A method and system for anomaly identification in underground utility tunnels based on digital twins

Abstract

This invention relates to the field of intelligent operation and maintenance technology, and particularly to a method and system for anomaly identification in underground utility tunnels based on digital twins. The method includes the following steps: acquiring edge-sensing data; performing edge fusion on the edge-sensing data to obtain edge-fused data; extracting local anomaly spectra from the edge-fused data to obtain local anomaly spectrum data; performing ubiquitous heterogeneous characterization and global modal mapping on the local anomaly spectrum data to obtain global modal mapping data; identifying anomalies based on the local anomaly spectrum data and the global modal mapping data to obtain anomaly identification data; and making intelligent repair decisions based on the anomaly identification data to obtain intelligent repair decision data. This invention, through coupled analysis combining local anomaly spectra and global modal mapping, achieves the identification of anomalies from local perception to global propagation and source tracing, improving the accuracy and stability of anomaly detection. Linking the anomaly identification results with repair path planning achieves integrated closed-loop processing of identification and decision-making.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance technology, and in particular to a method and system for anomaly identification of underground utility tunnels based on digital twins. Background Technology

[0002] With the continuous expansion of urban underground utility tunnels, various pipelines such as electricity, telecommunications, water supply and drainage, and gas are concentrated in underground spaces, making the system's operating environment increasingly complex. Significant structural coupling and operational condition correlations exist among these pipelines. Once problems such as pipeline leaks, abnormal gas accumulation, ventilation system failures, or abnormal structural vibrations occur, they can often trigger chain reactions, leading to significant safety risks. Existing anomaly monitoring methods largely rely on single-sensor threshold alarms, trend predictions based on simple statistical models, and confirmation through video surveillance or manual inspections. These methods generally suffer from problems such as reliance on localized information for anomaly judgment, high false alarm and false negative rates, and a disconnect between identification results and subsequent handling procedures, making it difficult to meet the needs of refined operation and management in utility tunnel environments. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention proposes a method and system for identifying anomalies in underground utility tunnels based on digital twins, thereby resolving at least one of the aforementioned technical issues.

[0004] This application provides a method for anomaly identification in underground utility tunnels based on digital twins, comprising the following steps: Acquire edge-aware data; perform edge fusion on the edge-aware data to obtain edge-fused data; Local heterogeneous spectrum extraction is performed on edge fusion data to obtain local heterogeneous spectrum data; ubiquitous heterogeneous characterization and global source tracing are performed on local heterogeneous spectrum data to obtain global modality mapping data; Anomaly identification data is obtained by performing anomaly identification based on local anomaly spectrum data and global modal mapping data. Intelligent repair decision-making is performed based on anomaly identification data to obtain intelligent repair decision data.

[0005] This invention reduces the noise and time-delay coupling effects of raw sensing data by performing data fusion processing at the edge, ensuring structural consistency and time alignment of the data before it enters the higher-order analysis stage, thus improving the stability and real-time performance of anomaly identification from the source. Through local heterogeneous spectrum extraction, the traditional anomaly judgment based on thresholds or simple statistical biases is upgraded to a high-dimensional expression based on spectral structure, entropy increase characteristics, and nonlinear geometric morphology. This enables the system to capture weak, gradual, or cross-modal coupled anomalies, significantly reducing false alarm and false negative rates. Through ubiquitous heterogeneous characterization and global source tracing, local anomalies are mapped from the spatial and topological levels to the global modal structure, achieving structural-level analysis of anomaly sources and propagation paths. This not only determines "whether it is an anomaly" but also reveals "how the anomaly forms and spreads," enhancing the interpretability and traceability of anomaly localization. The system's anomaly identification no longer relies on single-point data judgment but combines local spectral features with global modal mapping for collaborative inference, resulting in cross-scale consistency and structural stability in the identification results. By inputting the anomaly identification results into the intelligent repair decision-making module, an automated closed loop from anomaly detection to strategy generation is achieved. This can combine historical evolution data with the current modal structure to generate the optimal treatment path, thereby improving the response efficiency and resource utilization of underground utility tunnel operation and maintenance.

[0006] Preferably, edge blending specifically includes: Heterogeneous modal tensor quantization is performed on edge-sensing data to obtain heterogeneous modal data; Transient entropy increase feature anchoring is performed on heterogeneous modal data to obtain feature anchoring data; Cross-domain correlated manifold reconstruction is performed based on feature-anchored data to obtain manifold reconstruction data; Collaborative probabilistic fusion is performed on the manifold reconstruction data to obtain edge fusion data.

[0007] This invention employs heterogeneous modal tensor quantization to embed multi-source, multi-scale, and different physical dimensions of edge-aware data into a unified high-dimensional structural space, effectively eliminating the dimensional fragmentation problem caused by traditional splicing fusion. The system uses transient entropy increase feature anchoring, enabling it to use information increments rather than absolute values ​​as the core of anomaly representation, improving sensitivity to weak perturbations and gradual anomalies. Through cross-domain correlated manifold reconstruction, the nonlinear coupling relationships and implicit evolutionary paths between different modalities are revealed, enhancing structural expressive power. A collaborative probabilistic fusion mechanism is used to integrate multimodal confidence levels, forming stable and robust edge fusion data, improving the accuracy and anti-interference capability of anomaly identification.

[0008] Preferably, the transient entropy increase feature anchoring is specifically as follows: Modal-uniform resampling is performed on heterogeneous modal data to obtain in-phase modal data; Local background stripping is performed on the in-phase mode data to obtain transient data; Transient probability configuration data is obtained by performing transient probability configuration processing on transient data; The transient entropy increase rate is calculated from the transient probability configuration data to obtain the transient entropy increase rate data; Cross-modal entropy increase consistency coupling is performed based on transient entropy increase rate data to obtain entropy increase resonance data; Anchor points are fixed based on entropy increase resonance data to obtain feature anchoring data.

[0009] This invention eliminates structural biases in sampling frequency and time drift of heterogeneous data through modal uniformity resampling and phase alignment, making different physical modes comparable. After local background stripping, the system focuses on the true disturbance components, effectively suppressing the masking effect of steady-state trends on anomaly detection. By employing transient probability configuration and entropy increase rate calculation, the original numerical changes are transformed into incremental information expressions, improving the sensitivity to non-stationary and weakly coupled anomalies. Through cross-modal entropy increase uniform coupling and anchor point solidification, precise localization of anomaly events at the temporal, modal, and structural levels is achieved, enhancing the stability and interpretability of anomaly identification.

[0010] Preferably, the local heterogeneous spectrum extraction specifically involves: Local spatial kernels are generated from the edge-fused data to obtain local spatial kernel data; Variational mode decomposition is performed on the edge fusion data based on the local spatial kernel data to obtain heterogeneous mode spectrum data; Nonlinear manifold extraction is performed based on heterogeneous modal spectral data to obtain local heterogeneous spectral data.

[0011] In this invention, local spatial kernel generation is used to introduce the topology and reachability of the utility tunnel structure into the data representation space, enabling subsequent analysis to be based on the adjacency of the real structure and avoiding spatial distortion caused by pure temporal processing. Under this constraint, variational mode decomposition is performed to decouple the structures of different frequency bands and disturbance scales, thereby improving the ability to distinguish complex anomalies. Through nonlinear manifold extraction, high-dimensional spectral features are mapped to low-dimensional geometric space, revealing the implicit coupling relationship and evolution trajectory between modes, thus forming a local anomaly spectrum with spatial consistency and dynamic interpretability, improving the accuracy and robustness of anomaly identification.

[0012] Preferably, the local spatial kernel generation specifically involves: Local neighborhood data is obtained by selecting local neighborhood data from the edge-fused data; Structural reachability kernels and operational condition isomorphic kernels are constructed from local neighborhood data to obtain structural reachability kernel data and operational condition isomorphic kernel data, respectively. Local spatial kernel data is obtained by kernel synthesis based on structural reachability kernel data and operational condition isomorphic kernel data.

[0013] This invention limits the analysis scope to real structurally related units by selecting local neighborhoods, avoiding computational redundancy and the risk of misassociations caused by global data participation. The system constructs structural reachability kernels and operational condition isomorphism kernels respectively, enabling spatial connectivity and operational state similarity to be quantitatively expressed under the same metric framework, achieving dual constraints of physical structure and dynamic operating conditions. By synthesizing kernels to form a unified local spatial kernel, not only is the accuracy and stability of neighborhood modeling improved, but high-quality input data with structural consistency is also provided for modal decomposition and anomaly spectrum extraction.

[0014] Preferably, the structural reachability kernel is constructed as follows: The local neighborhood data is used to construct a multi-level local neighborhood index, which includes spatial index construction, topological index construction, and channel index construction. Reconstruct the reachability path graph from the local neighborhood index data to obtain the reachability path graph data; Edge impedance decomposition is performed based on the reachability path graph data to obtain graph impedance data; The path reachability distance is calculated based on the graph impedance data to obtain the path reachability distance data; Accessibility stability is estimated based on path reachability clustering data to obtain accessibility stability data; Joint kernel mapping is performed based on path reachability distance data and reachability stability data to obtain structural reachability kernel data.

[0015] This invention constructs a multi-layered index to uniformly encode spatial location, topological connectivity, and channel partitioning, achieving a structured representation of local neighborhoods. Based on this, a reachable path graph is reconstructed, and combined with an edge impedance decomposition mechanism, physical distance, structural obstructions, and environmental constraints are transformed into quantifiable graph costs, effectively representing real propagable paths. By solving for path reachability distances and estimating stability, connectivity strength and path redundancy are evaluated, avoiding the risk of misjudgment caused by single path dependence. A structural reachability kernel is formed through joint kernel mapping, allowing spatial reachability relationships to be expressed in a continuous kernel space.

[0016] Preferably, ubiquitous heterogeneous characterization and global source tracing specifically refers to: The local heteromorphic spectrum data is aligned with the heteromorphic spectrum manifold to obtain manifold aligned data; The transfer entropy is calculated on the manifold aligned data to obtain the transfer entropy data; Evolutionary chain data is obtained by generating evolutionary chains from manifold alignment data based on transfer entropy data; Global modal tensor fusion is performed based on evolutionary chain data to obtain global modal mapping data.

[0017] This invention employs heterogeneous spectrum manifold alignment to embed local anomalies at different spatial locations and modal scales into a unified geometric coordinate system, eliminating cross-regional data representation bias. The system calculates propagation entropy to quantify the causal transmission relationship of anomaly disturbances over time, improving the directionality and reliability of anomaly source identification. Anomaly propagation paths are constructed through evolutionary chain generation, enabling dynamic tracing from local disturbances to global impacts. Global modal tensor fusion integrates multi-link evolution results into a unified structural mapping, enhancing anomaly propagation analysis capabilities and system-level situational awareness accuracy.

[0018] Preferably, anomaly identification specifically includes: Spatiotemporal coupling is performed based on local heterogeneous spectrum data and global modal mapping data to obtain fully local coupled data; The spectral response matrix is ​​constructed from the fully locally coupled data to obtain the spectral response matrix data; Spectral shape matching is performed on the spectral response matrix data to obtain spectral shape data; Pattern recognition is performed on the spectral morphology data to obtain anomaly identification data.

[0019] This invention achieves unified modeling between local perturbations and global structures by spatiotemporally coupling local anomaly spectra with global modal mapping, avoiding misidentification caused by isolated judgments. The system constructs a spectral response matrix, transforming multimodal evolution relationships into quantifiable response structures, improving the observability of anomalous energy accumulation and propagation characteristics. Key structural patterns are extracted through spectral morphology matching, making complex high-dimensional spectral features comparable and discriminative, and anomaly detection is completed in conjunction with pattern recognition. The process enhances cross-scale correlation analysis capabilities, improving the accuracy and stability of anomaly identification.

[0020] Preferably, the intelligent repair decision-making specifically includes: Anomaly point data is obtained by extracting anomaly points from the anomaly identification data; Repair solution data is obtained by matching the repair solution library with the anomaly data; Repair path planning is performed based on the repair plan data to obtain intelligent repair decision data.

[0021] This invention extracts anomaly points from anomaly identification data to accurately locate key risk units, avoiding large-scale blind handling. Based on this, a remediation plan library is matched, creating a structured mapping between anomaly types and historical handling experience, improving the targeting and reusability of decisions. The system integrates remediation path planning, incorporating spatial topology, resource scheduling, and operational constraints into a unified optimization framework to generate executable intelligent remediation decisions.

[0022] Preferably, this application also provides a digital twin-based underground utility tunnel anomaly identification system for performing the digital twin-based underground utility tunnel anomaly identification method described above. The digital twin-based underground utility tunnel anomaly identification system includes: The edge-aware fusion module is used to acquire edge-aware data; and to perform edge fusion on the edge-aware data to obtain edge-fused data. The edge anomaly processing module is used to extract local anomalies from edge fusion data to obtain local anomaly spectrum data; and to perform ubiquitous heterogeneous characterization and global source tracing on the local anomaly spectrum data to obtain global modality mapping data. The anomaly detection module is used to identify anomalies based on local abnormal spectrum data and global modal mapping data to obtain anomaly detection data. The intelligent repair decision module is used to make intelligent repair decisions based on anomaly identification data, and obtain intelligent repair decision data.

[0023] The beneficial effects of this invention are as follows: This invention constructs a progressively layered anomaly processing chain, forming a closed-loop technical system from data perception to decision execution. Heterogeneous data fusion and probabilistic collaborative processing are completed at the edge, reducing the noise coupling and time drift effects of the original perceived data, providing a structurally consistent input foundation for higher-order analysis. Through modal decomposition under local spatial kernel constraints and nonlinear manifold embedding, a local anomaly spectrum with spectral decoupling and geometric structure expression capabilities is constructed, achieving a refined representation of multi-scale disturbances and hidden anomalies. Global modal mapping is completed through transfer entropy and evolution chain analysis, enabling anomalies not only to be detected but also to be traced and have their propagation paths analyzed, improving anomaly localization accuracy from a structural perspective. Anomaly identification is completed through spatiotemporal coupling and spectral response matrix modeling, strengthening the ability to determine cross-regional and cross-modal consistency. Combining anomaly point extraction and path planning to generate intelligent repair decisions, the direct transformation of anomaly identification results into engineering disposal strategies is achieved. The technical solution forms a synergistic enhancement effect at the spatial topology, spectral structure, and information causality levels, improving the accuracy, interpretability, and engineering feasibility of underground utility tunnel anomaly identification and operation and maintenance decisions. Attached Figure Description

[0024] Other features, objects, and advantages of this application will become more apparent from the following detailed description of the non-limiting embodiments, taken with reference to the accompanying drawings: Figure 1 A flowchart illustrating the steps of an underground utility tunnel anomaly identification method based on digital twins is shown in one embodiment. Figure 2 A flowchart illustrating the steps of an edge blending method according to one embodiment is shown. Figure 3 A flowchart illustrating the steps of a local heterogeneous spectrum extraction method according to an embodiment is shown. Figure 4 A flowchart illustrating the steps of an anomaly identification method according to one embodiment is shown. Figure 5 A flowchart illustrating the steps of an intelligent repair decision-making method according to one embodiment is shown. Detailed Implementation

[0025] The technical method of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0026] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. Functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor methods and / or microcontroller methods.

[0027] It should be understood that although the terms "first," "second," etc., may be used herein to describe various units, these units should not be limited by these terms. These terms are used merely to distinguish one unit from another. For example, without departing from the scope of the exemplary embodiments, a first unit may be referred to as a second unit, and similarly, a second unit may be referred to as a first unit. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0028] Please see Figures 1 to 5 This application provides a method for anomaly identification in underground utility tunnels based on digital twins, comprising the following steps: S1. Acquire edge-sensing data; perform edge fusion on the edge-sensing data to obtain edge-fused data; In one embodiment, multiple types of edge sensing nodes are deployed within the underground utility tunnel to continuously collect data on the structural environment and equipment operating status. These sensing nodes include temperature and humidity sensors, gas concentration sensors (for detecting methane, hydrogen sulfide, and carbon monoxide concentrations), vibration or strain sensors, cable current-carrying monitoring units, and operating status acquisition units for power equipment such as fans and drainage pumps. Each sensing node connects to a local network via an edge gateway, which collects data frames at preset sampling intervals (e.g., 1 second or 5 seconds). Each data frame contains at least a node identifier, timestamp, sensor type, collected data value, equipment operating status, and packet loss rate information. Basic data verification processing is performed at the edge, including discarding data exceeding the sensor's range; constructing a sliding judgment window based on three consecutive sampling points to identify and correct abrupt outliers; and marking the data frame as low-confidence when the communication packet loss rate exceeds a preset threshold. After the above processing, an edge sensing data stream is formed.

[0029] During the edge fusion phase, the system performs unified representation processing of heterogeneous modal data. Different types of sensor data are organized according to a spatial node—time series—modal category structure to construct a three-dimensional data structure, ensuring that all types of sensor data are aligned under the same time axis and spatial numbering system. The system uses a robust normalization method based on median and interquartile range for normalization processing to obtain heterogeneous modal tensor data. The system extracts local disturbance sequences within a sliding time window (e.g., 30 seconds). These sequences are used as a background baseline, with the median value of several preceding steady-state windows as the background. Each sample value within the current window is subtracted from this baseline to obtain the net disturbance value reflecting short-term fluctuations. This is a time subsequence obtained by slicing heterogeneous modal tensor data by node and modality dimensions. The system then constructs the probability distribution of the data within this window. Within the current time window, the local disturbance sequences are binned with equal width or equal frequency to calculate the frequency of samples in each interval and perform frequency normalization to obtain a discrete probability distribution. The system calculates the difference between the current window distribution / probability distribution and the system's steady-state background distribution (selecting time intervals where no anomalies are triggered, equipment operation is normal, and sensor health meets standards; calculating and normalizing the frequency distribution according to the same binning rules as the current window; and using this as the reference distribution for that node and mode). Finally, the system calculates the deviation trend / distribution deviation for continuous time windows to obtain the entropy increase rate. When multiple modalities exhibit synchronous entropy increases within a short time tolerance range, they are identified as multimodal resonance feature anchors, thus outputting feature anchoring data. The system constructs a weighted adjacency relationship based on the similarity between the feature vectors of each modality (each modality has been robustly normalized to dimensionless features (such as deviation, entropy increase rate, energy proportion, etc.) before entering the similarity calculation. The similarity is for "change patterns" or "perturbation morphologies"), forming a feature similarity matrix. Based on this similarity matrix, a graph structure is constructed, and the Laplacian representation of the graph structure is calculated. By performing low-dimensional feature extraction on the graph structure (after constructing the feature similarity matrix, the graph Laplacian matrix is ​​calculated first, and then eigenvalue decomposition is performed, selecting the feature vectors corresponding to the smallest non-zero eigenvalues ​​as low-dimensional embedding coordinates), manifold embedding results that can characterize the multimodal coupling relationship are obtained, forming cross-domain associated manifold reconstruction data. Based on the health status indicators and historical operational stability of each modality, information weights are assigned to different modalities (if the sensor health is higher than a preset threshold and the packet loss rate is lower than the threshold, a basic weight is assigned; if there is slight packet loss or minor drift, the weight is reduced according to a preset level; if the health is low or the long-term drift score is high, the weight is reduced to low; if there is continuous multi-window data loss, the weight is reset to zero). The independent judgment results of each modality on abnormal events are weighted to obtain the collaborative probability assessment result / edge fusion data / edge fusion data.

[0030] S2. Extract local heterogeneous spectra from the edge fusion data to obtain local heterogeneous spectrum data; perform ubiquitous heterogeneous characterization and global source tracing on the local heterogeneous spectrum data to obtain global modality mapping data; In one embodiment, the system constructs a local spatial kernel based on edge-fused data to limit the scope of anomaly analysis under spatial structural constraints. For any spatial node to be analyzed, the system selects its multi-hop neighborhood in the utility tunnel topology as the local analysis range, where the multi-hop neighborhood represents the set of adjacent nodes within a certain number of connectivity steps. Based on the connectivity relationships and access constraints of nodes within the utility tunnel, the system calculates the structural reachability distance between nodes in the neighborhood, where the structural reachability distance is the shortest reachable path length along the topological connection. Based on the structural reachability distance, the system constructs local spatial kernel weights, thereby obtaining local spatial kernel data, such as... , For local spatial kernel weights, It is a natural exponential function. For the reachable distance of the structure, This is the scale parameter.

[0031] The system performs heterogeneous mode decomposition (HMD) on weighted time-series signals within a local area to obtain spectral structure features. Specifically, the system weights the edge-fused data according to local spatial kernel weights to form locally weighted time-series signals for each node; it then performs variational mode decomposition (VMD) on these locally weighted time-series signals, decomposing the original signal into multiple modal components with different center frequencies. During VMD, the system aims to optimize the convergence of bandwidth for each modal component and the separability between modes, updating each modal component and its center frequency parameters through an iterative process until convergence is met, thereby outputting heterogeneous modal spectral data. , To optimize the target operator, For modal indexing, For time derivative operators, For time variables, For the Dirac function, The imaginary unit, Pi is a constant. For the k-th modal component, It is a complex exponential function. The k-th mode center frequency is denoted as . The heterogeneous modal spectral data is used to describe the heterogeneous characteristics such as energy distribution, frequency drift, and superposition of multimodal disturbances in different frequency bands within a local area.

[0032] The system performs nonlinear manifold extraction on heterogeneous modal spectral data to obtain low-dimensional embedding representations. For example, the system constructs a similarity graph structure from the modal spectral features / heterogeneous modal spectral data of each node. The edge weights of the similarity graph represent the degree of spectral similarity between different nodes or different time slices. Based on the similarity graph, the system uses nonlinear dimensionality reduction methods (such as isometric mapping or diffusion mapping) to map high-dimensional spectral features to a low-dimensional coordinate space, so that the spectral differences corresponding to local anomalies form a separable structural representation in the low-dimensional space. The resulting low-dimensional embedding coordinates constitute the local heterogeneous spectral data. After completing the extraction of local heterogeneous spectra, the system performs global source tracing processing for ubiquitous heterogeneous characterization. For example, the system aligns the local manifold embedding results obtained from different regions so that the low-dimensional coordinates of different regions can be mapped to the same embedding coordinate system. The system uses the selected reference region or reference embedding (selecting the region with the most main corridor or the region with the most anomalous samples as the benchmark region; or selecting the region embedding with the most stable historical data and the highest data quality as the reference; or determining a unified reference coordinate system by global principal component direction after preliminary splicing of all region embeddings) as the benchmark, and performs rigid or similar transformation alignment on the embeddings of other regions. By minimizing the shape difference of corresponding points between regions, the embedding coordinates of each region are comparable under the same coordinate frame, thereby obtaining manifold aligned data.

[0033] The system determines the propagation direction based on manifold alignment data to identify possible transmission relationships of anomalies between different nodes or regions. Specifically, the system performs an information transmission metric on the temporal embedding sequences of candidate node pairs or candidate region pairs. This information transmission metric represents whether "the historical information of the source sequence can improve the predictive ability of future changes in the target sequence," i.e. , To transmit entropy values, For joint probability distribution, Let represent the future state of the source sequence, indicating the state value of the source sequence at the next time step. This represents the current state of the source sequence, indicating the state value of the source sequence at the current time step. This represents the current state of the target sequence, indicating the state value of the target sequence at the current time step. For logarithmic operations, The joint conditional probability represents the probability of the future state given the known history of the source and the target. The self-conditional probability represents the probability of a future state only under conditions where its own history is known. When this propagation metric exceeds a preset threshold, the system determines that there is a significant propagation direction from the source node (or source region) to the target node (or target region) and outputs propagation relationship data. The system generates anomaly evolution chains. For example, the system sorts aligned local anomaly spectrum events by time and constructs a directed propagation relationship graph based on the propagation relationship data. The system extracts path segments that satisfy temporal and directional consistency from the directed propagation relationship graph and outputs them as anomaly propagation chains to represent the evolutionary order of the anomaly from the starting node to the diffusion node, key relay nodes, and potential diffusion range, thus obtaining evolution chain data. The system uniformly encapsulates the embedding vectors of each region (low-dimensional embedding coordinates obtained after nonlinear dimensionality reduction (e.g., fixed-dimensional coordinates output by diffusion mapping or isometric mapping) with the corresponding time index, spatial index, and modality index, and associates them with their respective evolution chain identifiers or propagation path information to form global modality mapping data. The system aggregates and stacks the low-dimensional representations of different regions under the same embedded coordinate system according to spatial nodes, time series, modality categories, and regional levels to obtain global modality mapping representations / global modality mapping data.

[0034] S3. Anomaly identification is performed based on local anomaly spectrum data and global modal mapping data to obtain anomaly identification data; In one embodiment, the system performs spatiotemporal coupling processing on local anomaly spectrum data and global modal mapping data to construct a basic representation for anomaly discrimination. The system uses the spectral morphology features reflected by the local anomaly spectrum as a local perturbation expression and the cross-regional embedding structure reflected by the global modal mapping as a global correlation expression. The two are aligned and integrated according to a unified spatial node index, time series index, and modal category index to construct a coupled data structure containing multidimensional information on spatial-temporal-modal-regional correlations. The system constructs a spectral response matrix. For example, based on the global modal mapping data, the system calculates the correlation or similarity of the spectral embedding changes of any two regions within a corresponding time period, forming a representation of the response intensity between regions. Higher response intensity indicates stronger coupling or linkage characteristics between the two regions during the anomaly evolution process. By statistically integrating multiple time slices, an inter-regional spectral response matrix is ​​formed. The system performs spectral morphology matching. For example, it performs time alignment matching on the abnormal spectral change curves between candidate anomalous sequences (response trajectory sequences extracted from sub-matrices or time segment sequences related to a candidate event in the spectral response matrix. The system first selects the response intensity change curve corresponding to a certain node or evolutionary chain in the spectral response matrix, or selects a time sequence segment composed of several high-response regions, and uses this time evolution sequence as a candidate anomalous sequence. Then, it performs time alignment and morphological matching on the response trajectory, rather than directly matching the entire static matrix). This allows for some scaling and misalignment on the time axis within a certain range. By calculating the matching distance or matching cost, the system determines the degree of spectral morphological similarity between different sequences. When the matching result meets a preset threshold condition, it is determined to be a candidate event with similar anomalous morphological characteristics.

[0035] After completing spectral morphology matching, the system performs pattern recognition processing to output anomaly identification results. For example, the system inputs the aforementioned spatiotemporal coupled representation and spectral response features into an anomaly classification model. This classification model can employ a gradient boosting tree model to fully utilize the topological relationships between regions; or it can use a supervised classification model such as a support vector machine or linear support vector machine to perform discrimination training based on labeled historical anomaly samples. The model output includes information such as anomaly category identifier, anomaly occurrence node, anomaly occurrence time period, and anomaly confidence level, thus forming anomaly identification data.

[0036] S4. Make intelligent repair decisions based on anomaly identification data to obtain intelligent repair decision data.

[0037] In one embodiment, the system performs anomaly point extraction processing on the anomaly identification data. For example, based on indicators such as anomaly category identifier, anomaly occurrence node, anomaly occurrence time period, and anomaly confidence level output during the anomaly identification stage, the system ranks each anomaly node by risk. For instance, after anomaly point extraction, the system performs risk classification processing based on the anomaly category and confidence level in the anomaly identification results. Nodes judged as severely anomaly are marked as Level 1 risk; nodes with minor anomalies are marked as Level 2 risk; results in normal states or with confidence levels below a preset threshold are not included in the subsequent risk ranking. The system combines the evolution chain information corresponding to the anomaly to determine the propagation impact. When the length of the evolution chain corresponding to the anomaly exceeds a preset hop count threshold (e.g., no less than three nodes), it is marked as a "diffusion-type" anomaly; otherwise, it is marked as a "local" anomaly. The system determines the equipment importance of the anomaly node based on a pre-established list of key equipment. If the anomaly node belongs to the scope of key equipment such as the main cable section, main fan, or main drainage node, it is marked as a "critical node." The system identifies the operational safety impact of the anomaly. When anomalies involve gas concentration exceeding limits, abnormal power load, or risks across fire compartments, they are marked as "safety-sensitive" anomalies. After marking, the system uses a fixed priority for ranking and comparison: first, the system compares the severity level of the anomaly; second, it compares whether it is a critical node; third, it compares whether it is a diffusion-type anomaly; and finally, it compares whether it is a safety-sensitive anomaly. The ranking process does not involve numerical weighting or superposition calculations, but only determines the level based on the above-mentioned classification labels, thereby forming a risk priority sequence. The risk score can combine anomaly intensity, propagation chain length, degree of involvement of critical equipment, and potential impact level on operational safety. The system filters according to the risk score from high to low, selecting the top few high-risk nodes as priority targets for handling, forming a set of anomaly nodes to be repaired.

[0038] The system queries and matches pre-established repair solution rule bases based on anomaly type identifiers, associated equipment categories, and operational status characteristics contained in the anomaly identification data. The rule base includes handling strategies, operational steps, required personnel and tool information, and safety constraints corresponding to different anomaly types. The system generates candidate repair solutions based on the matching results and performs applicability checks on these solutions to ensure they meet current environmental conditions and equipment operational constraints, thereby determining the recommended repair solution for each anomaly node. The system combines the underground utility tunnel's spatial topology, access restriction information, real-time equipment operational status, and the location of the node to be repaired to construct a maintenance path search space. Within this search space, the system uses a comprehensive optimization approach, considering factors such as path length, access safety, time cost, and the degree of interference with existing operating systems. Employing heuristic search methods or multi-objective optimization strategies, it calculates the optimal path from the maintenance personnel's current location or maintenance entrance to each node to be repaired. Specifically, the system constructs a path search space based on the path map data. Impassable edges (such as closed partition doors, construction occupancy, or unauthorized passages across fire compartments) are directly marked as unreachable and excluded from the search. For traversable edges, the actual passage length is used as the base cost, with additional fixed-level penalty segments added based on the current risk zone (safe zone, restricted zone, high-risk zone). The system prioritizes path searching within the safe zone subgraph, gradually expanding to restricted or high-risk zones when unreachable edges become inaccessible. Path costs are determined by accumulating edge costs. During the search, a heuristic shortest path algorithm generates candidate paths, selecting the path with the lowest total cost as the optimal path. When multiple abnormal nodes exist, the system can generate an optimal access order based on priority. The system outputs intelligent repair decision data, which includes at least: a list of high-risk abnormal nodes, corresponding recommended repair solutions, repair execution order, and optimal repair path information.

[0039] Preferably, edge blending specifically includes: S11. Perform heterogeneous modal tensor quantization on the edge-sensing data to obtain heterogeneous modal data; In one embodiment, the edge gateway collects multi-source data such as temperature and humidity, gas concentration, vibration, current, and device operating status according to a preset sampling period (e.g., 1 second), and completes a unified field for each data record, including node identifier, timestamp, sensor type, collected value, and data quality label. For data with different sampling rates, nearest neighbor interpolation is used and a maximum allowable time offset threshold is set for time alignment. When missing values ​​exist, if the consecutive missing values ​​do not exceed a preset number of steps, forward padding is performed; if they exceed the threshold, they are left empty and the quality label is reduced simultaneously. Robust normalization based on median and interquartile range is applied to the numerical data. The data is uniformly organized according to spatial nodes, time indexes, and modal categories, and written into a three-dimensional structure. At the same time, the quality label is written into the corresponding parallel structure to form heterogeneous modal data.

[0040] S12. Perform transient entropy increase feature anchoring on heterogeneous modal data to obtain feature anchoring data; In one embodiment, a fixed-duration sliding window is used for transient perturbation analysis for each spatial node and modal sequence. Using the sliding quantile as a local background trend, the original sequence is detrended to obtain a transient net perturbation sequence. The system calculates the probability distribution of the transient net perturbation sequence and uses a mixed distribution of several previous steady-state windows (time periods in which no abnormal events were triggered during historical operation, the posterior probability of the anomaly was in the low quantile range, and the data quality was marked as valid) as a reference distribution to calculate the degree of distribution difference between the current window and the steady state. The system calculates the rate of change of differences between adjacent windows. When the rate of change of differences exceeds a threshold determined based on historical averages and fluctuation ranges multiple times consecutively, candidate anomaly anchor points are generated. At the cross-modal level, if the same node triggers candidate anomaly anchor points simultaneously in no fewer than a preset number of modalities within a set time tolerance range, the time period is confirmed as a valid feature anchor point, and the node identifier, trigger time, peak time, maximum growth rate, and participating modal set are output, forming feature anchoring data.

[0041] S13. Perform cross-domain correlation manifold reconstruction based on feature anchoring data to obtain manifold reconstruction data; In one embodiment, feature-anchored data is used as a sample set, and each anchor point is encoded as a feature vector. The vector includes at least the following information: entropy increase peak value, half-peak width, modal consistency score, local energy index, and data quality score. (The entropy increase peak value is derived from the transient entropy increase rate curve, and the maximum entropy increase rate value is taken as the peak value within the triggering segment of the judgment threshold. The half-peak width refers to the time span corresponding to the half-peak height of the entropy increase rate curve, which is obtained by finding the half-peak height on both sides of the peak. The modal consistency score is derived from the cross-modal consistency coupling calculation result, which represents the comprehensive score / normalized representation of the overlap degree of multimodal peaks within the time tolerance range. The local energy index can be obtained from the energy statistics value or frequency band energy ratio of the modal signal within the corresponding time window, and is used to characterize the local disturbance intensity. The data quality score is derived from the quality assessment results such as sensor health, packet loss rate, and effective sample ratio, which are mapped by a preset mapping table to obtain a level mapping, or one of the parameters is obtained by a preset parameter level mapping, or one of the parameters is exponentialized.) Spatial or topological identifiers are attached, such as the neighborhood number or mileage segment information. Based on the aforementioned feature vectors, a similarity graph is constructed between samples. The similarity is determined according to the degree of feature difference, and only the closest neighbors of each sample are retained to form a sparse graph structure, thereby controlling computational complexity. The system constructs a graph structure representation based on the similarity graph and extracts its main feature directions to obtain low-dimensional embedded coordinates as a manifold representation. After embedding, the system uses a shape alignment method to unify to a common coordinate system and performs calibration with the main corridor samples as a reference. The system outputs anchor point identifiers, low-dimensional coordinates, adjacency relationships, and alignment errors, forming manifold reconstruction data.

[0042] S14. Perform collaborative probabilistic fusion based on the manifold reconstruction data to obtain edge fusion data.

[0043] In one embodiment, the system performs local aggregation analysis centered on each anchor point within the manifold embedding space to evaluate the anomaly posterior probability within its neighborhood. The system assigns a base confidence weight to each modality, the weight being determined based on sensor health, communication packet loss rate, and long-term drift score, such as... , For modal basis confidence weights, It is an exponentially decaying function. For communication packet loss rate, This is a drift penalty term. For each anchor point, the posterior probability of the modality for anomalies is calculated based on the entropy enhancement, modality consistency, and data quality score of the corresponding modality. For example, the system first discretizes the entropy enhancement, modality consistency, and data quality score, mapping them to level labels (e.g., high / medium / low or 0 / 1 / 2 / 3 levels). Then, gating rules are used to determine the posterior output of the modality: when the entropy enhancement reaches a high level and the data quality is not lower than a medium level, the modality is judged to have a high posterior; when the entropy enhancement is a medium level and the consistency reaches a threshold, it is judged to have a medium posterior; otherwise, a low posterior is output or no participation is considered. This yields the discrete posterior level or corresponding probability interval for each modality (e.g., high posterior corresponds to 0.7-0.9), which is then used for subsequent steps. The system employs a confidence-weighted fusion approach. The posterior results of each modality are weighted and fused according to their confidence weights, and then normalized to obtain the anchor-level anomaly posterior probability. This anomaly posterior probability is written back to the edge temporal raster data of the corresponding time interval, adding anomaly posterior fields and risk level identifiers divided by quantile intervals to the relevant data. The low-dimensional coordinates of the manifold corresponding to the anchor point are used as a context feature cache. The system generates edge-fused data containing heterogeneous modal data, anomaly posterior probabilities, risk levels, and low-dimensional coordinates of the manifold.

[0044] Preferably, the transient entropy increase feature anchoring is specifically as follows: Modal-uniform resampling is performed on heterogeneous modal data to obtain in-phase modal data; In one embodiment, a uniform time grid (e.g., 1 second) is selected at the edge as the alignment reference, and different modal data of each node are mapped to the corresponding grid index according to the timestamp. For modalities with a sampling frequency higher than the uniform grid frequency, a representative value is generated by median aggregation within each time window; for modalities with a sampling frequency lower than the uniform grid frequency, nearest neighbor interpolation is used for padding, and a maximum allowable time offset threshold is set. Records exceeding this threshold are marked as missing and their data quality is downgraded. For modalities with systematic time delays, the most likely time offset is identified by calculating the correlation between sequences within a preset time range, and phase compensation processing is performed. The system outputs a data structure and corresponding quality tag structure organized by spatial node—time index—modal category, forming in-phase modal data. The quality tag structure is derived from the data validity judgment results during the resampling and alignment process. During the unified time grid mapping process, the system generates corresponding data quality tags for each spatial node—time index—modal category unit, including: time matching quality tag: generated based on the offset between the original timestamp and the target grid time. When using nearest neighbor interpolation, if the time offset is less than a preset threshold, it is marked as normal; if it is close to the upper limit of the threshold, it is marked as low quality; if it exceeds the threshold, it is marked as missing; sampling sufficiency tag: for high-frequency modalities, it is marked as valid when the number of effective samples in the window reaches a preset proportion, otherwise it is marked as low coverage; phase compensation consistency tag: after time delay estimation, if the correlation peak is obvious and stable, it is marked as successful compensation; if the peak is not significant, the quality level of the modality in the window is reduced.

[0045] Local background stripping is performed on the in-phase mode data to obtain transient data; In one embodiment, for in-phase mode data, a fixed-duration sliding window is used to estimate the local background value. The background value is calculated using the median within the sliding window, and slow drift noise is suppressed by low-pass filtering to obtain a stationary background sequence. The original data is subtracted from the corresponding background value to obtain the transient component sequence. To avoid step changes caused by equipment start-up and shutdown being misjudged as anomalies, a short-term freeze interval is set near the time point when a state change of a fan or water pump is detected. Transient data within this interval is only recorded and stored and does not participate in subsequent probability calculations. For missing data, if the consecutive missing data does not exceed a preset number of steps, forward padding is performed; if it exceeds a threshold, it is identified by a mask. The system outputs the transient sequence and the corresponding mask data.

[0046] Transient probability configuration data is obtained by performing transient probability configuration processing on transient data; In one embodiment, a short window of fixed duration is used to segment the transient sequence, and a differential state vector containing the current value, first-order change, and second-order change is constructed at each time point. The differential state vector within the window is subjected to vector quantization, mapped using a pre-trained offline codebook, converting continuous features into discrete category statistics, thereby forming a discrete probability distribution within the window. Referring to the distribution (a probability distribution obtained by normalizing the discrete category statistics of the same modality within a time window during historical steady-state operation, without triggered abnormal events, and with a valid data mask), the normalized category frequency result of the modality within the historical steady-state window is selected. The steady-state window is the time period during which the mask is valid and no abnormal events have been triggered. When the proportion of valid samples within the window is lower than a preset threshold, the window is marked as a low-confidence state and its quality weight is reduced (the determination of the proportion of valid samples within the window and the data quality assessment result). When the proportion of valid samples participating in the statistics within the window is lower than the preset threshold, the system reduces the confidence level of the window according to the segmentation rules based on the proportion of valid samples, for example, mapping it to a low-quality label or assigning a smaller weight coefficient. The system outputs the probability distribution, reference distribution, window identifier, node identifier, modality identifier, and effective sample ratio of the current window, forming transient probability configuration data.

[0047] The transient entropy increase rate is calculated from the transient probability configuration data to obtain the transient entropy increase rate data; In one embodiment, for each time window, the degree of distribution difference between the current probability distribution and the reference distribution is calculated, preferably using a symmetric divergence index with numerical stability. The difference values ​​obtained from consecutive windows are sequenced in chronological order, and the ratio of the change in difference between adjacent windows to the window step size is calculated as the entropy increase rate. The entropy increase rate sequence is smoothed using a three-point median filter. The threshold is determined adaptively, using the median of historical entropy increase rates as a benchmark, and adding several times the absolute median difference to form a judgment threshold. When the entropy increase rate exceeds this threshold multiple times consecutively, it is determined to be a trigger segment, and the start time, peak time, maximum value, rising slope, and half-peak width, etc., are recorded. The system outputs the entropy increase rate sequence, the judgment threshold, and the corresponding morphological parameters to form transient entropy increase rate data.

[0048] Cross-modal entropy increase consistency coupling is performed based on transient entropy increase rate data to obtain entropy increase resonance data; In one embodiment, the triggering segments of each modality are time-aligned within the same spatial node, allowing propagation delays within a preset time tolerance range. The system uses the presence of an effective entropy increase peak near the peak time for each modality as a criterion, combining corresponding data quality (data quality originates from quality assessment results in edge fusion data, including sensor signal integrity, signal-to-noise ratio, drift detection results, and data continuity markers, etc. The system statistically analyzes the above indicators (one or more can be selected) within a time window to form a quality score for each modality) and sensor health (sensor health comes from the sensor operating status monitoring module, for example, generating a health score based on self-test results, calibration status, historical fault records, and stability assessments (one or more can be selected)) to assign weights to each modality (e.g., ...). , These are the modal weighting coefficients. The mean modal data quality score is the average data quality score of the m-th modality within the current time window, derived from the sensor data quality assessment module. It is an exponentially decaying function. Calculate the consistency score based on the data packet loss rate, i.e. , For consistency scoring, For modal indexing, These are the modal weighting coefficients. This serves as a modal trigger indicator variable. When the consistency score reaches a preset threshold and contains at least a specified key modal combination (a representative modal combination under specific abnormal scenarios, such as "gas concentration + wind speed," "vibration energy + structural displacement," or "temperature + humidity," etc.), the system pre-determines that at least one of these combinations is required for a valid resonance, based on historical statistics or safety regulations, to avoid false triggering by a single mode. This is then used as an event identifier for the node to determine if an entropy-increasing resonance event has occurred. Simultaneously, the time difference between the peak values ​​of different modes is calculated. If the peak values ​​appear in a stable order, this is recorded as a propagation signature. The system outputs the event identifier, node number, peak time, consistency score, set of participating modes, and propagation signature, forming the entropy-increasing resonance data.

[0049] Anchor points are fixed based on entropy increase resonance data to obtain feature anchoring data.

[0050] In one embodiment, for each entropy increase resonance event, three types of anchor points are determined: a start point, a peak point, and an end point. The start point anchor is the center moment of the window corresponding to the first time the entropy increase rate exceeds the judgment threshold; the peak point anchor is the time position after the peak moments of each participating mode are weighted according to their respective weights; and the end point anchor is the moment after the entropy increase rate falls back below the threshold and remains below it for several consecutive windows. The system binds the event to the corresponding spatial entity, records the node number and its corresponding pipe segment or compartment information, and attaches a preset neighborhood index list to support subsequent full-domain source tracing analysis. The system generates an entropy-increasing morphological fingerprint, which includes features such as peak intensity, half-maximum width, consistency score, propagation signature, and quality mean (the statistical mean of data quality indicators of each mode or sensor node participating in the entropy-increasing resonance event, used to characterize the overall reliability of the event data). The system determines the risk level based on historical quantile intervals or key modal threshold rules (key modal threshold rules refer to preset priority judgment rules for specific anomaly types. For example, when the peak value of the gas concentration mode exceeds the safety upper limit threshold, the risk level is directly increased; when the vibration energy mode exceeds the structural safety limit, it is marked as high-risk for the structure; when the multimodal consistency score exceeds the upper limit of the preset quantile interval, the risk level is increased). The system outputs anchor point identifiers, node numbers, start and end times, morphological fingerprints, and risk levels, forming feature anchoring data.

[0051] Preferably, the local heterogeneous spectrum extraction specifically involves: S21. Generate local spatial kernels from the edge-fused data to obtain local spatial kernel data; In one embodiment, a local neighborhood is selected centered on a spatial node in the edge-fused data, based on the topological multi-hop range and the reachable distance threshold along the corridor. For node pairs within the neighborhood, the structural reachability cost is calculated on the local path graph (the system performs unreachable gate control judgment: when there is a closed partition door, a construction occupation marked as impassable, or a path crossing a fire compartment without an authorized connecting passage, the corresponding edge or path is directly assigned an unreachable mark and does not participate in the subsequent shortest path calculation. For passable edges, the actual length of the passage or the mileage difference is used as the base cost. If the access control is in a semi-open, restricted, or manual confirmation state, a fixed penalty value is added to the base cost according to a preset level to distinguish between mild, moderate, or severe accessibility). The system imposes line restrictions without weighting the calculation based on length. Simultaneously, it adds tiered penalty values ​​to edges according to the current environmental risk level. High-risk areas, in addition to additional penalties, can also be marked with an "emergency-only" flag for subsequent path constraints. The system accumulates the costs of all valid edges and obtains the minimum total path cost through shortest path search, which serves as the structural reachability cost between nodes. This structural reachability cost, combined with factors such as channel length, access control status, and current risk level, assigns an unreachable flag to impassable paths and uses shortest path search to obtain the minimum cost. Simultaneously, within a preset time window, robust standardization is applied to features such as the anomaly posterior probability, environmental gradient, and vibration energy (the environmental gradient is obtained from the temperature / humidity / gas concentration differences between adjacent nodes (or upstream and downstream sampling points of the same node) within the same time window, used to characterize the spatial variation amplitude of the environmental field; the absolute value of adjacent differences or the difference along the corridor direction can be taken; vibration energy is calculated from the amplitude sequence of vibration sensors within the time window, and the mean square value, variance, or frequency band energy statistics within the window can be used (e.g., statistically analyzing the target frequency band energy after performing frequency domain analysis on the vibration sequence), used to characterize the structural vibration intensity), and the similarity of working conditions between nodes is calculated. Attenuation weights (e.g., based on structural reachability and working condition similarity) are constructed respectively. , As structural accessibility attenuation weight, It is a natural exponential function. The structural reachability cost is the minimum travel cost from node i to node j; if unreachable, it is marked as unreachable. The reachability cost scale parameter (used for normalization) The attenuation scale is taken from the neighborhood. (Median level and other robust statistics). For isomorphic attenuation weights under operating conditions, The cost of operating condition differences (the degree of difference in operating condition characteristics between node i and node j within the time window). State difference scaling parameter (used for normalization) The decay scale is combined with the mean quality of node data for weighted synthesis to form the kernel weight, i.e. , To synthesize kernel weights, This is a square root function used to calculate the joint confidence modulation factor by taking the geometric mean of the confidence levels of two nodes. For nodes The corresponding node credibility / quality mean (node ​​data credibility obtained from quality tags and packet loss, etc.). For nodes The corresponding node credibility / quality mean (the credibility of node data obtained from quality labels and packet loss, etc.). Only a few neighbors with the highest weight ranking are retained to construct a sparse local spatial kernel structure. The system outputs the neighborhood set (elements in the sparse local spatial kernel structure), sparse kernel weights (kernel weights corresponding to the few neighbors with the highest weight ranking) and related parameters, forming local spatial kernel data.

[0052] S22. Perform variational mode decomposition on the edge fusion data based on the local spatial kernel data to obtain heterogeneous mode spectrum data; In one embodiment, within a local neighborhood, the fusion sequences of the same mode at each node are weighted and aggregated according to the local spatial kernel weight to obtain a local representative signal, i.e. , For local representative signals, with nodes Centered on a specific point, the time-series signal obtained by spatial kernel weighted aggregation within a local neighborhood is used to represent the fusion response of that neighborhood. For local neighborhood The node index that participates in the aggregation. For a local neighborhood set, with the central node With this as the core, a set of neighboring nodes is selected through topological and distance rules. For local spatial kernel weights, nodes With nodes Spatial correlation strength between nodes, used for modulation nodes The degree of contribution to the central signal, For node fusion sequence, nodes In time The system performs variational mode decomposition (VMD) on the local signal, pre-setting the number of modes to cover the main frequency band and setting bandwidth constraint parameters to control mode convergence characteristics. Iterative optimization is used to solve for each modal component and its center frequency, stopping when the relative error meets a preset threshold or the maximum number of iterations is reached. To avoid over-decomposition, modes with energy proportions lower than a preset ratio are merged. The system outputs the time-domain components of each mode (the amplitude spectrum characteristics of each mode are obtained by frequency domain transformation of the decomposed time-domain components), corresponding center frequencies, amplitude spectrum characteristics, and their energy proportions (the energy proportions are obtained by calculating the ratio of the energy of each modal component to the total energy), and stability indices (the stability indices can be statistically obtained by statistically analyzing the changes in the center frequency or energy proportion of the same mode within adjacent time windows, indicating whether the modal structure is stable), forming heterogeneous modal spectrum data.

[0053] S23. Nonlinear manifold extraction is performed based on heterogeneous modal spectrum data to obtain local heterogeneous spectrum data.

[0054] In one embodiment, heterogeneous modal spectral data is used as input, and spectral fingerprint features are extracted within a fixed time window. These spectral fingerprint features include the center frequency of each mode (obtained from the dominant frequency position of each modal component in the frequency domain or frequency parameters output by the decomposition algorithm), bandwidth (determined by the effective energy distribution range in the modal spectrum, such as the frequency span of the statistical energy concentration interval), spectral peak shape indicators (obtained by analyzing the peak height, sharpness, or symmetry of the modal spectral curve), energy proportion (calculated from the proportion of energy of each modal time-domain component to the total energy), and the degree of cross-modal energy transfer (obtained by comparing the changing trends of energy proportions of different modes within adjacent time windows). The temporal variation is then superimposed (the temporal variation is obtained by differential calculation of the same spectral fingerprint feature within adjacent time windows, such as calculating the center frequency change amplitude, energy proportion change rate, or bandwidth change between the current window and the previous window; alternatively, a sliding window slope estimation method can be used to obtain the trend of feature evolution over time). A nearest-neighbor similarity graph is constructed based on the spectral fingerprint features, retaining only mutually nearest neighbor sample pairs to enhance stability, and edge weights are assigned according to the degree of feature differences. The system employs nonlinear dimensionality reduction (such as diffusion mapping or isomap) to obtain low-dimensional embedding representations, while simultaneously calculating geometric quantities such as the local curvature and tangent vector of the embedding trajectory. When the curvature increases or the trajectory deviates from the historical mainstream clustering region, it is identified as anomaly. The system encapsulates the low-dimensional coordinates, geometric indicators, corresponding time windows, and spectral fingerprint features to form locally anomalous spectral data.

[0055] Preferably, the local spatial kernel generation specifically involves: Local neighborhood data is obtained by selecting local neighborhood data from the edge-fused data; In one embodiment, spatial units in the edge-fused data are used as central nodes, and local neighborhoods are selected based on both topological relationships and corridor mileage. The system selects neighboring nodes within a preset multi-hop range in the corridor connectivity map and removes nodes that cross fire compartments or have closed partition doors; a distance threshold is set based on the reachable distance along the corridor centerline, prioritizing the retention of nodes with consistent channel indices. If the number of neighboring nodes exceeds the upper limit after filtering, the system sorts nodes based on recent anomaly posterior similarity (the degree of consistency in the trend and magnitude of the anomaly posterior probability sequences of two nodes within a preset recent time window (e.g., 30 seconds or 1 minute). The system reads the anomaly posterior probability and risk level sequences of each node from the edge-fused data within this time window, performs time alignment and standardization on the sequences, and then calculates the similarity index between the sequences of the two nodes, such as using the inverse vector degree of the mean squared error or the correlation coefficient) and spatial distance, retaining the nodes with higher rankings; if the number of nodes is below the lower limit, the topological range is appropriately expanded to make up for the difference. The system outputs the central node identifier, a list of neighboring nodes, connection relationships, and the element information of the pipe segment to which they belong, forming local neighborhood data.

[0056] Structural reachability kernels and operational condition isomorphic kernels are constructed from local neighborhood data to obtain structural reachability kernel data and operational condition isomorphic kernel data, respectively. In one embodiment, the system constructs a structural reachability kernel. Based on local neighborhood data, a path graph is reconstructed, and edge weights are set as a comprehensive impedance. This impedance is determined by the channel length, the gate or construction occupancy status, and the environmental risk level. For example, if the partition door corresponding to an edge is closed, or the construction occupancy is marked as impassable, the edge is directly assigned an inaccessible label and does not participate in the shortest path search. For passable edges, the channel length is used as the basic cost, which can be directly taken as meters or the difference in mileage, forming the basic passage cost. When the gate is half-open, restricted, or requires manual confirmation, the length is not weighted and superimposed; instead, a fixed penalty segment is added according to a preset level. For example, the status penalty is divided into three levels: "mild," "moderate," and "severe," with different fixed costs added to reflect the increased passage resistance. Fixed penalty segments are added according to the environmental risk level (such as safe zone / restricted zone / high-risk zone or corresponding over-limit level). When in a high-risk zone, in addition to adding penalties, a "emergency passage only" label can be triggered. Impassable edges are assigned an inaccessible label. The minimum reachability cost between nodes is calculated through shortest path search, and scaled according to the median level of the cost distribution within the neighborhood. The reachability cost is then mapped to decaying weights, resulting in the structural reachability kernel. , For structural reachability kernel weights, The starting node within the local neighborhood. For the target node within the local neighborhood, It is an exponentially decaying mapping function. To minimize reachability cost, from node in the path graph To the node The minimum travel cost is obtained by shortest path search; if the path is unreachable, it is marked as unreachable. The reachability cost scale parameter.

[0057] The system constructs a condition isomorphism kernel. Within a fixed time window, features such as environmental gradient, gas change rate, vibration frequency band energy, and anomalous posterior mean are extracted and standardized using a method based on median and interquartile range. The system calculates the degree of condition difference between nodes and performs scale normalization based on the median level of the difference distribution within the neighborhood, mapping the degree of difference to similarity weights to form the condition isomorphism kernel. , For isomorphic kernel weights under operating conditions, The starting node within the local neighborhood. For the target node within the local neighborhood, It is an exponentially decaying mapping function. The cost of different operating conditions For the state difference scale parameter, The parameters for scale normalization are taken from the neighborhood. The median level or its robust statistics. The system outputs sparse adjacency relation data for the two types of kernels.

[0058] Local spatial kernel data is obtained by kernel synthesis based on structural reachability kernel data and operational condition isomorphic kernel data.

[0059] In one embodiment, credibility modulation is applied to any pair of nodes within the same neighborhood. Node credibility is determined by a combination of data quality markers and packet loss data, and the joint value of the credibility of the two nodes is used as a modulation factor to suppress the impact of low-quality data on kernel weights. The system synthesizes a structural reachability kernel and a condition isomorphism kernel. , For synthesizing kernel weight coefficients, nodes With nodes Spatial correlation strength between them For structural reachability kernel weights, The starting node within the local neighborhood. For the target node within the local neighborhood, The structural weight index coefficient, For isomorphic kernel weights under operating conditions, The coefficient of the working condition weighting index. The joint credibility factor reflects spatial accessibility constraints and operational similarity constraints, respectively. When the number of structural blocking events (such as door state switching or construction occupancy) within the most recent preset time window exceeds 1.5 times the historical median, [the following factor will be considered]. Increase to between 1.5 and 2; when the operating condition fluctuation index (such as the variance of the abnormal posterior probability or the amplitude of energy fluctuation) exceeds the preset quantile threshold, Increase the value to between 1.5 and 2 to strengthen state consistency constraints; limit and The upper limit is no more than 2, and the lower limit is no less than 0.5. During synthesis, structural weight coefficients and operational condition weight coefficients are set, initially with equal weights. When recent structural blockages are frequent, the structural weight is increased to strengthen reachability constraints; when operational condition fluctuations are significant, the isomorphic weight is increased to enhance state consistency constraints. After synthesis, for each central node, only a few of its top-ranked neighbors are retained, and symmetry processing is performed to obtain sparse kernel weights. The system outputs the neighborhood set, sparse kernel weights, and related parameters, forming local spatial kernel data.

[0060] Preferably, the structural reachability kernel is constructed as follows: The local neighborhood data is used to construct a multi-level local neighborhood index, which includes spatial index construction, topological index construction, and channel index construction. In one embodiment, three types of structured indexes are established for each node within the local neighborhood. The system constructs a spatial index, recording the node's three-dimensional coordinates, mileage marker along the corridor, and section number, for location and distance calculation. The system constructs a topological index, recording the node's unique identifier in the utility tunnel connectivity diagram, a list of incoming and outgoing edges, multi-hop level information, and shortest path depth, for describing connectivity and hierarchical structure. The system constructs a channel index, marking the node's compartment, ventilation duct zone, fire compartment, and valve or partition door control domain, among other management attributes. These indexes are stored in a unified structure; if historical modifications or structural changes exist, a version number and effective date are appended. The system outputs a local neighborhood index table and an adjacent edge base table, forming the local neighborhood index data.

[0061] Reconstruct the reachability path graph from the local neighborhood index data to obtain the reachability path graph data; In one embodiment, a local pathway graph structure is constructed based on a local neighborhood index table and an adjacency edge base table. The node set in the graph consists of all nodes within the neighborhood, and the edge set is derived from basic adjacency relationships. For each candidate connection edge in the edge set, its validity is determined based on channel index information and real-time device status: when the partition door is closed or there is a construction occupancy mark, the edge is marked as temporarily invalid; when the connection crosses different fire compartments and there is no legal connecting channel, it is not included in the valid edge set. For situations with multiple propagation paths, such as personnel passage paths and airflow propagation paths, corresponding layer structures are constructed, and hierarchical labels and status indicators are added to the edges. The system outputs pathway graph data containing node sets, valid edge sets, hierarchical labels, and edge status information.

[0062] Edge impedance decomposition is performed based on the reachability path graph data to obtain graph impedance data; In one embodiment, the comprehensive impedance value is calculated for each valid connection edge in the path diagram. The comprehensive impedance consists of three parts: geometric impedance, structural impedance, and environmental impedance, which are directly added together. The geometric impedance is determined based on the actual length between nodes and multiplied by a preset geometric weight. , This is the geometric impedance value. These are the geometric weighting coefficients. For the side length parameter, This is the starting node in the path graph. This refers to the target node in the pathway diagram; the structural impedance is assigned different penalty values ​​based on the gate or isolation status, for example, a medium penalty is assigned to the half-open state, and the fully closed state is considered impassable. , This is the structural impedance value. For structural weight coefficients, The structural state penalty parameters are (door half-open = 0.5, closed = ∞); the environmental impedance is mapped according to the risk levels such as temperature, water depth, and gas concentration, and multiplied by an environmental weighting coefficient, i.e. , This is the environmental impedance value. For environmental weighting coefficients, The environmental risk level parameter represents the risk level value obtained by mapping indicators such as temperature, water depth, and gas concentration through a preset parameter table (constructed based on engineering experience or expert knowledge). The three types of impedance can also be superimposed according to a weighted ratio (1:2:1) to form a comprehensive impedance value. The weighting parameters can be set based on historical fault statistics. The system outputs the impedance matrix of each edge and a list of effective edges, forming graph impedance data.

[0063] The path reachability distance is calculated based on the graph impedance data to obtain the path reachability distance data; In one embodiment, based on the constructed graph impedance data, a shortest path search is performed on any pair of nodes within the local neighborhood to calculate the minimum reachable cost in the sense of comprehensive impedance, which is taken as the path reachable distance between nodes. While finding the optimal path, several suboptimal paths are calculated, and the total cost set corresponding to each path is recorded. When there is no effective path between a pair of nodes, it is marked as unreachable. The system outputs a path reachable distance matrix and the corresponding path set information.

[0064] Accessibility stability is estimated based on path reachability clustering data to obtain accessibility stability data; In one embodiment, the reachability stability between node pairs is evaluated based on a set of multiple candidate paths. The system calculates path redundancy, which is the proportion of candidate paths whose cost difference from the optimal path is within a preset range; higher redundancy indicates more alternative paths. The system also counts the proportion of critical edges that appear only once in the optimal path; a higher proportion of critical edges indicates a stronger dependence on a single path. The system then performs a weighted sum based on redundancy and the proportion of critical edges to obtain a stability index (i.e., ...). , The reachability stability coefficient, For redundant weighting coefficients, This is a path redundancy index. This represents the critical edge penalty coefficient. The system uses the proportion of critical edges as an indicator and normalizes the results to a preset range. When the number of paths is small and the proportion of critical edges is high, the stability is low. The system outputs a stability matrix between node pairs, forming reachability stability data.

[0065] Preferably, ubiquitous heterogeneous characterization and global source tracing specifically refers to: The local heteromorphic spectrum data is aligned with the heteromorphic spectrum manifold to obtain manifold aligned data; In one embodiment, the low-dimensional embedding results of local anomalies obtained from different compartments or zones are subjected to unified coordinate processing. The system selects a set of common anchor points, which satisfy conditions such as consistent anomaly type, similar occurrence time, and spatial topological adjacency. The manifold embedding results of the main corridor zone are used as the reference coordinate system. The system performs shape alignment on the embedding results of other zones by solving for rotation, scaling, and translation parameters to minimize the difference between the corresponding anchor points and the reference manifold after alignment. If the alignment error exceeds a preset threshold, outlier anchor points are removed and the system is recalculated. The system outputs the alignment parameters, alignment residuals, and aligned embedding coordinates for each zone, forming manifold alignment data.

[0066] The transfer entropy is calculated on the manifold aligned data to obtain the transfer entropy data; In one embodiment, a low-dimensional trajectory sequence after manifold alignment is used as the state variable, and a time series relationship is constructed for each pair of candidate regions or nodes. A preset historical order and prediction step size are selected, and the two sequences are discretized using equal-frequency binning within a sliding time window. The joint probability distribution of the current state and historical states is statistically analyzed. By comparing the difference between "prediction uncertainty under known historical conditions" and "prediction uncertainty under simultaneously known historical conditions," an information transmission strength index is calculated to represent the degree of influence from one side to the other. , To pass the entropy value, from the source node Point to target node Information transmission strength index Let be a probability function, and be the discrete joint probability or conditional probability distribution obtained statistically within a sliding time window. Joint conditional probability is the probability of a future state of the target sequence occurring, given the historical states of both the target and source sequences. Let be the target's future state variable, be the state value of the target node after the prediction step size Δ, and be the value of the low-dimensional embedded trajectory after manifold alignment at future time steps. Let be the target historical state vector, which is the combination vector of the k consecutive historical states of the target node before the current time step. The source historical state vector is a combination vector of the k consecutive historical states of the source node before the current time. The system constructs control samples by performing multiple random permutations on the original sequence and calculates the corresponding statistical significance level. When the transmission strength exceeds a preset threshold and the significance meets the requirements, the directional relationship is retained. The system outputs transmission entropy data containing the source node, target node, transmission strength value, significance level, and lag step size (the time offset corresponding to the influence of the source sequence on the target sequence when constructing the transmission relationship, specifically the time interval between the current prediction time of the target sequence and the historical state of the source sequence).

[0067] Evolutionary chain data is obtained by generating evolutionary chains from manifold alignment data based on transfer entropy data; In one embodiment, a directed graph is constructed based on the propagation entropy result, with nodes or regions as vertices and propagation intensity as edge weights. The earliest appearing anchor point within the event time window is used as the candidate source point set, and time constraints are applied to retain only directed edges that satisfy the condition that the peak time of the target node is later than that of the source node and falls within a preset time range, thereby avoiding backpropagation relationships. The system performs a maximum weight path search with time constraints on this directed graph to obtain a node sequence representing the anomaly propagation process. When a loop structure appears in the path, the connection with the smaller edge weight is disconnected to ensure the unidirectionality of the link. The system outputs evolution chain data containing link identifiers, node sequences, corresponding time sequences, and edge weight sequences.

[0068] Global modal tensor fusion is performed based on evolutionary chain data to obtain global modal mapping data.

[0069] In one embodiment, for each evolutionary chain, the trajectory data of each node in the manifold alignment space, the corresponding local heteromorphic spectral fingerprint features / local heteromorphic spectral data, and the time index are uniformly aggregated to construct a chain-level multidimensional data structure. The local heteromorphic spectral fingerprint features include spectral peak features, energy proportions, curvature indices, etc. During the aggregation process, the system aligns according to a unified time axis, and at the same time point, the embedding representations of multiple nodes are normalized and weighted according to their transmission intensity to obtain the chain-level state expression; at the same time, the statistics of the corresponding fingerprint features are superimposed, such as mean, maximum value, or gradient of change. The system forms a global tensor structure organized according to "chain identifier—time—modality or fingerprint channel—embedding dimension" and generates a mapping index to record the correspondence between the chain and the set of spatial nodes, key connection relationships, and source and sink node information. The above data constitutes global modality mapping data.

[0070] Preferably, anomaly identification specifically includes: S31. Spatiotemporal coupling is performed based on local heterogeneous spectrum data and global modal mapping data to obtain fully local coupled data; In one embodiment, anchor point information in the local heterogeneous spectral data is used as an index. The anchor point includes at least node identifier, start and end times, peak times, low-dimensional embedding representation, and spectral fingerprint features. Data fragments satisfying overlapping time windows and spatial reachability or being within the same evolutionary chain are retrieved from the global modal mapping tensor. The retrieved fragments are aligned: the time dimension is synchronized based on the anchor point peak time, and the spatial dimension is unified to the same node sequence according to the mapping index. The system concatenates and fuses the local spectral embedding representation, chain-level state representation, and spectral fingerprint features, and records the alignment error index. The system outputs the fused coupled representation and the corresponding event identifier, chain identifier, and node sequence information, forming fully local coupled data.

[0071] S32. Construct the spectral response matrix from the fully locally coupled data to obtain the spectral response matrix data; In one embodiment, within a fixed short time window, the response intensity between local and global parameters is calculated for fully coupled data. For each pair of nodes, local spectral energy features and corresponding global chain state change features are extracted, standardized, and then the correlation or frequency band consistency index between the two is calculated to characterize the response intensity of local anomalies to global situational changes. Simultaneously, propagation direction weights from the evolutionary chain are introduced to weight and adjust the response intensity, ensuring the results balance correlation and propagation reliability. , For the response intensity coefficient, For directional propagation of weights, from node Pointing to node Information transmission strength index The correlation consistency index is a function of the degree of correlation between local spectral features and global chain state changes. For local spectral energy vectors, nodes The spectral energy feature vector within the current time window includes features such as energy distribution in each frequency band, energy centroid, or principal component intensity. The global chain state change vector, node The chain-level state changes corresponding to the evolution chain tensor represent the temporal evolution magnitude of the global situation at that node. The obtained response intensity values ​​are filled into a matrix structure indexed by the node sequence, and items below a preset threshold are set to zero to form a sparse representation. The system outputs the response matrix, threshold parameters, time window settings, and matrix sparsity index, forming the spectral response matrix data.

[0072] S33. Perform spectral shape matching on the spectral response matrix data to obtain spectral shape data; In one embodiment, the system pre-establishes a typical anomaly template library, including templates for leakage, water accumulation, ventilation failure, and structural vibration, etc. Each template contains a morphological feature signature of the corresponding spectral response matrix, such as the main diagonal energy distribution, propagation bandwidth range, and peak evolution trend. For the current spectral response matrix, morphological feature indicators are extracted, including the overall energy principal components, energy distribution centroid, band structure width, and directional features, to form a morphological description of the current response. This morphological description is matched with various templates in the template library for similarity, which can be done using time warping matching or vector similarity calculation. If the matching score exceeds a preset threshold, the corresponding anomaly type and confidence level are output. The system generates spectral morphological data containing the current morphological features, candidate types, and matching scores.

[0073] S34. Perform pattern recognition on the spectral morphology data to obtain anomaly identification data.

[0074] In one embodiment, the morphological feature vectors and template matching scores from the spectral morphology data are used as input features and fed into a lightweight classification model for final determination. The classification model can employ a linear support vector machine or a gradient boosting tree model, suitable for edge deployment. Classification categories include normal state, mild anomaly, and severe anomaly, or further subdivided into specific anomaly types. Model training samples are derived from historical real-world event data and digital twin simulation samples, maintaining consistency with the current feature extraction process. During the determination phase, when the predicted probability of a severe anomaly category exceeds a preset threshold, or when the same event is identified as an anomaly in several consecutive determinations, an anomaly result is output. Simultaneously, the set of anomaly location nodes is determined based on the nodes with the highest response intensity ranking in the spectral response matrix, and corresponding propagation chain identifiers are associated with them. The system outputs anomaly identification data containing event identifier, anomaly type, risk level / anomaly probability, predicted probability, location node, and propagation chain information.

[0075] The classification model construction steps include building a training sample set during the offline training phase. Each training sample corresponds to a time window of a historical event or a digital twin simulation event. The sample input consists of two parts: one is a spectral morphological feature vector, containing morphological signatures extracted from the spectral response matrix (e.g., energy centroid, band width, directional index, principal component intensity, etc.); the other is a template matching score vector, containing the similarity score between the sample and each anomalous template. The system concatenates the above two types of features in a fixed order to form a unified feature record and adds sample labels, which are normal, mildly anomalous, severely anomalous, or specific anomalous type. Before training, the feature record undergoes consistency processing, including missing item imputation, outlier truncation, and normalization according to training set statistics. The processed results are then written into the feature matrix and label vector to form a trainable data structure. During model training, if a linear support vector machine is used, the system takes the feature matrix as input and iteratively solves for the classification hyperplane parameters to maximize the margin between samples of different classes in the feature space. If a gradient boosting tree model is used, the system takes the feature matrix and labels as input and iteratively generates weak classification trees round by round. Each round fits the residuals from the previous round, ultimately forming an ensemble model composed of multiple trees. After training, the system performs threshold calibration on the model, maps the class scores output by the model to predicted probabilities, and saves the model parameters, feature normalization parameters, and class mapping table as intermediate products for edge-side inference. On the output side, the model provides class prediction results and corresponding predicted probabilities for the input feature records. The system combines event numbers and time window indices to generate anomaly identification data and outputs event identifiers, anomaly types, risk levels or anomaly probabilities, predicted probabilities, location nodes, and associated propagation chain identifiers / propagation chain information.

[0076] Preferably, the intelligent repair decision-making specifically includes: S41. Extract anomaly points from the anomaly identification data to obtain anomaly point data; In one embodiment, the system performs dimensionless segmentation processing on three types of evidence: anomaly probability / predicted probability and level, response energy proportion, historical anomaly frequency, and sensor health. These are then mapped to discrete level labels based on preset thresholds or quantile intervals. Subsequently, the evidence is sorted lexicographically according to a fixed priority: severity level is compared first, then response proportion level, and finally historical risk level; no numerical aggregation is performed. The system outputs anomaly point data including event identifier, node identifier, sorting result, anomaly probability / risk level, anomaly type, time range, and evidence fields.

[0077] In one embodiment, the three types of evidence are treated as independent target dimensions. A set of candidate nodes that are simultaneously superior to other nodes across all dimensions is selected, and then truncated and ranked using a single auxiliary indicator. The system outputs anomaly data including event identifier, node identifier, ranking result, risk value, anomaly type, time range, and evidence fields.

[0078] S42. Match the repair solution database based on the anomaly point data to obtain repair solution data; In one embodiment, for primary anomalies (the highest-ranked or second-highest ranked anomaly data), high-response-speed or mandatory handling solutions are prioritized; for secondary anomalies (non-primary anomaly data), lower-risk or inspection-based solutions are prioritized. The system pre-establishes a repair solution library, organized using a combination of rules and parameter templates. The construction steps of the repair solution library involve the system extracting closed-loop handling records from historical maintenance work order data, including anomaly type, handling steps, resource configuration, execution time, and handling effect. Historical work orders are structurally parsed, breaking down the handling process into standard step units, such as "on-site confirmation," "risk isolation," "equipment reset," and "component replacement," forming reusable step templates. Anomaly types are mapped and categorized with equipment types, constructing a correspondence of "anomaly type—equipment category—handling step template." For each combination, parameterized rule entries are generated, including triggering conditions, applicable scope, resource requirements, safety constraints, and estimated duration range. Statistical learning is used to summarize historical success rates, average handling times, and risk levels, forming initial reference values ​​for rule parameters. Rule entries are stored in structured data format and support version number and effective time management. When new types of anomalies or new equipment appear, the solution library is expanded by adding templates or updating parameters. Each rule includes at least the anomaly type, triggering conditions, applicable equipment scope, required tools and personnel configuration, pre-check items, safety control measures, and estimated handling time. During matching, each anomaly is entered into the candidate solution set based on its anomaly type, and constraint filtering is performed sequentially, including accessibility of passages and fire compartments, availability of resources such as spare parts and personnel, security access control and work permit conditions, and whether there are conflicts with the current operating conditions. For the selected solutions, a scoring and ranking process is performed based on effectiveness, time cost, risk cost / risk level / risk value, and historical success rate (first, a hard stratification is applied based on risk cost, for example, prioritizing solutions with the lowest risk or those meeting mandatory safety constraints; within the same risk level, further ranking is based on estimated duration or resource consumption). The top-ranked solutions are then selected as recommendations. The system outputs remediation solution data including node identifiers, solution identifiers, a list of steps, required resources, estimated duration, risk warnings, and a score.

[0079] S43. Based on the repair plan data, a repair path is planned to obtain intelligent repair decision data.

[0080] In one embodiment, the node identifiers and corresponding priorities in the remediation plan data are used as the target node set for path planning. The system classifies the edges in the reachable path graph according to environmental risk levels, forming three sub-graph structures: safe zone, restricted zone, and high-risk zone. The environmental risk level is determined by the system converting various environmental monitoring indicators and anomaly identification results into graded risk identifiers based on preset mapping rules. When the gas concentration exceeds a set first-level threshold, the risk level of the corresponding spatial unit is increased by one level. When the water depth exceeds the safety control line, the area is marked as a restricted zone. When the anomaly posterior probability is higher than a preset high quantile threshold, the area is marked as a high-risk zone. When the same node triggers anomaly events multiple times within a continuous time window, its risk level is progressively increased. Based on the above rules, the system divides the spatial region into a three-level structure: the safe zone is an area with no indicators exceeding limits and anomaly posterior probabilities at a low level; the restricted zone is an area with slightly exceeded indicators or identified as a secondary anomaly point; and the high-risk zone is an area where severely anomaly nodes are located or where key indicators such as gas, high temperature, and low oxygen exceed limits. During path planning, reachable path searches are prioritized within the safe zone subgraph. If a target node is unreachable within the safe zone, the search is then expanded to the restricted zone subgraph. High-risk zones are only activated when no alternative paths exist, and mandatory security checkpoints and protection warnings are marked on the corresponding path nodes. After determining an available subgraph, a path search with the shortest time objective is performed within that subgraph, generating candidate reachable paths for each target node. In the task sequencing phase, priority-based bucket scheduling is used. Tasks are categorized according to their anomaly severity, with priority given to handling major anomalies. Within the same severity level, the system sorts tasks by nearest-neighbor distance from the current execution position to the target node, generating an execution sequence accordingly. The system outputs a path and task arrangement scheme that balances security constraints and handling priorities.

[0081] Preferably, this application also provides a digital twin-based underground utility tunnel anomaly identification system for performing the digital twin-based underground utility tunnel anomaly identification method described above. The digital twin-based underground utility tunnel anomaly identification system includes: The edge-aware fusion module is used to acquire edge-aware data; and to perform edge fusion on the edge-aware data to obtain edge-fused data. The edge anomaly processing module is used to extract local anomalies from edge fusion data to obtain local anomaly spectrum data; and to perform ubiquitous heterogeneous characterization and global source tracing on the local anomaly spectrum data to obtain global modality mapping data. The anomaly detection module is used to identify anomalies based on local abnormal spectrum data and global modal mapping data to obtain anomaly detection data. The intelligent repair decision module is used to make intelligent repair decisions based on anomaly identification data, and obtain intelligent repair decision data.

Claims

1. A method for anomaly identification in underground utility tunnels based on digital twins, characterized in that, Includes the following steps: Acquire edge-aware data; perform edge fusion on the edge-aware data to obtain edge-fused data; Local heterogeneous spectrum extraction is performed on edge fusion data to obtain local heterogeneous spectrum data; ubiquitous heterogeneous characterization and global source tracing are performed on local heterogeneous spectrum data to obtain global modality mapping data; Anomaly identification data is obtained by performing anomaly identification based on local anomaly spectrum data and global modal mapping data. Intelligent repair decision-making is performed based on anomaly identification data to obtain intelligent repair decision data.

2. The method for anomaly identification of underground utility tunnels based on digital twins according to claim 1, characterized in that, Edge blending specifically refers to: Heterogeneous modal tensor quantization is performed on edge-sensing data to obtain heterogeneous modal data; Transient entropy increase feature anchoring is performed on heterogeneous modal data to obtain feature anchoring data; Cross-domain correlated manifold reconstruction is performed based on feature-anchored data to obtain manifold reconstruction data; Collaborative probabilistic fusion is performed on the manifold reconstruction data to obtain edge fusion data.

3. The method for anomaly identification of underground utility tunnels based on digital twins according to claim 2, characterized in that, The transient entropy increase feature anchoring is specifically as follows: Modal-uniform resampling is performed on heterogeneous modal data to obtain in-phase modal data; Local background stripping is performed on the in-phase mode data to obtain transient data; Transient probability configuration data is obtained by performing transient probability configuration processing on transient data; The transient entropy increase rate is calculated from the transient probability configuration data to obtain the transient entropy increase rate data; Cross-modal entropy increase consistency coupling is performed based on transient entropy increase rate data to obtain entropy increase resonance data; Anchor points are fixed based on entropy increase resonance data to obtain feature anchoring data.

4. The method for anomaly identification of underground utility tunnels based on digital twins according to claim 1, characterized in that, Local heteromorphic spectrum extraction specifically involves: Local spatial kernels are generated from the edge-fused data to obtain local spatial kernel data; Variational mode decomposition is performed on the edge fusion data based on the local spatial kernel data to obtain heterogeneous mode spectrum data; Nonlinear manifold extraction is performed based on heterogeneous modal spectral data to obtain local heterogeneous spectral data.

5. The method for anomaly identification of underground utility tunnels based on digital twins according to claim 3, characterized in that, Local spatial kernel generation specifically involves: Local neighborhood data is obtained by selecting local neighborhood data from the edge-fused data; Structural reachability kernels and operational condition isomorphic kernels are constructed from local neighborhood data to obtain structural reachability kernel data and operational condition isomorphic kernel data, respectively. Local spatial kernel data is obtained by kernel synthesis based on structural reachability kernel data and operational condition isomorphic kernel data.

6. The method for anomaly identification of underground utility tunnels based on digital twins according to claim 5, characterized in that, The structural reachability kernel is constructed as follows: The local neighborhood data is used to construct a multi-level local neighborhood index, which includes spatial index construction, topological index construction, and channel index construction. Reconstruct the reachability path graph from the local neighborhood index data to obtain the reachability path graph data; Edge impedance decomposition is performed based on the reachability path graph data to obtain graph impedance data; The path reachability distance is calculated based on the graph impedance data to obtain the path reachability distance data; Accessibility stability is estimated based on path reachability clustering data to obtain accessibility stability data; Joint kernel mapping is performed based on path reachability distance data and reachability stability data to obtain structural reachability kernel data.

7. The method for anomaly identification of underground utility tunnels based on digital twins according to claim 1, characterized in that, The ubiquitous heterogeneous characterization and global source tracing specifically refers to: The local heteromorphic spectrum data is aligned with the heteromorphic spectrum manifold to obtain manifold aligned data; The transfer entropy is calculated on the manifold aligned data to obtain the transfer entropy data; Evolutionary chain data is obtained by generating evolutionary chains from manifold alignment data based on transfer entropy data; Global modal tensor fusion is performed based on evolutionary chain data to obtain global modal mapping data.

8. The method for anomaly identification of underground utility tunnels based on digital twins according to claim 1, characterized in that, Anomaly detection specifically includes: Spatiotemporal coupling is performed based on local heterogeneous spectrum data and global modal mapping data to obtain fully local coupled data; The spectral response matrix is ​​constructed from the fully locally coupled data to obtain the spectral response matrix data; Spectral shape matching is performed on the spectral response matrix data to obtain spectral shape data; Pattern recognition is performed on the spectral morphology data to obtain anomaly identification data.

9. The method for anomaly identification of underground utility tunnels based on digital twins according to claim 1, characterized in that, The intelligent repair decision-making process is as follows: Anomaly point data is obtained by extracting anomaly points from the anomaly identification data; Repair solution data is obtained by matching the repair solution library with the anomaly data; Repair path planning is performed based on the repair plan data to obtain intelligent repair decision data.

10. An anomaly identification system for underground utility tunnels based on digital twins, characterized in that, For executing the digital twin-based underground utility tunnel anomaly identification method as described in claim 1, the digital twin-based underground utility tunnel anomaly identification system comprises: The edge-aware fusion module is used to acquire edge-aware data; and to perform edge fusion on the edge-aware data to obtain edge-fused data. The edge anomaly processing module is used to extract local anomalies from edge fusion data to obtain local anomaly spectrum data; and to perform ubiquitous heterogeneous characterization and global source tracing on the local anomaly spectrum data to obtain global modality mapping data. The anomaly detection module is used to identify anomalies based on local abnormal spectrum data and global modal mapping data to obtain anomaly detection data. The intelligent repair decision module is used to make intelligent repair decisions based on anomaly identification data, and obtain intelligent repair decision data.

Images