Urban physical examination accurate analysis method and system based on AI multi-modal data collaboration

Abstract

The invention relates to the technical field of data analysis, in particular to a city physical examination accurate analysis method and system based on AI multi-modal data collaboration, and the method comprises the following steps: constructing a self-adaptive fractal space-time grid, and mapping a city multi-modal data source into a space-time coding vector with a fractal dimension; establishing a physical field driven cooperative resonance network, and extracting an abnormal cooperative mode exceeding a normal resonance threshold; calculating a modal entropy chain value according to the real-time metabolic rate of the urban system, and generating an optimal weight matrix through a non-equilibrium thermodynamic model; and fusing the abnormal cooperation mode and the dynamic weight matrix, and generating a space-time causal map for displaying a fault source and a propagation path by using a causal discovery algorithm. According to the method, the multi-modal data collaboration and causal analysis technology is utilized, the emergency response capability of an urban system in the case of sudden failures is improved, the stability of urban operation is guaranteed, and the space-time causal atlas provides comprehensive visual information for the failure propagation process.

Description

Technical Field

[0001] The present invention relates to the field of data analysis technology, and in particular to a method and system for accurate urban physical examination analysis based on AI multimodal data collaboration. Background Art

[0002] With the acceleration of urbanization, the complexity of urban infrastructure and the dynamic nature of its operating environment are gradually increasing, which brings challenges to urban management and infrastructure maintenance. Traditional urban system health monitoring and problem point diagnosis are limited to the status assessment of a specific area or equipment, lacking a comprehensive and cross-domain analysis of the entire urban system. This approach often has the following shortcomings: During operation, the health status of various infrastructures in urban systems is affected by multiple factors, including physical characteristics such as vibration spectrum, pressure waves, and thermal radiation. These characteristics are often manifested in different forms, and data from a single mode cannot fully reflect the operating status of the system.

[0003] Existing fault diagnosis systems are mainly based on static data analysis models, ignoring the complex coupling relationship between infrastructure and the fault propagation path. Even when a fault occurs, the diagnosis system often finds it difficult to quickly identify the fault source and propagation path, resulting in slow response during the processing process, which in turn increases repair costs and risks.

[0004] Traditional urban operation and maintenance management systems usually adopt a static operation and maintenance model, which lacks real-time performance and flexibility. They are unable to dynamically adjust operation and maintenance strategies based on the operating status of urban infrastructure and environmental changes. Urban systems may face different risks and loads in different time periods, which requires operation and maintenance management to have adaptive capabilities to monitor and predict potential failure risks in real time. Summary of the Invention

[0005] The present invention provides a method and system for accurate analysis of urban physical examinations based on AI multimodal data collaboration.

[0006] The urban health examination precision analysis method based on AI multimodal data collaboration includes the following steps: S1, Dynamic Fractal Space-Time Coding: Constructing an adaptive fractal space-time grid, mapping urban multimodal data sources into space-time coding vectors with fractal dimensions, generating a unified cross-modal representation. Urban multimodal data sources include urban infrastructure vibration spectra, underground pipe network pressure waves, and surface thermal radiation data. S2, multimodal cooperative resonance detection: establish a physical field-driven cooperative resonance network, calculate the data resonance intensity through the inter-modal energy transfer equation, and extract abnormal cooperative modes that exceed the normal resonance threshold; S3, dynamic weighting of metabolic entropy chain: Calculate the modal entropy chain value according to the real-time metabolic rate of the urban system, and generate the optimal weight matrix through the non-equilibrium thermodynamic model; S4. Causal Heterogeneous Graph Construction: By integrating the abnormal coordination model with the dynamic weight matrix, a causal discovery algorithm is used to generate a spatiotemporal causal graph that shows the origin and propagation path of the fault.

[0007] Optionally, the S1 specifically includes: S11, Dynamic Calculation of Fractal Dimension: Based on the spatiotemporal density distribution of multimodal data sources, the fractal dimension is calculated in real time to describe the self-similarity of data in space and time. The fractal dimension reflects the complexity and structure of the data point distribution. The grid distribution is dynamically adjusted by analyzing the density and spatial resolution of the data. The density of the data points determines the calculation of the fractal dimension, and the minimum spatial resolution is used to determine the accuracy and scale of the grid. This allows the model to adapt to different data sources and handle data point densities of different scales. S12, fractal grid generation: When generating a fractal grid, the city geographic coordinate system is used as the basis to construct a fractal grid structure with self-similar characteristics. The size of the grid unit is adjusted according to the fractal dimension, the grid side length is scaled according to the self-similarity characteristics of different data sources, and the network structure is divided into grids according to the characteristics of different regions; S13, Multimodal Data Mapping: During the data mapping phase, different types of urban data sources are processed differently to accommodate their unique physical characteristics. Specifically, the following are performed: For infrastructure vibration spectrum data, the spatiotemporal distribution characteristics of vibration are characterized by extracting the energy value of its main frequency band. A time-attenuation weighting function is applied to enhance the timeliness of recent data. The data weighted by the time-attenuation weighting function is mapped to the tangential component of the grid to reflect the spatial diffusion characteristics of vibration. For the pressure wave data of the underground pipe network, the spatiotemporal convolution value of the pressure gradient is calculated to obtain the temporal and spatial variation pattern of the pressure wave. The Gaussian smoothing method is used to remove noise in the data to ensure the smoothness of the pressure data. The processed pressure data is mapped to the radial component of the grid to reflect the conduction characteristics of the underground pressure wave. For surface thermal radiation data, wavelet packet decomposition is used to extract high-frequency abnormal temperature variations. These abnormal temperature variations are supplemented through spatial interpolation to obtain abnormal temperature components, providing a finer spatial resolution for the data. These abnormal temperature components are mapped to the axial components of the grid to capture the transfer characteristics of thermal radiation. S14, cross-modal fusion encoding: The tangential component of vibration data, the radial component of pressure wave data, and the axial component of thermal radiation data are fused according to rules, including nonlinear fusion methods, to combine the components into a unified spatiotemporal encoding vector. During fusion, the relative importance of different modal data sources is considered, and the contribution of each modality is dynamically adjusted through a trainable weight matrix. This fusion method not only ensures that the multidimensional information of the data is preserved, but also constrains the range of the final output vector through nonlinear functions, avoiding excessive growth or explosion of values.

[0008] Optionally, the S2 specifically includes: S21, physical field coupling modeling: Utilize the spatiotemporal encoding vectors generated in S1 and convert them into a multimodal physical field coupling tensor. The multimodal physical field coupling tensor is constructed based on a three-dimensional structure, where each dimension represents a different fractal grid cell, modal type, and spatiotemporal feature. The information from different modal data sources is integrated into a unified physical field representation, including mapping the spatiotemporal features extracted from multiple data sources into a high-dimensional tensor structure. S22, Energy Transfer Equation Construction: Construct an equation to describe inter-modal resonant energy transfer. This equation includes the coupling coefficient between modes, spatiotemporal convolution operations, and energy transfer characteristics. Energy transfer between multimodal data sources depends on spatiotemporal characteristics and spatial coupling relationships, energy propagation speed, and physical properties. It also introduces spatial smoothing under the action of the Laplace operator and the nonlinear characteristics of the Hadamard product processing to describe the energy propagation and interaction relationships between multimodal data sources in space and time. S23, dynamic resonance threshold calculation: Setting a dynamic threshold to determine which resonance events belong to abnormal multimodal data sources, including training based on historical normal data, calculating the energy mean and standard deviation of each pair of modal data sources under the baseline state, and marking energy events exceeding the standard deviation range as abnormal using the set dynamic threshold; S24. Abnormal collaborative mode extraction: When the resonance energy of a certain mode pair exceeds the set dynamic threshold, the abnormal collaborative mode extraction process will be triggered.

[0009] Optionally, the abnormal collaboration pattern extraction process includes: Backtracking the resonance source: Backtracking the resonance source along the space-time convolution kernel path, tracking the origin and propagation trajectory of energy, and drawing an energy propagation path map by tracing the backtracking path; Calculate the energy gradient and mark the high-risk impact area: During the backtracking process, calculate the resonance energy gradient, which represents the rate and direction of energy change. For areas where the gradient suddenly changes, mark them as "high-risk impact areas"; Fusion of multimodal resonance events: Fusion of resonance events from different modal data sources to generate a three-dimensional anomaly collaboration map, which includes the spatiotemporal coordinates, energy intensity, and energy conduction direction of each resonance event.

[0010] Optionally, the S3 specifically includes: S31, Multimodal Metabolic Flow Modeling: Based on the extracted abnormal synergy patterns, a metabolic flow tensor for an urban system is constructed. The dimensions of the metabolic flow tensor include multiple fractal grid units, time slices, and different types of metabolic channels. Metabolic channels include energy, matter, and information. Different types of metabolic flows are organized and quantified in a spatiotemporal grid. S32, dynamic calculation of entropy production rate: Based on the theory of non-equilibrium thermodynamics, entropy chain value calculation is performed for each metabolic mode. The entropy chain value calculation includes metabolic power and metabolic flux. By integrating the dynamic process of the metabolic system, the entropy production rate in the future time is evaluated. The entropy chain value reflects the non-equilibrium state of metabolism. S33, weight gradient field generation: Based on the calculation results of the entropy chain value, a weight gradient field is generated by constructing a weight distribution differential equation. The weight distribution is adjusted based on the entropy diffusion coefficient and the baseline weight regression factor. The cross-modal competition relationship is modeled through coupling parameters to adjust the weight distribution of each metabolic pathway. S34, dynamic matrix optimization: The weight matrix is ​​optimized and solved by entropy flow constraints, considering the diagnostic sensitivity and regularization terms of each modality to obtain the optimal weight matrix of metabolic flow.

[0011] Optionally, the metabolic flux tensor is expressed as: ,in, is the number of fractal grid cells, is the number of time slices, It is the metabolic channel type, including three types of metabolic flows: energy, matter, and information. The tangential / radial / axial components of the metabolic channel type correspond to the following: energy → vibration spectrum, matter → pressure wave, and information → thermal radiation.

[0012] Optionally, the dynamic calculation of the entropy production rate includes defining a modal entropy chain value according to non-equilibrium thermodynamics: ,in, Indicates the The amount of entropy generated by the change of modes (energy, matter, information) over time in a dynamic process is calculated by and flux Integrate and calculate the entropy chain value, represents the gradient, For the The metabolic power of each modality, is the characteristic relaxation time is the actual metabolic flux, is the equilibrium reference flux, Represents small increments of time, By resonance energy Obtained by Fourier transform.

[0013] Optionally, the S4 specifically includes: S41, heterogeneous data fusion: Combine the 3D anomaly collaborative map with the generated optimal weight matrix. The 3D anomaly collaborative map includes information about space-time coordinates, energy intensity, and conduction direction. The dynamic weight matrix assigns a weight value to each element. These are fused into an enhanced causal feature tensor through tensor operations. Specifically, the space-time coordinates and weight values ​​are combined through a tensor product operation to form a space-time causal map. S42, Spatiotemporal Causal Discovery: Constructing a dual-constrained causal structure learning model to constrain the inference of causal relationships through both temporal and spatial constraints; S43, spatiotemporal causal graph optimization: The optimization phase uses Bayesian structure learning and an enhanced pruning strategy. The Bayesian structure learning uses the Markov Chain Monte Carlo method to sample the posterior probabilities of causal edges and adaptively adjusts the sampling step size. The sampling results are used to screen causal edges with a posterior probability greater than 0.89. The enhanced pruning strategy includes: after causal edge screening, subgraph isomorphism detection is performed on redundant causal chains. By comparing the causal chain structures, spurious causal relationships that do not match the abnormal propagation path graph are removed. S44, Dynamic Graph Generation: The optimized spatiotemporal causal graph includes multiple attributes, including: Node attributes: the spatiotemporal encoding corresponding to each grid cell and the modal weight entropy of the node. It also includes the metabolic power of the node, which represents the energy consumed by the node during causal propagation. Edge attributes: Edge attributes include causal strength (reflecting the strength of the causal relationship), delay (describing the time delay in the occurrence of the causal relationship), and energy conduction efficiency.

[0014] Optionally, the delay constraint is used to capture the causal relationship between time series, using a cross-modal delay time window to determine whether the change of a node is affected by other nodes by calculating the variance change of the causal graph in different time windows; The spatial constraint is introduced in the causal reasoning process, and the adjacency matrix is ​​defined using the network topology of the fractal grid to ensure that causal edges are generated only between adjacent grid cells, thereby limiting the propagation range of the causal relationship.

[0015] The AI-based multimodal data collaboration-based urban health examination precision analysis system is used to implement the above-mentioned AI-based multimodal data collaboration-based urban health examination precision analysis method, including the following modules: Dynamic fractal space-time coding module: Based on the city's multimodal data sources, it constructs an adaptive fractal space-time grid and maps the multimodal data sources into space-time coding vectors with fractal dimensions; Multimodal collaborative resonance detection module: By establishing a physical field-driven collaborative resonance network, the energy transfer equation between multimodal data sources is used to calculate the resonance strength between different modal data, and abnormal collaborative modes that exceed the normal resonance threshold are extracted; Metabolic entropy chain dynamic weighting module: By monitoring and analyzing the real-time metabolic rate of the urban system, the corresponding modal entropy chain value is calculated, and the non-equilibrium thermodynamic model is used to generate the optimal dynamic weight matrix; Causal heterogeneous graph construction module: By integrating the abnormal collaborative patterns and dynamic weight matrices in multimodal data, a causal discovery algorithm is used to establish a spatiotemporal causal graph to show the fault origins and fault propagation paths of various facilities in the urban system.

[0016] Beneficial effects of the present invention: This invention maps multimodal data sources such as urban infrastructure, underground pipelines, and surface thermal radiation onto a fractal space-time grid to form a unified cross-modal representation, thereby enhancing the comprehensive perception of all levels of the city. It uses the fractal space-time grid to process urban data, converting various types of data into coding vectors with spatiotemporal characteristics, capturing the characteristics of different time and space scales in the urban system. The fractal structure can handle complex and irregular urban data, provides an efficient data representation method, and embodies the accuracy of urban physical examinations.

[0017] The present invention can identify the source and propagation path of faults in urban systems by generating a spatiotemporal causal graph of fault origins and propagation paths. By utilizing multimodal data collaboration and causal analysis technology, the emergency response capability of urban systems in the face of sudden faults is enhanced, thereby ensuring the stability of urban operations. The spatiotemporal causal graph provides comprehensive visual information on the fault propagation process, supporting decision makers to conduct precise resource scheduling based on the fault propagation paths, risk points and weak links in the graph, thus avoiding the waste of resources caused by the lack of data support in traditional operation and maintenance, improving the operation and maintenance efficiency and maintenance effect of urban infrastructure, ensuring the timely repair of key areas, and reducing long-term maintenance costs.

[0018] This invention introduces a calculation method for the dynamic weight coefficient matrix and metabolic entropy chain, which not only enhances the responsiveness of the urban system to dynamic changes, but also can identify key links and potential risk points in the urban system. This breakthrough enables urban management to make real-time adjustments in complex dynamic environments, improving the accuracy of prediction and emergency dispatch. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only for the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0020] Figure 1 Schematic diagram of a method flow in an embodiment of the present invention; Figure 2 Schematic diagram of system function modules according to an embodiment of the present invention. DETAILED DESCRIPTION

[0021] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. It is also noted that, to provide a more detailed description, the following embodiments are best and preferred embodiments, and those skilled in the art may employ alternative methods for implementing certain known technologies. Furthermore, the accompanying drawings are intended only to provide a more detailed description of the embodiments and are not intended to limit the present invention.

[0022] It should be noted that references in the specification to "one embodiment," "an embodiment," "exemplary embodiments," "some embodiments," etc. indicate that the described embodiments may include specific features, structures, or characteristics, but not necessarily every embodiment will include such specific features, structures, or characteristics. Furthermore, when specific features, structures, or characteristics are described in conjunction with an embodiment, it is within the knowledge of persons skilled in the relevant art to implement such features, structures, or characteristics in conjunction with other embodiments (whether or not explicitly described).

[0023] In general, terms can be understood, at least in part, from their use in context. For example, depending at least in part on the context, the term "one or more" as used herein can be used to describe any feature, structure, or characteristic in the singular sense, or can be used to describe a combination of features, structures, or characteristics in the plural sense. Additionally, the term "based on" can be understood as not necessarily intended to convey an exclusive set of factors, but can instead, depending at least in part on the context, allow for the presence of other factors that are not necessarily explicitly described.

[0024] like Figure 1 As shown in the figure, the urban physical examination precision analysis method based on AI multimodal data collaboration includes the following steps: S1, Dynamic Fractal Space-Time Coding: Constructing an adaptive fractal space-time grid, mapping urban multimodal data sources into space-time coding vectors with fractal dimensions, generating a unified cross-modal representation. Urban multimodal data sources include urban infrastructure vibration spectra, underground pipe network pressure waves, and surface thermal radiation data. S2, multimodal cooperative resonance detection: establish a physical field-driven cooperative resonance network, calculate the data resonance intensity through the inter-modal energy transfer equation, and extract abnormal cooperative modes that exceed the normal resonance threshold; S3, dynamic weighting of metabolic entropy chain: Calculate the modal entropy chain value according to the real-time metabolic rate of the urban system, and generate the optimal weight matrix through the non-equilibrium thermodynamic model; S4. Causal Heterogeneous Graph Construction: By integrating the abnormal coordination model with the dynamic weight matrix, a causal discovery algorithm is used to generate a spatiotemporal causal graph that shows the origin and propagation path of the fault.

[0025] By mapping multimodal data sources such as urban infrastructure vibration spectra, underground pipeline pressure waves, and surface thermal radiation data onto a fractal space-time grid, a unified cross-modal representation is formed. This fusion approach enhances comprehensive perception of all aspects of the city (infrastructure, environment, etc.) and is the foundation for precise analysis. Using a fractal space-time grid to process urban data, various data types are converted into encoding vectors with spatiotemporal characteristics, effectively capturing the characteristics of urban systems at different temporal and spatial scales. Fractal structures can handle complex and irregular urban data, providing an efficient data representation method and demonstrating the accuracy of urban health checks. A physical field-driven collaborative resonance network is established to analyze energy transfer between modes and detect anomalous collaborative patterns. The key to this process is to leverage the energy interaction between modals to reveal potential faults or anomalies in the system, enabling accurate identification of hidden dangers in urban systems and achieving early warning. Entropy chain values ​​are calculated using metabolic rates and a dynamic weight matrix is ​​generated. This approach considers the dynamic characteristics of urban systems and adjusts weights based on a non-equilibrium thermodynamic model. It can adapt to changing urban environments and ensure the real-time and accuracy of analytical results. Combining anomalous patterns with the dynamic weight matrix, a causal discovery algorithm is used to generate a spatiotemporal causal graph. Displaying the origin and propagation path of the fault helps to accurately locate the source of the problem and provide a scientific basis for fault repair.

[0026] S1 specifically includes: S11, Dynamic Calculation of Fractal Dimension: Based on the spatiotemporal density distribution of multimodal data sources, the fractal dimension is calculated in real time to describe the self-similarity of data in space and time. The fractal dimension reflects the complexity and structure of the data point distribution. The grid distribution is dynamically adjusted by analyzing the density and spatial resolution of the data. The density of the data points determines the calculation of the fractal dimension, and the minimum spatial resolution is used to determine the accuracy and scale of the grid. This allows the model to adapt to different data sources and handle data point densities of different scales. S12, Fractal Grid Generation: When generating a fractal grid, we use the city's geographic coordinate system as the basis to construct a fractal grid structure with self-similar properties. The size of the grid cells is adjusted according to the fractal dimension, and the grid side length is scaled according to the self-similarity characteristics of different data sources. An adaptive method is used to ensure that each data source is optimally matched in space. In this way, the network structure is divided into grids according to the characteristics of different regions, allowing complex data distribution to be more accurately represented. S13, Multimodal Data Mapping: During the data mapping phase, different types of urban data sources are processed differently to accommodate their unique physical characteristics. Specifically, the following are performed: For infrastructure vibration spectrum data, the spatiotemporal distribution characteristics of vibration are characterized by extracting the energy value of its main frequency band. A time-attenuation weighting function is applied to enhance the timeliness of recent data. The data weighted by the time-attenuation weighting function is mapped to the tangential component of the grid to reflect the spatial diffusion characteristics of vibration. For the pressure wave data of the underground pipe network, the spatiotemporal convolution value of the pressure gradient is calculated to obtain the temporal and spatial variation pattern of the pressure wave. The Gaussian smoothing method is used to remove noise in the data to ensure the smoothness of the pressure data. The processed pressure data is mapped to the radial component of the grid to reflect the conduction characteristics of the underground pressure wave. For surface thermal radiation data, wavelet packet decomposition is used to extract high-frequency abnormal temperature variations. These abnormal temperature variations are supplemented through spatial interpolation to obtain abnormal temperature components, providing a finer spatial resolution for the data. These abnormal temperature components are mapped to the axial components of the grid to capture the transfer characteristics of thermal radiation. Through these different types of data mapping, the model can comprehensively capture the various physical phenomena involved in the urban health check process, providing rich spatiotemporal data support for subsequent analysis; S14, cross-modal fusion encoding: The tangential component of vibration data, the radial component of pressure wave data, and the axial component of thermal radiation data are fused according to rules, including nonlinear fusion methods, to combine the components into a unified spatiotemporal encoding vector. During fusion, the relative importance of different modal data sources is considered, and the contribution of each modality is dynamically adjusted through a trainable weight matrix. This fusion method not only ensures that the multidimensional information of the data is preserved, but also constrains the range of the final output vector through nonlinear functions, avoiding excessive growth or explosion of values.

[0027] Fractal dimension is being dynamically calculated: the fractal dimension is calculated based on the spatiotemporal density distribution of multimodal data sources. , ,in, is the data point density, is the minimum spatial resolution, The recommended value is 1.2-1.8 (adjusted according to the complexity of the data); In the fractal grid generation, the city geographic coordinate system is used as the basis to generate a fractal grid with self-similar characteristics, and the grid unit side length is Perform dynamic scaling, where scaling is expressed as: , is the base side length, The data source type adjustment factor adjusts the change range of the grid unit size. Different data source types have different adjustment factors. Vibration data (high frequency): 0.8; thermal radiation data (low frequency): 1.2; underground pipe network data: 1.0; In multimodal data mapping: Extract the energy integral value of the main frequency band from infrastructure vibration spectrum data ,according to After time decay weighting, it is mapped to the grid tangential component. It's time, is the reference time point, which can be 0. is the time decay constant, Take a 72-hour period; Calculate pressure gradient based on underground pipe network pressure wave data The spatiotemporal convolution value of After Gaussian smoothing, it is mapped to the radial component of the grid. is the Gaussian smoothing kernel; For surface thermal radiation data, wavelet packet decomposition is used to extract high-frequency anomaly components ,through After spatial interpolation, it is mapped to the grid axial component. is the spatial frequency, is the spatial resolution; In cross-modal fusion coding, the tangential, radial, and axial components of the same grid unit are combined into . ) is fused into a spatiotemporal coding vector, where is the modal coupling coefficient. The recommended values ​​are: vibration data: 0.7; thermal radiation data: 0.3; underground pipe network data: 0.5. is the trainable weight matrix, It is a space-time coding vector, and the hyperbolic tangent function tanh is used to constrain the vector value range to (-1, 1) to avoid numerical explosion caused by dimensional differences between modes.

[0028] S2 specifically includes: S21, physical field coupling modeling: Utilize the spatiotemporal encoding vectors generated in S1 and convert them into a multimodal physical field coupling tensor. The multimodal physical field coupling tensor is constructed based on a three-dimensional structure, where each dimension represents a different fractal grid unit, modal type, and spatiotemporal feature. The information from different modal data sources is integrated into a unified physical field representation, including mapping the spatiotemporal features extracted from multiple data sources into a high-dimensional tensor structure. This allows for comprehensive modeling of the interactions and resonance phenomena between the modes, laying the foundation for energy transfer analysis. S22, Energy Transfer Equation Construction: Construct an equation to describe inter-modal resonant energy transfer. This equation includes the coupling coefficient between modes, spatiotemporal convolution operations, and energy transfer characteristics. Energy transfer between multimodal data sources depends on spatiotemporal characteristics and spatial coupling relationships, energy propagation speed, and physical properties. It also introduces spatial smoothing under the action of the Laplace operator and the nonlinear characteristics of the Hadamard product processing to describe the energy propagation and interaction relationships between multimodal data sources in space and time. S23, dynamic resonance threshold calculation: Setting a dynamic threshold to determine which resonance events belong to abnormal multimodal data sources, including training based on historical normal data, calculating the energy mean and standard deviation of each pair of modal data sources under the baseline state, and marking energy events exceeding the standard deviation range as abnormal using the set dynamic threshold; S24. Abnormal collaborative mode extraction: When the resonance energy of a certain mode pair exceeds the set dynamic threshold, the abnormal collaborative mode extraction process will be triggered.

[0029] The abnormal collaborative pattern extraction process includes: Backtracking the resonance source: Backtracking the resonance source along the space-time convolution kernel path, tracking the origin and propagation trajectory of energy, and drawing an energy propagation path map by tracing the backtracking path; Calculate the energy gradient and mark the high-risk impact area: During the backtracking process, calculate the resonance energy gradient, which represents the rate and direction of energy change. For areas where the gradient suddenly changes, mark them as "high-risk impact areas"; Fusion of multimodal resonance events: Fusion of resonance events from different modal data sources to generate a three-dimensional anomaly collaboration map. The anomaly collaboration map includes the spatiotemporal coordinates, energy intensity, and energy conduction direction of each resonance event. Physical field coupling modeling is based on the space-time encoding vector generated by S1 to construct a multi-modal physical field coupling tensor ,in, is the number of fractal grid cells, is the number of modal types, is the space-time characteristic dimension, the multimodal physical field coupling tensor It is directly composed of the stack of spatiotemporal coding vectors of S1; The energy transfer equation is constructed, including the definition of the inter-modal resonance energy equation: ,in, Represents a modal pair At the moment energy, is the modal coupling coefficient matrix, Indicates the coupling submode or subsystem index between modes, Used to represent the coupling relationship between multiple sub-modes or sub-systems. represents the spatiotemporal convolution kernel operation, is the Laplace operator, For Hadamard, is the cross-modal delay time window, spatiotemporal convolution kernel The size of the dynamic scaling side length L (kernel size = ⌈3L⌉), the delay time window ,in, is the effective speed of cross-modal energy transfer (such as the propagation speed of vibration waves in underground media). When the underground pipe network pressure wave is coupled with the surface vibration data, , automatically calculate the optimal delay window, represents the first mode (one of infrastructure vibration, underground pipe network pressure wave or surface thermal radiation), represents the second mode (one of infrastructure vibration, underground pipe network pressure wave or surface thermal radiation), and Pairing, which represents the interaction or coupling between two different modes, is an index pair representing a mode pair, representing the coupling relationship between different modes; The dynamic resonance threshold is calculated based on the historical normal data training benchmark resonance model. Set dynamic threshold, where Mode pairs The mean and standard deviation of energy in the baseline state; Abnormal collaborative pattern extraction: When When executing: a. Backtracking the resonance source along the spatiotemporal convolution kernel path to generate an anomaly propagation path map. The goal is to trace back to the source of the resonance event based on its propagation trajectory and depict the propagation path from the resonance source to the surrounding area. First, the spatiotemporal convolution kernel is used to analyze the energy transfer between different modes in the data. Through the backtracking process, the initial resonance source location is determined to be a relatively concentrated area in the system (such as a pipeline rupture or earthquake source). Then, these energy transfer paths are traced to generate a map that marks the path of energy transfer from the resonance source to other areas. This allows for intuitive identification of the anomaly's propagation method and the potential scope of impact expansion. b. Calculate the resonance energy gradient , marking the gradient mutation area as a high-risk impact area, calculating the gradient of the resonance energy. Specifically, it is to analyze the change rate of energy transfer between different areas. If the energy transfer speed changes very quickly, it means that the energy fluctuation in a certain part has suddenly changed, which is usually a sign of abnormality. Therefore, the resonance energy gradient, that is, the energy change rate, is calculated to detect whether there is an abnormal change area. When areas with sudden energy changes are detected, these areas are marked as "high-risk impact areas", indicating that these areas may be greatly affected and there are faults or risks. c. Fusion of multimodal resonance events to generate a three-dimensional anomaly collaborative map containing space-time coordinates, energy intensity, and transmission direction. Resonance events from different modes (vibration, pressure waves, thermal radiation) are integrated to form a comprehensive three-dimensional map. This map includes not only the space-time coordinates of each event (i.e., the specific location and time of occurrence), but also the energy intensity (i.e., the severity of the event) and the energy transmission direction (i.e., the propagation direction of the anomaly). Using this information, the map clearly demonstrates the overall spatial distribution, intensity, and propagation paths of various anomaly events. The judgment conditions for gradient mutation: ,in, represents the second norm of the gradient, which measures the amplitude of the resonance energy gradient, is the side length of the fractal grid cell, It is a fractal dimension; The specific implementation example is as follows: When the drainage network pressure is abnormal, the encoding vector , and road settlement data When the fractal grid G0512 unit resonates: 1. Calculation (Threshold ); 2. Tracing back along the spatiotemporal convolution kernel, the resonance source was found to be located in an old pipe section 300 m upstream. 3. Gradient calculations show that the maximum mutation zone covers six surrounding grid cells (4.8m in diameter); 4. The 3D anomaly map triggers S4 causal reasoning and locates it as "pipe section corrosion" Soil erosion The transmission chain of "roadbed collapse".

[0030] S3 specifically includes: S31, Multimodal Metabolic Flow Modeling: Based on the extracted abnormal synergy patterns, a metabolic flow tensor for an urban system is constructed. The dimensions of the metabolic flow tensor include multiple fractal grid units, time slices, and different types of metabolic channels. Metabolic channels include energy, matter, and information. Different types of metabolic flows are organized and quantified in a spatiotemporal grid. S32, dynamic calculation of entropy production rate: Based on the theory of non-equilibrium thermodynamics, entropy chain value calculation is performed for each metabolic mode. The entropy chain value calculation includes metabolic power and metabolic flux. By integrating the dynamic process of the metabolic system, the entropy production rate in the future time is evaluated. The entropy chain value reflects the non-equilibrium state of metabolism. S33, Weight Gradient Field Generation: Based on the calculation results of the entropy chain value, a weight gradient field is generated by constructing a weight distribution differential equation. The weight distribution is adjusted based on the entropy diffusion coefficient and the baseline weight regression factor. The cross-modal competition relationship is modeled through coupling parameters, and the weight distribution of each metabolic channel is adjusted to more accurately reflect the complex dynamic behavior of the urban system. S34, dynamic matrix optimization: The weight matrix is ​​optimized and solved by entropy flow constraints, considering the diagnostic sensitivity and regularization terms of each modality to obtain the optimal weight matrix of metabolic flow.

[0031] Multimodal metabolic flow modeling is based on the abnormal collaborative pattern extracted by S2 to construct the metabolic flow tensor of the urban system. The metabolic flow tensor is expressed as: ,in, is the number of fractal grid cells, is the number of time slices, It is the metabolic channel type, including three types of metabolic flows: energy, matter, and information. The tangential / radial / axial components of the metabolic channel type correspond to the following: energy → vibration spectrum, matter → pressure wave, and information → thermal radiation.

[0032] The dynamic calculation of entropy production rate includes the definition of modal entropy chain value according to non-equilibrium thermodynamics: ,in, Indicates the The amount of entropy generated by the change of modes (energy, matter, information) over time in a dynamic process is calculated by and flux Integrate and calculate the entropy chain value, represents the gradient, For the The metabolic power of each modality, is the characteristic relaxation time, the characteristic relaxation time According to the fractal dimension according to Dynamic adjustment, is the actual metabolic flux, is the equilibrium reference flux, Obtained from historical normal data training, Represents small increments of time, By resonance energy Obtained by Fourier transform, the specific derivation is as follows: 1. Modal energy aggregation: for each mode , aggregated with all other modalities The resonance energy of: ,in It is the space-time integral domain, reflecting the space-time range of the abnormal propagation path; 2. Time-frequency domain conversion: Perform short-time Fourier transform to extract the frequency domain characteristics of energy distribution: ,in is the Hanning window function, is the time window sliding step; 3. Metabolic power calculation: identifying dominant frequencies in the frequency domain , calculate its corresponding power spectral density: , is the time slice length of the fractal grid to ensure alignment with the spatiotemporal encoding; Construct the weight distribution differential equation and generate the weight gradient field: ;in, Indicates the The amount of entropy generated by the change of mode over time in the dynamic process, is the entropy diffusion coefficient, is the benchmark weight regression factor, is the cross-modal competition parameter, indicating the modality With modal The competitive relationship between them affects the weight update. By the modal coupling coefficient pass Dynamic association, For the The weight of each mode, For the The reference weight of each mode is represented by the reference weight of each mode. is the Laplace operator, which represents the second-order gradient operator and is used to describe the curvature of the function in space, indicating the spatial diffusion of the entropy chain value and reflecting the diffusion rate of entropy; In dynamic matrix optimization, the optimal weight matrix is ​​solved by the optimal transmission model under entropy flow constraints : ,in, is the diagnostic sensitivity vector of each mode, is the total variation regularization term, is the sparsification coefficient, is the optimal weight matrix.

[0033] Specific implementation examples are as follows: When heavy rain causes the underground drainage system to be overloaded, the metabolic flux During a surge: 1. Calculate the material modal delinking value (base value 1.3); 2. Obtained by weighted differential equation ; 3. The optimal transmission model changes the drainage system weight from Upgrade to , triggering the causal graph of S4 to focus on analyzing the pipeline network nodes.

[0034] S4 specifically includes: S41, heterogeneous data fusion: Combine the 3D anomaly collaborative map with the generated optimal weight matrix. The 3D anomaly collaborative map includes information about space-time coordinates, energy intensity, and conduction direction. The dynamic weight matrix assigns a weight value to each element. These are fused into an enhanced causal feature tensor through tensor operations. Specifically, the space-time coordinates and weight values ​​are combined through a tensor product operation to form a space-time causal map. S42, Spatiotemporal Causal Discovery: Constructing a dual-constrained causal structure learning model to constrain the inference of causal relationships through both temporal and spatial constraints; S43, spatiotemporal causal graph optimization: The optimization phase uses Bayesian structure learning and enhanced pruning strategies. Bayesian structure learning samples the posterior probabilities of causal edges through the Markov Chain Monte Carlo method, adaptively adjusting the sampling step size. The sampling results are used to screen causal edges with posterior probabilities higher than 0.89. The enhanced pruning strategy includes: After causal edge screening, subgraph isomorphism detection is performed on redundant causal chains. By comparing the causal chain structures, pseudo causal relationships that do not match the abnormal propagation path graph are removed, and false or irrelevant causal relationships are removed to ensure that the optimized spatiotemporal causal graph contains only valid and reliable causal reasoning paths. S44, Dynamic Graph Generation: The optimized spatiotemporal causal graph includes multiple attributes, including: Node attributes: the spatiotemporal encoding corresponding to each grid cell and the modal weight entropy of the node. It also includes the metabolic power of the node, which represents the energy consumed by the node during causal propagation. Edge attributes: Edge attributes include causal strength (reflecting the strength of the causal relationship), delay (describing the time delay of the causal relationship), and energy conduction efficiency (used to evaluate the efficiency of information or energy transmission between different nodes).

[0035] Delay constraints are used to capture causal relationships between time series. Cross-modal delay time windows are used to calculate the variance of the causal graph within different time windows to determine whether changes at a certain node are affected by other nodes. Based on the classic Granger causality detection principle and introducing a delay parameter, causal reasoning can adaptively handle causal relationships at different time intervals. Spatial constraints are introduced in the causal reasoning process. The adjacency matrix is ​​defined using the network topology of the fractal grid to ensure that causal edges are only generated between adjacent grid cells, limiting the propagation range of the causal relationship.

[0036] The three-dimensional abnormal collaborative map is expressed as: (space-time coordinates, energy intensity, conduction direction), and compare it with the dynamic weight matrix Perform tensor fusion to construct enhanced causal feature tensor: ,in, is the enhanced causal feature tensor, represents the tensor product, is the diagonalization operation, is the gradient enhancement factor, is a smoothing constant to prevent division by zero errors in logarithmic operations, represents element-wise multiplication, The gradient of the three-dimensional anomaly synergy map represents its rate of change; In spatiotemporal causal discovery, a dual-constrained causal structure learning model is established as follows: a. Delay constraint: based on cross-modal delay time window , construct a time-delayed Granger causal detector: , Represents node-based The variance calculated from the historical data, Represents node-based Historical data and nodes The variance calculated from the lagged data of ; b. Spatial constraints: Using the network topology of the fractal grid to define the spatial adjacency matrix , 0 means no adjacency, 1 means adjacency, limiting the causal edge to be generated only between adjacent grid cells; Using Bayesian structure learning and enhanced pruning strategies to optimize causal graphs: The posterior probability of the causal edge is calculated by NUTS (No-U-TurnSampler) Perform MCMC sampling and retain The NUTS Markov Chain Monte Carlo (MCMC) method adaptively adjusts the step size during the sampling process to avoid the "U-turn" in traditional MCMC methods, that is, the chain goes back and forth in circles. Through NUTS, sampling is performed in the posterior probability space of the causal edge. First, the initial state is set and a candidate causal edge of the causal graph is selected. , that is, from the node To Node The causal relationship is obtained by using the NUTS algorithm from the causal edge The posterior probability distribution of MCMC sampling is performed in NUTS, and NUTS automatically selects the appropriate step size based on the current sample and explores the posterior space. After multiple samplings, the NUTS algorithm will adaptively adjust the sampling process to ensure that it converges to the posterior distribution of the causal edge. Finally, all posterior probabilities are retained. The causal edges of the causal graph are considered to have significant causal relationships, which can effectively screen out those edges with high posterior probabilities (i.e., strong causal relationships) and optimize the causal graph. Subgraph isomorphism detection is performed on redundant causal chains to remove pseudo causal relationships that do not match the abnormal propagation path graph. Subgraph isomorphism detection is a process used to identify identical or similar structures in a graph, especially for identifying redundant causal chains. In a causal graph, redundant causal chains will produce pseudo causal relationships, which cannot find corresponding matches in the abnormal propagation path graph. To this end, it is necessary to extract redundant causal chains. By analyzing the redundant causal chains in the causal graph (that is, those paths that are repeated multiple times or have the same causal structure), potential pseudo causal relationships are identified. Subgraph isomorphism detection is performed on redundant causal chains, that is, checking whether the redundant causal chains match the paths in the abnormal propagation path graph. Subgraph isomorphism detection compares the causal structures of redundant causal chains with those in the abnormal propagation path graph, identifies those causal chains that are inconsistent with the abnormal path graph, removes the redundant causal chains that do not match the abnormal propagation path graph, and retains those paths that can effectively reflect the actual causal relationships. The optimized causal structure is mapped onto a fractal space-time grid to generate a causal graph containing the following elements: Node attributes: grid unit spatiotemporal encoding, modal weight entropy , metabolic power ; Edge attribute: causal strength , delay , energy conduction efficiency , Represents nodes respectively and nodes The weight matrix, Representation node and nodes The efficiency of energy transfer between them; Risk thermodynamic layer: Integrates the gradient mutation zone and the weight gradient field to generate a risk propagation probability cloud map.

[0037] The specific implementation examples are as follows: When the heating pipe leaks When a road collapse occurs: 1. Tensor fusion computing ; 2. Latency Detection (Interval ), triggering the generation of causal edges; 3. NUTS sampling confirmation , retaining the causal relationship; 4. The final graph shows: Leakage source: fractal grid G0709 (coordinate accuracy to 0.6m); Main propagation path: G0709 G0710 G0801 (conduction efficiency ); Risk thermal layer coverage radius: 8.4m (matching the actual sinkhole area).

[0038] The spatiotemporal causal graph can identify the origin of faults in the system and track the path of fault propagation. This allows us to quickly locate the source of the problem and determine its scope of expansion when an abnormality occurs in urban infrastructure or systems, thereby improving the efficiency and accuracy of fault diagnosis.

[0039] By analyzing the propagation paths in the causal graph, we can identify potential high-risk areas and system weaknesses. Based on the structural characteristics of the graph, we can predict possible future failures or abnormal propagation, issue early warnings, and help city managers take preventive measures to reduce the impact of emergencies on city operations.

[0040] like Figure 2 As shown, the urban physical examination precision analysis system based on AI multimodal data collaboration is used to implement the above-mentioned urban physical examination precision analysis method based on AI multimodal data collaboration, including the following modules: Dynamic fractal space-time coding module: Based on the city's multimodal data sources, it constructs an adaptive fractal space-time grid and maps the multimodal data sources into space-time coding vectors with fractal dimensions; Multimodal collaborative resonance detection module: By establishing a physical field-driven collaborative resonance network, the energy transfer equation between multimodal data sources is used to calculate the resonance strength between different modal data, and abnormal collaborative modes that exceed the normal resonance threshold are extracted; Metabolic entropy chain dynamic weighting module: By monitoring and analyzing the real-time metabolic rate of the urban system, the corresponding modal entropy chain value is calculated, and the non-equilibrium thermodynamic model is used to generate the optimal dynamic weight matrix; Causal heterogeneous graph construction module: By integrating the abnormal collaborative patterns and dynamic weight matrices in multimodal data, a causal discovery algorithm is used to establish a spatiotemporal causal graph to show the fault origins and fault propagation paths of various facilities in the urban system.

[0041] The present invention encompasses any alternatives, modifications, equivalents, and solutions that fall within the spirit and scope of the present invention. To provide a thorough understanding of the present invention, specific details are described in detail below in connection with the preferred embodiments of the present invention, but those skilled in the art will be able to fully understand the present invention without these detailed descriptions. Furthermore, to avoid unnecessary confusion regarding the essence of the present invention, well-known methods, processes, procedures, components, and circuits have not been described in detail.

[0042] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A precise urban health examination analysis method based on AI multimodal data collaboration, characterized by: The following steps are involved: S1, Dynamic Fractal Space-Time Coding: Constructing an adaptive fractal space-time grid, mapping urban multimodal data sources into space-time coding vectors with fractal dimensions, generating a unified cross-modal representation. Urban multimodal data sources include urban infrastructure vibration spectra, underground pipe network pressure waves, and surface thermal radiation data. S2, multimodal cooperative resonance detection: establish a physical field-driven cooperative resonance network, calculate the data resonance intensity through the inter-modal energy transfer equation, and extract abnormal cooperative modes that exceed the normal resonance threshold; S3, dynamic weighting of metabolic entropy chain: Calculate the modal entropy chain value according to the real-time metabolic rate of the urban system, and generate the optimal weight matrix through the non-equilibrium thermodynamic model; S4. Causal Heterogeneous Graph Construction: By integrating the abnormal coordination model with the dynamic weight matrix, a causal discovery algorithm is used to generate a spatiotemporal causal graph that shows the origin and propagation path of the fault.

2. The urban physical examination precision analysis method based on AI multimodal data collaboration according to claim 1 is characterized in that: Said S1 specifically includes: S11, Dynamic Calculation of Fractal Dimension: Based on the spatiotemporal density distribution of multimodal data sources, the fractal dimension is calculated in real time to describe the self-similarity of data in space and time. The fractal dimension reflects the complexity and structure of the data point distribution. The grid distribution is dynamically adjusted by analyzing the density and spatial resolution of the data. The density of the data points determines the calculation of the fractal dimension, and the minimum spatial resolution is used to determine the accuracy and scale of the grid. This allows the model to adapt to different data sources and handle data point densities of different scales. S12, fractal grid generation: When generating a fractal grid, the city geographic coordinate system is used as the basis to construct a fractal grid structure with self-similar characteristics. The size of the grid unit is adjusted according to the fractal dimension, the grid side length is scaled according to the self-similarity characteristics of different data sources, and the network structure is divided into grids according to the characteristics of different regions; S13, Multimodal Data Mapping: During the data mapping phase, different types of urban data sources are processed differently to accommodate their unique physical characteristics. Specifically, the following are performed: For infrastructure vibration spectrum data, the spatiotemporal distribution characteristics of vibration are characterized by extracting the energy value of its main frequency band. A time-attenuation weighting function is applied to enhance the timeliness of recent data. The data weighted by the time-attenuation weighting function is mapped to the tangential component of the grid to reflect the spatial diffusion characteristics of vibration. For the pressure wave data of the underground pipe network, the spatiotemporal convolution value of the pressure gradient is calculated to obtain the temporal and spatial variation pattern of the pressure wave. The Gaussian smoothing method is used to remove noise in the data to ensure the smoothness of the pressure data. The processed pressure data is mapped to the radial component of the grid to reflect the conduction characteristics of the underground pressure wave. For surface thermal radiation data, high-frequency abnormal temperature change parts are extracted through wavelet packet decomposition. The abnormal temperature change parts are supplemented by spatial interpolation to obtain abnormal temperature components. The abnormal temperature components will be mapped to the axial components of the grid to capture the transfer characteristics of thermal radiation. S14, cross-modal fusion coding: The tangential component of the vibration data, the radial component of the pressure wave data, and the axial component of the thermal radiation data are fused according to the rules, including combining the components into a unified spatiotemporal coding vector through nonlinear fusion methods. When fusing, the relative importance of different modal data sources is considered, and the contribution of each modality is dynamically adjusted through a trainable weight matrix.

3. The urban physical examination precision analysis method based on AI multimodal data collaboration according to claim 1 is characterized in that: The S2 specifically includes: S21, physical field coupling modeling: Utilize the spatiotemporal encoding vectors generated in S1 and convert them into a multimodal physical field coupling tensor. The multimodal physical field coupling tensor is constructed based on a three-dimensional structure, where each dimension represents a different fractal grid cell, modal type, and spatiotemporal feature. The information from different modal data sources is integrated into a unified physical field representation, including mapping the spatiotemporal features extracted from multiple data sources into a high-dimensional tensor structure. S22, Energy Transfer Equation Construction: Construct an equation to describe inter-modal resonant energy transfer. This equation includes the coupling coefficient between modes, spatiotemporal convolution operations, and energy transfer characteristics. Energy transfer between multimodal data sources depends on spatiotemporal characteristics and spatial coupling relationships, energy propagation speed, and physical properties. It also introduces spatial smoothing under the action of the Laplace operator and the nonlinear characteristics of the Hadamard product processing to describe the energy propagation and interaction relationships between multimodal data sources in space and time. S23, dynamic resonance threshold calculation: Setting a dynamic threshold to determine which resonance events belong to abnormal multimodal data sources, including training based on historical normal data, calculating the energy mean and standard deviation of each pair of modal data sources under the baseline state, and marking energy events exceeding the standard deviation range as abnormal using the set dynamic threshold; S24. Abnormal collaborative mode extraction: When the resonance energy of a certain mode pair exceeds the set dynamic threshold, the abnormal collaborative mode extraction process will be triggered.

4. The urban physical examination precision analysis method based on AI multimodal data collaboration according to claim 3 is characterized in that: The abnormal collaborative pattern extraction process includes: Backtracking the resonance source: Backtracking the resonance source along the space-time convolution kernel path, tracking the origin and propagation trajectory of energy, and drawing an energy propagation path map by tracing the backtracking path; Calculate energy gradients and mark high-risk impact areas: During the backtracking process, calculate the resonance energy gradient, which represents the rate and direction of energy change. Areas where the gradient suddenly changes are marked as "high-risk impact areas." Fusion of multimodal resonance events: Fusion of resonance events from different modal data sources to generate a three-dimensional anomaly collaboration map, which includes the spatiotemporal coordinates, energy intensity, and energy conduction direction of each resonance event.

5. The urban physical examination precision analysis method based on AI multimodal data collaboration according to claim 1 is characterized in that: The S3 specifically includes: S31, Multimodal Metabolic Flow Modeling: Based on the extracted abnormal synergy patterns, a metabolic flow tensor for an urban system is constructed. The dimensions of the metabolic flow tensor include multiple fractal grid units, time slices, and different types of metabolic channels. Metabolic channels include energy, matter, and information. Different types of metabolic flows are organized and quantified in a spatiotemporal grid. S32, dynamic calculation of entropy production rate: Based on the theory of non-equilibrium thermodynamics, entropy chain value calculation is performed for each metabolic mode. The entropy chain value calculation includes metabolic power and metabolic flux. By integrating the dynamic process of the metabolic system, the entropy production rate in the future time is evaluated. The entropy chain value reflects the non-equilibrium state of metabolism. S33, weight gradient field generation: Based on the calculation results of the entropy chain value, a weight gradient field is generated by constructing a weight distribution differential equation. The weight distribution is adjusted based on the entropy diffusion coefficient and the baseline weight regression factor. The cross-modal competition relationship is modeled through coupling parameters to adjust the weight distribution of each metabolic pathway. S34, dynamic matrix optimization: The weight matrix is ​​optimized and solved by entropy flow constraints, considering the diagnostic sensitivity and regularization terms of each modality to obtain the optimal weight matrix of metabolic flow.

6. The urban physical examination precision analysis method based on AI multimodal data collaboration according to claim 1 is characterized in that: The metabolic flux tensor is expressed as: ,in, is the number of fractal grid cells, is the number of time slices, It is the metabolic channel type, including three types of metabolic flows: energy, matter, and information. The tangential / radial / axial components of the metabolic channel type correspond to the following: energy → vibration spectrum, matter → pressure wave, and information → thermal radiation.

7. The urban physical examination precision analysis method based on AI multimodal data collaboration according to claim 1 is characterized in that: The dynamic calculation of entropy production rate includes defining the modal entropy chain value according to non-equilibrium thermodynamics: ,in, Indicates the The entropy generated by the change of mode over time in the dynamic process is calculated by the metabolic power and flux Integrate and calculate the entropy chain value, represents the gradient, For the The metabolic power of each modality, is the characteristic relaxation time, is the actual metabolic flux, is the equilibrium reference flux, Represents small increments of time, By resonance energy Obtained by Fourier transform.

8. The urban physical examination precision analysis method based on AI multimodal data collaboration according to claim 4 is characterized in that: The S4 specifically includes: S41, heterogeneous data fusion: Combine the 3D anomaly collaborative map with the generated optimal weight matrix. The 3D anomaly collaborative map includes information about space-time coordinates, energy intensity, and conduction direction. The dynamic weight matrix assigns a weight value to each element. These are fused into an enhanced causal feature tensor through tensor operations. Specifically, the space-time coordinates and weight values ​​are combined through a tensor product operation to form a space-time causal map. S42, Spatiotemporal Causal Discovery: Constructing a dual-constrained causal structure learning model to constrain the inference of causal relationships through both temporal and spatial constraints; S43, spatiotemporal causal graph optimization: The optimization phase uses Bayesian structure learning and an enhanced pruning strategy. The Bayesian structure learning uses the Markov Chain Monte Carlo method to sample the posterior probabilities of causal edges and adaptively adjusts the sampling step size. The sampling results are used to screen causal edges with a posterior probability greater than 0.

89. The enhanced pruning strategy includes: after causal edge screening, subgraph isomorphism detection is performed on redundant causal chains. By comparing the causal chain structures, spurious causal relationships that do not match the abnormal propagation path graph are removed. S44, Dynamic Graph Generation: The optimized spatiotemporal causal graph includes multiple attributes, including: Node attributes: the spatiotemporal encoding corresponding to each grid cell and the modal weight entropy of the node. It also includes the metabolic power of the node, which represents the energy consumed by the node during causal propagation. Edge attributes: Edge attributes include causal strength, delay, and energy conduction efficiency.

9. The urban physical examination precision analysis method based on AI multimodal data collaboration according to claim 8 is characterized in that: The delay constraint is used to capture the causal relationship between time series. It uses a cross-modal delay time window to determine whether the change of a node is affected by other nodes by calculating the variance change of the causal graph in different time windows. The spatial constraint is introduced in the causal reasoning process, and the adjacency matrix is ​​defined using the network topology of the fractal grid to ensure that causal edges are generated only between adjacent grid cells, thereby limiting the propagation range of the causal relationship.

10. A city physical examination precision analysis system based on AI multimodal data collaboration, used to implement the city physical examination precision analysis method based on AI multimodal data collaboration as described in any one of claims 1-9, characterized in that: Includes the following modules: Dynamic fractal space-time coding module: Based on the city's multimodal data sources, it constructs an adaptive fractal space-time grid and maps the multimodal data sources into space-time coding vectors with fractal dimensions; Multimodal collaborative resonance detection module: By establishing a physical field-driven collaborative resonance network, the energy transfer equation between multimodal data sources is used to calculate the resonance strength between different modal data, and abnormal collaborative modes that exceed the normal resonance threshold are extracted; Metabolic entropy chain dynamic weighting module: By monitoring and analyzing the real-time metabolic rate of the urban system, the corresponding modal entropy chain value is calculated, and the non-equilibrium thermodynamic model is used to generate the optimal dynamic weight matrix; Causal heterogeneous graph construction module: By integrating the abnormal collaborative patterns and dynamic weight matrices in multimodal data, a causal discovery algorithm is used to establish a spatiotemporal causal graph to show the fault origins and fault propagation paths of various facilities in the urban system.

Images