Engineering anomaly identification method, system, and storage medium

Abstract

The application discloses an engineering anomaly identification method and system and a storage medium, and relates to the technical field of structural health monitoring and intelligent identification. The method comprises the following steps: acquiring a three-dimensional data body of a target engineering area collected by a distributed optical fiber sensor; performing feature extraction on the three-dimensional data body to obtain a multi-physical field parameter data body; performing space-time alignment processing and channel splicing on the multi-physical field parameter data body to obtain a multi-dimensional input tensor; inputting the multi-dimensional input tensor into a pre-trained anomaly identification model to obtain an anomaly score matrix corresponding to different structural anomaly types of each spatial measurement point; and comprehensively judging according to the anomaly score matrix, the multi-physical field parameter data body and a preset rule library to obtain the structural anomaly type of the target engineering area. The application can more comprehensively utilize multi-source sensing information of the target engineering area, reduce the misjudgment risk caused by a single detection result, and is beneficial to improving the automation degree, continuity and reliability of engineering structural anomaly identification.

Description

Technical Field

[0001] This application relates to the field of structural health monitoring and intelligent identification technology, and in particular to an engineering anomaly identification method, system and storage medium. Background Technology

[0002] The geological and structural conditions of various engineering facilities and their surrounding areas directly affect the safety of engineering operations, public safety and property, and the stability of socio-economic activities. Engineering projects can include roads, railways, tunnels, underground passages, integrated utility tunnels, slopes, foundation pits, dams, mines, goaf areas, bridge foundations, pile foundations, culverts, areas surrounding underground pipelines, foundations, and other engineering objects or areas requiring long-term monitoring or safety assessment. During the long-term service, construction, or operation of these projects, factors such as vehicle loads, train loads, mechanical vibrations, construction disturbances, rainwater infiltration, groundwater activity, material aging, differences in construction quality, excavation disturbances, mining impacts, uneven foundation deformation, and changes in the natural environment can easily cause engineering anomalies such as cavities, voids, loosening, fissures, cracks, abnormal water content, abnormal seepage, uneven settlement, slippage, goaf development, structural interface debonding, or localized damage within or around the target engineering area. The aforementioned engineering anomalies are often concealed in their early stages, frequently located below the surface, inside the structure, or in hidden areas around the project, making them difficult to detect in a timely manner through manual inspections alone. If they are not identified and addressed promptly, they may further develop into safety accidents such as cavity collapse, structural voiding, surface subsidence, slope instability, tunnel deformation, dam leakage, and foundation instability.

[0003] Existing methods for detecting engineering anomalies mainly include manual inspection, core sampling, ground-penetrating radar (GPR), falling hammer deflection testing, transient surface wave detection, and point sensor monitoring. Among these, manual inspection is simple to operate but highly subjective and struggles to promptly detect hidden defects underground or within structures. Core sampling is a destructive testing method with a limited detection range, making it unsuitable for continuous inspection of large-scale, long-distance, or complex engineering areas. Non-destructive testing technologies such as GPR and surface wave detection can identify geological or structural anomalies within engineering areas to a certain extent; however, these methods typically require on-site inspection vehicles, manual excitation equipment, or specialized equipment. The inspection process is significantly affected by site conditions, frequency of inspection, site environment, and personnel operation, making long-term continuous monitoring of the engineering condition difficult. Furthermore, for long-term operational safety management of large-scale, long-distance, or key engineering areas, periodic inspection methods often suffer from long time intervals and delayed anomaly detection. Point-based sensor monitoring can only reflect the status information near a limited number of measuring points, with limited spatial coverage, making it difficult to achieve continuous perception of the target engineering area. When the location of the disease is far from the sensor deployment point, monitoring blind spots are likely to occur.

[0004] Therefore, existing engineering anomaly detection technologies still have shortcomings in areas such as identifying latent anomalies, providing continuous coverage over large areas or long distances, real-time monitoring, and comprehensive discrimination of multiple types of anomalies, making it difficult to fully meet the needs of refined, intelligent, and real-time operation and maintenance management of engineering facilities. How to provide a technical solution capable of continuous sensing, intelligent identification, and comprehensive discrimination of engineering anomalies has become an urgent problem to be solved in the field of engineering safety monitoring. Summary of the Invention

[0005] This application aims to address at least one of the technical problems existing in the prior art. To this end, this application proposes an engineering anomaly identification method, system, and storage medium, which can continuously sense, intelligently identify, and comprehensively judge engineering anomalies within a target engineering area.

[0006] In a first aspect, embodiments of this application provide a method for identifying engineering anomalies.

[0007] An engineering anomaly identification method according to a first aspect of this application includes: acquiring a three-dimensional data volume of a target engineering area collected by a distributed optical fiber sensor, the three-dimensional data volume including acoustic wave data, temperature field data, and strain field data; performing feature extraction on the three-dimensional data volume to obtain a multi-physics parameter data volume; the multi-physics parameter data volume including at least one of a three-dimensional velocity model data volume, a three-dimensional density data volume, a wave impedance data volume, an instantaneous feature tensor, a temperature gradient anomaly index data volume, and a strain curvature data volume; performing spatiotemporal alignment processing and channel stitching on the multi-physics parameter data volume to obtain a multi-dimensional input tensor; inputting the multi-dimensional input tensor into a pre-trained anomaly identification model to obtain an anomaly score matrix corresponding to different structural anomaly types for each spatial measurement point within the target engineering area; and performing comprehensive discrimination based on the anomaly score matrix and the multi-physics parameter data volume, using a preset rule base, to obtain the structural anomaly type of the target engineering area.

[0008] The engineering anomaly identification method according to the first aspect of this application has at least the following beneficial effects: It collects acoustic wave data, temperature field data, and strain field data of the target engineering area based on distributed optical fiber sensors, further extracts multi-physics parameter data volumes, and inputs them into an anomaly identification model after spatiotemporal alignment and channel stitching to obtain an anomaly score matrix. This matrix is ​​then combined with a rule base for comprehensive discrimination, thus forming a structural anomaly identification process from data acquisition, feature extraction, model recognition to rule interpretation. Compared with traditional road defect detection methods that rely on manual inspection, periodic detection, or localized point monitoring, this application can more comprehensively utilize multi-source sensing information of the target engineering area, reduce the risk of misjudgment caused by a single detection result, and improve the automation, continuity, and reliability of engineering anomaly identification. It also provides more timely judgment basis for engineering operation and maintenance, disaster early warning, and safety handling.

[0009] Secondly, embodiments of this application provide an engineering anomaly identification system.

[0010] An engineering anomaly identification system according to a second aspect of this application includes: a data acquisition module for acquiring a three-dimensional data volume of a target engineering area collected by a distributed optical fiber sensor, the three-dimensional data volume including acoustic wave data, temperature field data, and strain field data; a feature extraction module for performing feature extraction based on the three-dimensional data volume to obtain a multi-physics parameter data volume, the multi-physics parameter data volume including at least one of a three-dimensional velocity model data volume, a three-dimensional density data volume, a wave impedance data volume, an instantaneous feature tensor, a temperature gradient anomaly index data volume, and a strain curvature data volume; a tensor construction module for performing spatiotemporal alignment processing and channel splicing on the multi-physics parameter data volume to obtain a multi-dimensional input tensor; a model inference module for inputting the multi-dimensional input tensor into a pre-trained anomaly identification model to obtain an anomaly score matrix corresponding to different structural anomaly types for each spatial measurement point within the target engineering area; and a rule interpretation module for performing comprehensive discrimination based on the anomaly score matrix and the multi-physics parameter data volume, based on a preset rule base, to obtain the structural anomaly type of the target engineering area.

[0011] The engineering anomaly identification system according to the second aspect of this application has at least the following beneficial effects: Through the coordinated cooperation between the data acquisition module, feature extraction module, tensor construction module, model reasoning module, and rule interpretation module, it can process the target engineering area data collected by distributed fiber optic sensors and complete the structural anomaly identification process from multi-physics parameter extraction, multi-dimensional input tensor construction, anomaly identification model reasoning to rule base comprehensive judgment. Compared with traditional road defect detection methods that rely on manual inspection, periodic detection, or local point monitoring, this system can more comprehensively utilize multi-source sensing information of the target engineering area, reduce the risk of misjudgment caused by a single detection result, and help improve the automation, continuity, and reliability of engineering anomaly identification, and provide more timely judgment basis for engineering operation and maintenance, disaster early warning, and safety handling.

[0012] Thirdly, embodiments of this application provide a computer-readable storage medium. The computer-readable storage medium stores a computer program thereon, which, when executed by a processor, implements the steps of the engineering anomaly identification method of any embodiment of the first aspect. The principles and beneficial effects of the computer-readable storage medium are derived from the engineering anomaly identification method of the first aspect, and will not be repeated here. Attached Figure Description

[0013] The present application will be further described below with reference to the accompanying drawings and embodiments, wherein: Figure 1 A flowchart illustrating the engineering anomaly identification method provided in this application embodiment; Figure 2 A flowchart illustrating the multidimensional input tensor construction method provided in this application embodiment; Figure 3 A flowchart illustrating the anomaly recognition model reasoning method provided in this application embodiment; Figure 4 This is a flowchart illustrating a method for outputting an anomaly score matrix using a multi-task branch network, as provided in an embodiment of this application. Figure 5 This is a flowchart illustrating the method for comprehensive structural anomaly type discrimination based on a rule base provided in an embodiment of this application. Figure 6 A flowchart illustrating the disaster risk level determination method provided in this application embodiment. Detailed Implementation

[0014] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain this application, and should not be construed as limiting this application.

[0015] In the description of this application, it should be understood that the orientation descriptions, such as up, down, front, back, left, right, etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application.

[0016] In the description of this application, "several" means one or more, "multiple" means two or more, "greater than," "less than," and "exceeding" are understood to exclude the stated number, while "above," "below," and "within" are understood to include the stated number. The use of "first" and "second" in the description is merely for distinguishing technical features and should not be construed as indicating or implying relative importance, or implicitly indicating the number of indicated technical features, or implicitly indicating the order of the indicated technical features.

[0017] In the description of this application, unless otherwise expressly defined, terms such as "setup," "installation," and "connection" should be interpreted broadly, and those skilled in the art can reasonably determine the specific meaning of the above terms in this application in conjunction with the specific content of the technical solution.

[0018] In the description of this application, the terms "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0019] It should be noted that in all specific embodiments of this application, when processing data related to user identity or characteristics, such as user information, user behavior data, user historical data, and user location information, user permission or consent is obtained first. Furthermore, the collection, use, and processing of this data comply with relevant laws, regulations, and standards. In addition, when embodiments of this application require access to sensitive personal information of users, separate permission or consent from the user is obtained through pop-ups or redirection to confirmation pages. Only after obtaining the user's separate permission or consent is the necessary user-related data required for the proper functioning of these embodiments acquired.

[0020] In a first aspect, embodiments of this application provide a method for identifying engineering anomalies.

[0021] In this application embodiment, "engineering" can refer to an engineering object, engineering facility, or engineering area that requires monitoring of its geological condition, structural condition, or operational safety condition. For example, engineering may include roads, railways, tunnels, underground passages, integrated utility tunnels, slopes, foundation pits, dams, mines, mining subsidence areas, bridge foundations, pile foundations, culverts, areas surrounding underground pipelines, foundations, and other engineering objects or engineering areas that require continuous monitoring or safety assessment. Correspondingly, the target engineering area can be the internal area, surface area, adjacent stratum area, foundation area, or surrounding influence area of ​​the aforementioned engineering object.

[0022] In the embodiments of this application, engineering anomalies may include geological anomalies, structural anomalies, and composite anomalies formed by the coupling of changes in geological state and structural state. For example, engineering anomalies may include cavities, voids, development of goaf areas, loose areas, weak interlayers, fissures, cracks, water-bearing anomalies, seepage anomalies, uneven settlement, slippage, slope instability, loosening of surrounding rock, lining voids, dam leakage, foundation instability, structural interface debonding, local structural damage, pipeline leakage, and other abnormal states that can cause abnormal changes in sound waves, temperature fields, or strain fields.

[0023] It should be noted that road structure anomalies are merely one example application scenario of engineering anomalies in this application. The embodiments of this application are not limited to road scenarios, but can also be applied to the identification of engineering anomalies in engineering scenarios such as tunnels, slopes, dams, underground spaces, mines, pipe corridors, bridge foundations, and ground foundations.

[0024] The engineering anomaly identification method of the embodiment includes, but is not limited to, steps S100 to S500: S100: Acquire a three-dimensional data volume of the target engineering area collected by a distributed optical fiber sensor. The three-dimensional data volume includes acoustic data, temperature field data and strain field data. S200. Perform feature extraction on the three-dimensional data volume to obtain a multi-physics parameter data volume; the multi-physics parameter data volume includes at least one of the following: three-dimensional velocity model data volume, three-dimensional density data volume, wave impedance data volume, instantaneous feature tensor, temperature gradient anomaly index data volume, and strain curvature data volume. S300: Perform spatiotemporal alignment and channel splicing on the multiphysics parameter data volume to obtain a multidimensional input tensor; S400. Input the multidimensional input tensor into the pre-trained anomaly recognition model to obtain the anomaly score matrix corresponding to different structural anomaly types for each spatial measurement point within the target engineering area. S500: Based on the anomaly score matrix and multiphysics parameter data, a comprehensive judgment is made according to the preset rule base to obtain the structural anomaly type of the target engineering area.

[0025] In step S100, a three-dimensional data volume of the target engineering area is acquired through distributed optical fiber sensors. The target engineering area can be the internal area, surface area, adjacent stratum area, foundation area, or surrounding influence area of ​​the engineering object to be monitored. For example, the target engineering area can be a road, railway, tunnel, underground passage, integrated pipe gallery, slope, foundation pit, dam, mine, goaf, bridge foundation, pile foundation, culvert, area surrounding underground pipeline, foundation, or other engineering area. The target engineering area can be a linear engineering area, a planar engineering area, or a three-dimensional spatial engineering area, and its interior or surrounding area may contain engineering anomalies such as cavities, voids, looseness, cracks, fissures, abnormal water content, abnormal seepage, uneven settlement, slippage, goaf development, or local structural damage. The distributed optical fiber sensors can be deployed along the extension direction, boundary direction, structural surface, structural interior, foundation area, or surrounding stratum of the target engineering area, or can be deployed in single-line, double-line, grid, serpentine, circumferential, longitudinal, transverse, or multi-layer combination methods according to the structural characteristics of the target engineering area and monitoring needs. Distributed fiber optic sensors can include single-mode optical fibers, optical cables, sensing fiber units, demodulation equipment, and data acquisition equipment embedded within or near road structures. Distributed fiber optic sensors can form continuous or quasi-continuous spatial measurement points along the length of the fiber, thereby achieving distributed sensing of the target engineering area. Compared to point sensors, distributed fiber optic sensors can acquire continuous monitoring data over a longer distance with smaller spatial intervals.

[0026] The three-dimensional data volume includes acoustic wave data, temperature field data, and strain field data. For example, the three-dimensional data volume can be represented as: ,in, r This indicates the planar location of data points within the target engineering area. The planar location can include the location along the road extension direction and the lateral location. t Indicates the acquisition time; c i Indicates a data channel; i Channel numbering is assigned. The three-dimensional data volume includes acoustic wave data, temperature field data, and strain field data. Acoustic wave data can correspond to one acoustic wave channel, temperature field data can correspond to one temperature channel, and strain field data can correspond to one strain channel.

[0027] Understandably, acoustic data refers to data related to the propagation of vibration, sound waves, or elastic waves collected by distributed fiber optic sensors. Acoustic data can originate from natural or engineering excitations generated within or near the target engineering area. For example, acoustic data can originate from vehicle movement, train operation, machinery operation, construction disturbance, water erosion, wind load, environmental micro-vibration, artificially induced vibration, or a combination thereof. That is, this application is applicable to passive source monitoring scenarios as well as active source or combined active-passive source monitoring scenarios. It should be noted that the embodiments of this application do not limit the acoustic data to be induced by vehicle loads. For different engineering scenarios, acoustic data can be generated by vehicle loads, train loads, machinery operation, construction disturbances, water flow, wind loads, artificial excitation, environmental micro-vibration, or other natural excitations. As long as the acoustic data can propagate within the target engineering area and carry the physical property information of the medium or structure of the target engineering area, it can be used for multiphysics parameter extraction and engineering anomaly identification in the embodiments of this application.

[0028] Understandably, temperature field data refers to the temperature distribution information along the target engineering area obtained by distributed fiber optic sensors. Temperature field data can reflect the thermal conductivity, heat capacity, and the impact of groundwater seepage on temperature distribution within the target engineering area. When cavities, voids, mining subsidence areas, fissures, water-bearing areas, seepage channels, or other thermophysical anomalies exist within or around the target engineering area, the heat exchange conditions in that area may differ from those in the surrounding healthy areas. Therefore, temperature field data can serve as auxiliary information for identifying structural anomalies. Understandably, strain field data refers to data related to road structure deformation collected by distributed fiber optic sensors. Strain field data can reflect the tensile and compressive deformation state of the target engineering area under vehicle loads, train loads, mechanical vibration, construction disturbance, temperature changes, settlement development, slope slippage, structural voids, foundation deformation, or other natural and engineering actions. For anomalies such as roadbed voids, local settlement, or cracks, strain field data typically exhibits characteristics such as local abrupt changes, gradient anomalies, or continuous accumulation.

[0029] Understandably, during data acquisition, three-dimensional data volumes can be acquired according to preset sampling frequencies and preset spatial sampling intervals. The preset sampling frequency can be determined based on the frequency range of vehicle load-induced vibrations, road structure type, monitoring objectives, and data processing capabilities; the preset spatial sampling interval can be determined based on the spatial resolution of the distributed fiber optic sensor, the length of the target engineering area, and the required anomaly identification accuracy. Generally, a smaller spatial sampling interval is more conducive to identifying local anomalies; a higher sampling frequency is more conducive to preserving the time-frequency characteristics of the vibration signal. To improve data quality, basic preprocessing can be performed after acquiring the three-dimensional data volume. Basic preprocessing can include outlier removal, bad road repair, denoising, detrending, time synchronization, coordinate mapping, and matching of fiber optic mileage with road mileage. The purpose of the above basic preprocessing is to reduce the impact of acquisition noise, equipment drift, poor fiber coupling, and environmental disturbances on subsequent feature extraction and model recognition.

[0030] In step S200, after obtaining the three-dimensional data volume, feature extraction is performed on the three-dimensional data volume to obtain a multiphysics parameter data volume. The multiphysics parameter data volume refers to a set of parameters with clear physical meaning formed by signal processing, geophysical inversion, mathematical transformation, or statistical calculation of the original acoustic wave data, temperature field data, and strain field data.

[0031] In some implementations, the multiphysics parameter data volume includes at least one of a three-dimensional velocity model data volume, a three-dimensional density data volume, a wave impedance data volume, an instantaneous characteristic tensor, a temperature gradient anomaly index data volume, and a strain curvature data volume. In different application scenarios, one or more physical field parameters can be selected for anomaly identification based on the monitoring conditions of the target engineering area, sensor configuration, data quality, and computing resources. Preferably, to improve the reliability of anomaly identification, multiple physical field parameter data volumes can be used simultaneously to form multi-source evidence fusion.

[0032] The three-dimensional velocity model data volume is used to characterize the spatial distribution of elastic wave propagation velocity in the medium within the target engineering area and its variation over time. For example, dense, intact, and stiff areas within the target engineering area have relatively high wave velocities; while loose, voided, water-softened, or fractured areas typically exhibit lower wave velocities. Therefore, the three-dimensional velocity model data volume can reflect the stiffness, density, and integrity within the target engineering area.

[0033] Three-dimensional density data volumes are used to characterize the spatial distribution of medium density within the target engineering area and its changes over time. Density variations are related to factors such as medium porosity, water content, filler loss, and void formation. For voided or degraded areas, the density may be significantly lower than that of the surrounding healthy subgrade; for water-bearing or saturated loose areas, density changes may be related to increased water content.

[0034] The wave impedance data volume is used to characterize the impedance characteristics of a medium to the propagation of elastic waves. In one embodiment, the wave impedance data volume can be calculated based on a three-dimensional velocity model data volume and a three-dimensional density data volume, and can be: ,in, Z r,t Indicates spatial location r In time t The corresponding wave impedance value; r r,t Indicates spatial location r In time t The corresponding density value; V r,t Indicates spatial location r In time t The corresponding speed value; r This indicates the spatial location within the target engineering area, which can include along the line, laterally, and at depth. Wave impedance is highly sensitive to medium interfaces, void boundaries, crack surfaces, and strongly scattering interfaces, thus serving as an important parameter for identifying engineering anomalies.

[0035] The instantaneous feature tensor is used to characterize the time-frequency features of acoustic wave data, such as instantaneous amplitude, instantaneous frequency, and instantaneous phase. Instantaneous amplitude reflects the strength of wave field energy and scattering intensity; instantaneous frequency reflects the absorption and attenuation characteristics of the propagation medium for different frequency components; and instantaneous phase can be used to help determine the direction of wave field propagation, phase abrupt changes, and the location of discontinuities. By combining instantaneous amplitude, instantaneous frequency, and instantaneous phase, the instantaneous feature tensor can be obtained, thus providing richer acoustic wave feature information for anomaly identification models.

[0036] Temperature gradient anomaly index data volumes are used to characterize the degree of anomaly in the spatial temperature gradient of a target engineering area relative to the normal background state. For anomalies such as cavities, aquifers, fracture zones, seepage channels, or localized voids, the temperature gradient may exhibit localized anomalies due to differences in thermal conductivity, heat capacity, and heat exchange conditions compared to the surrounding medium. Therefore, temperature gradient anomaly index data volumes can serve as supplementary evidence for identifying thermal and hydrogeological anomalies.

[0037] Strain curvature data volumes are used to characterize the second-order spatial variation of strain field data. Compared with simple strain values, strain curvature can amplify the spatial variation characteristics caused by local bending, local voids, differential deformation on both sides of cracks, and uneven settlement in engineering structures. For normal media or structural regions, strain curvature usually exhibits a relatively smooth distribution; however, for regions with voids, settlement, cracks, or slippage, strain curvature may show local peaks, antisymmetric variations, continuous accumulation, or large-scale low-frequency fluctuations.

[0038] Feature extraction from three-dimensional data volumes can convert acoustic wave data, temperature field data, and strain field data collected by distributed fiber optic sensors into multiphysics parameters that reflect the physical state of engineering structures. In this way, subsequent anomaly identification models can not only rely on the original signal patterns but also utilize parameters with clear engineering and geophysical implications, thereby improving the accuracy and interpretability of the identification results.

[0039] In step S300, the multiphysics parameter data volume undergoes spatiotemporal alignment and channel stitching to obtain a multidimensional input tensor. Understandably, because the three-dimensional velocity model data volume, three-dimensional density data volume, wave impedance data volume, instantaneous feature tensor, temperature gradient anomaly index data volume, and strain curvature data volume have different sources, update frequencies, spatial resolutions, and dimensions, directly inputting them into the anomaly recognition model may lead to inaccurate correspondence between different data, thus affecting the model's recognition performance. Therefore, this step performs spatiotemporal alignment and channel stitching on the multiphysics parameter data volume to obtain a multidimensional input tensor.

[0040] In some implementations, spatiotemporal alignment includes alignment in both the spatial and temporal dimensions. Spatial alignment refers to unifying different multiphysics parameter data volumes to the same or corresponding spatial measurement points. For example, the spatial measurement points of distributed fiber optic sensors in the target engineering area can be used as a reference, or a regular grid formed by a preset spatial resolution can be used as a reference to map each multiphysics parameter data volume to a unified spatial coordinate system. For parameter data volumes with low spatial resolution, spatial alignment can be performed using interpolation, resampling, nearest neighbor mapping, kriging interpolation, spline interpolation, or spatial reconstruction methods based on physical constraints; for parameter data volumes with high spatial resolution, downsampling or local statistical methods can be used for matching. Temporal alignment refers to unifying data with different update frequencies or different sampling frequencies to the same time axis. For real-time updated data, statistical characteristics can be calculated according to a preset time step; for data updated within a fixed time window, it can be mapped to the corresponding time window. Statistical characteristics can include mean, standard deviation, maximum value, minimum value, median, quantiles, peak value, energy value, or rate of change, etc. Time alignment ensures that different physical field parameters within the same time window describe the state of the target engineering area during the same or similar monitoring periods.

[0041] Understandably, since different physical field parameters have different dimensions and numerical ranges, some implementations require standardization or normalization of the spatiotemporally aligned multiphysics parameter data. Standardization or normalization is used to reduce the impact of dimensional differences on model training and inference, enabling the anomaly detection model to learn the relationships between different parameters more stably. In one implementation, a hybrid normalization strategy can be adopted, that is, different standardization methods are used for different types of physical field parameters. For example, for three-dimensional velocity model data, three-dimensional density data, and wave impedance data, a robust normalization method based on historical normal baseline values, median, quantiles, or interquartile ranges can be used; for temperature gradient anomaly index data and strain curvature data, a normalization method based on mean and standard deviation can be used. The baseline values ​​required for normalization can be derived from historical healthy periods, manually confirmed normal road sections, design data, geological survey data, or long-term monitoring statistics.

[0042] After completing spatiotemporal alignment and normalization, the multiphysics parameter data volume is concatenated along the channel dimension to obtain a multidimensional input tensor. The multidimensional input tensor can be represented as: ,in, X Represents a multidimensional input tensor; N r Indicates the number of spatial measurement points; T Indicates the number of time steps; C This indicates the number of feature channels. A feature channel can correspond to one or more of the following: a 3D velocity model data volume, a 3D density data volume, a wave impedance data volume, an instantaneous feature tensor, a temperature gradient anomaly index data volume, and a strain curvature data volume. It should be noted that when the instantaneous feature tensor includes instantaneous amplitude, instantaneous frequency, and instantaneous phase, it can be used as a single composite input channel or split into multiple independent feature channel inputs; this application does not impose any restrictions on this. By constructing a multidimensional input tensor, information about the target engineering area in the spatial, temporal, and physical field dimensions can be uniformly expressed, enabling the anomaly identification model to simultaneously utilize the correlation between different spatial measurement points, the temporal evolution of the same spatial measurement point, and the coupling relationship between different physical field parameters.

[0043] In step S400, after obtaining the multidimensional input tensor, the multidimensional input tensor is input into a pre-trained anomaly recognition model to obtain anomaly score matrices corresponding to different structural anomaly types for each spatial measurement point within the target engineering area. The anomaly recognition model includes machine learning models, deep learning models, time-series models, graphical models, Transformer models, or combinations thereof, capable of recognizing engineering anomalies based on the multidimensional input tensor. The anomaly recognition model can be trained using historical monitoring data, manual inspection records, geological survey results, road maintenance records, numerical simulation data, model test data, or combinations thereof. After training, the anomaly recognition model can output the anomaly degree corresponding to different structural anomaly types for each spatial measurement point based on the input multidimensional input tensor. In one embodiment, the anomaly recognition model can employ a deep learning model with time-series modeling capabilities to capture the evolution of the engineering structure's state over time. For example, settlement anomalies typically exhibit slow accumulation characteristics, slip anomalies may have continuous displacement or low-frequency disturbance characteristics, and the development of cavities or voids may manifest as coordinated changes in wave velocity, wave impedance, temperature gradient, and strain curvature. Therefore, anomaly detection models need to have the ability to comprehensively analyze long-term series, multi-channel data, and spatially correlated data.

[0044] In one implementation, the anomaly identification model can employ a deep learning model with temporal modeling capabilities to capture the evolution of the engineering structure's state over time. For example, settlement anomalies typically exhibit slow accumulation characteristics, slip anomalies may show continuous displacement or low-frequency disturbance characteristics, and the development of cavitation or voids may manifest as coordinated changes in wave velocity, wave impedance, temperature gradient, and strain curvature. Therefore, the anomaly identification model needs to have the ability to comprehensively analyze long-term series, multi-channel data, and spatially correlated data. In another implementation, the anomaly identification model can include a network structure for extracting spatial or channel relationships to characterize the correlation between different spatial measurement points or the dependency between different physical field parameters. For example, the spatial co-location relationship between a velocity reduction region and a strain curvature increase region may indicate voids or cavitation; the overlap between a temperature gradient anomaly region and a density anomaly region may indicate a water-bearing anomaly or a void anomaly; the correspondence between a sudden change in wave impedance and a region with a high instantaneous amplitude may indicate a crack surface or a strong scattering interface. By introducing modeling of spatial or channel relationships into the anomaly identification model, the model's ability to identify complex anomalies can be enhanced.

[0045] Understandably, the anomaly score matrix is ​​used to represent the anomaly intensity of different structural anomaly types corresponding to each spatial measurement point within the target engineering area. For example, the anomaly score matrix can be represented as: ,in, Q r,c Indicates spatial location r Corresponding structural anomaly type cThe overall abnormal score; r Indicates a spatial measurement point within the target engineering area or the spatial location corresponding to a spatial measurement point; c This indicates the preset structural anomaly type. Preset engineering anomaly types can include one or more of the following: cavities, voids, goaf areas, loose areas, weak interlayers, uneven settlement, fissures, cracks, slippage, slope instability, abnormal seepage, pipeline leakage, freeze-thaw damage, structural interface debonding, and localized structural damage. For road applications, engineering anomaly types can include roadbed cavities, pavement voids, loose roadbeds, uneven settlement, and roadbed slippage; for tunnels or underground spaces, engineering anomaly types can include loose surrounding rock, lining voids, fissure development, seepage channels, and structural cracks; for dams, slopes, or foundation pits, engineering anomaly types can include seepage channels, slip surfaces, weak interlayers, locally loose areas, and abnormal deformation of support structures; for mines or goaf areas, engineering anomaly types can include goaf development, loose surrounding rock, roof fissures, slope slippage, and surface subsidence.

[0046] Understandably, the comprehensive anomaly score can be a probability value, confidence value, normalized anomaly score, or other numerical values ​​that characterize the intensity of anomalies. Generally, the higher the comprehensive anomaly score, the higher the probability that the corresponding spatial measurement point contains the corresponding structural anomaly type. The range of values ​​for the comprehensive anomaly score can be determined according to the output method of the anomaly identification model; for example, it can be normalized to between 0 and 1, or other numerical ranges can be used depending on the model training method. Understandably, after the anomaly identification model outputs the anomaly score matrix, it does not directly use the anomaly score matrix as the final diagnostic result, but rather uses it as an important input for subsequent rule base comprehensive discrimination. This avoids false alarms or unexplainable problems caused by relying solely on data-driven models, while also ensuring that the identification results can be corroborated with engineering and geophysical laws.

[0047] In step S500, after obtaining the anomaly score matrix, a comprehensive discrimination is performed based on the anomaly score matrix and the multiphysics parameter data volume, using a pre-set rule base, to obtain the structural anomaly type of the target engineering area. Understandably, the rule base includes a pre-established set of discrimination rules describing the correspondence between different structural anomaly types and multiphysics parameters. The rule base can be established based on geophysical principles, road engineering theory, historical disease cases, field testing experience, numerical simulation results, indoor test results, or a combination thereof. Each structural anomaly type in the rule base can correspond to one or more sets of discrimination rules. Each set of discrimination rules can include one or more of the following: anomaly score conditions, physical field parameter conditions, spatial morphology conditions, temporal evolution conditions, and physical-logical consistency conditions.

[0048] In some specific implementations, a comprehensive judgment is performed based on a preset rule base, including fusing the anomaly score matrix output by the anomaly identification model with physical evidence in the multiphysics parameter data volume. For example, for a certain spatial measurement point or a certain spatial region, if the anomaly score matrix shows that it has a high comprehensive anomaly score for a certain structural anomaly type, and the multiphysics parameter data volume of that location or region also exhibits anomaly characteristics that conform to the physical cause mechanism of that structural anomaly type, then it is determined that the structural anomaly type exists in the target engineering area.

[0049] For example, anomalies such as roadbed voids are typically associated with low velocity, low wave impedance, low density, abnormal temperature gradients, and increased strain curvature. Anomalies in loose areas may manifest as a moderate decrease in velocity and wave impedance, along with a slight decrease in density. Anomalies involving cracks may exhibit enhanced local scattering, linear temperature anomalies, and differential strain curvature on both sides of the crack. Anomalies involving uneven settlement may present as large-scale changes in strain curvature and abnormal spatial wave velocity gradients. Anomalies involving roadbed slippage may manifest as continuous strain changes, low-frequency vibration events, and changes in physical properties related to water content. It should be noted that the correspondence between the above anomaly types and physical field parameters is not an isolated judgment but should be comprehensively assessed using the anomaly fraction matrix and multi-physics parameter data. Single-parameter anomalies may be caused by noise, local construction disturbances, vehicle load changes, temperature changes, or changes in sensor coupling states; however, when multiple parameters exhibit anomalies that conform to a physical cause mechanism, the reliability of structural anomaly type determination can be significantly improved.

[0050] For example, the rule base comprehensive judgment can be performed as follows: First, based on the anomaly score matrix, select spatial measurement points or spatial regions whose comprehensive anomaly scores exceed the corresponding thresholds; then, call the rules for the corresponding candidate structural anomaly types to check whether the multi-physics parameter data volume of the spatial measurement point or spatial region meets the corresponding physical parameter anomaly conditions; finally, determine whether the physical logical consistency exists between the met physical parameter anomaly conditions, and output the structural anomaly type when the rule base requirements are met.

[0051] The structural anomaly types of the target engineering area can be output in the form of points, sections, regions, or three-dimensional volumes. The output results can include the anomaly type name, spatial location of the anomaly, the range of influence of the anomaly, the time of occurrence or time window of the anomaly, the corresponding comprehensive anomaly score, and the anomaly characteristics of relevant physical field parameters. For the same spatial measurement point or the same area, one or more structural anomaly types may exist simultaneously. For example, the development of roadbed cavities may be accompanied by uneven settlement, and cracked areas may be accompanied by local loosening or seepage anomalies. Therefore, this application allows the output of multi-label structural anomaly identification results.

[0052] Through steps S100 to S500, the engineering anomaly identification method of this embodiment realizes a complete technical process from distributed fiber optic sensor data acquisition, multi-physics parameter extraction, multi-dimensional input tensor construction, anomaly identification model reasoning, to rule base comprehensive judgment. This process not only utilizes the long-distance, continuous sensing capabilities of distributed fiber optic sensors, but also uses multi-physics parameters to physically interpret the state of engineering structures, and captures complex spatiotemporal evolution patterns through anomaly identification models, thereby achieving automatic identification of structural anomaly types in the target engineering area.

[0053] In some embodiments, in step S200, feature extraction is performed on the three-dimensional data volume to obtain a multiphysics parameter data volume, which may further include, but is not limited to, steps S210 to S250: S210. Perform inversion processing on the acoustic wave data to obtain a three-dimensional velocity model data volume and a three-dimensional density data volume; S220. Based on the three-dimensional velocity model data volume and the three-dimensional density data volume, the wave impedance data volume is calculated. S230. Based on the acoustic wave data, extract the instantaneous amplitude, instantaneous frequency, and instantaneous phase through Hilbert transform, and combine the instantaneous amplitude, instantaneous frequency, and instantaneous phase to obtain the instantaneous feature tensor; S240. Based on the temperature field data, the temperature gradient anomaly index data volume is calculated using the preset temperature gradient anomaly index formula. S250. Based on the strain field data, the strain curvature data volume is obtained by calculating the local strain curvature.

[0054] It should be noted that steps S210 to S250 are not necessarily required to be executed in strict numerical order. In practical applications, the processing based on acoustic wave data, temperature field data, and strain field data can be executed in parallel or asynchronously depending on data arrival time, computing resources, and business requirements. As long as at least one of the following can be obtained—a three-dimensional velocity model data volume, a three-dimensional density data volume, a wave impedance data volume, an instantaneous characteristic tensor, a temperature gradient anomaly index data volume, and a strain curvature data volume—it can be used as a component of the multiphysics parameter data volume.

[0055] In step S210, the acoustic wave data is inverted to obtain a three-dimensional velocity model data volume and a three-dimensional density data volume. The acoustic wave data includes data related to the propagation of sound waves, vibration waves, or elastic waves, collected by distributed fiber optic sensors within the three-dimensional data volume. The acoustic wave data can be induced by vehicle loads, environmental vibrations, artificial excitation, or a combination thereof. When acoustic wave data propagates in the medium of the target engineering area, its propagation speed, amplitude attenuation, scattering characteristics, and phase changes are affected by the stiffness, density, porosity, water content, and structural integrity of the medium within the engineering structure. Therefore, a three-dimensional velocity model data volume and a three-dimensional density data volume can be obtained by inverting the acoustic wave data.

[0056] Understandably, inversion processing involves inferring the spatial distribution of medium properties in the target engineering area based on observed acoustic wave data. Inversion processing can employ methods such as travel-time tomography, surface wave dispersion inversion, full waveform inversion, passive source seismic interferometry, noise cross-correlation imaging, wavefield imaging, machine learning-assisted inversion, or a combination of these methods. For road monitoring scenarios where vehicle loads serve as the passive source, continuous vibrations generated during vehicle movement can be used as environmental excitation. By cross-correlation, superposition, or time-frequency analysis of acoustic wave data from different spatial measurement points, information related to medium propagation characteristics can be extracted, thereby constructing a three-dimensional velocity model data volume. For scenarios with both artificial excitation and active source acoustic wave data, active source acoustic wave data can also be used to correct or fuse the passive source inversion results.

[0057] For example, a three-dimensional velocity model data volume can be represented as: ,in, V r,t Indicates spatial location r In time t The corresponding speed value; r It indicates the spatial location of the target engineering area. Taking a road engineering scenario as an example, it can include the direction of road extension, the lateral direction, and the depth direction. t This indicates time or a time window. The velocity value can be volume wave velocity, surface wave velocity, equivalent wave velocity, shear wave velocity, compression wave velocity, or other velocity parameters reflecting the elastic wave propagation capability of the road structure medium. The specific velocity parameter selected depends on the type of acoustic wave data, sensor deployment method, inversion method, and engineering application requirements.

[0058] For example, a three-dimensional density data volume can be represented as: ,in, r r,t Indicates spatial location r In time tThe corresponding density value. The density value can be obtained through inversion from acoustic data, or it can be estimated by combining acoustic data, velocity information, amplitude attenuation information, prior geological data, road structure design data, or historical monitoring data.

[0059] In some implementation methods, the inversion process can be performed according to a fixed time window. For example, for the first... i Within a time window, the three-dimensional velocity model data volume can be represented as: ,in, t win Indicates the width of the time window; t i Indicates the first i Each time window corresponds to a specific moment or time period. n This represents the time window number. The width of the time window can be determined based on the vehicle passing frequency, the signal-to-noise ratio of the acoustic data, the rate of change of target anomalies, and computational resources. For example, for scenarios with high real-time requirements, a shorter time window can be used; for scenarios requiring improved inversion stability, a longer time window can be used for data overlay. It should be noted that the above time window expression is only used to illustrate that the 3D velocity model data volume can be updated according to the time window, and does not limit the inversion processing to a specific time discretization form. The 3D density data volume can also use the same or different time window update strategies as the 3D velocity model data volume.

[0060] In step S220, after obtaining the three-dimensional velocity model data volume and the three-dimensional density data volume, the wave impedance data volume is calculated based on the three-dimensional velocity model data volume and the three-dimensional density data volume. In some embodiments, the wave impedance data volume can be represented as: ,in, Z r,t Indicates spatial location r In time t The corresponding wave impedance value; r r,t Indicates spatial location r In time t The corresponding density value; V r,t Indicates spatial location r In time t The corresponding velocity values. In actual calculations, if the 3D velocity model data volume and the 3D density data volume have different spatial resolutions or time update frequencies, they can be spatiotemporally matched before calculating the wave impedance data volume, or a unified spatiotemporal alignment process can be performed after the calculation. The wave impedance data volume can be updated according to the same time window as the 3D velocity model data volume, or it can be periodically updated according to the availability of the 3D density data volume.

[0061] Wave impedance is used to characterize the ability of a medium to resist the propagation of elastic waves. When sound waves propagate at the interface of media with different wave impedances, reflection, transmission, or scattering will occur. Therefore, wave impedance data is particularly sensitive to void boundaries, crack surfaces, filler interfaces, soft and hard interlayers, voided layers, and other interfaces with abrupt changes in physical properties in road structures.

[0062] In step S230, based on the acoustic wave data, the instantaneous amplitude, instantaneous frequency, and instantaneous phase are extracted using the Hilbert transform, and then combined to obtain the instantaneous feature tensor. Understandably, the Hilbert transform is a signal processing method used to construct analytic signals and extract instantaneous features. For acoustic wave data at a specific spatial measurement point, the measured signal and its Hilbert transform result can be used to construct an analytic signal, from which the instantaneous amplitude, instantaneous phase, and instantaneous frequency can be further obtained.

[0063] For example, the instantaneous feature tensor can be represented as: ,in, H Represents the instantaneous feature tensor; A Indicates instantaneous amplitude; F Indicates instantaneous frequency; F Instantaneous phase represents the instantaneous amplitude. Instantaneous amplitude reflects the energy intensity of acoustic data at a corresponding time and spatial location. For abrupt changes in media interfaces, void boundaries, crack surfaces, or other areas with strong scattering, acoustic signals may exhibit significant amplitude enhancement or abnormal attenuation. Therefore, instantaneous amplitude can serve as an important feature for identifying scattering and energy anomalies. Instantaneous frequency reflects the change of acoustic data frequency components over time. Different media have different attenuation and filtering effects on different frequency components. When the target engineering area contains loose areas, water-bearing areas, weak interlayers, or areas with developed cracks, high-frequency components may be more easily attenuated, resulting in a decrease in instantaneous frequency or abnormal frequency distribution. Therefore, instantaneous frequency can be used to assist in identifying abnormal states such as enhanced media attenuation, softening, or loosening. Instantaneous phase reflects the phase change characteristics of acoustic data. Instantaneous phase can assist in identifying the wave propagation direction, phase continuity, phase jumps, and media discontinuities. When cracks, void boundaries, or strong reflective interfaces exist in the road structure, the phase of the acoustic signal may abruptly change or exhibit local discontinuities. Therefore, instantaneous phase can serve as an auxiliary feature for interface identification and anomaly localization.

[0064] Understandably, before extracting instantaneous amplitude, instantaneous frequency, and instantaneous phase using Hilbert transform, bandpass filtering, denoising, detrending, normalization, or time-windowing can be performed on the acoustic data to improve the stability of the instantaneous features. The instantaneous feature tensor can be directly input as a composite parameter in the multiphysics parameter data volume into subsequent steps; alternatively, the instantaneous amplitude, instantaneous frequency, and instantaneous phase can be input as independent channels into subsequent steps.

[0065] In step S240, the temperature gradient anomaly index data volume is calculated based on the temperature field data using a preset temperature gradient anomaly index formula. The temperature field data includes temperature distribution data collected by distributed fiber optic sensors in the target engineering area. Taking a road engineering scenario as an example, the temperature distribution inside the road structure is affected by factors such as solar radiation, air temperature changes, road surface materials, groundwater seepage, pore structure, cavities, cracks, and water content. For a healthy road structure, the temperature field usually exhibits a relatively continuous, smooth, or diurnal variation spatial distribution; when cavities, cracks, seepage channels, or water content anomalies exist, local heat capacity and thermal conductivity conditions change, which may lead to anomalies in the temperature gradient.

[0066] In some specific implementations, the temperature gradient anomaly index data volume can be represented as: ,in, G r,t Indicates planar position r In time t The corresponding temperature gradient anomaly index; r It indicates the planar location of data points in the target engineering area. Taking a road engineering scenario as an example, it can include the location along the road extension direction and the lateral location. t This indicates the acquisition time or time window. The temperature gradient anomaly index can be calculated using the following formula:

[0067] in, Indicates planar position r In time t Temperature field data at the location; c 2 indicates the data channel corresponding to the temperature field data; Indicates planar position r The reference temperature field data can be the average background temperature under no abnormal conditions, the average temperature during historical healthy periods, the average temperature of similar normal road sections, or the background temperature after seasonal correction. This represents the temperature gradient along the spatial direction of the measured temperature field data; This represents the temperature gradient along the spatial direction of the reference temperature field data; l fThis represents the seepage heat exchange coefficient. The seepage heat exchange coefficient characterizes the influence of groundwater seepage, aquifer state, or heat exchange conditions on the temperature gradient. The seepage heat exchange coefficient can be determined based on the water content of the target project area, groundwater activity, seasonal temperature variations, thermal parameters of road materials, field calibration data, or historical monitoring data. Typically, in areas with strong groundwater seepage or significant water content variations, the seepage heat exchange coefficient can be appropriately adjusted to enhance the sensitivity of the temperature gradient anomaly index to seepage-related anomalies. If prior data on seepage is lacking for the target project area, the seepage heat exchange coefficient can be initially set as an empirical value and corrected using subsequent monitoring data.

[0068] In step S250, a strain curvature data volume is obtained by calculating the local strain curvature based on the strain field data. The strain field data includes data related to the deformation of the target engineering area collected by distributed fiber optic sensors. For example, the strain curvature data volume can be represented as: ,in, K r,t Indicates planar position r In time t The corresponding strain curvature; r Indicates the planar location of data points within the target engineering area; t Indicates the collection time or time window.

[0069] In some implementations, strain curvature can be calculated from the spatial second derivative of the strain field data. Taking a road engineering scenario as an example, for discrete spatial measurement points, strain curvature can be calculated using the following expression:

[0070] in, Indicates planar position r In time t Strain field data at the location; c 3 indicates the data channel corresponding to the strain field data; Δ r It represents the spatial interval between adjacent spatial measurement points or the spatial step size used to calculate local second-order differences; Indicates position r +Δ r Strain field data at the location; Indicates position r-Δr Strain field data at the location. Spatial step Δ rThe spatial step size can be determined based on the spatial resolution of the distributed fiber optic sensor, the fiber optic deployment method, and the anomaly scale. When identifying small-scale anomalies such as cracks and local voids, a smaller spatial step size can be selected; when identifying larger-scale anomalies such as uneven settlement and slope slippage, a larger spatial step size or multi-scale spatial step sizes can be used to calculate the strain curvature separately. Through multi-scale calculation, the ability to identify both sharp local anomalies and large-scale gradually changing anomalies can be balanced.

[0071] In practical calculations, to reduce the amplification effect of noise on strain curvature calculations, the strain field data can be smoothed, outlier removed, baseline corrected, or fitted with a moving window before calculating the local strain curvature. Alternatively, a more stable strain curvature data volume can be obtained by fitting a local polynomial and then calculating the second derivative. For scenarios with high noise levels, time accumulation, event counting, peak statistics, or threshold filtering can be combined to enhance the anomaly response.

[0072] Taking road engineering scenarios as an example, strain-curvature data volumes are mainly used to reflect the local bending deformation and spatial discontinuous deformation of road structures. In healthy road structures, the strain field data usually shows a relatively smooth spatial change when vehicle loads pass over it, with corresponding strain-curvature values ​​being small or continuously changing. When there are cavities, voids, or weak areas in the roadbed, vehicle loads may cause local deflection of the pavement, and the strain-curvature data volume will show obvious peaks or alternating positive and negative changes at the corresponding locations. When cracks are present, the strain response on both sides of the crack may show abrupt changes or antisymmetric distributions, and the strain-curvature data volume can amplify this local difference. When uneven settlement exists, the strain-curvature data volume may exhibit large-scale low-frequency spatial fluctuations.

[0073] Through steps S210 to S250, three-dimensional velocity model data volumes, three-dimensional density data volumes, wave impedance data volumes, instantaneous feature tensors, temperature gradient anomaly index data volumes, and strain curvature data volumes can be obtained, respectively. These parameters describe the structural state of the target engineering area from different physical perspectives. In one embodiment, all of these parameters can be combined into a multiphysics parameter data volume to improve the accuracy of engineering structure anomaly identification. In another embodiment, some parameters can be selected based on the on-site sensor configuration and data availability. By performing the above feature extraction on the three-dimensional data volume, the original distributed fiber optic sensing data can be converted into a multiphysics parameter data volume with engineering and geophysical significance. This enables subsequent anomaly identification models and rule bases to identify structural anomalies based on multi-source physical evidence, thereby improving the reliability and interpretability of the identification results.

[0074] In some embodiments, step S300 performs spatiotemporal alignment and channel stitching on the multiphysics parameter data volume to obtain a multidimensional input tensor, which may further include, but is not limited to, steps S310 to S340: S310. Based on the preset spatial resolution, perform spatial interpolation on the multiphysics parameter data volume to complete spatial alignment. S320. Using a preset time step as a reference, perform time statistics calculations on the multiphysics parameter data volume to complete time alignment. S330, the aligned multiphysics parameter data volume is standardized; S340. The standardized multiphysics parameter data volume is spliced ​​in the channel dimension to obtain a multidimensional input tensor.

[0075] Because the 3D velocity model data volume, 3D density data volume, wave impedance data volume, instantaneous characteristic tensor, temperature gradient anomaly index data volume, and strain curvature data volume in the multiphysics parameter data volume originate from different physical fields, different processing methods, and different update mechanisms, these data volumes may differ in spatial resolution, temporal update frequency, numerical range, and dimensions. Without unified processing, the anomaly identification model will struggle to learn the correspondence between different physical field parameters at the same spatial measurement point and the same time step, potentially leading to inaccurate anomaly type identification.

[0076] In step S310, spatial interpolation is performed on the multiphysics parameter data volume based on a preset spatial resolution to complete spatial alignment. The preset spatial resolution includes the uniform spatial sampling interval used when constructing the multidimensional input tensor. This preset spatial resolution is used to determine the uniform set of spatial measurement points participating in anomaly identification within the target engineering area, so that different multiphysics parameter data volumes can be expressed on the same set of spatial measurement points. In one embodiment, the preset spatial resolution can be determined based on the spatial sampling interval of the distributed fiber optic sensor. For example, taking a road engineering scenario as an example, when the spatial sampling interval of the distributed fiber optic sensor along the road direction is 0.5m, 0.5m can be used as the preset spatial resolution; when the engineering scenario has high computational efficiency requirements, or when the scale of the anomaly to be identified is large, 1m, 2m, or other suitable distances can also be used as the preset spatial resolution. The preset spatial resolution can be determined based on the length of the target engineering area, the laying method of the distributed fiber optic sensor, the thickness of the road structure layer, the spatial scale of the structural anomaly type to be identified, and the computational resources of the anomaly identification model. After determining the preset spatial resolution, a uniform set of spatial measurement points can be established along the target engineering area. For example, the uniform set of spatial measurement points can be represented as: ,in, R Represents a unified set of spatial measurement points; r 1, r 2,…, Nr This represents the spatial measurement points obtained by dividing the space according to a preset spatial resolution. NrThis indicates the number of spatial measurement points within the target engineering area that participate in anomaly identification.

[0077] For example, taking a road engineering scenario, the spatial location of the three-dimensional velocity model data volume, the three-dimensional density data volume, and the wave impedance data volume can be represented as follows: r ,in r This can include longitudinal, transverse, and depth coordinates of the road. For temperature gradient anomaly index data volumes and strain curvature data volumes, their spatial location can be represented as... r '',in r '' can represent the planar position corresponding to the distributed fiber optic sensor within the target engineering area. This is to unify the data from the different spatial representations mentioned above into a single set of spatial measurement points. R It can perform spatial interpolation processing on various multiphysics parameter data volumes.

[0078] Spatial interpolation processing includes calculating the parameters of each multiphysics parameter data volume at its original spatial location and then applying these parameters to a unified set of measurement points in space. R The parameter values ​​at each spatial measurement point are used. Spatial interpolation processing can employ linear interpolation, nearest neighbor interpolation, spline interpolation, inverse distance weighted interpolation, Kriging interpolation, grid resampling, or projection mapping methods based on road structure layers. For example, for temperature gradient anomaly index data volumes and strain curvature data volumes distributed along the road direction, one-dimensional spatial interpolation can be performed along the road mileage direction; for planar data with longitudinal and transverse road distribution, two-dimensional spatial interpolation can be performed; for three-dimensional velocity model data volumes, three-dimensional density data volumes, and wave impedance data volumes, the three-dimensional spatial location can be determined based on the distributed fiber optic sensor burial depth, target identification depth, or road structure layer. r The parameter values ​​at that location are mapped to the corresponding unified spatial measurement points.

[0079] In step S310, all multiphysics parameter data volumes are converted to the same set of spatial measurement points. R Above. That is to say, at the spatial measurement point. ri At this point, the corresponding parameter values ​​in the three-dimensional velocity model data volume, three-dimensional density data volume, wave impedance data volume, instantaneous characteristic tensor, temperature gradient anomaly index data volume, and strain curvature data volume can be obtained simultaneously, thereby completing spatial alignment.

[0080] In step S320, time statistics are calculated on the multiphysics parameter data volume based on a preset time step to complete time alignment. The preset time step includes the uniform time interval used when constructing the multidimensional input tensor. This preset time step is used to establish a unified time axis, so that multiphysics parameter data volumes with different sampling frequencies, different update time windows, and different calculation cycles can all be mapped to the same set of time indices.

[0081] In one implementation, the unified timeline can be represented as: ,in, The Indicates a unified timeline; t 1, t 2,…, t T This represents the time index obtained by dividing the time according to a preset time step. T This indicates the number of time steps in the input anomaly detection model. The interval between two adjacent time indices is the preset time step size Δ. t .

[0082] Preset time step Δ t The time step can be determined based on the data volume with the longest update cycle in the multiphysics parameter data volume, the input window length of the anomaly identification model, the development speed of structural anomalies in the target engineering area, and the real-time requirements of the system. For example, the three-dimensional velocity model data volume and the three-dimensional density data volume are usually obtained by inverting acoustic data through a fixed time window, and their update cycles are relatively long; the instantaneous feature tensor, temperature gradient anomaly index data volume, and strain curvature data volume can usually be calculated at a higher frequency. To avoid time mismatch between different data volumes in the same input sample, a preset time step Δ that can cover the update cycles of the main parameters can be selected. t Alternatively, a uniform time window can be set according to the model input requirements.

[0083] After establishing a unified timeline, each time index can be... t k This corresponds to a specific time interval: ,in, I k Indicates the first k The time intervals corresponding to each time index; Δ t This indicates the preset time step. Subsequently, the data volumes of each multiphysics parameter fall within this time interval. I k Data mapping or aggregation within a time index t k Place.

[0084] For data volumes with sampling frequencies higher than the preset time step, such as instantaneous feature tensors, temperature gradient anomaly exponent data volumes, and strain curvature data volumes, sampling can be performed within the time interval. I k Time statistics are calculated within a given time interval. These statistics can include one or more of the following: mean, maximum, minimum, standard deviation, median, quantiles, peak count, outlier count, cumulative value, or trend slope. By calculating time statistics, a time interval can be... I k Multiple sampled values ​​are aggregated into one or more corresponding feature values.

[0085] Understandably, in summary, for any multiphysics parameter data volume... Y The calculation of time statistics can be expressed as: ,in, Y ri,tk Indicates parameters Y At the spatial measurement point r i and time index t k The aligned value at the specified location; Y ri,t Indicates parameters Y At the spatial measurement point r i and original time t The value at that location; I k Indicates the first k Stat( time intervals; This represents a function for calculating time statistics. The function for calculating time statistics can be based on parameters. Y The physical meaning is determined. For example, for strain curvature data, the maximum value, mean, standard deviation, or anomaly count can be taken; for temperature gradient anomaly index data, the mean, maximum value, or duration of exceeding the limit can be taken; for instantaneous amplitude in instantaneous characteristic tensors, the mean, maximum value, or energy statistics can be taken.

[0086] For data volumes with sampling frequencies lower than or equal to a preset time step, such as 3D velocity model data volumes, 3D density data volumes, and wave impedance data volumes updated with a fixed time window, the update time window and time interval can be determined accordingly. I k Time alignment is performed based on the overlapping relationships between time intervals. Specifically, this can be done by aligning the time intervals with the time range. I k The inversion result with the largest overlap ratio is assigned to the time index. t k ; or the coverage time range I k The most recent inversion result is retained up to this time index; when two adjacent inversion results are located at time indices... t k When comparing different times, time interpolation can also be used to obtain the time index. t k The corresponding parameter values.

[0087] Through the above processing, each type of multiphysics parameter data volume is aligned on a unified time axis. T a Each time index on t kEach location has a corresponding parameter value. In other words, at the same spatial measurement point... r i and the same time index t k Under these conditions, the corresponding feature values ​​in the three-dimensional velocity model data volume, three-dimensional density data volume, wave impedance data volume, instantaneous feature tensor, temperature gradient anomaly index data volume, and strain curvature data volume can be obtained simultaneously.

[0088] In step S330, the spatiotemporally aligned multiphysics parameter data volume is standardized. Standardization involves converting multiphysics parameter data volumes with different physical dimensions, numerical ranges, and statistical distributions into standardized feature values ​​suitable for the anomaly identification model input. Since the three-dimensional velocity model data volume is typically expressed in velocity dimensions, the three-dimensional density data volume in density dimensions, and the wave impedance data volume in impedance dimensions, and instantaneous feature tensors include parameters with different dimensions or numerical ranges such as amplitude, frequency, and phase, and the temperature gradient anomaly index data volume and strain curvature data volume also have their own physical or mathematical dimensions, directly concatenating these data volumes into the anomaly identification model would result in significant differences in numerical scales between different channels, affecting model convergence and anomaly identification stability.

[0089] In some specific implementations, a hybrid normalization strategy is employed to standardize the spatiotemporally aligned multiphysics parameter data. This hybrid normalization strategy does not forcibly convert different physics parameters into the same physical units. Instead, it selects an appropriate normalization method based on the physical meaning and data distribution of the different physics parameters, ensuring that they all ultimately represent dimensionless, similarly scaled model input features capable of characterizing relative anomalies. In other words, on the one hand, different physics parameters can use different normalization benchmarks or formulas; on the other hand, the normalized physics parameters should enter a unified numerical representation space so that the anomaly identification model can perform fusion calculations.

[0090] In one implementation, this can be applied to any spatiotemporally aligned physical field parameters. Y r,t Construct its standardized features .in, Y r,t It can represent the velocity values ​​in a three-dimensional velocity model data volume. V r,t Density values ​​in a three-dimensional density data volume r r,t Wave impedance values ​​in wave impedance data body Z r,t Sub-features in the instantaneous feature tensor, temperature gradient anomaly index G r,t or strain curvature Kr,t ; This represents the dimensionless eigenvalue after standardization.

[0091] For example, parameters reflecting the absolute level of medium properties, such as three-dimensional velocity model data volumes, three-dimensional density data volumes, and wave impedance data volumes, can be normalized using a robust normalization method based on historical reference period quantiles. For instance: ,in, MY Represents physical field parameters Y r,t The median during the historical baseline period; I Y Represents physical field parameters Y r,t Interquartile range within the historical baseline period; e This represents the stability coefficient, used to avoid a denominator that is zero or too small. The historical reference period can be selected from a historical time period in which the target engineering area is in normal service, with no obvious structural anomalies recorded, and where environmental conditions are relatively stable.

[0092] For example, for parameters that primarily reflect the degree of local anomalous changes or deviations, such as temperature gradient anomaly index data and strain curvature data, z-score standardization based on the mean and standard deviation of the baseline period can be used. For example: ,in, m Y Represents physical field parameters Y r,t The mean value during the reference period; s Y Represents physical field parameters Y r,t Standard deviation over the reference period; e This represents the stability coefficient. In this way, the standardized temperature gradient anomaly index and strain curvature can represent the degree of deviation from the normal fluctuation range.

[0093] For example, for instantaneous feature tensors, standardization can be performed based on the specific data characteristics of instantaneous amplitude, instantaneous frequency, and instantaneous phase. Instantaneous amplitude can be normalized after amplitude compression or logarithmic transformation to reduce the impact of extreme strong vibration events on the numerical scale; instantaneous frequency can be standardized based on the normal frequency band or baseline statistics; instantaneous phase can be processed using phase expansion, interval scaling, or sine / cosine encoding to avoid discontinuous effects of phase period jumps on the model input. The instantaneous features processed in the above way are also converted into dimensionless features.

[0094] It should be noted that although standardization converts each physical field parameter into dimensionless features, it does not mean that their physical meaning is lost. On the contrary, the standardized eigenvalues ​​typically represent the degree of deviation of the corresponding physical field parameter from the normal baseline state. For example, the standardized velocity feature can represent the degree of decrease in velocity relative to the speed baseline of a healthy road section; the standardized strain curvature feature can represent the degree of enhancement of local bending deformation relative to the normal fluctuation range; and the standardized temperature gradient anomaly index can represent the degree of deviation of thermal anomalies relative to the background temperature gradient. Therefore, standardization can eliminate dimensional differences while retaining the physical information needed for structural anomaly identification.

[0095] In step S330, the different physical field parameters in the multiphysics parameter data volume are converted into dimensionless standardized features under a unified numerical scale, thereby reducing the impact of differences in the dimensions and orders of magnitude of different physical quantities on the convergence and inference of the anomaly identification model.

[0096] In step S340, the standardized multiphysics parameter data volumes are concatenated along the channel dimension to obtain a multidimensional input tensor. The channel dimension includes dimensions used to distinguish different physics parameters or different feature components. After processing in steps S310 to S330, each multiphysics parameter data volume has a unified spatial index, a unified temporal index, and a unified or similar dimensionless numerical scale. Therefore, it can be used as the input for concatenating different feature channels to construct an anomaly detection model.

[0097] For example, a multidimensional input tensor can be represented as: ,in, X Represents a multidimensional input tensor; Nr Indicates the number of spatial measurement points involved in the identification within the target engineering area; T Indicates the number of time steps on a unified timeline; C This indicates the number of feature channels. Feature channels consist of standardized multiphysics parameter data volumes. For example, a feature channel may include a standardized 3D velocity model data volume, a standardized 3D density data volume, a standardized wave impedance data volume, a standardized instantaneous feature tensor, a standardized temperature gradient anomaly exponent data volume, and a standardized strain curvature data volume. If the instantaneous feature tensor is decomposed into three sub-features—instantaneous amplitude, instantaneous frequency, and instantaneous phase—these three sub-features can participate in the stitching as independent channels, or as a combined channel, depending on the input structure of the anomaly recognition model.

[0098] During channel stitching, for any spatial measurement point and any time step, the channel values ​​in the multidimensional input tensor should correspond to different standardized physical field parameters at the same spatiotemporal location. Thus, the anomaly identification model can learn the synergistic relationships between different physical field parameters. For example, in a road engineering scenario, when the velocity normalization feature decreases, the wave impedance normalization feature decreases, and the strain curvature normalization feature increases simultaneously within the same spatial region, it may collectively indicate an anomaly in the road structure's support state; when the temperature gradient anomaly index normalization feature increases and is accompanied by density normalization feature anomalies, it may indicate anomalies related to water content or cavities.

[0099] In some implementations, if some multiphysics parameter data volumes are missing due to sensor failure, insufficient acquisition conditions, or computational failure, missing value imputation, masking, neighborhood estimation, or adaptive model input channels can be used to address this. Missing value imputation can use baseline values, statistical values ​​from neighboring spatial measurement points, or valid values ​​from the previous time step; masking can be used as an additional channel input to the anomaly detection model to indicate that some physical parameters at that location or time step are unavailable.

[0100] The multidimensional input tensor obtained through step S340 corresponds to spatial measurement points within the target engineering area in the spatial dimension, temporal variations under a uniform time step in the temporal dimension, and different standardized physical field parameters in the channel dimension. This multidimensional input tensor not only preserves the physical anomaly information of the multi-physics field parameters but also eliminates the differences in the dimensions and orders of magnitude of different physical quantities, making it a unified input for the anomaly identification model.

[0101] Through steps S310 to S340, the engineering anomaly identification method of the embodiment can convert multi-physics parameter data volumes from multiple sources, at multiple scales, and with multiple dimensions into multi-dimensional input tensors under a unified spatial reference, a unified temporal reference, and a unified numerical scale. This process not only solves the problem of inconsistent spatial locations and temporal update frequencies among different data volumes, but also converts different physical field parameters into dimensionless standardized features through a hybrid normalization strategy, thereby eliminating the impact of dimensional differences on the convergence and inference stability of the anomaly identification model and improving the accuracy and reliability of identifying structural anomaly types in the target engineering area.

[0102] In some embodiments, the anomaly detection model may include a graph neural network module, an input projection layer, a Transformer encoder, and a multi-task branch network. The graph neural network module performs spatial dimension feature aggregation on the multi-dimensional input tensor to extract inter-channel dependencies and obtain spatially enhanced features. The input projection layer converts the spatially enhanced features into a fragment embedding sequence adapted to the input format of the Transformer encoder. The Transformer encoder extracts temporal features from the fragment embedding sequence to obtain an intermediate layer feature representation that characterizes the spatiotemporal evolution of the structural state of the target engineering area. The multi-task branch network outputs anomaly classification results and anomaly localization results based on the intermediate layer feature representation, and further obtains an anomaly score matrix.

[0103] The anomaly identification model in this application embodiment can use a pre-trained time-series model as the backbone model. Pre-training can be a process that utilizes historical data from engineering structure monitoring, data fragments corresponding to manual inspection records, indoor experimental data, numerical simulation data, publicly available time-series datasets, or a combination of the above. Through pre-training, the anomaly identification model can learn the temporal variation patterns of multiphysics parameter data volumes under normal and abnormal states, thereby enabling it to infer the structural anomaly types at various spatial measurement points within the target engineering area during actual monitoring.

[0104] The Transformer encoder in this embodiment can be a general Transformer encoder or a Transformer encoder pre-trained with time-series data; it can be based on a MOMENT-type time-series pre-trained model or other pre-trained time-series models capable of processing long-term time-series data. To avoid unnecessary restrictions on specific model names, this embodiment focuses on an anomaly detection model that includes a graph neural network module, an input projection layer, a Transformer encoder, and a multi-task branch network, and is capable of outputting an anomaly score matrix based on a multi-dimensional input tensor.

[0105] In some embodiments, step S400 inputs the multidimensional input tensor into a pre-trained anomaly recognition model to obtain anomaly score matrices corresponding to different structural anomaly types for each spatial measurement point within the target engineering area, further including but not limited to steps S410 to S440: S410. The multidimensional input tensor is subjected to spatial dimension feature aggregation through the graph neural network module to extract inter-channel dependencies and obtain spatial enhancement features. S420. The spatial enhancement features are grouped and temporally segmented along the spatial dimension by the input projection layer to obtain the fragment embedding sequence. S430. Input the embedded sequence of the fragments into the Transformer encoder for temporal feature extraction to obtain the intermediate layer feature representation; wherein, the Transformer encoder includes a self-attention module, and an Adapter module for fine-tuning is provided after the self-attention module; S440. Input the intermediate layer feature representation into the multi-task branch network and output the anomaly score matrix.

[0106] In step S410, the multidimensional input tensor is subjected to spatial dimension feature aggregation through the graph neural network module to extract inter-channel dependencies and obtain spatial enhancement features.

[0107] The graph neural network module includes a neural network module for feature propagation and aggregation based on a graph structure. The graph structure can include nodes and edges. Nodes can correspond to spatial measurement points within the target engineering area, different physical field parameter channels, or combinations of spatial measurement points and physical field parameter channels; edges represent the relationships between nodes. These relationships can include spatial proximity, engineering topology, adjacency along fiber optic routes, statistical correlations between physical field parameters, or coupling relationships based on engineering mechanisms. The graph neural network module can employ graph convolutional networks, graph attention networks, graph sampling aggregation networks, spatiotemporal graph neural networks, or other network structures capable of graph structure feature aggregation. The number of layers, hidden dimensions, activation functions, and adjacency relationships of the graph neural network module can be set according to the number of training samples, the length of the target engineering area, computational resources, and recognition accuracy requirements.

[0108] In one implementation, spatial measurement points within the target engineering area can be used as nodes in a graph structure, and edges can be established based on the distances between these points. Taking a road engineering scenario as an example, for distributed fiber optic sensors deployed along the road direction, adjacent or near spatial measurement points can be connected to form a chain-like or strip-like graph structure; for two-dimensional or three-dimensional target engineering areas, adjacency relationships can be constructed based on spatial distance, structural strata, or geological units. This approach is beneficial for graph neural network modules to capture the spatial continuity and extensibility of structural anomalies. In another implementation, different feature channels in the multidimensional input tensor can be used as nodes in a graph structure, and edges can be established based on the correlation or physical coupling relationships between different physical field parameters. For example, the three-dimensional velocity model data volume and the wave impedance data volume are both related to the elasticity and density of the medium, and there is a physical correlation between them; the temperature gradient anomaly index data volume and the three-dimensional density data volume may jointly reflect water content or void anomalies; the strain curvature data volume and the three-dimensional velocity model data volume may jointly reflect changes in the roadbed support state. Through this channel relationship graph, the graph neural network module can extract the dependencies between channels.

[0109] In one implementation, spatial augmentation features are obtained after processing by the graph neural network module: ,in, X ′ represents spatial augmentation features; N r Indicates the number of spatial measurement points; T Indicates the number of time steps; C ′ represents the number of feature channels output by the graph neural network module. C ′ can be equal to C It can also be greater than C .when C ′ greater than C When this occurs, it indicates that the graph neural network module generates a higher-dimensional fused feature representation based on the original feature channels.

[0110] Through step S410, the anomaly detection model can first use graph structure to fuse the spatial relationships and inter-channel dependencies in the multidimensional input tensor before entering the temporal modeling process, thereby obtaining spatially enhanced features that are more suitable for subsequent temporal feature extraction.

[0111] In step S420, the spatial enhancement features are grouped and temporally segmented along the spatial dimension by the input projection layer to obtain the fragment embedding sequence.

[0112] The input projection layer includes a network layer or processing module for converting spatial augmentation features into an input format that can be processed by the Transformer encoder. Since the Transformer encoder typically takes a sequence as input, the spatial augmentation features need to be divided into segments in a certain way, and each segment needs to be mapped to an embedding vector with a fixed dimension.

[0113] In one implementation, spatial enhancement features X Grouping can be performed along spatial dimensions. This involves dividing spatial measurement points within the target engineering area into several spatial groups, each group comprising one or more adjacent spatial measurement points. Spatial grouping can be determined based on engineering mileage sections, fiber optic measurement point neighborhoods, structural zones, geological units, or sliding spatial windows. Spatial grouping reduces the data size of a single input to the Transformer encoder while preserving local spatial continuity. For example, several consecutive spatial measurement points can be grouped together, and the spatial enhancement features within each group can be stitched together, average pooled, max pooled, or attention-weighted converged to form the group features corresponding to that group. The length of the spatial group can be determined based on the spatial scale of the structural anomaly to be identified.

[0114] Temporal segmentation involves dividing spatially enhanced features or spatial group features into several time segments along the time dimension. A sliding window approach can be used for segmentation. Let the window length be...L Step size is P Then each time segment contains continuous L Each time step is a time interval between adjacent time segments. P Each time step. Window length. L Step size is used to control the historical time range of model observations. P Used to control the degree of overlap between adjacent segments. L and P The timing of the segmentation can be determined based on the duration of the vehicle load event, the rate of anomaly development, model input requirements, and computational resources. After temporal segmentation, the input projection layer maps each time segment to a segment embedding. Segment embedding involves converting a local data block containing spatial, temporal, and channel features into a fixed-dimensional vector representation. The input projection layer can be implemented using linear mapping, one-dimensional convolution, two-dimensional convolution, multilayer perceptron, or pooling combined with linear mapping. To enable the Transformer encoder to perceive temporal order and spatial location, temporal location encoding, spatial location encoding, or engineering odometer encoding can also be added to the segment embedding.

[0115] The segment embedding sequence obtained through step S420 contains feature information from the multiphysics parameter data volume, as well as local relationships between spatial measurement points and dynamic changes within time segments, and can be used as input for the subsequent Transformer encoder.

[0116] In step S430, the embedded sequence of the fragments is input into the Transformer encoder for temporal feature extraction to obtain the intermediate layer feature representation. The Transformer encoder includes a self-attention module, followed by an Adapter module for fine-tuning.

[0117] A Transformer encoder comprises a temporal feature extraction network consisting of one or more Transformer encoding layers. Each Transformer encoding layer typically includes a self-attention module and a feedforward network. The self-attention module is used to calculate the degree of correlation between different time segments in the fragment embedding sequence, thereby capturing long-term temporal dependencies; the feedforward network is used to perform nonlinear transformations on the features output by the self-attention module, enhancing the feature representation capability.

[0118] Temporal feature extraction involves extracting feature representations from the embedded sequence of segments that reflect the temporal changes in the structural state of the target engineering area. For example, structural anomalies may manifest as short-term sudden changes or long-term slow evolution. Voids or detached areas may exhibit localized strong vibrations and sudden increases in strain curvature when vehicles pass through; settlement zones may show slow changes in strain curvature or velocity parameters; and roadbed slippage may manifest as low-frequency deformation and persistent parameter anomalies. The Transformer encoder, through its self-attention mechanism, can model these patterns of change over a longer time period.

[0119] The Transformer encoder includes a self-attention module, followed by an Adapter module for fine-tuning. The Adapter module is a lightweight, trainable module inserted within the Transformer encoder to adapt the anomaly detection model to engineering structure anomaly detection tasks, even with the Transformer encoder's backbone parameters frozen or partially frozen. By using the Adapter module, the number of parameters requiring training can be reduced with a limited number of samples, mitigating the risk of overfitting and reducing the computational resources required for training and deployment.

[0120] In one implementation, the Adapter module may include a dimensionality reduction layer, a non-linear activation layer, and a dimensionality increase layer. The dimensionality reduction layer compresses the high-dimensional features output by the self-attention module to a lower dimension, the non-linear activation layer enhances the feature transformation capability, and the dimensionality increase layer restores the compressed features to their original dimension. The bottleneck dimension of the Adapter module can be set according to the hidden dimension of the Transformer encoder, the number of training samples, and computational resources. For example, the bottleneck dimension can be set to a value much smaller than the hidden dimension of the Transformer encoder to achieve efficient parameter fine-tuning.

[0121] During training, most of the backbone parameters of the Transformer encoder can be frozen, and only the parameters in the Adapter module and the multi-task branch network can be trained; alternatively, some layers of the Transformer encoder can be unfrozen and trained according to the scale of the training data and the degree of task difference. Fine-tuning refers to adapting the anomaly recognition model to the multiphysics parameter distribution and anomaly type characteristics of the engineering structure anomaly recognition task in this application, while retaining the ability to extract general temporal features.

[0122] After processing by the Transformer encoder, intermediate layer feature representations are obtained. These intermediate layer feature representations can be the hidden states output from the last layer of the Transformer encoder, or features obtained by fusing one or more hidden states. These intermediate layer feature representations contain multiphysics parameter information, spatial neighborhood relationships, and long-term temporal dependencies from the spatial augmentation features, and can be used to output anomaly detection results in subsequent multi-task branch networks.

[0123] In step S440, the intermediate layer feature representation is input into the multi-task branch network, and the anomaly score matrix is ​​output.

[0124] Multi-task branch networks are network structures that simultaneously perform multiple related identification tasks based on shared intermediate layer feature representations. Compared to models that only output a single classification result, multi-task branch networks can simultaneously learn information such as structural anomaly type, anomaly location, and anomaly temporal distribution, thereby improving the accuracy and interpretability of anomaly identification results.

[0125] In some embodiments, the multi-task branch network includes an anomaly classification branch and an anomaly localization branch. In step S440, the intermediate layer feature representation is input into the multi-task branch network, and the output is an anomaly score matrix. This may further include, but is not limited to, steps S441 to S443: S441. Process the intermediate layer feature representation through the anomaly classification branch, and output the probability distribution of each spatial measurement point in the target engineering area belonging to each preset anomaly category. S442. Decode and restore the intermediate layer feature representation through the anomaly localization branch, and output the anomaly activation map of each spatial measurement point in the target engineering area on the time axis. S443. Based on the probability distribution and the anomaly activation map, calculate the anomaly score of each spatial measuring point within the target engineering area to obtain the anomaly score matrix.

[0126] In step S441, the intermediate layer feature representation is processed through the anomaly classification branch, and the probability distribution of each spatial measurement point in the target engineering area belonging to each preset anomaly category is output.

[0127] The anomaly classification branch is used to determine the probability that each spatial measuring point within the target engineering area belongs to different preset anomaly categories within the current time range or current time window. Preset anomaly categories can include one or more of the following: roadbed voids, loose areas, settlement areas, cracks, roadbed slippage, and normal categories; they can also be expanded to include voiding, water-bearing anomalies, piping, slope instability, frost heave, or other structural anomaly categories depending on the actual monitored object. Spatial measuring points include spatially aligned spatial measurement points within the target engineering area that participate in anomaly identification. These can correspond to measurement locations along the distributed fiber optic sensor line or to locations on a unified spatial grid after interpolation or resampling.

[0128] In one implementation, the anomaly classification branch may include a pooling layer and a fully connected classifier. The pooling layer is used to aggregate the temporal segment features or local spatial features in the intermediate layer feature representation to form the classification features of the corresponding spatial measurement points. The pooling layer may use average pooling, max pooling, attention pooling, or weighted pooling. The fully connected classifier is used to map the classification features to the probability distribution of each preset anomaly category.

[0129] For example, the probability distribution can be represented as: ,in, p r,c Representing spatial measurement points r Belongs to the preset exception category c The probability or confidence level; r Indicates spatial measurement points; c This indicates a preset anomaly category. If multiple structural anomaly types are allowed to coexist at the same spatial measurement point, a multi-label output method can be used, meaning that the preset anomaly categories are not required to be mutually exclusive. If the preset anomaly categories are set as mutually exclusive categories in specific applications, a single-label classification method can also be used. In the multi-label output method, the anomaly classification branch can output probability values ​​for each preset anomaly category. For example, the same spatial measurement point may simultaneously exhibit crack and settlement-related features, or a loose area may accompany the development of a cavity; therefore, using multi-label output is beneficial for representing complex anomaly scenarios.

[0130] In step S442, the intermediate layer feature representation is decoded and restored through the anomaly localization branch, outputting anomaly activation maps of each spatial measurement point within the target engineering area on the time axis. The anomaly localization branch is used to determine the spatial location of anomalies within the target engineering area and their activation degree on the time axis. Unlike the anomaly classification branch, which mainly outputs category probabilities, the anomaly localization branch focuses more on which spatial measurement points, time steps, or time segments the anomalies occur at.

[0131] The decoding and restoration process involves restoring or mapping the intermediate layer feature representations output by the Transformer encoder to the spatial and temporal resolution corresponding to the input data. Since the input projection layer and the Transformer encoder may have performed spatial grouping, temporal segmentation, and embedding mapping on the original spatial augmentation features, the decoding and restoration process is required to convert the high-dimensional features back into the anomalous activation results of each spatial measurement point in the target engineering area on the time axis.

[0132] In one implementation, the anomaly localization branch can be implemented using an upsampling layer, a deconvolution layer, an interpolation layer, a convolutional decoder, an attention decoder, or a multilayer perceptron decoder. For spatially grouped data, the localization results at the spatial group level can be restored to the spatial measurement point level; for temporally segmented data, the localization results at the segment level can be restored to the time step level on a unified time axis.

[0133] For example, an abnormal activation graph can be represented as: ,in, m r,t Representing spatial measurement points r In time t Abnormal activation intensity at the location; r Indicates spatial measurement points; t This represents a time step on a unified time axis. The anomalous activation intensity can be a normalized value, such as one ranging from 0 to 1; or it can be an unnormalized anomalous response value. Typically, m r,t The larger the value, the more likely it is to be a spatial measuring point. r In time t The higher the probability or intensity of an anomaly.

[0134] By outputting an anomaly activation graph through the anomaly localization branch, it is possible to determine the time period and spatial range of anomalies, providing a location basis for subsequent rule base comprehensive judgment, disaster level assessment, and engineering response.

[0135] In step S443, based on the probability distribution and the anomaly activation map, the anomaly score of each spatial measuring point within the target engineering area is calculated to obtain the anomaly score matrix.

[0136] The anomaly score matrix is ​​used to represent the comprehensive anomaly degree of different structural anomaly types corresponding to each spatial measuring point within the target engineering area. For example, the anomaly score matrix can be represented as: Qr , c ,in, Qr , c Representing spatial measurement points r Corresponding to preset exception categories c The overall abnormal score; r Indicates spatial measurement points;c Indicates the preset exception category.

[0137] In one implementation, the anomaly score matrix can be calculated based on the probability distribution of the anomaly classification branch output and the maximum value of the anomaly activation map on the time axis of the anomaly localization branch output: ,in, p r,c Representing spatial measurement points r Belongs to the preset exception category c The probability or confidence level; m r,t Representing spatial measurement points r In time t Abnormal activation intensity at the location; This indicates the spatial measurement points within the current input time range. r This corresponds to the maximum value of the anomalous activation map on the time axis. This maximum value characterizes the strongest activation level of the anomalous activity at that spatial measurement point within the current time frame.

[0138] The anomaly classification branch provides a classification judgment of "which type of structural anomaly it belongs to," while the anomaly localization branch provides a location and intensity judgment of "whether the spatial measurement point shows obvious anomalous activation on the time axis." Multiplying the two results in a high overall anomaly score only when a spatial measurement point has both a high category probability and strong anomalous activation on the time axis. This reduces false alarms caused solely by high category probabilities or localized noise activation.

[0139] Abnormal score matrix Q r,c This can serve as a preliminary identification result of structural anomaly types within the target engineering area, and can be further input into a subsequent rule base for comprehensive judgment. Each row in the anomaly score matrix corresponds to a spatial measurement point, and each column corresponds to a preset anomaly category. By viewing the anomaly score matrix, it can be determined which spatial measurement points in the target engineering area have a high membership strength to which structural anomaly types.

[0140] Through steps S410 to S443, the anomaly recognition model in this embodiment first utilizes a graph neural network module to fuse spatial relationships and inter-channel dependencies in the multi-dimensional input tensor. Then, it constructs a fragment embedding sequence through an input projection layer and extracts long-term temporal features using a Transformer encoder with an Adapter module. Finally, it outputs an anomaly score matrix through anomaly classification and anomaly localization branches. This approach fully utilizes the collaborative information between multi-physics parameters and models the evolution of engineering structural anomalies over time, thereby improving the recognition accuracy and localization capability of different structural anomaly types at various spatial measurement points within the target engineering area.

[0141] In some embodiments, step S500, based on the anomaly score matrix and the multiphysics parameter data volume, performs a comprehensive judgment based on a preset rule base to obtain the structural anomaly type of the target engineering area, further including but not limited to steps S510 to S530: S510. For each preset candidate structural anomaly type, determine whether the corresponding comprehensive anomaly score in the anomaly score matrix meets the preset score threshold condition. S520. Determine whether the number of physical field parameters in the multi-physics parameter data volume that meet the physical property parameter anomaly conditions corresponding to the candidate structural anomaly type meets the preset number threshold condition. S530. If the comprehensive anomaly score meets the preset score threshold condition, the number of physical field parameters that meet the physical property parameter anomaly condition meets the quantity threshold condition, and the combination relationship of the met physical property parameter anomaly conditions conforms to the physical cause mechanism of the candidate structural anomaly type, it is determined that there is a candidate structural anomaly type among the structural anomaly types of the target engineering area.

[0142] A rule base is a pre-established set of discrimination rules used to describe the correspondence between different candidate structural anomaly types and multiphysics parameter anomaly patterns. The rule base can be established based on engineering structural mechanics principles, geophysical exploration principles, historical inspection records, manual inspection conclusions, field verification results, indoor test data, numerical simulation results, or expert experience. The role of the rule base is to constrain the anomaly score matrix output by the anomaly identification model with physical mechanisms and verify it with engineering logic, thereby avoiding misjudgments caused by relying solely on the model output. Candidate structural anomaly types include the pre-defined structural anomaly categories in the rule base that need to be identified and discriminated. Candidate structural anomaly types can include one or more of the following: roadbed voids, loose areas, settlement areas, cracks, and roadbed slippage.

[0143] Anomaly score matrix is ​​used to characterize the degree of anomaly for different candidate structural anomaly types corresponding to each spatial measuring point within the target engineering area. Multiphysics parameter data volume is used to characterize the physical state of each spatial measuring point or spatial region within the target engineering area. By combining the two, the engineering anomaly identification method of the embodiment can simultaneously utilize the identification capability of the data-driven model and the interpretability of physical rules to obtain more reliable structural anomaly type identification results.

[0144] In step S510, for each preset candidate structural anomaly type, it is determined whether the corresponding comprehensive anomaly score in the anomaly score matrix meets the preset score threshold condition.

[0145] The comprehensive anomaly score is a numerical value in the anomaly score matrix used to represent the degree of anomaly for a specific candidate structural anomaly type corresponding to a given spatial measurement point. For example, the score threshold condition can be expressed as: ,in, θc i Indicates the candidate structural anomaly type c i The corresponding score threshold can be determined based on validation set results, receiver operating characteristic (ROC) curves, precision-recall curves, historical false positive rates, historical false negative rates, engineering grade, safety risk requirements, or engineering experience. For candidate structural anomaly types with high safety risks or rapid development, a lower score threshold can be set to improve sensitivity; for candidate structural anomaly types easily affected by noise or with high false positive costs, a higher score threshold can be set to improve discrimination stability. In one specific implementation, different score thresholds can be set for different candidate structural anomaly types. For example, roadbed voids and roadbed slippage have high safety risks, and their score thresholds can be appropriately reduced according to safety management requirements; for anomaly types such as loose areas or minor cracks that develop relatively slowly, their score thresholds can be set in conjunction with on-site treatment strategies. The score threshold can also be dynamically adjusted based on season, rainfall, traffic load, engineering grade, or historical anomaly distribution.

[0146] In step S510, if a certain spatial measuring point r Corresponding candidate structural anomaly type c i Comprehensive abnormal score Qr , c i If the preset score threshold is not met, it can be determined that the spatial measurement point will not trigger the rule-based comprehensive judgment of this candidate structural anomaly type, or it can be marked as a low-confidence anomaly candidate. Qr , c i If the preset score threshold condition is met, proceed to subsequent steps S520 and S530 to further determine whether the multiphysics parameter data volume supports the candidate structural anomaly type.

[0147] Step S510 allows the anomaly score matrix output by the anomaly identification model to be used as a preliminary screening basis, reducing invalid calculations of the rule base for low-risk or low-confidence areas and improving system processing efficiency.

[0148] In step S520, it is determined whether the number of physical field parameters in the multiphysics parameter data volume that satisfy the physical property parameter anomaly conditions corresponding to the candidate structural anomaly type meets the preset number threshold condition.

[0149] Anomaly conditions for physical property parameters include pre-defined conditions for a specific candidate structural anomaly type, used to determine whether multiphysics parameters exhibit anomalies consistent with that candidate structural anomaly type. Different candidate structural anomaly types have different physical causes, therefore their corresponding anomaly conditions for physical property parameters may also differ. Physical field parameters include parameters in the multiphysics parameter data volume, which may include parameters in the three-dimensional velocity model data volume, three-dimensional density data volume, wave impedance data volume, instantaneous characteristic tensor, temperature gradient anomaly index data volume, and strain curvature data volume. For ease of identification, the physical field parameters at a spatial measurement point or within its neighborhood can be compared with the corresponding normal reference value, background value, spatial neighborhood statistical value, or time history value to determine whether the anomaly conditions for physical property parameters are met.

[0150] For example, for roadbed cavities, abnormal physical property conditions may include decreased velocity, decreased wave impedance, decreased density, increased temperature gradient anomaly index, and increased strain curvature; for loose areas, abnormal physical property conditions may include moderately decreased velocity, moderately decreased wave impedance, and slightly decreased density; for settlement areas, abnormal physical property conditions may include anomalies in the spatial velocity gradient, large-scale fluctuations in strain curvature, and a systematic decrease in wave impedance; for cracks, abnormal physical property conditions may include a linear or banded distribution of the temperature gradient anomaly index, antisymmetric changes in strain curvature on both sides of the crack, and strong scattering response in the instantaneous characteristic tensor; for roadbed slippage, abnormal physical property conditions may include continuously increasing strain curvature, decreased velocity, and changes in density.

[0151] Normal baseline values ​​can be determined based on historical monitoring data of the target engineering area in a healthy state, statistical values ​​of adjacent normal road sections, design parameters, engineering survey results, on-site testing results, or expert experience. Normal baseline values ​​can be fixed values ​​or dynamically updated values ​​that change with seasons, temperature, traffic load, water content, or engineering structure type. By setting normal baseline values, the current state of multiphysics parameters can be compared with the normal state, thereby identifying abnormal deviations.

[0152] It should be noted that the "number of physical field parameters that satisfy the anomaly conditions of the physical property parameters corresponding to the candidate structural anomaly type" in step S520 is not limited to simply counting the number of data volume types. In some implementations, one physical field parameter can correspond to multiple anomaly conditions. For example, an instantaneous feature tensor can include three sub-features: instantaneous amplitude, instantaneous frequency, and instantaneous phase; a strain curvature data volume can simultaneously correspond to peak anomalies, spatial morphology anomalies, and duration anomalies. In this case, it can be counted as multiple conditions according to the rule base settings, or it can be counted by merging according to the physical field parameter type. The specific counting method can be determined according to the rule base configuration.

[0153] In step S520, if the number of entities meeting the physical property parameter anomaly conditions does not meet the preset threshold, even if the overall anomaly score is high, the candidate structural anomaly type may not be determined to exist in the target engineering area. Instead, it may be marked as an anomaly to be verified, a low-confidence anomaly, or a noise interference event. If the number of entities meeting the physical property parameter anomaly conditions reaches the preset threshold, then step S530 is performed for physical logic consistency verification. Through step S520, this embodiment of the application can avoid misjudgment caused by a single physical field parameter anomaly, and make the final structural anomaly type determination based on the joint support of multiple sources of evidence.

[0154] In step S530, if the comprehensive anomaly score meets the preset score threshold condition, the number of physical field parameters that meet the physical property parameter anomaly condition meets the quantity threshold condition, and the combination relationship of the met physical property parameter anomaly conditions conforms to the physical cause mechanism of the candidate structural anomaly type, then it is determined that there is a candidate structural anomaly type among the structural anomaly types of the target engineering area.

[0155] The physical mechanisms of formation include the physical causes of the generation and development of a candidate structural anomaly type and its manifestation in multiple physics parameters. While different candidate structural anomaly types may exhibit similarities in certain individual parameters, their multi-parameter combinations are usually different. For example, both roadbed cavities and loose zones may show reduced velocity and wave impedance, but roadbed cavities are usually accompanied by a more significant decrease in density, strong scattering characteristics, or strain curvature anomalies; anomalies caused by water content may show abnormal temperature gradients and density changes, while dry cavities may show a significant decrease in density. Verifying the physical mechanisms of formation can improve the ability to distinguish between different anomaly types.

[0156] Understandably, a combination relationship that conforms to the physical causal mechanism of the candidate structural anomaly type can include at least one of the following meanings: consistent parameter change direction, i.e., the satisfied physical property parameter anomaly conditions are consistent with the physical causal mechanism of the candidate structural anomaly type in the direction of change; consistent parameter spatial location, i.e., multiple physical field parameter anomalies should appear at the same spatial measurement point or in adjacent spatial ranges; consistent parameter temporal evolution, i.e., multiple physical field parameter anomalies should appear within the same or related time windows, or their development trend should conform to the evolution law of the candidate structural anomaly type; reasonable parameter strength relationship, i.e., the degree of anomaly of different physical field parameters should conform to the typical manifestation of the candidate structural anomaly type; and spatial morphology conforms to engineering characteristics, i.e., the spatial morphology of multiple physical field parameter anomalies should conform to the geometric or engineering manifestation of the candidate structural anomaly type.

[0157] In one implementation, step S530 can be implemented using a rule engine. The rule engine pre-configures score threshold conditions, physical property parameter anomaly conditions, quantity threshold conditions, and physical-logical consistency verification conditions for each candidate structural anomaly type. When a spatial measurement point or spatial region simultaneously meets the above conditions, the candidate structural anomaly type is output as one of the structural anomaly types of the target engineering area. In another implementation, step S530 can also be implemented using a decision tree, expert system, fuzzy inference, Bayesian rule fusion, or an interpretable machine learning model. These implementations can take the comprehensive anomaly score, the number of satisfied conditions, the degree of deviation of physical field parameters, spatial continuity, and temporal persistence as inputs, and output a judgment result indicating whether the candidate structural anomaly type is valid.

[0158] In practical engineering applications, the structural anomaly types of the target engineering area can be output in the form of spatial measurement points, spatial segments, or regions. For example, when multiple consecutive spatial measurement points are determined to have the same candidate structural anomaly type, these spatial measurement points can be merged into an anomaly segment, and the start and end positions, length, candidate structural anomaly type, and comprehensive confidence score of the anomaly segment can be output. When multiple candidate structural anomaly types simultaneously meet the discrimination criteria, multiple structural anomaly types can be output, or they can be sorted according to the comprehensive anomaly score, the number of physical property parameter anomaly conditions met, the degree of deviation of physical field parameters, and engineering priority.

[0159] Through step S530, the engineering anomaly identification method of the embodiment can introduce engineering physical logic verification on the basis of model identification results, so that the final structural anomaly type not only has data-driven statistical basis, but also has multi-physics field parameter support and physical cause explanation, thereby improving the credibility of the diagnostic results.

[0160] In summary, through steps S510 to S530, the engineering anomaly identification method of this embodiment can, for each preset candidate structural anomaly type, sequentially determine whether the comprehensive anomaly score in the anomaly score matrix meets the score threshold condition, whether the number of physical field parameters that meet the physical property parameter anomaly condition meets the quantity threshold condition, and whether the combination relationship of the met physical property parameter anomaly conditions conforms to the physical causal mechanism of the corresponding candidate structural anomaly type. Only when all the above conditions are met is it determined that the candidate structural anomaly type exists in the structural anomaly type of the target engineering area. This comprehensive discrimination method can effectively reduce false alarms caused by a single model output or a single physical field parameter anomaly, and improve the accuracy, stability, and interpretability of the engineering structural anomaly identification results.

[0161] In some embodiments, taking a road engineering scenario as an example, the preset candidate structural anomaly types include roadbed cavities. Roadbed cavities include low-density or unsupported areas in the roadbed, base course, or underlying medium formed due to filler loss, groundwater erosion, pipeline leakage, local voids, or long-term dynamic loads. The following section uses roadbed cavities as a specific structural anomaly type to explain how to comprehensively determine the structural anomaly type of the target engineering area based on the anomaly score matrix and multiphysics parameter data volume, using a preset rule base.

[0162] In one implementation, the conditions for determining the presence of roadbed cavities in the abnormal structural types of the target engineering area include multiple of the following conditions, preferably all of which are met simultaneously or at least a preset number of them: (1) The overall abnormal score is greater than or equal to the corresponding score threshold; (2) The velocity value in the three-dimensional velocity model data volume is less than the corresponding velocity reference value; (3) The wave impedance value in the wave impedance data body is less than the corresponding wave impedance reference value; (4) The density value in the three-dimensional density data volume is less than the corresponding density reference value; (5) The temperature gradient anomaly index in the temperature gradient anomaly index data volume is greater than the corresponding temperature gradient baseline value; (6) The strain curvature in the strain curvature data volume is greater than the corresponding strain curvature reference value.

[0163] The comprehensive abnormal score can be represented as an abnormal score matrix. Qr , c The fraction corresponding to roadbed voids. For example, it can be represented as Q r,cavity ,in, cavity This indicates that the candidate anomaly type is roadbed voids. The corresponding score threshold can be denoted as... i cavity In one implementation, the following conditions can be used to determine whether the model identification results support roadbed voids: The score threshold can be determined based on the validation set test results, historical detection samples, manual review results, false alarm rate requirements, and road safety level.

[0164] The reference values ​​for velocity, wave impedance, density, temperature gradient, and strain curvature can be determined based on historical monitoring data of the target engineering area under normal conditions, statistical values ​​of adjacent healthy road sections, road design parameters, engineering survey data, or on-site calibration results. These reference values ​​can be fixed or dynamically updated based on seasons, temperature, traffic load, water content, and road structure type.

[0165] In specific judgments, the rule base can require all the above conditions to be met, or it can require the comprehensive anomaly score to meet a score threshold condition, and at least a preset number of items among velocity, wave impedance, density, temperature gradient anomaly index, and strain curvature to meet the corresponding physical property parameter anomaly conditions. For example, when the comprehensive anomaly score meets the score threshold condition, and at least three or four physical property parameter anomaly conditions are met simultaneously, and these anomalies overlap or are adjacent in spatial location, it can be determined that there are roadbed cavities in the anomaly structure type of the target engineering area. Using the "at least a preset number of items to be met" approach can maintain the applicability of the rule base even when some data is missing or there is noise interference.

[0166] Understandably, for road engineering scenarios, in addition to roadbed voids, the rule base can also set corresponding judgment conditions for other candidate anomaly types.

[0167] Loose zones are typically caused by insufficient compaction, softening from water immersion, or structural loosening due to long-term vehicle vibration. Loose zones still retain some solid filling, and compared to roadbed cavities, their density and wave impedance reduction is usually less severe. Therefore, in one embodiment, the criteria for determining a loose zone may include: a comprehensive anomaly score greater than or equal to the score threshold corresponding to the loose zone; a moderate decrease in velocity value in the three-dimensional velocity model data volume relative to the velocity reference value; a moderate decrease in wave impedance value in the wave impedance data volume relative to the wave impedance reference value; a slight decrease in density value in the three-dimensional density data volume relative to the density reference value; and the aforementioned anomalies are spatially distributed in patches or continuously.

[0168] Settlement zones are typically caused by the compression of subgrade fill, consolidation of weak underlying layers, development of underground cavities, or uneven support. The physical characteristics of settlement zones tend to be large-scale, slowly developing deformation anomalies. Therefore, in one implementation, the criteria for determining a settlement zone may include: a comprehensive anomaly score greater than or equal to the corresponding score threshold for the settlement zone; anomalies in the velocity values ​​or velocity spatial gradients in the three-dimensional velocity model data volume; low-frequency continuous fluctuations in the strain curvature data volume over a large spatial range; a systematic decreasing trend in the wave impedance values ​​in the wave impedance data volume within the settlement area; and the aforementioned anomalies being persistent or cumulative over time.

[0169] Cracks are typically caused by temperature fatigue, vehicle load fatigue, flexural failure, or uneven settlement. Cracks often manifest as local discontinuities, easily leading to wave field scattering, phase abrupt changes, and local strain abrupt changes. Therefore, in one implementation, the criteria for crack determination may include: a comprehensive anomaly score greater than or equal to the score threshold corresponding to the crack; a local enhancement of the instantaneous amplitude or a sudden change in the instantaneous phase in the instantaneous characteristic tensor; a linear or banded distribution of the temperature gradient anomaly index in the temperature gradient anomaly index data volume; a peak or antisymmetric change in the strain curvature in the strain curvature data volume on both sides of the crack; and a slight decrease in velocity values ​​in the three-dimensional velocity model data volume, to distinguish them from deep loose areas or cavities.

[0170] Subgrade slippage is typically caused by rainfall infiltration, increased pore water pressure, decreased slope stability, or the development of potential slip surfaces. Subgrade slippage exhibits a certain directionality and continuous evolution. Therefore, in one implementation, the criteria for determining subgrade slippage may include: a comprehensive anomaly score greater than or equal to the corresponding threshold for subgrade slippage; a continuous increase in strain curvature in the strain curvature data volume within the subgrade or slope toe region; a decrease in velocity values ​​in the three-dimensional velocity model data volume; an abnormal increase in density values ​​in the three-dimensional density data volume due to increased water content or abnormal spatial distribution; an anomaly in the temperature gradient anomaly index in the temperature gradient anomaly index data volume related to seepage or water content changes; and the aforementioned anomalies are persistent in time or correlated with external factors such as rainfall or changes in groundwater level.

[0171] In some implementations, the rule base can also set different risk levels or confidence scores for different candidate anomaly types. For example, for roadbed cavities, if the comprehensive anomaly score meets the score threshold condition, and all five physical property parameters (velocity, wave impedance, density, temperature gradient anomaly index, and strain curvature) are also anomaly conditions, it can be determined as a high-confidence roadbed cavity; if only three or four of the physical property parameters are anomaly conditions, it can be determined as a medium-confidence or low-confidence roadbed cavity, and a manual review or subsequent follow-up monitoring will be prompted.

[0172] In some embodiments, after obtaining the structural anomaly type of the target engineering area by performing step S500, further steps including but not limited to S610 to S640 may be performed to perform risk quantification evaluation on the identified structural anomaly type and obtain the disaster risk level of the target engineering area.

[0173] S610. Extract the deviation of each physical field parameter from the normal reference value in the multiphysics parameter data volume; S620: Obtain the preset anomaly score weights and the deviation weights corresponding to each physical field parameter; S630. Based on the comprehensive anomaly score, anomaly score weight, deviation of each parameter and corresponding deviation weight in the anomaly score matrix, perform weighted summation to obtain the comprehensive disaster risk index. S640. Based on the preset numerical range of the comprehensive disaster risk index, the disaster risk level of the target project area is obtained.

[0174] In step S610, the parameter deviation of each physical field parameter in the multiphysics parameter data volume relative to the normal reference value is extracted.

[0175] The physical field parameters may include one or more of the following: velocity values ​​in the three-dimensional velocity model data volume, density values ​​in the three-dimensional density data volume, wave impedance values ​​in the wave impedance data volume, instantaneous amplitude, instantaneous frequency and instantaneous phase in the instantaneous characteristic tensor, temperature gradient anomaly index in the temperature gradient anomaly index data volume, and strain curvature in the strain curvature data volume.

[0176] Normal baseline values ​​include reference values ​​used to represent the target project area in a normal, healthy, or structurally normal state. Normal baseline values ​​can be determined in at least one of the following ways: based on historical health period monitoring data of the target project area; based on statistical results from adjacent normal road sections or normal spatial monitoring points; based on engineering design parameters, engineering survey data, or on-site testing data; based on data from time periods confirmed by manual inspection to be free of abnormalities; or dynamically updated based on season, temperature, traffic load, and moisture content.

[0177] Parameter deviation is used to characterize the degree of deviation of the current physical field parameters from the normal reference value. Parameter deviation can be an absolute deviation, a relative deviation, a standardized deviation, or a normalized anomaly intensity value.

[0178] In step S620, the preset abnormal score weights and the deviation weights corresponding to each physical field parameter are obtained.

[0179] The anomaly score weight is used to characterize the contribution of the overall anomaly score in the anomaly score matrix to the overall disaster risk index. The deviation weight is used to characterize the contribution of the parameter deviation corresponding to each physical field parameter to the overall disaster risk index.

[0180] In one implementation, the weight of the outlier score can be represented as: w Q The deviation weights corresponding to each physical field parameter can be expressed as follows: w V , w ρ , w Z , w H, w G and w K .in, w V This indicates the deviation weight corresponding to the speed value; w ρ This indicates the deviation weight corresponding to the density value; w Z This indicates the deviation weight corresponding to the wave impedance value; w H This represents the deviation weight corresponding to the instantaneous feature tensor; w G The deviation weights corresponding to the temperature gradient anomaly index; w K This represents the deviation weight corresponding to the strain curvature. The above weights can be determined in at least one of the following ways: based on the regression optimization results of the validation set; based on the contribution of different physical field parameters to the disaster level in historical disease samples; based on expert experience or engineering maintenance specifications; set separately according to the physical causal mechanisms of different structural anomaly types; or dynamically calibrated based on on-site verification results.

[0181] In some implementations, the weights can satisfy normalization constraints, for example: In this system, all weights are non-negative. This normalization constraint helps to control the comprehensive disaster risk index within a relatively stable numerical range, facilitating subsequent disaster risk level classification. It should be noted that whether the weights are strictly normalized can be determined based on the actual risk index design method and does not constitute a limitation on this application.

[0182] In step S630, the comprehensive disaster risk index is obtained by performing a weighted summation calculation based on the comprehensive anomaly score, anomaly score weight, deviation of each parameter and corresponding deviation weight in the anomaly score matrix.

[0183] The comprehensive disaster risk index is used to comprehensively characterize the degree of danger of a specific spatial measurement point or spatial segment for a particular type of structural anomaly. The comprehensive disaster risk index can be expressed as: F r,c Where Fr,c represents the spatial measuring point r Corresponding structural anomaly type c The comprehensive disaster risk index; r 'c' represents a spatial measuring point; 'c' represents the structural anomaly type.

[0184] In one implementation, the comprehensive disaster risk index can be calculated as follows:

[0185] in, Qr,c Represents spatial measurement points in the anomaly score matrix r Corresponding structural anomaly type c The overall abnormal score; Δ V r,t Δ represents the deviation of velocity values ​​in a three-dimensional velocity model data volume from normal reference values. r r,t Δ represents the deviation of density values ​​in a three-dimensional density data volume from normal reference values. Z r,t Δ represents the deviation of the wave impedance value in the wave impedance data body from the normal reference value. H r,t Δ represents the deviation of the instantaneous characteristic tensor from the normal baseline state. G r,t The parameter Δ represents the degree of deviation of the temperature gradient anomaly index from the normal baseline value. K r,t This parameter indicates the degree of deviation of the strain curvature from the normal reference value. w Q Indicates the weight of outlier scores; w V , w ρ , w Z , w H , w G and w K These represent the deviation weights of the corresponding physical field parameters.

[0186] It should be noted that if the determination of a certain structural anomaly type does not depend on all physical parameters, or if some physical parameters are unavailable in the current scenario, then only physical parameters related to or available for that structural anomaly type can be selected for weighted summation. For unavailable parameters, their deviation weights can be reset to zero, or the weights can be renormalized based on available parameters.

[0187] In step S640, the disaster risk level of the target engineering area is obtained based on the preset numerical range to which the comprehensive disaster risk index belongs. The preset numerical range includes a pre-defined range of values ​​used to map the comprehensive disaster risk index to different disaster risk levels. For example, taking a road engineering scenario, the disaster risk level can be set to Level 1 High Risk, Level 2 Medium Risk, Level 3 Low Risk, and Normal, or it can be set to Red Alert, Orange Alert, Yellow Alert, and No Alert, etc., according to road management needs. In one implementation, the disaster risk level can be divided as follows: when F r,cA value ≥0.75 is considered Level 1 high risk; a value ≤0.50 is considered high risk. F r,c When <0.75, it is judged as Level II, medium risk; when 0.30≤ F r,c When <0.50, it is judged as Level III low risk; when F r,c When the value is less than 0.30, it is considered normal or no warning is issued. Among these, F r,c Representing spatial measurement points r Corresponding structural anomaly type c The comprehensive disaster risk index. The above numerical range can be adjusted based on validation set test results, historical disaster cases, road grade, traffic flow, maintenance standards, and safety management requirements.

[0188] Through steps S610 to S640, the engineering anomaly identification method of this embodiment can, after obtaining the structural anomaly type of the target engineering area, further calculate a comprehensive disaster risk index based on the anomaly score matrix and multiphysics parameter data volume, and determine the disaster risk level based on the comprehensive disaster risk index. This method can not only identify "whether an anomaly exists" and "what type of anomaly it is," but also evaluate "the degree of anomaly risk and the priority of handling," thereby improving the application value of engineering structural anomaly monitoring results in engineering operation and maintenance.

[0189] Secondly, embodiments of this application provide an engineering anomaly identification system. The system is used to execute the engineering anomaly identification method of any embodiment of the first aspect, and is capable of sensing, identifying, and interpreting geological or structural anomalies within a target engineering area based on multi-source data collected by distributed fiber optic sensors.

[0190] The engineering anomaly identification system of the embodiment includes a data acquisition module, a feature extraction module, a tensor construction module, a model reasoning module, and a rule interpretation module.

[0191] The data acquisition module is used to acquire three-dimensional data volumes of the target engineering area collected by distributed fiber optic sensors. These three-dimensional data volumes include acoustic wave data, temperature field data, and strain field data. The target engineering area can be a road, railway, tunnel, underground passage, integrated utility tunnel, slope, foundation pit, dam, mine, goaf, bridge foundation, pile foundation, culvert, area surrounding underground pipelines, foundation, or other engineering area. The target engineering area can be a linear, planar, or three-dimensional spatial engineering area, and its interior or periphery may contain engineering anomalies such as cavities, voids, looseness, cracks, fissures, abnormal water content, abnormal seepage, uneven settlement, slippage, goaf development, or localized structural damage. The distributed fiber optic sensors can be deployed along the extension direction, edge areas, or key structural parts of the target engineering area to acquire continuous or quasi-continuous monitoring data from multiple spatial measurement points along the route.

[0192] A three-dimensional data volume can be represented as: ,in, r Indicates the location of the spatial measurement point corresponding to the distributed optical fiber sensor; c i Indicates a data channel; t This indicates the acquisition time. Data channels include channels for acoustic wave data, temperature field data, and strain field data. Acoustic wave data reflects the acoustic or elastic wave response induced by vehicle loads, environmental vibrations, active excitation, or other vibration sources within the target engineering area; temperature field data reflects the temperature distribution within the target engineering area; and strain field data reflects the strain response of the target engineering area under load or deformation.

[0193] In some implementations, the data acquisition module may further include a data receiving unit, a time synchronization unit, a spatial positioning unit, and a data buffering unit. The time synchronization unit is used to time-stamp data from different channels; the spatial positioning unit is used to correlate the acquired data with the spatial measurement point locations of the distributed fiber optic sensor; and the data buffering unit is used to temporarily store the continuously acquired three-dimensional data volume.

[0194] The feature extraction module is used to extract features from the three-dimensional data volume to obtain a multiphysics parameter data volume. The multiphysics parameter data volume includes at least one of the following: a three-dimensional velocity model data volume, a three-dimensional density data volume, a wave impedance data volume, an instantaneous feature tensor, a temperature gradient anomaly exponent data volume, and a strain curvature data volume.

[0195] In one implementation, the feature extraction module can perform inversion processing on the acoustic wave data to obtain a three-dimensional velocity model data volume and a three-dimensional density data volume. The three-dimensional velocity model data volume can be represented as follows: V r,t Used to characterize spatial location r In timet Velocity distribution; the three-dimensional density data volume can be represented as r r,t Used to characterize spatial location r In time t The density distribution. The feature extraction module can also calculate the wave impedance data volume based on the three-dimensional velocity model data volume and the three-dimensional density data volume: ,in, Z r,t Indicates spatial location r In time t The wave impedance value.

[0196] In some implementations, the feature extraction module can also extract instantaneous amplitude, instantaneous frequency, and instantaneous phase from the acoustic wave data using Hilbert transform, and combine the instantaneous amplitude, instantaneous frequency, and instantaneous phase to obtain an instantaneous feature tensor. ,in, A Indicates instantaneous amplitude. F Indicates instantaneous frequency. F Indicates the instantaneous phase.

[0197] The feature extraction module can also calculate the temperature gradient anomaly index data volume based on the temperature field data. G r,t It is used to characterize the degree of anomaly of the measured temperature gradient relative to the background temperature gradient; it can also calculate strain curvature data based on strain field data. K r,t It is used to characterize the local bending or discontinuous changes of the strain field in space.

[0198] The tensor construction module is used to perform spatiotemporal alignment and channel stitching on multiphysics parameter data volumes to obtain multidimensional input tensors. Since different multiphysics parameter data volumes may have different spatial resolutions, time update frequencies, and physical dimensions, the tensor construction module can first perform spatial interpolation on the multiphysics parameter data volume based on a preset spatial resolution to complete spatial alignment; then, based on a preset time step, it can perform time statistics calculation on the multiphysics parameter data volume to complete time alignment.

[0199] After completing spatiotemporal alignment, the tensor construction module can employ a hybrid normalization strategy to standardize the multiphysics parameter data volume. This hybrid normalization strategy transforms different physics parameters into dimensionless, standardized features with similar numerical scales, thereby eliminating the impact of differences in the dimensions and orders of magnitude of different physical quantities on model inference.

[0200] Subsequently, the tensor construction module concatenates the standardized multiphysics parameter data volume along the channel dimension to obtain a multidimensional input tensor: ,in,X Represents a multidimensional input tensor; N r Indicates the number of spatial measurement points involved in the identification within the target engineering area; T Indicates the number of time steps on a unified timeline; C Indicates the number of feature channels.

[0201] The model inference module is used to input multidimensional input tensors into a pre-trained anomaly recognition model to obtain anomaly score matrices corresponding to different structural anomaly types for each spatial measurement point within the target engineering area.

[0202] In one implementation, the anomaly detection model may include a graph neural network module, an input projection layer, a Transformer encoder, and a multi-task branch network. The graph neural network module is used to aggregate spatial dimension features of the multidimensional input tensor to extract inter-channel dependencies and obtain spatially enhanced features; the input projection layer is used to group and temporally segment the spatially enhanced features along the spatial dimension to obtain a segment embedding sequence; the Transformer encoder is used to extract temporal features from the segment embedding sequence to obtain an intermediate layer feature representation.

[0203] In some implementations, the Transformer encoder includes a self-attention module, followed by an adapter module for fine-tuning. The adapter module is used to efficiently fine-tune the parameters of the anomaly detection model with a limited number of labeled samples, adapting the model to engineering structure anomaly detection scenarios.

[0204] The multi-task branch network can include an anomaly classification branch and an anomaly localization branch. The anomaly classification branch outputs the probability distribution of each spatial measurement point within the target engineering area belonging to each preset anomaly category; the anomaly localization branch outputs the anomaly activation map of each spatial measurement point within the target engineering area on the time axis. The model inference module can calculate the anomaly score matrix based on the probability distribution and the anomaly activation map. Q r,c ,in, Q r,c Representing spatial measurement points r Corresponding structural anomaly type c The overall abnormal score.

[0205] The rule interpretation module is used to make a comprehensive judgment based on the anomaly score matrix and multiphysics parameter data volume, and on a preset rule base, to obtain the structural anomaly type of the target engineering area.

[0206] The rule base describes the correspondence between different structural anomaly types and multiphysics parameter anomaly patterns. The rule interpretation module can, for each candidate structural anomaly type, determine whether the corresponding comprehensive anomaly score in the anomaly score matrix meets a preset score threshold condition; determine whether the number of physical parameters in the multiphysics parameter data volume that meet the corresponding physical property parameter anomaly condition meets a quantity threshold condition; and determine whether the combination relationship of the met physical property parameter anomaly conditions conforms to the physical causal mechanism of the candidate structural anomaly type.

[0207] In one implementation, the structural anomaly type may include one or more of the following: roadbed voids, loose zones, settlement zones, cracks, and roadbed slippage. For example, for roadbed voids, the rule interpretation module can make a comprehensive judgment by combining the integrated anomaly score, velocity value, wave impedance value, density value, temperature gradient anomaly index, and strain curvature; for cracks, it can make a judgment by combining the instantaneous characteristic tensor, temperature gradient anomaly index, and the spatial morphology of strain curvature; for roadbed slippage, it can make a judgment by combining continuous changes in strain curvature, velocity reduction, density anomalies, and external environmental factors.

[0208] In some implementations, the rule interpretation module can further calculate a comprehensive disaster risk index based on the anomaly score matrix and the multiphysics parameter data volume, and output the disaster risk level according to the preset numerical range to which the comprehensive disaster risk index belongs. The disaster risk level can include Level 1 High Risk, Level 2 Medium Risk, Level 3 Low Risk, and Normal, or it can correspond to Red Alert, Orange Alert, Yellow Alert, and No Alert.

[0209] In actual operation, the data acquisition module first acquires the three-dimensional data volume of the target engineering area and sends it to the feature extraction module. The feature extraction module generates a multiphysics parameter data volume based on the three-dimensional data volume and sends it to the tensor construction module. The tensor construction module performs spatiotemporal alignment, standardization, and channel splicing on the multiphysics parameter data volume to obtain a multidimensional input tensor. The model inference module inputs the multidimensional input tensor into a pre-trained anomaly recognition model and outputs an anomaly score matrix. The rule interpretation module performs rule base comprehensive discrimination based on the anomaly score matrix and the multiphysics parameter data volume to obtain the structural anomaly type of the target engineering area.

[0210] Therefore, the engineering anomaly identification system described in this example can realize a complete processing flow from distributed fiber optic data acquisition, multi-physics parameter extraction, multi-dimensional input tensor construction, anomaly identification model reasoning to rule base comprehensive interpretation. This system can improve the continuity, accuracy, and interpretability of structural anomaly identification in target engineering areas, and is suitable for long-term online monitoring of projects such as roads, railways, tunnels, underground passages, and integrated utility tunnels.

[0211] Thirdly, embodiments of this application provide a computer-readable storage medium. The computer-readable storage medium stores computer-executable instructions that are executed by a processor or controller. When executed by the processor, the computer program implements the steps of the engineering anomaly identification method of any embodiment of the first aspect.

[0212] It will be understood by those skilled in the art that all or some of the steps and systems in the methods disclosed above can be implemented as software, firmware, hardware, and suitable combinations thereof. Some or all of the physical components can be implemented as software executed by a processor, such as a central processing unit, digital signal processor, or microprocessor, or as hardware, or as an integrated circuit, such as an application-specific integrated circuit. Such software can be distributed on a computer-readable medium, which can include computer storage media (or non-transitory media) and communication media (or transient media). As is known to those skilled in the art, the term computer storage media includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information (such as computer-readable instructions, data structures, program modules, or other data). Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technologies, CD-ROM, digital versatile disc (DVD) or other optical disc storage, magnetic cartridges, magnetic tape, disk storage or other magnetic storage devices, or any other medium that can be used to store desired information and is accessible to a computer. Furthermore, as is known to those skilled in the art, communication media typically contain computer-readable instructions, data structures, program modules, or other data in modulated data signals such as carrier waves or other transmission mechanisms, and may include any information delivery medium.

[0213] The embodiments of this application have been described in detail above with reference to the accompanying drawings. However, this application is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of this application. Furthermore, unless otherwise specified, the embodiments and features described in the embodiments of this application can be combined with each other.

Claims

1. A method for identifying engineering anomalies, characterized in that, include: Acquire a three-dimensional data volume of the target engineering area collected by a distributed optical fiber sensor, the three-dimensional data volume including acoustic wave data, temperature field data and strain field data; Feature extraction is performed on the three-dimensional data volume to obtain a multi-physics parameter data volume; the multi-physics parameter data volume includes at least one of the following: three-dimensional velocity model data volume, three-dimensional density data volume, wave impedance data volume, instantaneous feature tensor, temperature gradient anomaly index data volume, and strain curvature data volume. The multiphysics parameter data volume is spatiotemporally aligned and spliced ​​to obtain a multidimensional input tensor. The multidimensional input tensor is input into a pre-trained anomaly recognition model to obtain anomaly score matrices corresponding to different structural anomaly types for each spatial measurement point within the target engineering area. Based on the anomaly score matrix and the multiphysics parameter data, a comprehensive judgment is made according to a preset rule base to obtain the structural anomaly type of the target engineering area.

2. The engineering anomaly identification method according to claim 1, characterized in that, The step of extracting features from the three-dimensional data volume to obtain a multiphysics parameter data volume includes: The acoustic wave data is inverted to obtain the three-dimensional velocity model data volume and the three-dimensional density data volume; The wave impedance data volume is calculated based on the three-dimensional velocity model data volume and the three-dimensional density data volume; Based on the acoustic wave data, instantaneous amplitude, instantaneous frequency, and instantaneous phase are extracted using Hilbert transform, and the instantaneous amplitude, instantaneous frequency, and instantaneous phase are combined to obtain the instantaneous feature tensor; Based on the temperature field data, the temperature gradient anomaly index data volume is calculated using a preset temperature gradient anomaly index formula. Based on the strain field data, the strain curvature data volume is obtained by calculating the local strain curvature.

3. The engineering anomaly identification method according to claim 1, characterized in that, The process of performing spatiotemporal alignment and channel stitching on the multiphysics parameter data volume to obtain a multidimensional input tensor includes: Based on a preset spatial resolution, spatial interpolation processing is performed on the multiphysics parameter data volume to complete spatial alignment; Based on a preset time step, time statistics are calculated on the multiphysics parameter data volume to complete time alignment; The spatiotemporally aligned multiphysics parameter data volume is then standardized. The standardized multiphysics parameter data volume is spliced ​​along the channel dimension to obtain the multidimensional input tensor.

4. The engineering anomaly identification method according to claim 1, characterized in that, The anomaly detection model includes a graph neural network module, an input projection layer, a Transformer encoder, and a multi-task branch network; The multidimensional input tensor is input into a pre-trained anomaly recognition model to obtain anomaly score matrices for different structural anomaly types corresponding to each spatial measurement point within the target engineering area, including: The graph neural network module performs spatial dimension feature aggregation on the multidimensional input tensor to extract inter-channel dependencies and obtain spatially enhanced features. The spatial enhancement features are grouped and temporally segmented along the spatial dimension by the input projection layer to obtain a fragment embedding sequence. The embedded sequence of the fragment is input into the Transformer encoder for temporal feature extraction to obtain the intermediate layer feature representation; wherein, the Transformer encoder includes a self-attention module, and an Adapter module for fine-tuning is provided after the self-attention module; The intermediate layer feature representation is input into the multi-task branch network, and the anomaly score matrix is ​​output.

5. The engineering anomaly identification method according to claim 4, characterized in that, The multi-task branch network includes an anomaly classification branch and an anomaly localization branch; The step of inputting the intermediate layer feature representation into the multi-task branch network and outputting the anomaly score matrix includes: The intermediate layer feature representation is processed by the anomaly classification branch to output the probability distribution of each spatial measurement point in the target engineering area belonging to each preset anomaly category; The intermediate layer feature representation is decoded and restored by the anomaly localization branch, and the anomaly activation map of each spatial measurement point in the target engineering area on the time axis is output. Based on the probability distribution and the anomaly activation map, the anomaly score of each spatial measuring point within the target engineering area is calculated to obtain the anomaly score matrix.

6. The engineering anomaly identification method according to claim 1, characterized in that, The structural anomaly type of the target engineering area is obtained by comprehensively judging based on the anomaly score matrix and the multiphysics parameter data volume according to a preset rule base, including: For each preset candidate structural anomaly type, determine whether the corresponding comprehensive anomaly score in the anomaly score matrix meets the preset score threshold condition; Determine whether the number of physical field parameters in the multiphysics parameter data volume that satisfy the physical property parameter anomaly conditions corresponding to the candidate structural anomaly type meets a preset number threshold condition. If the comprehensive anomaly score meets the preset score threshold condition, the number of physical field parameters that meet the physical property parameter anomaly condition meets the quantity threshold condition, and the combination relationship of the met physical property parameter anomaly conditions conforms to the physical causal mechanism of the candidate structural anomaly type, then it is determined that the candidate structural anomaly type exists in the structural anomaly type of the target engineering area.

7. The engineering anomaly identification method according to claim 6, characterized in that, The preset candidate anomaly types include roadbed cavities; The conditions for determining the presence of roadbed cavities in the structural anomaly types of the target engineering area include: The overall abnormal score is greater than or equal to the corresponding score threshold; The velocity value in the three-dimensional velocity model data volume is less than the corresponding velocity reference value; The wave impedance value in the wave impedance data body is less than the corresponding wave impedance reference value; The density values ​​in the three-dimensional density data volume are less than the corresponding density reference values; The temperature gradient anomaly index in the temperature gradient anomaly index data body is greater than the corresponding temperature gradient benchmark value. The strain curvature in the strain curvature data volume is greater than the corresponding strain curvature reference value.

8. The engineering anomaly identification method according to claim 1, characterized in that, After obtaining the structural anomaly type of the target engineering area, the method further includes: Extract the deviation of each physical field parameter from the normal reference value in the multiphysics parameter data volume; Obtain the preset anomaly score weights and the deviation weights corresponding to each of the physical field parameters; The comprehensive disaster risk index is obtained by performing a weighted summation based on the comprehensive anomaly score in the anomaly score matrix, the anomaly score weight, the deviation of each parameter, and the corresponding deviation weight. The disaster risk level of the target project area is obtained based on the preset numerical range to which the comprehensive disaster risk index belongs.

9. An engineering anomaly identification system, characterized in that, include: The data acquisition module is used to acquire a three-dimensional data volume of the target engineering area collected by a distributed optical fiber sensor. The three-dimensional data volume includes acoustic data, temperature field data and strain field data. The feature extraction module is used to extract features based on the three-dimensional data volume to obtain a multi-physics parameter data volume, wherein the multi-physics parameter data volume includes at least one of the following: three-dimensional velocity model data volume, three-dimensional density data volume, wave impedance data volume, instantaneous feature tensor, temperature gradient anomaly index data volume, and strain curvature data volume. The tensor construction module is used to perform spatiotemporal alignment and channel splicing on the multiphysics parameter data volume to obtain a multidimensional input tensor. The model inference module is used to input the multidimensional input tensor into a pre-trained anomaly recognition model to obtain anomaly score matrices corresponding to different structural anomaly types for each spatial measurement point within the target engineering area. The rule interpretation module is used to perform comprehensive discrimination based on the anomaly score matrix and the multiphysics parameter data volume, and on a preset rule base, to obtain the structural anomaly type of the target engineering area.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the engineering anomaly identification method according to any one of claims 1 to 8.

Images