechanisms, the system not only enhances the robustness and stability of multimodal information fusion but also ensures the credibility of risk identification results.

[0063] 3. This invention proposes a federated online learning and canary release mechanism, supporting edge nodes to trigger incremental learning with small samples when data distribution drifts, and uploading the updated results to the cloud for aggregation. Combined with distillation optimization, this enhances the generalization ability of the global model. During the update process, the cloud employs consistency verification and batch promotion to ensure the new model's performance is stable before full application; if the conditions are not met, it rolls back to a stable version. This mechanism guarantees the model's long-term adaptability and secure updates, enabling the monitoring system to continuously cope with the complex and ever-changing tunnel operating environment, achieving a virtuous cycle of "real-time monitoring—intelligent analysis—dynamic optimization." Attached Figure Description

[0064] Figure 1 This is a schematic diagram of the overall structure of the edge-cloud integrated tunnel structure health monitoring and early warning system according to Embodiment 1 of the present invention;

[0065] Figure 2 This is a schematic diagram of the end-side sensing subsystem of Embodiment 1 of the present invention;

[0066] Figure 3 This is a schematic diagram of the edge computing subsystem of Embodiment 1 of the present invention;

[0067] Figure 4 This is a schematic diagram of the cloud-based analysis and model governance subsystem of Embodiment 1 of the present invention;

[0068] Figure 5This is an overall schematic diagram of the end-edge-cloud integrated tunnel structure monitoring data fusion and analysis method according to Embodiment 2 of the present invention;

[0069] Figure 6 This is an overall flowchart of the end-edge-cloud integrated tunnel structure monitoring data fusion and analysis method of Embodiment 2 of the present invention. Detailed Implementation

[0070] To facilitate understanding of this application, a more complete description will be provided below with reference to the accompanying drawings, which illustrate embodiments of the present application. However, the present application can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that the disclosure of this application will be thorough and complete.

[0071] Unless otherwise defined, 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 application belongs. The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the application.

[0072] Example 1:

[0073] refer to Figure 1-4 An integrated edge-cloud tunnel structure health monitoring and early warning system includes:

[0074] The end-side sensing subsystem is deployed at different monitoring sections and key components of the tunnel structure to collect multimodal raw data such as stress, strain, displacement, crack images, temperature and humidity, leakage, vibration, gas concentration and environmental conditions. The sampling time accuracy is controlled within ±5 milliseconds through the local synchronization unit.

[0075] The edge computing subsystem is communicatively connected to the end-side sensing subsystem. It is used to perform semantic parsing, feature extraction, and quality scoring on received data based on cross-layer data semantics and data lineage models. It also performs traffic priority scheduling and congestion control on feature data according to a preset delay budget, and performs initial risk assessment and local linkage control.

[0076] The cloud-based analysis and model governance subsystem communicates with the edge computing subsystem and is used to perform causal reasoning and global risk assessment on feature data uploaded by multiple edge nodes based on the risk state evolution diagram, generate disposal strategies and distribute them to the edge side, and at the same time perform federated online learning, version canary release and consistency verification on the risk identification model on the edge side.

[0077] The cross-layer data fusion and linkage control module spans the edge, cloud, and terminal layers. It is used to realize multimodal data fusion, uncertainty quantification, and conflict evidence processing, and dynamically adjust the early warning thresholds and control strategies of each layer to meet real-time and reliability requirements.

[0078] Preferably, the end-side sensing subsystem is set at different monitoring sections and key component locations of the tunnel structure to acquire and preprocess multimodal monitoring information.

[0079] Specifically, the end-side sensing subsystem includes multiple sensor units: First, a mechanical sensor unit, including stress gauges, strain gauges, and displacement gauges, used to acquire stress, strain, and displacement changes in the tunnel surrounding rock and lining structure, respectively. The measurement accuracy of the mechanical sensors is better than 0.5%FS to ensure high-precision acquisition of structural response parameters. Second, an image acquisition unit, including a high-definition crack monitoring camera and a matching laser stripe light source, used for high-definition imaging and wide-span scanning of cracks on the tunnel lining surface. The camera has 4K resolution and an acquisition capability of no less than 25 frames per second to meet the requirements of continuity and clarity. Third, an environmental monitoring unit, including temperature and humidity sensors and anemometers, used to collect data on temperature and humidity changes and airflow within the tunnel. Fourth, a hydrological monitoring unit, including pore water pressure gauges and leakage detection devices, used to monitor groundwater pressure fluctuations and lining leakage.

[0080] To ensure the time consistency of the collected data, each node of the edge sensing subsystem is equipped with a timing unit. The timing unit performs time synchronization based on the GPS or Beidou satellite system and combines the IEEE 1588 precision clock protocol to achieve unified time synchronization of various sensor data, thereby controlling the cross-sensor sampling time deviation within ±5 milliseconds.

[0081] In addition, the edge sensing subsystem also includes a data preprocessing module for preprocessing raw data of different modalities. Specifically, it performs distortion correction, region of interest (ROI) cropping, and JPEG2000 compression on image data to improve transmission efficiency and reduce storage load; and it performs low-pass filtering on strain and displacement signals to suppress high-frequency noise interference, thereby obtaining higher-quality structural response data.

[0082] Through the above design, the edge sensing subsystem can achieve high-precision perception of the multi-dimensional state of the tunnel structure, and complete time synchronization and data preprocessing at the source end, providing standardized and highly reliable input data for the edge computing subsystem.

[0083] Preferably, the edge computing subsystem is connected to the end-side sensing subsystem for real-time analysis, quality evaluation, scheduling optimization and risk assessment of the collected multimodal monitoring data, and serves as the data and control hub between the end and the cloud.

[0084] Specifically, the edge computing subsystem includes a data parsing and feature extraction module, which performs semantic parsing and cross-modal mapping on the multimodal raw data uploaded from the edge side. It also combines cross-layer data semantics and data lineage models to generate feature vectors with data source identifiers, sampling benchmark information, and model version numbers, thereby ensuring the traceability and verifiability of the data processing process.

[0085] Furthermore, the edge computing subsystem includes a data quality scoring module, which performs a comprehensive quality assessment of the data based on sampling accuracy, time synchronization, and integrity indicators, and calculates the quality index according to the following formula:

[0086] ;

[0087] in, Score for sampling accuracy. To achieve time synchronization scoring, This is the integrity score. This quality score is used to guide subsequent data scheduling and fusion priority allocation.

[0088] The edge computing subsystem also includes a traffic scheduling and congestion control module. This module prioritizes and dynamically schedules feature data based on a preset delay budget. The delay budget consists of three parts: end-to-edge transmission delay, edge processing delay, and control feedback delay, and is expressed by the formula:

[0089] ;

[0090] Constraints are imposed to ensure that the real-time performance of the entire processing chain does not exceed 3 seconds.

[0091] In addition, the edge computing subsystem also includes a risk assessment and local linkage control module, which is used to make real-time judgments on the tunnel structure status based on feature data and a pre-set risk discrimination model, and to update the probability of newly emerging abnormal events in conjunction with a risk status evolution diagram. When the judgment result exceeds a preset risk threshold, this module can directly trigger an early warning and link the control actuators locally to achieve rapid response in the event of network latency or cloud unavailability.

[0092] Through the above design, the edge computing subsystem can achieve low-latency parsing and processing of multi-source data, taking into account data traceability, transmission efficiency and real-time risk response, thereby providing reliable input for cloud analysis and model governance, and realizing rapid closed-loop control under the collaboration of end, edge and cloud.

[0093] As the preferred option, the cloud-based analytics and model governance subsystem serves as the core decision-making and global optimization unit of the edge-cloud architecture. Its functions include global risk inference, cross-edge data aggregation, model governance, and security updates. The specific technical solutions are as follows:

[0094] Global Risk Analysis and Causal Reasoning: Based on feature data and event records uploaded from multiple edge nodes, the cloud constructs a risk state evolution diagram. The evolution diagram includes tunnel structure component nodes, monitoring event nodes, sensor nodes, and equipment status nodes. Each node is connected by edges through load-response relationships, environment-structure coupling relationships, and event causal relationships, with each edge carrying a relationship weight parameter. With timeliness parameters This is used to characterize the intensity and delay characteristics of risk propagation. The cloud utilizes causal inference algorithms to analyze target risk events. In the collection of evidence The conditional probability is updated in real time:

[0095] ;

[0096] Among them, prior probability The likelihood function is obtained based on historical monitoring data statistics. Dynamically combine real-time sensor status with relational weight adjustments. When the judgment result satisfies... And the risk growth rate At that time, the system generates risk management strategies across components or tunnel sections and distributes them to the edge computing subsystem for execution.

[0097] Model Governance and Federated Online Learning: The cloud-based analytics and model governance subsystem collects local model update parameters and performance metadata reported by multiple edge nodes. It then aggregates global parameters based on a federated online learning mechanism, optimizing the objective function as follows:

[0098] ;

[0099] in, These are global model parameters. For the first The percentage of data volume per node This is the local loss function. After aggregation, the cloud uses a distillation optimization strategy to improve model compression and generalization capabilities, and calculates the accuracy difference between the updated model and the original model on a unified test set. Difference with inference delay Perform consistency checks only if the following conditions are met. and Under these conditions, the gray-scale release phase begins.

[0100] Security Updates and Canary Releases: During canary releases, the cloud analytics and model governance subsystem distributes the new model to selected edge nodes in batches for execution, and collects runtime performance and consistency metrics. If continuous... If the new model performs stably and does not trigger an abnormal rollback within a monitoring period, then global promotion will be performed; if the conditions are not met, the model will be automatically rolled back to the previous stable version, and the reasons for the model update failure and the differences in data characteristics will be recorded for future optimization reference.

[0101] Dynamic strategy adjustment and threshold write-back: The cloud subsystem dynamically adjusts the early warning thresholds, data scheduling strategies and control parameters at the edge and end sides according to the global risk evolution trend, so as to realize cross-level closed-loop linkage control, thereby ensuring the real-time performance, robustness and reliability of the tunnel monitoring system.

[0102] As a preferred option, the cross-layer data fusion and linkage control module serves as the core bridge between the end-side sensing subsystem, the edge computing subsystem, and the cloud-based analysis and model governance subsystem. It is mainly used to realize the cross-layer fusion of multimodal monitoring data, uncertainty quantification, and conflict evidence processing, and dynamically adjust the warning thresholds and control strategies of the three layers accordingly, thereby forming a closed-loop control of the entire tunnel structure monitoring chain.

[0103] In terms of data fusion, the cross-layer data fusion and linkage control module uses an improved Dempster-Shafer evidence theory to synthesize multi-source data from the edge, cloud, and endpoints. The fusion rules are as follows:

[0104] ;

[0105] The conflict coefficient K is defined as follows:

[0106] ;

[0107] When K exceeds the preset threshold, the cross-layer data fusion and linkage control module automatically performs weight reduction processing on conflicting data sources and corrects the fusion result by combining the confidence interval compression method to ensure the stability and robustness of cross-layer fusion.

[0108] In terms of uncertainty quantification, the cross-layer data fusion and linkage control module adopts the Monte Carlo Dropout method. By randomly deactivating neural network nodes multiple times during the inference phase, the predicted mean μ and variance σ² are obtained, and the risk credibility is determined using the following conditions:

[0109]

[0110] in, The uncertainty threshold is determined through statistical analysis of historical monitoring data. This design effectively suppresses overfitting and abnormal fluctuations in model predictions, improving the overall stability and interpretability of the system.

[0111] In terms of coordinated control, the cross-layer data fusion and coordinated control module dynamically adjusts the early warning strategies of the end, edge, and cloud layers based on the cross-layer policy orchestration engine:

[0112] ① When a high-risk state is identified in the cloud, the cross-layer data fusion and linkage control module issues a cross-section or cross-component handling strategy and adjusts the sampling frequency and transmission priority on the edge side;

[0113] ② When a short-term emergency is detected at the edge but the global risk conditions have not yet been met, the cross-layer data fusion and linkage control module allows the edge to directly execute local early warning and temporary control.

[0114] ③ When data drift or failure occurs in the edge sensor, the cross-layer data fusion and linkage control module triggers the end-edge-cloud verification and compensation mechanism to ensure the continuity of the data link.

[0115] In addition, the cross-layer data fusion and linkage control module is equipped with a closed-loop strategy evaluation unit. By comparing the risk level, response delay and early warning accuracy before and after intervention, the fusion weight and control parameters are dynamically adjusted to achieve traceable continuous optimization.

[0116] Preferably, the cross-layer data fusion and linkage control module is deployed across the edge, cloud, and endpoint layers. Based on a cross-layer policy orchestration engine, it dynamically adjusts early warning strategies to ensure that risk events can be identified at multiple levels, issued tiered early warnings, and trigger linkage control. Specifically, it includes the following mechanisms:

[0117] (1) High-risk linkage of the entire cloud

[0118] When the cloud identifies a tunnel structure as being in a high-risk state (such as the global risk index) The module issues cross-section or cross-component handling strategies and dynamically adjusts the edge-side sampling frequency and transmission priority. The adaptive update of the sampling frequency can be expressed as:

[0119] ;

[0120] in, This represents the overall risk index of cloud computing. To adjust the coefficients and ensure more intensive sampling under high-risk conditions.

[0121] (2) Criteria and triggering logic for graded early warning

[0122] The system sets three tiered thresholds: edge, device, and cloud, forming a hierarchical triggering mechanism.

[0123]

[0124] in, Risk thresholds for the terminal, edge, and cloud layers, respectively. This is a hysteresis quantity used to avoid false alarms caused by frequent fluctuations.

[0125] (3) Handling of short-term emergencies on the edge

[0126] When a short-term, sudden event is detected at the edge but has not yet reached the global risk threshold, it is allowed to directly trigger a temporary early warning and execute local control measures. Edge detection can employ sequential statistics:

[0127] ;

[0128] in For the observed values, This is the historical average. For smoothing terms, To determine the threshold.

[0129] (3.1) End-to-Edge-to-Cloud Verification and Compensation

[0130] When the end-side sensor drifts or fails, the module triggers a step-by-step verification and compensation to ensure the integrity of the data link.

[0131] End-side detection: when ;

[0132] If these symptoms occur consecutively, the sensor is determined to be drifting or malfunctioning.

[0133] Edge compensation: Reconstruction using interpolation from nearby sensors. This result is then used to replace the failed data, ensuring the stability of the local fusion results.

[0134] Cloud-based review: Passed consistency score verification.

[0135] ;

[0136] If inconsistency is insufficient, a global data comparison and secondary correction will be performed.

[0137] (4) Evaluation and optimization of closed-loop strategy

[0138] After intervention, the system compares the risk reduction rate with the accuracy of the early warning:

[0139] ;

[0140] By combining response latency metrics, the module dynamically adjusts thresholds and control parameters to achieve traceable and continuous optimization.

[0141] Example 2:

[0142] refer to Figure 5-6This embodiment presents an integrated edge-cloud tunnel structure monitoring data fusion and analysis method. The tunnel structure monitoring adopts an edge-cloud collaborative architecture to address the issues of delayed response, data silos, and untimely risk handling inherent in existing single-monitoring methods. The method mainly includes the following steps:

[0143] S1 deploys multimodal sensors at different monitoring sections of the tunnel structure, and collects raw data on stress, strain, displacement, crack images, temperature and humidity, leakage and environmental conditions through the end-side sensing subsystem, and performs high-precision time synchronization.

[0144] S2 transmits the raw data to the edge computing subsystem, performs semantic parsing and feature extraction, records the data source, processing steps, model version and strategy adjustment information based on the data semantics and data lineage model, and performs priority scheduling and packet loss retransmission control on the feature data according to the delay budget.

[0145] S3 uses a preset model to make a preliminary risk assessment at the edge and uploads the feature data and assessment results to the cloud analysis and model governance subsystem.

[0146] S4 cloud performs causal reasoning and global risk assessment based on the risk state evolution diagram, generates cross-component or cross-tunnel segment handling strategies, and distributes them to the edge side for execution;

[0147] The S5 cloud collects incremental update results of the model from multiple edge nodes, performs federated online learning and distillation optimization, and achieves secure model updates through canary release and consistency verification.

[0148] The S6 cross-layer data fusion and linkage control module fuses multimodal data from the edge, cloud, and terminal to process conflict evidence and dynamically adjusts the early warning thresholds of each layer to achieve real-time closed-loop risk handling.

[0149] Preferably, multimodal sensing is synchronized with time, and multimodal sensors are deployed in the tunnel lining rings, arch, sidewalls, invert, and interface sections, including:

[0150] Mechanics: Stress gauges, strain gauges, and displacement gauges are used to collect structural stress and deformation parameters;

[0151] Image-based: A 4K high-definition crack camera combined with a laser stripe light source is used to acquire crack images and contour features;

[0152] Environmental sensors: temperature and humidity sensors, anemometers, used to reflect tunnel ventilation and microclimate environment;

[0153] Hydrological: Pore pressure gauges and leakage sensors are used to identify seepage and abnormal water pressure.

[0154] Each end node is equipped with a GPS / BeiDou timing module and uses IEEE 1588 Precision Clock Protocol (PTP) alignment to ensure that the time error of the acquired data is controlled within ±5 ms.

[0155] Preferably, in edge computing and delay budget scheduling, sensor data is preprocessed locally (image distortion correction, ROI cropping, JPEG2000 compression, stress / displacement signal low-pass filtering noise reduction) before being transmitted to the edge computing subsystem. Edge nodes perform the following operations:

[0156] Data is parsed based on semantic and data lineage models, recording the source, processing steps, and model version;

[0157] Extract structural risk-related feature vectors and perform quality scoring;

[0158] A hierarchical delay budget mechanism is introduced during transmission, with the following formula:

[0159] ;

[0160] Among them, end-to-edge link delay Edge-to-cloud link latency Cloud-to-edge / device feedback latency ,satisfy .

[0161] Bandwidth and priority are allocated to different characteristics, and the scheduling weight is determined by the latency sensitivity coefficient. calculate:

[0162] ;

[0163] in, The larger the value, the higher the real-time requirement for that feature. The larger the value, the higher the priority it is to obtain transmission resources.

[0164] As a preferred approach, the edge subsystem performs preliminary risk assessment and uploading based on its built-in risk identification model. If a sudden increase in displacement, abnormal seepage pressure, or accelerated crack propagation occurs, a temporary early warning can be triggered locally, and on-site execution units (such as reinforced support or temporary precipitation) can be activated. The assessment results and feature data are then simultaneously uploaded to the cloud.

[0165] As a preferred approach, the cloud-based global analysis and model governance subsystem centrally processes the feature data uploaded from multiple edge nodes.

[0166] Construct a risk state evolution diagram, perform causal reasoning, and identify coupling risks between different components;

[0167] Perform a global risk assessment, generate response strategies for cross-sections or tunnel sections, and distribute them to the edge side for implementation;

[0168] Collect the model update results uploaded by each edge node, and perform federated online learning and distillation optimization;

[0169] Dynamic and secure model upgrades are achieved through canary releases and consistency verification mechanisms, preventing unstable versions from affecting the entire network operation.

[0170] As a preferred approach, cross-layer data fusion and conflict evidence processing fuse data from the endpoint, edge, and cloud using an improved Dempster-Shafer evidence theory. The fusion formula is as follows:

[0171] ;

[0172] Among them, the conflict coefficient Calculated by the following formula:

[0173] ;

[0174] like If a preset threshold is reached, then conflicting data sources will be downgraded:

[0175] ;

[0176] This mechanism can reduce the interference of a single abnormal sensor on the fusion result, and improve the overall stability and accuracy of the system.

[0177] As a preferred option, the dynamic threshold adjustment and risk closed-loop management cross-layer data fusion and linkage control module dynamically adjusts the early warning thresholds of the terminal, edge, and cloud layers based on the fusion results, achieving closed-loop risk management. When a significant risk trend is detected, the system can automatically link with the tunnel construction or operation control unit to achieve intelligent prevention and emergency response.

[0178] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An integrated end-edge-cloud tunnel structure health monitoring and early warning system, characterized in that, include: The end-side sensing subsystem is deployed at different monitoring sections and key components of the tunnel structure to collect multimodal raw data and control the sampling time accuracy to ±5 milliseconds through a local synchronization unit. The edge computing subsystem is communicatively connected to the end-side sensing subsystem. It is used to perform semantic parsing, feature extraction, and quality scoring on received data based on cross-layer data semantics and data lineage models. It also performs traffic priority scheduling and congestion control on feature data according to a preset delay budget, and performs initial risk assessment and local linkage control. The cloud-based analysis and model governance subsystem communicates with the edge computing subsystem and is used to perform causal reasoning and global risk assessment on feature data uploaded by multiple edge nodes based on the risk state evolution diagram, generate disposal strategies and distribute them to the edge side, and at the same time perform federated online learning, version canary release and consistency verification on the risk identification model on the edge side. The cross-layer data fusion and linkage control module spans the edge, cloud, and terminal layers. It is used to realize multimodal data fusion, uncertainty quantification, and conflict evidence processing, and dynamically adjust the early warning thresholds and control strategies of each layer to meet real-time and reliability requirements.

2. The system according to claim 1, characterized in that, The cross-layer data semantics and data lineage model includes a monitoring event object model, data sampling benchmark, quality scoring system, timestamp and version information, and calculates the comprehensive quality index using the following formula: ； in, Score for sampling accuracy. To achieve time synchronization scoring, To score data integrity, .

3. The system according to claim 1, characterized in that, The latency budget comprises three parts: the end-to-edge link, the edge-to-cloud link, and the cloud-to-edge / end feedback link, and is expressed by the formula: ； in, For end-to-edge transmission delay, To reduce latency in edge and cloud data processing, To control the delay in issuing commands, and .

4. The system according to claim 3, characterized in that, The nodes of the risk state evolution diagram include: tunnel structure component nodes, monitoring event nodes, sensor nodes, and equipment status nodes; the edges include load-response relationships, environment-structure coupling relationships, and event causal relationships, with each edge having a corresponding weight. Timeliness parameters This is used to describe the intensity and delay characteristics of risk propagation; the cloud-based analysis and model governance subsystem constructs a causal reasoning model based on the evolutionary graph to analyze target risk events. In the collection of evidence The probability of occurrence of the following event is updated in real time according to the following formula: ； Among them, prior probability The likelihood function is obtained from historical monitoring data. Combining the current sensor state with dynamic adjustments to edge weights, new events uploaded by edge nodes trigger cascading updates of the probabilities of the corresponding node and its neighboring nodes; when continuous prediction results satisfy... And the risk growth rate At the same time, risk management strategies are generated in the cloud across components or tunnel sections and distributed to relevant edge nodes through the linkage control module to achieve early warning of risks and regional collaborative management.

5. The system according to claim 4, characterized in that, The multimodal data fusion employs an improved Dempster-Shafer evidence theory, with the following fusion rules: ； in, For the conflict coefficient, in When the conflicting data source exceeds a preset threshold, a weight reduction process is applied.

6. The system according to claim 4, characterized in that, The uncertainty quantification is based on the Monte Carlo Dropout method, which calculates the prediction mean by randomly deactivating neurons multiple times during the inference phase. With variance ,when When the preset threshold is exceeded, multi-source data review and manual intervention are triggered.

7. The system according to claim 5 or 6, characterized in that, The federated online learning includes: edge nodes continuously monitoring the difference between the local data distribution and the historical baseline distribution during operation, and when the distribution offset index... satisfy At this time, a small-sample incremental learning process is triggered, using the current window data to update local model parameters and generate an updated model; the cloud analysis and model governance subsystem periodically receives updated model parameters and performance metadata from multiple edge nodes, based on the data volume ratio of each node. loss function of local model Perform global model parameter aggregation, with the optimization objective being: ； in, These are global model parameters. The number of participating nodes; after aggregation, the cloud uses a distillation optimization strategy to compress and enhance the generalization ability of the global model, and calculates the accuracy difference between the updated model and the original model on a unified test set. and inference delay difference ,when and At this point, the gray-scale release phase begins, where the new model is distributed to select edge nodes in batches for execution, and performance and consistency metrics are collected during runtime. If continuous... If the new model performs stably and does not trigger an abnormal rollback within a monitoring period, the model will be promoted to all edge nodes in batches. If the above conditions are not met, a rollback to the previous stable version will be performed, and the reasons for the model update failure and the differences in data characteristics will be recorded to optimize subsequent training and deployment strategies.

8. A method for integrated end-edge-cloud tunnel structure health monitoring and early warning, characterized in that, Includes the following steps: Multimodal sensors are deployed at different monitoring sections of the tunnel structure. Raw data on stress, strain, displacement, crack images, temperature and humidity, leakage and environmental conditions are collected through the end-side sensing subsystem and synchronized with high precision in time. The raw data is transmitted to the edge computing subsystem, where semantic parsing and feature extraction are performed. Based on the data semantics and data lineage model, the data source, processing steps, model version and strategy adjustment information are recorded. The feature data is then prioritized and retransmitted in case of packet loss according to the delay budget. At the edge, a pre-set model is used to make a preliminary risk assessment, and the feature data and assessment results are uploaded to the cloud analysis and model governance subsystem. The cloud-based system performs causal reasoning and global risk assessment based on the risk state evolution diagram, generates cross-component or cross-tunnel segment handling strategies, and distributes them to the edge side for execution. The cloud collects incremental update results of the model from multiple edge nodes, performs federated online learning and distillation optimization, and achieves secure model updates through canary release and consistency verification. The cross-layer data fusion and linkage control module fuses multimodal data from the edge, cloud, and terminal to process conflict evidence and dynamically adjusts the early warning thresholds of each layer to achieve real-time closed-loop risk management.

9. The method according to claim 8, characterized in that, The delayed budget adopts a hierarchical constraint approach, including: End-to-edge link delay This represents the time required for the edge sensing subsystem to transmit the raw data to the edge computing subsystem. Edge-to-cloud link latency This represents the time required for the edge computing subsystem to transmit the processed data to the cloud analysis and model governance subsystem. Cloud-to-edge / device feedback latency This indicates the time required for the cloud to distribute the processing strategy to the edge or endpoint. The total delay budget Calculate according to the following formula: ; The constraints are as follows: When prioritizing feature data scheduling, the latency sensitivity coefficients of various features are used as a basis. Calculate scheduling weights The formula for its calculation is: ; in, The total number of feature types, A larger value indicates that the feature is more sensitive to latency, and the weight... The higher the value, the more transmission bandwidth will be allocated to ensure the real-time nature of risk closed-loop handling.

10. The method according to claim 8 or 9, characterized in that, The cross-layer data fusion employs an improved Dempster-Shafer evidence theory for multimodal information fusion, and its fusion rules are as follows: ; in, and These represent the basic probability assignment functions from different data sources. The conflict coefficient is calculated as follows: ; when When the (preset conflict threshold) is reached, the data sources involved in the conflict are downweighted, and the basic probability allocation function after downweighting is updated as follows: ; in, This is the conflict reduction factor, with a range of values. , This method, by reducing the weight of highly conflicting data sources, can reduce the interference of a single abnormal sensor on the fusion result, thereby improving the reliability and stability of the fusion decision.

Images