Urban lifeline risk assessment and early warning method and system based on multi-source monitoring fusion

Abstract

The application discloses a kind of urban lifeline risk assessment early warning method and system based on multi-source monitoring fusion, it is related to facility safety warning technical field.The application is repaired by unified data identification and quality weighting, effectively suppresses multi-source data anomaly and absence, improves data availability;Multi-relationship graph structure and credible edge modeling are constructed, environment adaptive dynamic graph modeling is realized, and the topological structure of facility, space and physical coupling relationship are accurately represented;Cross-modal consistency constraint and quality-uncertainty gate control fusion are introduced, stable state representation is generated, reasoning is carried out in combination with health index and calibration failure probability, and the accuracy of risk assessment is significantly improved;False alarm is inhibited by using cost-sensitive bayesian grading and persistence trigger mechanism, dominant risk chain and optimization disposal suggestion are output, a closed loop from data governance to operation decision is formed, and the intelligentialization and engineering practicability of urban lifeline risk management are improved.

Description

Technical Field

[0001] This invention relates to the field of facility safety early warning technology, and in particular to a method and system for risk assessment and early warning of urban lifelines based on multi-source monitoring fusion. Background Technology

[0002] Urban lifeline infrastructure is the core support system for ensuring the normal operation of a city and safeguarding public safety. It encompasses critical infrastructure networks such as water supply, drainage, gas, electricity, communications, and transportation, and is deeply coupled with urban population density, economic vitality, and emergency response capabilities. These facilities are characterized by their wide coverage, complex network structure, and strong cross-sectoral interconnectivity. Once a part of them fails, it can easily trigger cascading failures through network topology or functional chains, leading to supply disruptions, traffic paralysis, and the spread of secondary disasters, causing significant social impact. Therefore, the safe and stable operation of urban lifelines is not only related to the order of daily production and life, but also directly determines the city's resilience, emergency response capabilities, and governance effectiveness under major disasters.

[0003] With rapid urbanization and extended infrastructure service life, urban lifelines face increasingly complex and uncertain risks. On the one hand, deterioration issues such as material aging, fatigue accumulation, hidden engineering defects, and long-term uneven settlement continue to accumulate. On the other hand, external disasters such as extreme rainfall, floods, earthquakes, ground subsidence, soil instability caused by pipeline leaks, overloaded operation, and construction disturbances occur frequently. These multiple factors are driving risks from single-point failures to systemic risks, placing higher demands on monitoring, early warning, and risk assessment. It is necessary not only to identify anomalies in individual facilities but also to analyze the risk propagation and coupling effects across facilities, regions, and disciplines.

[0004] Currently, urban lifeline safety management still relies primarily on manual inspections, periodic testing, and post-incident handling, supplemented only by limited online monitoring methods. Although technologies such as the Internet of Things, video surveillance, drones, lidar, GNSS, fiber optic sensing, and remote sensing are gradually being implemented, generating massive amounts of multi-source heterogeneous data, four core pain points still exist in engineering applications. First, while source monitoring data can be collected, it is difficult to utilize. Data exhibits inconsistencies in sampling frequency, significant differences in noise levels, frequent data gaps and anomalies, and inconsistent spatiotemporal benchmarks across sources, resulting in insufficient data availability and comparability. Second, there is a lack of a unified correlation modeling mechanism between multi-source data. Alarms are often based on single-sensor thresholds or local judgments of single systems, making it difficult to handle facility topology connections, spatial adjacencies, and physical coupling relationships, leading to false alarms, missed alarms, and delayed risk identification. Third, risk assessments rely heavily on empirical rules or single indicators, lacking a system that jointly quantifies structural state representation with failure probability and uncertainty, making it difficult to support tiered response and optimal resource allocation. Fourth, the early warning and response process lacks a closed-loop link from data to decision-making. Alarm information often only indicates the occurrence of an abnormal event but fails to explain the source of the anomaly, key driving factors, and risk evolution trends, thus affecting the efficiency of operation and maintenance decisions and the effectiveness of emergency response. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for risk assessment and early warning of urban lifelines based on multi-source monitoring fusion, so as to solve the technical problems of poor quality and heterogeneous data of multi-source monitoring, which leads to inaccurate risk identification and inability to automatically generate executable operation and maintenance decisions.

[0006] To achieve the above-mentioned objectives, the embodiments of the present invention provide the following technical solutions: A method for urban lifeline risk assessment and early warning based on multi-source monitoring fusion includes: Acquire multi-source monitoring data of the city's lifeline infrastructure to be evaluated, generate peer samples and data quality characterization information, process the peer samples by combining quality time series constraint objectives and event consistency alignment loss, and generate a unified monitoring dataset; Based on a unified monitoring dataset, graph nodes and edges are determined, and a comprehensive adjacency matrix is ​​constructed by combining the credibility coefficient. Based on the comprehensive adjacency matrix, the infrastructure status representation vector of the unified monitoring dataset is extracted, and the corresponding infrastructure health index and failure probability are calculated. Based on the infrastructure health index and failure probability, the risk level is calculated, and an early warning trigger result is generated.

[0007] In the above-mentioned scheme, this invention effectively solves the problem of data unavailability caused by missing, abnormal, noisy, and time misaligned multi-source monitoring data by constructing a unified data identification and quality-weighted joint repair mechanism, thereby improving data continuity and comparability. On this basis, an environment-adaptive comprehensive adjacency matrix is ​​constructed based on a multi-relationship graph structure and credibility coefficient, which can dynamically represent facility topological connections, spatial neighborhoods, and physical coupling relationships, overcoming the shortcomings of traditional methods in modeling risk propagation paths across facilities and regions. Furthermore, a cross-modal consistency constraint and quality-uncertainty gating fusion strategy are introduced to generate a stable and unified state representation vector, which significantly suppresses single-modal false alarms and risk level mutations, improving the accuracy and reliability of risk assessment. By combining the health index and calibration failure probability for probabilistic reasoning, and combining cost-sensitive Bayesian decision-making and a persistent triggering mechanism, a closed loop from "data governance - correlation modeling - probabilistic reasoning - hierarchical early warning - disposal suggestions" is realized, solving the problems of existing technologies being unable to automatically generate executable operation and maintenance suggestions and having poor interpretability of early warnings. This reduces reliance on manual inspections and improves the intelligence, refinement, and engineering practicality of risk management for urban lifeline infrastructure.

[0008] A city lifeline risk assessment and early warning system based on multi-source monitoring fusion includes: The data generation module is used to acquire multi-source monitoring data of the city's lifeline infrastructure to be evaluated, generate peer samples and data quality characterization information, process the peer samples in combination with quality time series constraint objectives and event consistency alignment loss, and generate a unified monitoring dataset. The matrix construction module is used to determine graph nodes and edges based on a unified monitoring dataset and to construct a comprehensive adjacency matrix by combining the credibility coefficient. The calculation module is used to extract the infrastructure status representation vector of the unified monitoring dataset based on the comprehensive adjacency matrix, and calculate the corresponding infrastructure health index and failure probability. The early warning module is used to calculate the risk level based on the infrastructure health index and failure probability, and generate early warning trigger results. Attached Figure Description

[0009] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a flowchart of the method in an embodiment of the present invention; Figure 2 This is a system structure diagram in an embodiment of the present invention. Detailed Implementation

[0011] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the 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.

[0012] Please see Figure 1 This embodiment provides a method for risk assessment and early warning of urban lifelines based on multi-source monitoring fusion. Figure 1 The execution entity of the method shown can be a software and / or hardware device. The execution entity of this application can include, but is not limited to, at least one of the following: user equipment, network equipment, etc. User equipment can include, but is not limited to, computers, smartphones, personal digital assistants (PDAs), and the aforementioned electronic devices. Network equipment can include, but is not limited to, a single network server, a server group consisting of multiple network servers, or a cloud based on cloud computing consisting of a large number of computers or network servers. Cloud computing is a type of distributed computing, consisting of a super virtual computer composed of a group of loosely coupled computers. This embodiment does not impose any limitations on this.

[0013] A method for urban lifeline risk assessment and early warning based on multi-source monitoring fusion includes: S1. Obtain multi-source monitoring data of the city's lifeline infrastructure to be evaluated, generate peer samples and data quality characterization information, process the peer samples in combination with quality time series constraint objectives and event consistency alignment loss, and generate a unified monitoring dataset. S1 includes: S1-1. Obtain multi-source monitoring data of the city's lifeline infrastructure to be evaluated, and generate peer samples by combining the data representation system; The multi-source monitoring data refers to the multi-source data generated during the operation of the urban lifeline infrastructure to be evaluated, including structural response monitoring data, environmental monitoring data, inspection image data, and operation and maintenance management data.

[0014] The structural response monitoring data includes settlement, displacement, tilt angle, strain, pressure, flow rate, and liquid level of lifeline infrastructure; environmental monitoring data may include rainfall, groundwater level, temperature, humidity, and gas concentration of the city to be assessed; inspection image data may include fixed monitoring images, mobile inspection images, or drone inspection images of lifeline infrastructure; and operation and maintenance management data may include inspection records, maintenance records, historical alarm records, and handling ledgers of lifeline infrastructure.

[0015] It needs to be explained that the encoding of each multi-source monitoring data is constructed in the form of "facility ID-component / section ID-monitoring point ID-data type-timestamp". That is, the data representation system determines which object the data belongs to, what type it is, and when it was collected, so as to realize the unique index and retrieval of cross-source data.

[0016] Based on the preset time window length, multi-source monitoring data are organized into a single window sample, with the corresponding formula being: ; in, Indicates the first A sample of classmates within a time window. , , , They represent the first The data collected includes structural response monitoring data, environmental monitoring data, inspection image data, and operation and maintenance management data within a specific time window. During implementation, a uniform time window length is set based on actual monitoring needs, and data from different sources falling within the same time window are organized into the same processing unit.

[0017] S1-2. Based on multi-source monitoring data, calculate the corresponding quality characterization values ​​and generate data quality characterization information; To enhance the reliability of subsequent fusion inference, data quality characterization information is synchronously generated for multi-source monitoring data within each time window at the acquisition or data access end. Specifically, for the four modalities of structural response, environmental monitoring, inspection images, and operation and maintenance management, their respective quality values ​​are calculated and finally combined into a quality vector, with the corresponding expression as follows: ; in, Indicates the time window index. Indicates the first Data quality characterization information for each time window , , , They represent the first The quality sub-indicators corresponding to the structural response monitoring data, environmental monitoring data, inspection image data, and operation and maintenance management data for each time window.

[0018] Different modalities have different data characteristics, so the quality sub-indicators used also differ, but the overall approach remains the same. First, the raw data for that modality within the time window is extracted. Then, several sub-indicators reflecting data availability are calculated. Finally, a weighted fusion is used to obtain the quality value for that modality. All sub-indicators are normalized to the [0, 1] interval, and a larger value indicates higher quality.

[0019] The quality characterization values ​​for structural response data include data integrity (the ratio of actual data collection points to theoretical data collection points), average transmission delay (the degree of lag relative to the theoretical timestamp), noise level (a normalized estimate of signal variance or high-frequency components), and device self-test status (whether sensors report normal operation). The quality characterization values ​​of the structural response data are weighted and summed to obtain the quality sub-indicators. .

[0020] Similar to structural response data, environmental monitoring data typically does not involve image sharpness. Therefore, a weighted average of three quality characteristics—data integrity, latency, and noise level—is used to calculate the quality sub-indicators. .

[0021] The quality characteristics of inspection image data include image presence (whether an image exists within the time window, 1 if present, 0 if absent), image sharpness (e.g., Laplacian variance normalized value), and occlusion ratio (the proportion of effective area in the image). These three quality characteristics are weighted and fused to obtain the quality sub-indicators. .

[0022] The quality characterization value of operation and maintenance management data (inspection records, maintenance logs, etc.) is determined by record completeness and timeliness. Record completeness reflects whether required fields are missing; timeliness indicates whether the record was reported within the specified inspection cycle. The quality sub-indicator is obtained by weighting the two. .

[0023] For each time window, after calculating four quality sub-indicators in sequence, the first... The data includes samples from the same time window, which are stored along with the original data in the infrastructure monitoring database. These samples are used for anomaly repair weight control in subsequent preprocessing, edge credibility modeling in graph structures, and modality gating factor calculation in multi-source feature fusion. If a modality has no data at all during the entire time window, its quality value is set to 0.

[0024] S1-3. Based on the facility physical constraint penalty term and data quality characterization information, construct the quality time series constraint target and perform anomaly detection and joint repair on the peer samples to generate peer repair samples. The physical constraints and penalties for the facilities include range constraints, upper limits for settlement and displacement rates, non-negative pore pressure constraints, and unexplainable jump constraints.

[0025] It needs to be explained that, in order to avoid the information destruction and error propagation caused by the traditional method of first deleting outliers and then separately imputing missing values, outlier suppression, missing value repair and engineering constraints are unified into the quality-weighted reconstruction and optimization process, and a quality time-series constraint target is constructed.

[0026] Specifically, anomaly detection is performed on peer samples, and those with anomalies are designated as samples to be repaired. Anomaly detection employs a combined approach of engineering threshold rules, statistical deviation characteristics, and time-series change characteristics to identify anomalous samples. First, based on sensor range, design allowable range, and historical normal operating conditions, it is determined whether the monitored values ​​exceed the physical upper or lower limits. Second, the mean and standard deviation of each monitored variable are calculated using historical stable operating data, and statistical deviation anomalies are identified through standard scores (e.g., exceeding 3 times the standard deviation) or Mahalanobis distance. Third, the rate of change of adjacent time windows is analyzed; if it exceeds a preset maximum rate of change and there is no external event explanation, it is determined to be a sudden anomaly, and abnormal fluctuations or continuous drifts are detected using local trend regression residuals. When any monitored variable is judged to be anomalous under any rule and cannot be reasonably explained by data quality (e.g., delay, rounding errors), the peer samples within that time window are marked as anomalous peer samples, and an anomaly indication matrix is ​​generated (used to reduce the weight of anomalous data positions in the quality weight matrix, guiding the repair process), which is then incorporated into the subsequent joint repair and optimization process.

[0027] Then, according to the formula: ; ; ; Construct quality time-series constraint objectives. Among them, Describes the minimum value function. Indicates the total number of time windows. This represents an abnormal sample set among classmates. Represents norm 2, Indicates the first Samples to be repaired within a time window. , They represent the first The first time window, the first The peer repair sample within a time window (optimization variable). , These represent the temporal smoothing constraint coefficient and the physical constraint penalty coefficient, respectively. This indicates the physical constraint penalty for the facility. Represents the quality weight matrix. , , , These represent the structural response quality weight, environmental quality weight, inspection image quality weight, and operation and maintenance management quality weight, respectively. Represents a diagonal matrix. This represents the summation function. Indicates the first The first time window The quality weights corresponding to each mode , They represent the first The first time window Data quality characterization information and noise intensity estimation for each modality This represents a small constant used to prevent the denominator from being zero.

[0028] Facility physical constraint penalty items The corresponding formula is: ; in, , , , These represent the range weight, the upper limit weight of settlement and displacement rate, the non-negative weight of pore pressure, and the unexplained jump weight, respectively.

[0029] Range constraints This is used to limit the repair value to a physically feasible range jointly determined by the sensor's nominal range, the engineering allowable range, and the historical normal operating condition statistical interval. The corresponding formula is: ; Upper limit constraints on settlement and displacement rates This is used to limit the rate of settlement or displacement change between adjacent time windows to not exceed the allowable upper limit, so as to ensure the temporal continuity of deformation evolution and engineering rationality. The corresponding formula is: ; Non-negative pore pressure constraint This is used to limit the repair result to no less than zero in scenarios where the monitored object is saturated medium pore water pressure or hydraulic pressure defined as non-negative gauge pressure, in order to avoid negative values ​​that do not conform to physical meaning. The corresponding formula is: ; Unexplainable sudden jump constraint This method is used to suppress abnormal deviations of repair results from local trend predictions when there are no construction disturbances, environmental changes, maintenance operations, or other known external triggering events. It aims to reduce unrealistic spikes caused by communication glitches, local distortions, single-point acquisition anomalies, or compensation errors. The corresponding formula is: ; in, Indicates the total number of modes. Indicates the first The number of data dimensions in each modality Represents the maximum value function. Represents absolute value. Indicates the first The first time window The first mode A sample of peer-to-peer repair data across multiple data dimensions. , They represent The upper and lower bounds (the upper and lower bounds are preferably determined based on the sensor range parameters, facility design or operation and maintenance control indicators, and historical normal operating condition statistical characteristics). This represents the set of data dimensions to which rate constraints apply. , They represent the first The first time window, the first The first time window A sample of peer-to-peer repair data (optimization variables) across multiple data dimensions. Indicates the interval between adjacent time windows. Indicates the first The maximum allowable rate of change for each data dimension corresponds to the physical monitoring variable (determined by combining the design control value, alarm threshold, and the statistical upper limit of the rate of change under historical normal operating conditions). This represents the set of monitoring variables related to pore pressure. Indicates the first The pore pressure variable at the th [ ] ...]th [ ]th ]th ] ] ] ] ] ] ] ] ] ] ] ] ] ] ] ] ] ] Repair values ​​under a time window This represents the set of data dimensions to which the jump constraint applies. Indicates an explanatory indicator quantity. This represents the predicted value obtained based on the local trend estimation of the preceding time window. This indicates the permissible deviation from the threshold. If no explainable external event exists within the current time window... Conversely The allowable deviation threshold is preferably determined based on the statistical characteristics of the increments of adjacent time windows under historical normal operating conditions.

[0030] Furthermore, based on the quality time-series constraint objective, joint repair is performed on the peer anomaly samples corresponding to the anomaly indicator matrix to generate peer repair samples. Specifically, the original peer anomaly samples are used as the initial optimization values, and the data dimensions with missing values ​​are initialized using historical mean, neighborhood interpolation, or forward preservation methods. Subsequently, under the conditions of the quality weight matrix and facility physical constraints, the objective function is iteratively optimized. Preferably, gradient descent, alternating iterative optimization, or constraint optimization methods can be used to solve the quality time-series constraint objective. In each iteration, the objective function value is calculated based on the current repair result, and the peer repair samples are updated by combining data consistency terms, time-series continuity terms, and physical constraint penalty terms. Among them, higher-quality data has higher weights in the update process, and data with large anomaly fluctuations have relatively lower weights, thereby reducing the impact of anomaly noise propagation on the repair results. Furthermore, when the change in the objective function between two adjacent iterations is less than a preset threshold, or when the preset maximum number of iterations is reached, the iteration stops, and the current optimization result is used as the final peer repair sample.

[0031] Through the above-mentioned joint repair process, anomaly suppression, missing data repair and engineering constraints can be integrated and optimized, thereby improving the reliability of the unified monitoring dataset and the stability of subsequent risk inference.

[0032] S1-4. Standardize the peer repair samples and peer samples (i.e. peer samples without anomalies), combine them with the event consistency alignment loss to perform cross-source time series alignment, and generate a unified monitoring dataset.

[0033] S1-4 includes: S1-4-1. Standardize the peer repair sample and the peer sample to generate a peer standard sample; After anomaly detection and joint remediation are completed, to eliminate scale differences between data from different sources, with different dimensions, and with different value ranges, and to improve the consistency of subsequent cross-source analysis and fusion calculations, it is necessary to standardize the same-window remediation samples and same-window samples. Specifically, for same-window remediation samples and same-window samples corresponding to structural response monitoring data and environmental monitoring data, normalization or standardization processing can be performed; for same-window remediation samples and same-window samples corresponding to inspection image data, their size, grayscale distribution, or embedded feature representation can be unified; for same-window remediation samples and same-window samples corresponding to operation and maintenance management data, category attributes, event records, or statistical characteristics can be standardized.

[0034] S1-4-2. Define the temporal neighborhood and extract cross-source event features corresponding to structural response monitoring data and environmental monitoring data through the event feature extraction function; In S1-4-2 to S1-4-5, the objects of cross-source time-series alignment are the standard samples of the same window corresponding to structural response monitoring data and environmental monitoring data.

[0035] The event feature extraction function is used to extract cross-source event features (peak value, inflection point, rate of change, acceleration, upward trend, and stage transition, etc.) from standard samples corresponding to structural response monitoring data and environmental monitoring data. The corresponding formula is: ; ; in, Represents the transpose matrix. , These represent the event feature extraction functions corresponding to structural response monitoring data and environmental monitoring data, respectively. , They represent the first The time window ends with a time window length of [number]. The structural response monitoring data and environmental monitoring data, corresponding to the same standard sample, , , , , , These represent the local peak detection function, inflection point detection function, first-order difference function, second-order difference function, rate of rise calculation function, and change point detection function, respectively.

[0036] The method of determining the peak significance using a combination of neighborhood comparison and peak comparison is defined as follows: ; ; in, Indicates length is Structural response monitoring data / environmental monitoring data univariate local time window series, Represents a local peak candidate set. Represents the relative time position index in the event feature sequence. Indicates the event feature sequence at the 1st The value at a relative time position , These represent the event feature sequences at the th, respectively. The feature values ​​corresponding to the consecutive time positions. Indicates the number of neighbors within the range. The feature values ​​corresponding to each position Indicates the peak significance threshold. Indicates the peak detection neighborhood width. This represents the empty set.

[0037] The determination is based on the change in the sign of the second-order difference and the amplitude of the curvature change, and is defined as the inflection point intensity index within the local time window, namely: ; in, Indicates an indicator function, , These represent the event feature sequences at the th, respectively. At the relative time position, the first The feature values ​​corresponding to each relative time position Indicates the sampling time interval. This indicates the threshold for determining the inflection point.

[0038] The rate of change is preferably defined as the maximum absolute value of the first-order difference within the local time window, i.e.: ; To characterize changes in acceleration or curvature, it is preferably defined as the maximum value of the absolute value of the second-order difference within the local time window, i.e.: ; It is used to characterize the average rise intensity over a fixed time span, and is defined as follows: ; in, This indicates the width of the window used to calculate the rate of increase. Indicates the event feature sequence at the 1st The feature value at the i-th relative time position, i.e., the i-th The position is the endpoint, and the forward interval is... The feature values ​​corresponding to each time step.

[0039] (Normalized change point location) is used to identify the location of the dominant stage transition within a local time window. The preferred method for determining the location of the main change point is the single-change point minimum piecewise residual method, i.e.: ; ; in, This represents the value of the independent variable that minimizes the objective function. This indicates the location of the principal point of change that minimizes the sum of squares of the piecewise residuals. Indicates the event feature sequence at the 1st Feature values ​​at relative time positions , Representing intervals Interval The mean.

[0040] S1-4-3. Calculate the alignment weight based on temporal neighborhood and data quality characterization information; It should be explained that the alignment weight is used to comprehensively characterize the degree of temporal proximity and the reliability of data quality, and it is defined as follows: ; ; in, This represents the index of candidate time windows within the alignable time neighborhood. This represents the index of the candidate time window traversed during the normalization summation process. Indicates the first The first time window and environmental monitoring data Alignment weights between candidate time windows This represents an exponential function with base e. This represents the data quality characterization information corresponding to the modalities of structural response monitoring data. , These represent the environmental monitoring data corresponding to the mode in the 1st... The candidate time window, the first Data quality characterization information for each time window Indicates the time decay scale parameter, Indicates the alignable temporal neighborhood. This indicates the maximum allowed time offset range for alignment. In S1-4, the... The time window is the target time window for the optimization process.

[0041] S1-4-4. Construct an event consistency alignment loss based on standard samples from the same window, alignment weights, and characteristics of each cross-source event; Traditional interpolation methods force the alignment of different modal data according to the original timestamps. However, different sensors have different sampling times, transmission delays, and response lags, which can easily force physical events that should occur sequentially to the same moment, disrupting causal logic. While DTW (Dynamic Time Warping) allows for time axis scaling, its alignment results are mainly based on numerical shape similarity and are easily affected by noise, local jitter, and amplitude differences, leading to incorrect matching or omission of key events (such as peaks, inflection points, and abrupt change start and end points). To avoid event misalignment caused by relying solely on interpolation or DTW, this embodiment constructs an event consistency alignment loss, so that the alignment results truly reflect the inherent causal temporal relationship between structural response and environmental disturbances.

[0042] Event Consistency Alignment Loss The corresponding formula is: .

[0043] S1-4-5. Optimize the standard samples of the same window based on the event consistency alignment loss to generate a unified monitoring dataset.

[0044] Specifically, the first corresponding to the structural response monitoring data Using a time window as the target time window, within the alignable time neighborhood of the environmental monitoring data. Internal search There are several candidate time windows, and the alignment cost of each candidate time window is determined based on the differences in event features and alignment weights. ; in, The first part representing the structural response monitoring data The first time window and environmental monitoring data Alignment cost between candidate time windows.

[0045] Calculate the normalized alignment coefficient of the candidate time window based on the alignment cost: ; in, The first part representing environmental monitoring data The candidate time window is relative to the structural response monitoring data. Normalized alignment coefficients for each time window; The first structural response monitoring data The first candidate time window and the environmental monitoring data Alignment cost between candidate time windows; The temperature coefficient represents the alignment cost and is used to adjust the concentration of the candidate time window weight distribution.

[0046] Subsequently, the environmental monitoring data were weighted and fused according to the normalized alignment coefficient to obtain the result that is consistent with the structural response monitoring data. The formula for the aligned environmental monitoring samples matched within a time window is: ; in, This represents the aligned environmental monitoring samples. This indicates that environmental monitoring data is in the first... A standard sample of classmates under a candidate time window.

[0047] Finally, the standard sample of the structural response monitoring data in the t-th time window was used. Aligned environmental monitoring samples Standard samples of classmates corresponding to inspection image data And the corresponding standard sample of operation and maintenance management data. Together they constitute the samples in the unified monitoring dataset : ; Among them, the inspection image data and operation and maintenance management data are low-frequency discrete event data. Their temporal correlation has already been mapped through a unified time window during the co-organization phase. Therefore, continuous event feature alignment will not be performed in this step. This generates a unified monitoring dataset. : ; Through the above process, structural response monitoring data and environmental monitoring data can be aligned at the event level, enabling the unified monitoring dataset to more accurately reflect the causal and temporal relationship between structural response and environmental disturbances.

[0048] S2. Based on the unified monitoring dataset, determine the graph nodes and edges, and construct a comprehensive adjacency matrix by combining the credibility coefficient; S2 includes: S2-1. Take each monitoring point and component (facility unit) of the urban lifeline infrastructure to be evaluated as a graph node, and determine the edge and multi-relationship basic adjacency matrix according to the physical spatial relationship between the graph nodes; the physical spatial relationship includes topological connection relationship, spatial adjacency relationship and physical coupling relationship.

[0049] Specifically, based on the monitoring deployment method, structural composition, and functional unit division of the urban lifeline infrastructure to be evaluated, the actual monitoring objects are uniformly abstracted as graph nodes, and a graph node set is constructed. The graph nodes can be monitoring points such as settlement monitoring points, displacement monitoring points, pressure monitoring points, and liquid level monitoring points, or they can be pipe sections, cross-sections, manholes, road units, bridge components, or other facility units with independent functional attributes and risk representation significance. Through this abstraction process, the original multi-source monitoring objects are uniformly mapped to the graph structure expression space, providing a unified carrier for subsequent edge relationship construction, spatial association modeling, and risk propagation analysis.

[0050] If a graph node corresponds to a single monitoring point (such as a settlement point, displacement point, or pressure point), then that monitoring point is directly used as the graph node, and its monitoring data is used as the node attribute. If a node corresponds to a component-level object (such as a section of pipeline or a manhole) or a facility unit-level object (such as a bridge or a road section), then all relevant monitoring quantities within that component or unit (such as multi-point settlement, strain, pressure, etc.) are aggregated and represented (e.g., weighted average, statistical features, or vector concatenation) and used as the attribute input for that graph node. This allows for flexible establishment of graph structure models at the "monitoring point level," "component level," or "facility unit level" according to actual engineering scenarios, thereby improving the applicability to different types of urban lifeline infrastructure.

[0051] To characterize the multidimensional relationships between different nodes in urban lifeline infrastructure, a corresponding multi-relationship basic adjacency matrix is ​​determined based on the physical spatial relationships of graph node sets. ,Right now: ; in, Indicates the total number of relation types. Indicates the first The basic adjacency matrix corresponding to class relationships. In this embodiment, The value is 3, which represents topological connectivity, spatial adjacency, and physical coupling.

[0052] Topological connectivity is used to characterize the objectively existing structural connections, transmission connections, upstream and downstream connections, support connections, or functional transfer connections between facilities. For example, in lifeline networks such as water supply, drainage, and transportation, when the topological connectivity is... The graph node and the first When there are direct connections, sequential connections, or explicit functional transfer relationships between nodes in a graph, topological connection edges are constructed. For directional transport networks, topological connection relationships can also be constructed as directed adjacency relationships to reflect the characteristics of risk propagation along the network direction.

[0053] Spatial adjacency is used to characterize the proximity of graph nodes in terms of spatial location. For example, the geometric center of a monitoring point, component, or facility unit can be used as the spatial representative point of the node. The graph node and the first The center distance of each graph node is used as Spatial adjacency relationships are constructed based on center distance. In this embodiment, spatial adjacency relationships can be determined using a distance threshold method or a distance decay method, for example, based on: ; The basic edge weights for calculating the spatial adjacency relationships between any two distinct nodes .in, This represents the distance attenuation scale parameter. When... The closer the edges are, the larger their basic edge weights; conversely, the farther apart they are, the smaller their basic edge weights.

[0054] Physical coupling relationships are used to characterize the potential mutual influence between graph nodes under structural response, operational status, or environmental disturbances. The physical coupling strength is constructed based on the correlation, statistical dependence, and engineering mechanism association of monitoring response sequences between pairs of graph nodes. ,Right now: ; in, , , They represent the first The graph node and the first The graph nodes' normalized correlation indices (correlation indices can be calculated using Pearson correlation coefficient, Spearman rank correlation coefficient, or time-delay correlation coefficient, etc., and obtained after normalization), normalized mutual information indices, and mechanistic correlation coefficients determined according to engineering mechanism rules. , , Represents the weight coefficient and .

[0055] Topological connectivity, spatial adjacency, and physical coupling are used as edge weights between corresponding graph nodes to construct the edges. All edges and their weights are then organized to construct a multi-relationship basic adjacency matrix. .

[0056] S2-2. Based on data quality characterization information, graph nodes, and the basic adjacency matrix of multiple relations, calculate the confidence coefficient and the confidence edge matrix; To utilize the aforementioned data quality characterization information for graph structure reliability modeling, this embodiment aggregates the monitoring points, component attributes, and corresponding modal data quality values ​​associated with graph nodes into the data quality characterization information of the graph nodes. If a graph node corresponds to a single monitoring point, the data quality value of the modality to which that monitoring point belongs can be used as the data quality characterization information of that graph node; if a graph node corresponds to a component or facility unit, a weighted average, minimum value constraint, or importance-weighted fusion of the data quality values ​​of multiple monitoring points or multiple modalities within it can be performed to obtain the data quality characterization information of that graph node.

[0057] To reduce the interference of noise, missing values, outliers, or local distortions on the graph structure modeling results, the first... In the graph of class relationships, a confidence coefficient is introduced into the basic edge weights between nodes to correct the edges, resulting in a confidence edge matrix. Specifically, for the ... Normalize the basic adjacency matrix of multiple relations corresponding to nodes in the pairwise graph under class relations to obtain the normalized relation strength. Then, based on the relationship sensitivity coefficients of the graph nodes and the data quality characterization information, according to: ; Calculate the first The graph node and the first Reliability coefficient of each graph node .in, This represents the Sigmoid function. Indicates the first Sensitivity coefficient under class relationship , They represent the first The first graph node, the first Data quality characterization information for each graph node and Let and represent the adjustment coefficients for the edge reliability of the data quality of the two nodes at the two ends of the edge under the r-th type of relation, respectively, used to control the edge reliability. The graph node and the first Data quality characterization value of each graph node , Its contribution to credibility calculation. For undirected relations, we can let... = For directed relationships, different values ​​can be set based on upstream / downstream relationships or the direction of risk propagation, and can be calibrated using historical event samples. The sensitivity coefficient is used to characterize the explanatory power and importance of different relationship types in risk propagation: the higher the sensitivity coefficient, the stronger the characterization ability of a certain type of relationship in risk propagation; the lower the sensitivity coefficient, the more auxiliary the relationship, or the less stable the relationship.

[0058] Then, according to the formula: ; ; Calculate the reliable edge matrix .in, Indicates the first under class relation The graph node and the first Trusted edge weights between graph nodes Indicates the first under class relation The graph node and the first The original edge weights between nodes in a graph based on a multi-relational adjacency matrix.

[0059] S2-3. Calculate the relation weights and combine them with the reliable edge matrix to construct a comprehensive adjacency matrix.

[0060] To adapt to the dynamic changes in the risk propagation mechanism of urban lifeline infrastructure under different environmental conditions, a comprehensive adjacency matrix is ​​constructed by weighting and fusing the trust edge matrices corresponding to various relationships based on the trust edge matrix and environmental conditions. The corresponding formula is: ; in, Indicates the first Environmental condition feature vectors under a time window (which can be composed of environmental monitoring indicators such as rainfall intensity, cumulative rainfall, groundwater level, temperature, humidity, water immersion status, and gas concentration). Indicates the first The reliable edge matrix of class relationships, Indicates the relation weight.

[0061] Preferably, the relationship weights are determined based on environmental conditions. First, the weights are calculated based on the current values ​​or classification intervals of each environmental factor. Class Relationship Environmental response score under each time window Then, the environmental response scores for each relationship are normalized, and finally, the result is calculated according to the formula: ; Calculate the relation weights. Wherein, Indicates the first Class relationship at the current moment The environmental response score is as follows.

[0062] In S2-3, instead of performing isolated secondary corrections on individual edges based on the various trusted edge matrices obtained in S2-2, the trusted edge matrices are directly used as input. The relationships are dynamically weighted and fused according to the strength of the current environmental conditions' influence on different relationship types (such as topological connections, spatial adjacency, and physical coupling). The generated comprehensive adjacency matrix can be used for subsequent spatial message passing, health index calculation, and failure probability inference, enabling a shift from "static multi-relationship modeling" to "trustworthy and time-varying adaptive graph modeling."

[0063] S3. Based on the comprehensive adjacency matrix, extract the infrastructure status representation vector of the unified monitoring dataset, and calculate the corresponding infrastructure health index and failure probability. S3 includes: S3-1. Based on the comprehensive adjacency matrix, multi-source feature encoding is performed on the unified monitoring dataset to obtain multi-source encoded features; the multi-source encoded features include temporal features, image features, and spatial correlation features; The formula corresponding to S3-1 is: ; ; ; in, , , These represent temporal features, image features, and spatial correlation features, respectively. , They represent the first The time window ends with a time window length of [number]. The standard sample corresponding to the substructure response monitoring data and environmental monitoring data. , , These represent the temporal feature extraction function, the image feature extraction function, and the spatial relationship feature extraction function, respectively. Indicates the first Each time window represents a standard sample of the same time window corresponding to the inspection image data at the end of the period. Indicates the first The time window ends with a time window length of [number]. The standard sample of classmates. Preferably, the time-series feature extraction function. It can be implemented using long short-term memory networks, gated recurrent units, one-dimensional convolutional networks, or temporal Transformer encoders to extract temporal evolution features from structural response monitoring data and environmental monitoring data; image feature extraction function. Convolutional neural networks, ResNet networks, or visual Transformer networks can be used to extract the destructive features of cracks, leaks, subsidence, and damage from inspection images; spatial relationship feature extraction functions. It can be implemented using graph convolutional networks, graph attention networks, or graph spatiotemporal networks, and is used to synthesize adjacency matrices. Extract spatial association features between different graph nodes under constraints.

[0064] S3-2. Construct cross-modal consistency constraints, perform feature semantic alignment on multi-source encoded features, and generate multi-source semantic temporal features; S3-2 includes: S3-2-1. Project the multi-source encoded features onto the common semantic space to construct cross-modal consistency constraints; Specifically, the temporal features in multi-source semantic temporal features Image features By mapping each element to the common semantic space using a projection function, we obtain: ; ; in, , These represent the projected temporal features and image features, respectively. , These represent the time-series projection function and the image projection function, respectively.

[0065] Then, according to the formula: ; Constructing cross-modal consistency constraints .in, This indicates the event pairing indicator, used to characterize whether the temporal features and image features under the current time window belong to the same facility object and meet the consistency constraint triggering condition. When the same facility object has both matchable structural response anomalies and image defects within the current time window or the aligned time neighborhood, then... ,otherwise . The consistency constraint weight represents the reliability of cross-modal consistency constraints within the current time window. Preferably, it is determined jointly based on the data quality and modal uncertainty of both the time-series and image modalities. The expression is: ; in, , These represent the data quality representation information corresponding to time-series features and image features, respectively. This represents a stable term to prevent the denominator from being zero. , These represent the modal uncertainty estimates corresponding to time-series features and image features, respectively.

[0066] S3-2-2. Using cross-modal consistency constraints as the optimization objective, iterate the multi-source coding features to generate multi-source semantic temporal features.

[0067] Specifically, with cross-modal consistency constraints as the optimization objective, the Adam gradient descent optimization algorithm can be used to iteratively update the network parameters of the temporal projection function and the image projection function, mapping temporal features and image features to the same common semantic space, so that temporal features and image features under the same time window maintain semantic consistency in this space; after the model training converges, the projection parameters are fixed, and the semantically aligned bimodal features (aligned temporal features and image features) are directly output during the inference stage, providing stable and consistent feature input for subsequent multimodal weighted fusion based on data quality and modal uncertainty, reducing false alarms caused by single-modal noise or semantic contradictions from the source.

[0068] The multi-source semantic temporal features include spatial association features, aligned temporal features, and image features.

[0069] S3-3. Set the gating factor, calculate the fusion weight under different modalities, and fuse the multi-source semantic temporal features in combination with data quality representation information to generate a unified state representation vector. Specifically, considering that data from different modalities may contain missing values, delays, noise interference, local anomalies, or output fluctuations within the same time window, data quality constraints and modal uncertainty constraints are introduced into the features of each modality to ensure that the fusion result reflects the availability and stability of each modality at the current moment. This embodiment defines the first... The mode in the th ... Gating factors within a time window for: ; in, Indicates the first The first time window Modal uncertainty estimation for each mode, Indicates the first The first time window Data quality characterization information for each modality denoted as a constant. Modal uncertainty estimation can be determined by the sampling variance, residual fluctuation, confidence dispersion, or other statistics that can characterize the stability of the feature under the current time window for the corresponding mode, and preferably obtained after normalization to reflect the fluctuation and confidence level of the feature output of the mode.

[0070] According to the formula: ; ; Generate a unified state representation vector .in, Indicates the first The first time window The fusion weights of each modality Indicates the total number of modes. , They represent the first The modality, the first The key vector of each modality Indicates the first The first time window Multi-source semantic temporal features of each modality Indicates the first The mode in the th ... Gating factors within a time window.

[0071] In this step, data quality characterization information and modal uncertainty are used simultaneously to adjust the fusion weights because different modalities in real-world engineering scenarios not only differ in data collection completeness and timeliness but also in output stability and reliability. By jointly introducing these two indicators, the fusion process can assign higher weights to modalities with high quality and stable output, while appropriately reducing the weights of modalities with missing data, large fluctuations, local distortions, or short-term anomalies. This helps suppress false alarms caused by single-modal anomalies, local noise, instantaneous spikes, short-term occlusion, or distortions, and avoids sudden changes in risk levels due to short-term anomalies in a single modality. This improves the stability, consistency, and engineering applicability of the unified state characterization in continuous monitoring scenarios, and provides more reliable input for subsequent health index calculations, failure probability inferences, and early warning determinations.

[0072] S3-4. Calculate the anomaly distance between the unified state representation vector and the normal baseline, and enhance the consistency of the graph structure space to generate the infrastructure health index. Specifically, to ensure that the infrastructure health index has engineering significance regarding the degree of deviation from the normal baseline, a normal representation distribution is learned based on historical data of stable operating conditions, and anomaly distance is defined. for: ; in, This represents the mean of the normal representation distribution. Let represent the inverse covariance matrix of the constant characterization distribution.

[0073] Considering that urban lifeline infrastructure focuses more on clustered or propagated anomalies than isolated noise point anomalies, graph structure spatial consistency enhancement is applied to anomaly distances to obtain graph smoothness anomaly degree, with the corresponding formula being: ; Then, according to the formula: ; Calculate the infrastructure health index A lower infrastructure health index value indicates a more significant deviation from the normal baseline in facility condition and a higher level of risk. Among these, Indicates the smoothness anomaly of the graph. , They represent the first The first graph node, the first The abnormal distances of each graph node from the normal baseline. Indicates the smoothing enhancement coefficient. Represents the comprehensive adjacency matrix The Middle The graph node and the first The values ​​between the nodes of the graph This represents the mapping coefficient.

[0074] S3-5. Infer and calibrate the infrastructure health index through a probabilistic reasoning model to generate the failure probability.

[0075] The processing steps of the probabilistic reasoning model include: S3-5-1. Construct risk inference input vectors for different graph nodes, i.e., based on: ; Construct the first Risk inference input vector for each graph node . , They represent the first The first time window Unified state representation vector and modal uncertainty estimation for each graph node This indicates the characteristics of operation and maintenance management.

[0076] S3-5-2. Based on the input vectors of each risk inference, calculate the residual resistance and effect of different graph nodes. The corresponding formula is: ; ; in, , These represent the resistance estimation function (used to estimate remaining resistance based on the current state of the facility, its health level, operational history, and data reliability) and the effect estimation function (used to estimate current risk demand based on environmental disturbances, anomalies, uncertainties, and the propagation effect from neighboring nodes), respectively. Preferably, the resistance estimation function... and effect estimation function It can be implemented using linear regression models, ridge regression models, Bayesian regression models, random forest models, gradient boosting tree models, or multilayer perceptron models, and can be trained or calibrated based on historical monitoring data, historical risk events, operation and maintenance records, or expert-annotated samples. , They represent the first The residual resistance (residual safety margin) and effect (risk requirement) of each graph node.

[0077] S3-5-3. Based on the residual resistance and action effect of different graph nodes, construct the corresponding limit state function values, and the corresponding formula is: ; in, This represents the limit state function value.

[0078] Furthermore, to reflect the amplifying effect of anomaly propagation from adjacent nodes and modal uncertainty on failure risk, the limit state function value can be extended to determine an improved limit state function value, with the corresponding formula being: ; in, , These represent the anomaly propagation penalty coefficient and the uncertainty penalty coefficient, respectively, used to characterize the weakening effect of network propagation anomalies and modal uncertainties on the security margin of the current node. Indicates the first Graph smoothness anomaly of each graph node.

[0079] S3-5-4. Calculate the failure probability of each node in the diagram based on the value of each limit state function and the reliability index.

[0080] Calculate the mean value of the limit state function. and standard deviation Then, according to the formula: ; Calculate the first Reliability indicators of individual graph nodes .

[0081] Reliability indicators are processed using the cumulative distribution function of the standard normal distribution, i.e., based on: ; Calculate the first Initial failure probability of each graph node .in, The cumulative distribution function represents the standard normal distribution.

[0082] Since model bias and engineering data bias can affect the probability, this embodiment calibrates the failure probability using historical event samples or expert-annotated data to obtain the failure probability. The corresponding formula is: ; in, , All represent probability calibration parameters. This represents the inverse function of the cumulative distribution function of the standard normal distribution. Probability calibration parameters can be determined by fitting data based on historical event samples, handling results, or expert annotation information.

[0083] S4. Based on the infrastructure health index and failure probability, calculate the risk level and generate early warning trigger results.

[0084] Specifically, the infrastructure health index and failure probability are processed using a Bayesian risk model to generate risk level results; To align with the "cost-benefit-risk" decision-making logic in urban lifeline management, the Bayesian risk model employs cost-sensitive Bayesian decision-making for risk classification, with the corresponding formula being: ; in, Indicates the first The graph node at the th Risk level results within a specific time window Indicates the probability of failure. This represents the value of the independent variable that minimizes the objective function. This indicates the resource cost of initiating a response level. Indicates the first The cost-sensitive risk loss function corresponding to the response level is used to characterize the application of the first response level under the current health index and failure probability conditions. The remaining risk or losses due to inadequate handling may arise after each response level. Indicates the first The graph node at the th Infrastructure health index under a specific time window Indicates the first The graph node at the th Failure probability within a time window.

[0085] Furthermore, after obtaining the risk levels of each node in the graph, the overall risk level of the infrastructure to be assessed can be generated based on the worst-case scenario principle or the node importance weight. Preferably, the worst-case scenario principle is used to determine the overall risk level: ; in, The city's lifeline infrastructure to be evaluated is in the [number]th [year]. The overall risk level under each time window can prevent local high-risk nodes from being masked by averaging.

[0086] To filter out sensor noise, single-modal false alarms, or transient risk level jumps caused by local fluctuations, and to avoid overwhelming maintenance personnel with invalid warnings, warnings should only be triggered for continuous and stable real risk events. This improves the system's engineering practicality and warning reliability. Therefore, it is necessary to trigger warnings in conjunction with preset conditions, introducing a continuous or hysteresis mechanism. Preferably, warnings should only be triggered when continuous... Triggered when a threshold condition is met within a time window, i.e.: ; in, Indicates time The risk level results below, Indicates the preset trigger threshold level. This represents the minimum number of times the trigger condition must be met.

[0087] When the above preset conditions are met, a corresponding early warning trigger result is generated. The early warning trigger result includes at least the early warning object, the overall risk level, the trigger time window, and the scope of risk impact, which is used to provide a basis for the generation of structured early warning information, the dominant risk chain, and operation and maintenance handling suggestions in S5.

[0088] In addition, this application also includes S5, which generates a scenario-based operation and maintenance handling process based on the early warning trigger result and performs closed-loop feedback.

[0089] S5 includes: S5-1. Generate a response scenario card based on the early warning trigger result; Specifically, after an early warning trigger result is generated in S4, the system does not directly output the handling action. Instead, it converts the early warning trigger result into a handling scenario card. The handling scenario card describes the specific handling context of the current risk event in engineering operation and maintenance, and includes at least the risk object, facility type, risk level, risk type, trigger time window, spatial location, scope of impact, key abnormal indicators, dominant risk chain, on-site operating status, and management constraint information.

[0090] The risk object is determined by the graph node that triggers the early warning or the set of adjacent high-risk graph nodes; the facility type is determined based on the facility attributes in the infrastructure monitoring database; and the risk type is determined based on key anomaly indicators and the dominant risk chain. In specific implementation, the system can generate risk types according to preset judgment rules. For example, when rainfall or groundwater level rises, pipeline liquid level or pressure is abnormal, and adjacent node settlement anomalies occur simultaneously in environmental monitoring data, the current event is classified into the "leakage and water accumulation - settlement linkage" scenario; when settlement, displacement, or tilt angle exceeds the rate of change threshold for multiple consecutive time windows, and adjacent components have synchronous anomalies, the current event is classified into the "structural deformation development" scenario; when pressure, flow, or liquid level changes abruptly and is accompanied by abnormal propagation from upstream and downstream nodes, the current event is classified into the "abnormal operating pressure" scenario.

[0091] Furthermore, the system retrieves engineering constraint information corresponding to the risk object from the infrastructure monitoring database and operation and maintenance management data. This engineering constraint information includes the facility's ownership unit, operating status, accessibility, location of nearby valves or switches, traffic impact range, confined space operation attributes, whether gas or electrical safety is involved, available personnel and equipment, historical maintenance records, and applicable operation and maintenance specifications. This generates a response scenario card. Scenario Card for Handling Risks = {Risk Object, Facility Type, Risk Level, Risk Type, Dominant Risk Chain, Spatial Scope, Operational Status, Safety Constraints, Resource Constraints, Review Requirements}.

[0092] By using the above method, the aforementioned risk assessment results are transformed from simple risk levels into engineering scenarios that can be identified by operation and maintenance rules, avoiding a direct jump from risk levels to abstract handling actions.

[0093] S5-2. Match handling rules based on handling scenario cards and perform hard constraint verification; Specifically, the system pre-establishes a scenario-based handling rule library. Each handling rule includes at least the applicable facility type, applicable risk type, applicable risk level, triggering conditions, prohibition conditions, prerequisite safety requirements, resource requirements, recommended handling process, responsible department, and completion criteria.

[0094] When performing rule matching, the system uses the disposal scenario card generated by S5-1 as input, sequentially retrieving disposal rules that meet the conditions of facility type, risk type, and risk level, and further determining whether the triggering conditions of the rule are consistent with the dominant risk chain and key anomaly indicators. For example, for the scenario of "drainage network - leakage and water accumulation - settlement linkage", if the risk level is medium, the disposal rule of "on-site verification - pipeline video detection - intensified settlement monitoring - drainage diversion - local repair" will be matched first; if the risk level is high, disposal steps such as "impact area warning, traffic diversion, emergency linkage and emergency repair resource scheduling" will be added before the above process. For the scenario of "gas pipeline network - pressure anomaly - suspected leakage", the system does not directly output the "excavation and emergency repair" action, but first matches the safety priority process of "warning and isolation - leakage detection - valve positioning - shutdown approval or emergency shutdown - ventilation detection - emergency repair and testing".

[0095] Furthermore, before outputting the rules, the system performs hard constraint verification on the candidate disposal rules. This hard constraint verification includes at least safety procedure verification, work permit verification, resource availability verification, and site condition verification. Specifically, safety procedure verification determines whether the disposal rule meets safety requirements such as gas, electricity, confined space, and traffic control; work permit verification determines whether a permit for power outage, road closure, excavation, or confined space work is required; resource availability verification determines whether personnel, equipment, materials, and collaborating departments have the necessary conditions for execution; and site condition verification determines whether the risk object is accessible, whether there are secondary risks, and whether manual review is required first.

[0096] If a candidate disposal rule meets the aforementioned hard constraints, it enters the disposal process orchestration; otherwise, the disposal rule is not directly output, but rather the pre-processing requirements are output. For example, when a high-risk gas scenario does not meet the conditions for shut-off or work permits, the pre-processing steps of "on-site warning, leak verification, emergency response coordination, and shutdown approval" are generated first; when road subsidence risk disposal requires road closure but traffic diversion conditions are not met, the system first generates the pre-processing steps of "traffic impact assessment, temporary fencing, and coordination with traffic management departments." In this way, disposal recommendations are constrained by safety procedures, resource allocation, and on-site conditions, rather than being mathematically selected solely based on implementation costs and residual risks.

[0097] S5-3. Generate a work order-based processing flow and update the rule base based on feedback results; Specifically, the system will organize the handling rules verified by hard constraints into a work order-based handling process. The work order-based handling process is not a single action, but consists of several handling steps with sequential order and triggering conditions. Each handling step includes at least the handling name, execution object, execution location, responsible unit, prerequisites, required resources, completion deadline, safety requirements, and acceptance criteria.

[0098] Preferably, the work order-based handling process is generated in the order of "review and confirmation - risk control - engineering handling - effect retesting - resolution or escalation". Specifically, the review and confirmation stage is used to confirm the authenticity of the risk through manual inspection, video review, drone verification, special testing, or third-party testing; the risk control stage is used to control the risk from escalating by taking measures such as increased monitoring, temporary fencing, drainage diversion, pressure regulation, load and speed limits, segmented valve closure, or traffic diversion; the engineering handling stage is used to perform leak sealing and repair, temporary reinforcement, component replacement, pressure relief and hazard mitigation, emergency repairs, or special rectification; the effect retesting stage is used to review the health index, failure probability, key abnormal indicators, and on-site review results; and the resolution or escalation stage is used to determine whether the risk has been closed-loop.

[0099] In practice, if the overall risk level is below the warning threshold for several consecutive time windows after the incident, and the key abnormal indicators have recovered to the allowable range, and the on-site verification results meet the acceptance criteria, the system will mark the event as risk resolved. If the risk level does not decrease after the incident, the probability of failure continues to increase, or the dominant risk chain continues to expand, the system will automatically generate an upgraded handling process, including raising the response level, expanding the scope of impact for verification, adding coordinating departments, dispatching emergency repair resources, or switching to emergency response.

[0100] After the handling is completed, the system writes back the actual handling measures, execution time, executing unit, resource consumption, on-site verification results, risk resolution status, subsequent monitoring changes, and manual review conclusions to the infrastructure monitoring database. This feedback data is used to update historical risk event samples, correct risk classification thresholds, calibrate failure probability parameters, and adjust the applicable conditions, completion criteria, and priorities of corresponding rules in the scenario-based handling rule base. Through this method, a closed-loop handling process is formed from "early warning triggering - scenario identification - rule matching - constraint verification - process execution - effect feedback," enabling early warning results to be transformed into executable instructions that conform to on-site operation and maintenance specifications.

[0101] In summary, this invention effectively suppresses anomalies and missing data in multi-source monitoring data by constructing a unified data identification system and a quality-weighted joint repair target, thereby improving data continuity and availability. Building upon this, it introduces a multi-relationship graph structure and trusted edge modeling, combined with adaptive fusion based on environmental conditions to form a comprehensive adjacency matrix, enabling dynamic representation of facility topology, spatial neighborhood, and physical coupling relationships. Furthermore, it employs a cross-modal consistency constraint and quality-uncertainty gating fusion mechanism to generate a stable and unified state representation vector, and combines health indices and calibration failure probabilities for probabilistic reasoning, significantly improving the accuracy and interpretability of risk assessment. Simultaneously, it utilizes cost-sensitive Bayesian decision-making for risk classification, combined with a persistent triggering mechanism to suppress false alarms, ultimately outputting early warning information containing the dominant risk chain and optimized handling suggestions, forming a closed loop from data governance to operation and maintenance decision-making. This effectively reduces reliance on manual inspections and enhances the intelligence, refinement, and engineering practicality of risk management for urban lifeline infrastructure.

[0102] like Figure 2 As shown, an urban lifeline risk assessment and early warning system based on multi-source monitoring fusion includes: The data generation module is used to acquire multi-source monitoring data of the city's lifeline infrastructure to be evaluated, generate peer samples and data quality characterization information, process the peer samples in combination with quality time series constraint objectives and event consistency alignment loss, and generate a unified monitoring dataset. The matrix construction module is used to determine graph nodes and edges based on a unified monitoring dataset and to construct a comprehensive adjacency matrix by combining the credibility coefficient. The calculation module is used to extract the infrastructure status representation vector of the unified monitoring dataset based on the comprehensive adjacency matrix, and calculate the corresponding infrastructure health index and failure probability. The early warning module is used to calculate the risk level based on the infrastructure health index and failure probability, and generate early warning trigger results.

[0103] The data generation module includes: The sample construction unit is used to acquire multi-source monitoring data of the lifeline infrastructure of the city to be evaluated, and generate peer samples by combining the data representation system. The quality calculation unit is used to calculate the corresponding quality characterization value based on multi-source monitoring data and generate data quality characterization information. The repair unit is used to construct quality time-series constraint targets based on facility physical constraint penalty terms and data quality characterization information, and to perform anomaly detection and joint repair on peer samples to generate peer repair samples. The alignment unit is used to standardize the peer repair samples and perform cross-source time-series alignment by combining the event consistency alignment loss to generate a unified monitoring dataset.

[0104] The urban lifeline risk assessment and early warning system also includes a closed-loop response module, which is used to generate response scenario cards based on the early warning trigger results, match scenario-based response rules and perform hard constraint verification, generate work order-style response processes, and update historical risk event samples, risk classification thresholds, failure probability calibration parameters and scenario-based response rule library according to the response feedback results.

[0105] It should be noted that the specific methods by which each module performs operations in the system described in the above embodiments have been described in detail in the embodiments related to the method, and will not be elaborated here.

[0106] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0107] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for risk assessment and early warning of urban lifelines based on multi-source monitoring fusion, characterized in that, include: Acquire multi-source monitoring data of the city's lifeline infrastructure to be evaluated, generate peer samples and data quality characterization information, process the peer samples by combining quality time series constraint objectives and event consistency alignment loss, and generate a unified monitoring dataset; Based on a unified monitoring dataset, graph nodes and edges are determined, and a comprehensive adjacency matrix is ​​constructed by combining the credibility coefficient. Based on the comprehensive adjacency matrix, the infrastructure status representation vector of the unified monitoring dataset is extracted, and the corresponding infrastructure health index and failure probability are calculated. Based on the infrastructure health index and failure probability, the risk level is calculated, and an early warning trigger result is generated.

2. The urban lifeline risk assessment and early warning method based on multi-source monitoring fusion according to claim 1, characterized in that, The generation of the unified monitoring dataset includes: Acquire multi-source monitoring data of the city's lifeline infrastructure to be evaluated, and generate peer samples by combining the data representation system; Based on multi-source monitoring data, the corresponding quality characterization values ​​are calculated, and data quality characterization information is generated. Based on the facility physical constraint penalty term and data quality characterization information, a quality time series constraint target is constructed, and anomaly detection and joint repair are performed on peer samples to generate peer repair samples. Standardize the peer repair samples and peer samples, and combine them with the event consistency alignment loss to perform cross-source time series alignment to generate a unified monitoring dataset.

3. The urban lifeline risk assessment and early warning method based on multi-source monitoring fusion according to claim 2, characterized in that, The method of combining event consistency alignment loss for cross-source time series alignment to generate a unified monitoring dataset includes: Standardize the peer repair samples and peer samples to generate peer standard samples; Define a temporal neighborhood and extract cross-source event features corresponding to structural response monitoring data and environmental monitoring data through an event feature extraction function; Alignment weights are calculated based on temporal neighborhood and data quality characterization information. Based on standard samples from the same class, alignment weights, and characteristics of each cross-source event, an event consistency alignment loss is constructed. The standard sample of the same class is optimized based on the event consistency alignment loss to generate a unified monitoring dataset.

4. The urban lifeline risk assessment and early warning method based on multi-source monitoring fusion according to claim 2, characterized in that, The physical constraints and penalties for the facilities include range constraints, upper limits for settlement and displacement rates, non-negativity of pore pressure constraints, and unexplainable jump constraints. The range constraint is used to limit the values ​​of the unified monitoring dataset to a physically feasible range jointly determined by the sensor's nominal range, the engineering allowable range, and the historical normal operating condition statistical interval. The upper limit constraint on settlement and displacement rate is used to limit the settlement or displacement change rate between adjacent time windows from not exceeding the allowable upper limit. The non-negative pore pressure constraint is used to limit the repair result to no less than zero in scenarios where the monitored object is the pore water pressure of a saturated medium or the hydraulic pressure defined by non-negative gauge pressure, so as to avoid negative values ​​that do not conform to physical meaning. The unexplainable jump constraint is used to suppress abnormal deviations of the repair results from the local trend prediction values.

5. The urban lifeline risk assessment and early warning method based on multi-source monitoring fusion according to claim 1, characterized in that, The construction of the comprehensive adjacency matrix by combining the credibility coefficient includes: Each monitoring point and component of the urban lifeline infrastructure to be evaluated is taken as a graph node, and the edge and multi-relationship basic adjacency matrix is ​​determined based on the physical spatial relationship between the graph nodes. Based on data quality characterization information, graph nodes, and multi-relational basic adjacency matrix, the credibility coefficient and credibility edge matrix are calculated. Calculate the relation weights and combine them with the trusted edge matrix to construct a comprehensive adjacency matrix.

6. The urban lifeline risk assessment and early warning method based on multi-source monitoring fusion according to claim 1, characterized in that, The calculation of the corresponding infrastructure health index and failure probability includes: Based on the comprehensive adjacency matrix, multi-source feature encoding is performed on the unified monitoring dataset to obtain multi-source encoded features; Construct cross-modal consistency constraints, perform feature semantic alignment on multi-source encoded features, and generate multi-source semantic temporal features; Set a gating factor, calculate the fusion weights under different modalities, and fuse multi-source semantic temporal features in combination with data quality representation information to generate a unified state representation vector. Calculate the anomalous distance between the unified state representation vector and the normal baseline, and enhance the consistency of the graph structure space to generate an infrastructure health index; The infrastructure health index is inferred and calibrated using a probabilistic inference model to generate failure probabilities.

7. The urban lifeline risk assessment and early warning method based on multi-source monitoring fusion according to claim 6, characterized in that, The generation of failure probabilities includes: Construct risk inference input vectors for different graph nodes; Based on the input vectors of each risk inference, calculate the residual resistance and effect of different graph nodes; Based on the residual resistance and effects of different graph nodes, the corresponding limit state function values ​​are constructed. The failure probability of each node in the graph is calculated based on the value of each limit state function and the reliability index.

8. A city lifeline risk assessment and early warning system based on multi-source monitoring fusion, used to implement the city lifeline risk assessment and early warning method based on multi-source monitoring fusion as described in any one of claims 1 to 7, characterized in that, include: The data generation module is used to acquire multi-source monitoring data of the city's lifeline infrastructure to be evaluated, generate peer samples and data quality characterization information, process the peer samples in combination with quality time series constraint objectives and event consistency alignment loss, and generate a unified monitoring dataset. The matrix construction module is used to determine graph nodes and edges based on a unified monitoring dataset and to construct a comprehensive adjacency matrix by combining the credibility coefficient. The calculation module is used to extract the infrastructure status representation vector of the unified monitoring dataset based on the comprehensive adjacency matrix, and calculate the corresponding infrastructure health index and failure probability. The early warning module is used to calculate the risk level based on the infrastructure health index and failure probability, and generate early warning trigger results.

9. A city lifeline risk assessment and early warning system based on multi-source monitoring fusion as described in claim 8, characterized in that, The data generation module includes: The sample construction unit is used to acquire multi-source monitoring data of the lifeline infrastructure of the city to be evaluated, and generate peer samples by combining the data representation system. The quality calculation unit is used to calculate the corresponding quality characterization value based on multi-source monitoring data and generate data quality characterization information. The repair unit is used to construct quality time-series constraint targets based on facility physical constraint penalty terms and data quality characterization information, and to perform anomaly detection and joint repair on peer samples to generate peer repair samples. The alignment unit is used to standardize the peer repair samples and perform cross-source time-series alignment by combining the event consistency alignment loss to generate a unified monitoring dataset.

10. A city lifeline risk assessment and early warning system based on multi-source monitoring fusion as described in claim 8, characterized in that, It also includes a closed-loop handling module, which is used to generate handling scenario cards based on the early warning triggering results, match scenario-based handling rules and perform hard constraint verification, generate work order-style handling processes, and update historical risk event samples, risk classification thresholds, failure probability calibration parameters and scenario-based handling rule library according to the handling feedback results.

Images