Real-time three-dimensional city modeling and dynamic updating method and system based on deep learning

Abstract

The invention provides a real-time three-dimensional city modeling and dynamic updating method and system based on deep learning. According to the method, building structure static load parameters of a target building are obtained, a micro-seismic sensor array is deployed at a node of a bearing wall and a frame beam column, and stress wave propagation path data are collected at a preset sampling frequency. And synchronously receiving electrical load data uploaded by the intelligent electric meter, and extracting a current abrupt change characteristic section in a time window corresponding to the abnormal vibration event. Building structure static load parameters, stress wave frequency domain energy distribution and the phase angle offset of the current abrupt change characteristic section are input into the deep learning model, the association rule of structural mechanical response and electrical load disturbance is coupled, and a three-dimensional model updating instruction is generated. And dynamically adjusting the grid vertex coordinates of the target building in the initial three-dimensional city model based on the instruction, and overlapping and displaying the space influence range boundary of the current abrupt change feature segment. The technical scheme provided by the invention is based on real-time three-dimensional city modeling and dynamic updating of deep learning.

Description

Technical Field

[0001] The present application relates to the technical field of three-dimensional city modeling, and in particular to a real-time three-dimensional city modeling and dynamic updating method and system based on deep learning. Background Art

[0002] In the fields of smart city construction and building safety monitoring, there is an urgent need to achieve dynamic, high-precision updates of urban three-dimensional models to address complex risk scenarios such as structural deformation, abnormal vibration events, and sudden load changes. In particular, in super-high-rise buildings, historic buildings, and large public facilities, it is necessary to capture the micromechanical response of load-bearing structures and the load disturbance characteristics of power distribution systems in real time. This requires the establishment of a dynamic assessment system for building safety status through spatiotemporal correlation analysis of multi-source heterogeneous data (structural stress waves, power load). Such requirements require technical solutions that can integrate multimodal data, analyze high-frequency signals, and reconstruct the topology of three-dimensional models in real time to ensure precise decision-making throughout the lifecycle of urban infrastructure management.

[0003] The current mainstream solution utilizes static model update technology based on multi-sensor fusion, such as a combined monitoring system based on LiDAR and vibration sensors. This approach deploys a distributed fiber-optic sensor (FOS) network to collect building strain data. This data is combined with load time series data from the supervisory control and data acquisition (SCADA) system, and then uses a finite element analysis (FEA) model to regularly update the 3D mesh model of the building structure. The core process involves periodically collecting structural vibration spectra and current fluctuation data, calculating the structural stiffness degradation coefficient using an offline batch processing algorithm, and finally manually triggering batch corrections to the coordinates of key nodes in the 3D model.

[0004] However, the sampling frequency discrepancy between vibration sensors and the power system (typically in the kHz and Hz ranges) results in millisecond-level time window offsets in correlation analysis of abnormal events, making it impossible to accurately capture the causal temporal relationship between stress wave propagation paths and current mutations. Furthermore, finite element models rely on a preset static mesh topology. Dynamic adjustments to mesh vertex coordinates require recalculation of the global stiffness matrix, resulting in model update delays of up to tens of minutes, making it difficult to meet real-time warning requirements. Summary of the Invention

[0005] The embodiments of the present application provide a real-time three-dimensional city modeling and dynamic updating method and system based on deep learning, which are used to solve the problems of low processing efficiency and poor accuracy of three-dimensional city modeling in the prior art.

[0006] In a first aspect, embodiments of the present application provide a real-time 3D city modeling and dynamic updating method based on deep learning, comprising: Obtaining static load parameters of the target building's structure, including beam and column cross-sectional dimensions, material elastic modulus, and foundation settlement history data; Deploy a microseismic sensor array at the nodes between the load-bearing walls and frame beams of the target building, wherein the microseismic sensor array collects data on the propagation path of stress waves generated inside the building structure at a preset sampling frequency; Synchronously receiving power load data uploaded by a smart meter connected to the target building's power distribution system, the power load data including hourly current fluctuation characteristics and a three-phase imbalance index; Extracting a current mutation feature segment within a corresponding time window in the power load data according to a start timestamp of the abnormal vibration event in the stress wave propagation path data; Inputting the static load parameters of the building structure, the frequency-domain energy distribution of the stress wave corresponding to the abnormal vibration event, and the phase angle offset of the current mutation characteristic segment into a pre-built deep learning model, so as to generate a model update instruction by coupling the association rules of the mechanical response characteristics in the stress wave propagation path data and the load disturbance characteristics in the power load data through the deep learning model; Based on the model update instruction, the grid vertex coordinates associated with the load-bearing walls and frame beam-column nodes of the target building are dynamically adjusted in the initial three-dimensional city model, and the spatial influence range boundary of the current mutation feature segment is superimposed and displayed.

[0007] Optionally, the static load parameters of the building structure, the frequency-domain energy distribution of the stress wave corresponding to the abnormal vibration event, and the phase angle offset of the current mutation characteristic segment are input into a pre-built deep learning model, so as to generate a model update instruction by coupling the association rules of the mechanical response characteristics in the stress wave propagation path data and the load disturbance characteristics in the power load data through the deep learning model, including: Perform equivalent load mapping on the input static load parameters of the building structure, the stress wave frequency domain energy distribution corresponding to the abnormal vibration event, and the phase angle offset of the current mutation characteristic segment, generate an energy density vector and a phase angle offset sequence, and output a static load parameter matrix; Based on the static load parameter matrix, the stress wave energy density vector and the phase angle offset sequence are time-frequency aligned, and the structural mechanical field distribution diagram and the time-frequency joint characteristic matrix are output based on the time-frequency alignment results; Convert the structural mechanics field distribution map into a high-dimensional gridded tensor, superimpose the energy information in the time-frequency joint feature matrix, establish a mapping relationship between the three-dimensional space grid and the frequency domain energy, and output a multi-dimensional feature tensor; The multidimensional feature tensor is input into a pre-built deep learning model, and the association rules between the mechanical response characteristics in the stress wave propagation path data and the load disturbance characteristics in the power load data are extracted through an adaptive convolution kernel and coupled with the deep learning model, and a model update instruction is generated based on the coupling result.

[0008] Optionally, the multidimensional feature tensor is input into a pre-built deep learning model, and the association rules between the mechanical response characteristics in the stress wave propagation path data and the load disturbance characteristics in the power load data are extracted through an adaptive convolution kernel, and the association rules are coupled with the deep learning model. Based on the coupling results, a model update instruction is generated, including: Input a multidimensional feature tensor, use the sliding window segmentation method to segment the continuous time series signal, and output a standardized multidimensional spatiotemporal feature tensor based on the segmentation results; Based on the multidimensional spatiotemporal feature tensor, a multidimensional spatiotemporal channel attention mechanism is constructed to quantify the coupling relationship between the mechanical response characteristics and the load disturbance characteristics, and an association rule matrix is generated based on the coupling relationship.

[0009] The association rule matrix is coupled with the weight matrix of the deep learning model, and the bidirectional coupling between the rules and the model is achieved through a layer-by-layer parameter fusion mechanism, and the coupled joint parameter tensor is output; According to the characteristic distribution of the joint parameter tensor, residual connection and nonlinear activation function are used to spatially decouple the deformation instructions, and model update instructions are generated based on the spatial decoupling results.

[0010] Optionally, the method of constructing an inter-channel attention mechanism based on the multidimensional spatiotemporal feature tensor to quantify the coupling relationship between the mechanical response feature and the load disturbance feature, and generating an association rule matrix based on the coupling relationship, includes: Based on the multi-dimensional spatiotemporal feature tensor, an inter-channel attention mechanism is constructed, and an inter-channel attention weight matrix is generated by calculating the correlation weights between different channels; The coupling relationship between the mechanical response characteristics and the load disturbance characteristics is quantified using the inter-channel attention weight matrix, and the contribution of each channel to the coupling relationship is obtained to generate the quantitative results of the coupling relationship. Based on the quantitative results of the coupling relationship, an association rule matrix is constructed, and the association rule matrix is generated by describing the association strength between the mechanical response characteristics and the load disturbance characteristics.

[0011] Optionally, extracting a current mutation feature segment within a corresponding time window in the power load data according to a start timestamp of the abnormal vibration event in the stress wave propagation path data includes: According to the start timestamp of the abnormal vibration event in the stress wave propagation path data, the transmission delay between the start timestamp of the abnormal vibration event and the current signal receiving end is calculated, and a current signal trigger timestamp sequence is generated based on the transmission delay; Taking the current signal trigger timestamp sequence as the reference point, the forward extension window length and the backward retention window length are dynamically calculated, and the current mutation feature segment containing the complete vibration feature is intercepted before and after the reference point based on the window length.

[0012] Optionally, after intercepting the current mutation feature segment containing the complete vibration feature before and after the reference point based on the window length, the method further includes: Multi-resolution time-frequency decomposition is performed on the intercepted current mutation feature segment to separate the narrowband high-frequency transient component feature subset that is strongly correlated with the abnormal vibration event; A sliding time window is used to perform time-warping matching on a narrowband high-frequency transient component feature subset, and a current mutation feature segment within a corresponding time window in the power load data is screened out based on the time-warping matching result.

[0013] Optionally, based on the model update instruction, dynamically adjusting the grid vertex coordinates associated with the load-bearing walls and frame beam-column nodes of the target building in the initial three-dimensional city model, and overlaying and displaying the spatial influence range boundary of the current mutation characteristic segment, including: Based on the model update instruction, extract the geometric topological relationship between the load-bearing walls and the frame beam-column nodes in the initial three-dimensional city model, traverse the mesh vertex data of the geometric topological relationship and establish a vertex index table, and generate a vertex mapping relationship table based on the vertex index table; Based on the vertex mapping table of the target building structure, the dynamic offset vector is extracted and superimposed point by point with the local coordinate values of the vertices in the vertex mapping table, the vertex coordinate values are updated, and an updated vertex coordinate set is output based on the vertex coordinate values; According to the current mutation feature segment, a vertex subset affected by the offset is extracted from the updated vertex coordinate set, a geometric outline of the spatial influence range boundary is generated by a convex hull algorithm, and the spatial influence range boundary of the current mutation feature segment is output based on the geometric outline.

[0014] In a second aspect, the embodiments of the present application provide a real-time 3D city modeling and dynamic updating system based on deep learning, including: An acquisition module is used to acquire static load parameters of the building structure of the target building, wherein the static load parameters of the building structure include beam and column cross-sectional dimensions, material elastic modulus, and foundation settlement history data; A deployment module deploys a microseismic sensor array at the nodes between the load-bearing walls and frame beams of the target building, wherein the microseismic sensor array collects data on the propagation path of stress waves generated inside the building structure at a preset sampling frequency; a receiving module for synchronously receiving power load data uploaded by a smart meter connected to the power distribution system of the target building, wherein the power load data includes hourly current fluctuation characteristics and a three-phase imbalance index; an extraction module, which extracts a current mutation characteristic segment within a corresponding time window in the power load data according to a start timestamp of the abnormal vibration event in the stress wave propagation path data; A generation module inputs the static load parameters of the building structure, the stress wave frequency domain energy distribution corresponding to the abnormal vibration event, and the phase angle offset of the current mutation characteristic segment into a pre-built deep learning model, so as to generate a three-dimensional model update instruction by coupling the association rules between the structural mechanical response and the electrical load disturbance through the deep learning model; The overlay module dynamically adjusts the grid vertex coordinates associated with the load-bearing walls and frame beam-column nodes of the target building in the initial three-dimensional city model based on the three-dimensional model update instruction, and overlays and displays the spatial influence range boundary of the current mutation feature segment.

[0015] In a third aspect, an embodiment of the present application provides a computing device comprising a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are used to be called and executed by the processing component to implement a real-time three-dimensional urban modeling and dynamic update method based on deep learning as described in the first aspect above.

[0016] In a fourth aspect, an embodiment of the present application provides a computer storage medium storing a computer program. When the computer program is executed by a computer, it implements a real-time three-dimensional city modeling and dynamic update method based on deep learning as described in the first aspect.

[0017] In the embodiment of the present application, the static load parameters of the building structure of the target building are obtained, and the static load parameters of the building structure include the cross-sectional dimensions of beams and columns, the elastic modulus of the materials, and the historical data of foundation settlement; a microseismic sensor array is deployed at the load-bearing walls and the frame beam-column nodes of the target building, and the microseismic sensor array collects the stress wave propagation path data generated inside the building structure at a preset sampling frequency; the power load data uploaded by the smart meter connected to the power distribution system of the target building is synchronously received, and the power load data includes the hourly current fluctuation characteristics and the three-phase imbalance index; according to the starting timestamp of the abnormal vibration event in the stress wave propagation path data, Extract the current mutation characteristic segment within the corresponding time window in the power load data; input the static load parameters of the building structure, the stress wave frequency domain energy distribution corresponding to the abnormal vibration event, and the phase angle offset of the current mutation characteristic segment into a pre-built deep learning model, so as to generate a three-dimensional model update instruction by coupling the association rules of the structural mechanical response and the electrical load disturbance through the deep learning model; based on the three-dimensional model update instruction, dynamically adjust the grid vertex coordinates associated with the load-bearing walls and frame beam-column nodes of the target building in the initial three-dimensional city model, and superimpose and display the spatial influence range boundary of the current mutation characteristic segment.

[0018] The technical solution of this application has the following beneficial effects: By extracting data on beam and column cross-sectional dimensions, material elastic moduli, and foundation settlement history, fundamental physical parameters are provided for subsequent structural mechanical response analysis, ensuring the comprehensiveness and accuracy of model input data. An array of microseismic sensors is deployed at the joints between load-bearing walls and frame beams and columns, collecting data on the internal stress wave propagation paths of the building structure at a preset sampling frequency. This provides high-precision spatiotemporal features for detecting abnormal vibration events. Hourly current fluctuation characteristics and three-phase imbalance indicators are captured, providing key data support for analyzing the impact of electrical load disturbances on the building structure and enabling simultaneous collection and correlation analysis of multi-source data. Based on the start timestamp of the abnormal vibration event, the current mutation feature segments within the corresponding time window are extracted to accurately capture the temporal correlation between the electrical load disturbance and the building structural vibration, providing high-quality feature data for subsequent model input. The static load parameters of the building structure, the frequency-domain energy distribution of the stress wave, and the phase angle offset of the current mutation feature segments are input into a deep learning model. By coupling the association rules between the structural mechanical response and the electrical load disturbance, accurate 3D model update instructions are generated, providing a scientific basis for dynamic adjustment of the building model. Based on the 3D model update instructions, the grid vertex coordinates of the load-bearing walls and frame beam-column nodes are dynamically adjusted in the initial 3D city model, and the spatial influence range boundaries of the current mutation characteristic segments are superimposed and displayed to intuitively reflect the actual impact of electrical load disturbances on the building structure.

[0019] Furthermore, at the target building's load-bearing wall and frame beam-column nodes, a spatial grid modeling method was used to establish the initial deployment locations of the microseismic sensor array based on the building's three-dimensional geometric parameters. Redundant measurement points were eliminated through iterative optimization, retaining microseismic sensor array locations sensitive to structural vibration characteristics. Based on the optimized locations, the microseismic sensor array was installed using a drilling fixation technique, with the drilling depth dynamically adjusted based on the elastic modulus and stress wave attenuation characteristics of the node material. Based on the installed locations, an orthogonal matching pursuit algorithm was used to reconstruct high-precision waveform data and extract preprocessed stress wave propagation path data. A dynamic velocity-space mapping table was generated based on the preprocessed data, and the stress wave propagation path data was output in a standardized format. Through spatial grid modeling and iterative optimization, the optimal deployment locations of the microseismic sensor array were determined to ensure comprehensive and representative monitoring data. A drilling fixation technique and dynamic adjustment of the drilling depth were used to improve sensor installation stability and data acquisition accuracy. An orthogonal matching pursuit algorithm was used to reconstruct high-precision waveform data, extract preprocessed stress wave propagation path data, and generate a dynamic velocity-space mapping table, providing a standardized, high-precision data foundation for subsequent building structural health monitoring and safety assessment.

[0020] These and other aspects of the present application will become more readily apparent from the description of the following embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0022] Figure 1 A flowchart of a real-time three-dimensional city modeling and dynamic updating method based on deep learning provided by the present application is shown; Figure 2 A schematic diagram of the structure of a real-time three-dimensional city modeling and dynamic updating system based on deep learning provided by the present application is shown; Figure 3 A schematic structural diagram of a computing device provided by the present application is shown. DETAILED DESCRIPTION

[0023] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0024] In some of the processes described in the specification and claims of this application and the above-mentioned figures, multiple operations that appear in a specific order are included, but it should be clearly understood that these operations may not be executed in the order in which they appear in this document or may be executed in parallel. The serial numbers of the operations, such as 101, 102, etc., are only used to distinguish between different operations, and the serial numbers themselves do not represent any order of execution. In addition, these processes may include more or fewer operations, and these operations may be executed in sequence or in parallel. It should be noted that the descriptions of "first", "second", etc. in this document are used to distinguish different messages, devices, modules, etc., and do not represent a sequential order, nor do they limit "first" and "second" to being different types.

[0025] This application obtains the static load parameters of the target building's building structure, combines the stress wave propagation path data collected by the microseismic sensor array deployed at the load-bearing wall and frame beam-column nodes, and the power load data uploaded by the smart meter received synchronously, to extract the current mutation characteristic segment within the time window corresponding to the abnormal vibration event; inputs the static load parameters of the building structure, the stress wave frequency domain energy distribution and the phase angle offset of the current mutation characteristic segment into a pre-built deep learning model, couples the association rules of the structural mechanical response and the electrical load disturbance, and generates a three-dimensional model update instruction; based on the instruction, the grid vertex coordinates of the target building in the initial three-dimensional city model are dynamically adjusted, and the spatial influence range boundary of the current mutation characteristic segment is superimposed and displayed, thereby realizing the building structure health monitoring and the accurate assessment of the impact of the electrical load disturbance.

[0026] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without making creative efforts are within the scope of protection of this application.

[0027] Figure 1 A flowchart of a real-time three-dimensional city modeling and dynamic updating method based on deep learning is provided for the embodiment of the present application. Figure 1 As shown, the method includes: 101. Obtain the static load parameters of the building structure of the target building; In this step, the static load parameters of the building structure refer to the inherent property parameters of the building under static loads (such as its own weight and the weight of fixed equipment), including the cross-sectional dimensions of beams and columns (geometric shape data, such as width and height), material elastic modulus (a quantitative indicator of the material's ability to resist deformation, such as the elastic modulus of concrete is 30GPa), and foundation settlement history data (a record of the cumulative settlement of the foundation caused by long-term loads).

[0028] The historical data of foundation settlement is time series data collected through long-term installation of settlement monitoring points (such as static levels), which reflects the dynamic relationship between the distribution of building loads and the compression characteristics of the foundation soil.

[0029] In this application example, design parameters are first directly derived through a Building Information Model (BIM) or point cloud modeling of the existing building is performed using a 3D laser scanner to extract beam and column cross-sectional dimensions. Next, core samples are drilled from key building components (e.g., first-floor frame columns) and pressure tested in the laboratory. Stress-strain curves are recorded and the material elastic modulus is calculated. Alternatively, the elastic modulus is estimated using nondestructive testing equipment (e.g., a rebound hammer) combined with empirical formulas. Next, historical foundation settlement data is recorded in real time using pre-embedded electronic settlement sensors (e.g., a static level) or by exporting historical records from an existing monitoring system (e.g., time-based settlement data in Excel spreadsheets). Finally, data from various sources (BIM models, laboratory reports, sensor data) is converted into a database with a unified format. For example, beam and column dimensions are stored as millimeters, elastic moduli are standardized in GPa, and settlement data is aligned by timestamp.

[0030] For example, engineers at a high-rise office building first extracted beam and column cross-sectional dimensions from the BIM model (for example, the first-floor frame columns were 600mm x 800mm). They then drilled and cored the concrete columns, and laboratory measurements revealed an elastic modulus of 32.5 GPa. They also exported settlement data from the foundation monitoring system over the past three years, which showed a maximum cumulative settlement of 15mm. All this data was then consolidated into a structured table, serving as input for subsequent analysis.

[0031] 102. Deploy a microseismic sensor array at the joints between the load-bearing walls and frame beams of the target building, wherein the microseismic sensor array collects data on the propagation path of stress waves generated inside the building structure at a preset sampling frequency; In this step, the microseismic sensor array, a monitoring network consisting of multiple highly sensitive vibration sensors (such as piezoelectric sensors), is deployed at key force transmission nodes of the building (such as the junction of beams and columns and the bottom of load-bearing walls) to capture tiny vibration signals generated by load changes or external disturbances inside the structure.

[0032] Stress wave propagation path data is the vibration waveform data recorded by the sensor, which includes the stress wave propagation time, amplitude attenuation characteristics and frequency components, and is used to infer the vibration source location and propagation path.

[0033] In this application example, sensors with a frequency response range of 0.1Hz to 1kHz (such as PCB Piezotronics393B12) are selected to ensure the capture of low-frequency structural vibrations and high-frequency local impact signals. Based on the analysis of the force transmission path of the building structure (such as finite element model simulation), the measurement point locations are determined at the nodes between the load-bearing walls and beams and columns. For example, four sensors are deployed on each floor to form a grid layout. The sampling frequency is set to 2kHz, and the signals of each sensor are recorded synchronously using a multi-channel data acquisition card. A low-pass filter (cutoff frequency 500Hz) is used to suppress high-frequency noise. The raw waveform data is baseline corrected (to eliminate temperature drift) and amplitude normalized before the microseismic sensor array is deployed.

[0034] Engineers deployed microseismic sensors at eight key locations within the steel frame of a gymnasium, sampling at a 2kHz frequency for 48 hours. After data preprocessing, they captured stress wave signals caused by the start and stop of air conditioning units. The propagation path revealed that the vibrations traveled from the equipment floor along the columns to the roof, with a propagation time of 12ms.

[0035] 103. Synchronously receive power load data uploaded by a smart meter connected to the power distribution system of the target building; In this step, the power load data includes hourly current fluctuation characteristics and three-phase imbalance indicators. The power load data is a sequence of current, voltage, power and other parameters recorded by the smart meter over time, including hourly current fluctuation characteristics (such as peak and valley value changes) and three-phase imbalance (reflecting the percentage of current difference between each phase of the distribution system).

[0036] In this application example, high-precision clock modules are deployed in smart meters and microseismic sensor arrays to ensure that their timestamps are aligned within ±1ms. The system time of the data acquisition server is calibrated using NTP (Network Time Protocol) to ensure a consistent time base at the data receiving end. Based on the Modbus TCP or IEC 61850 protocol, raw sampled data (sampling frequency 1kHz, i.e., 1000 data points per second) of three-phase current, voltage, and power are acquired from the smart meters in real time. A message queue is used to establish a data pipeline, streaming the meter data to edge computing nodes. Time windows are defined to segment the current time series data. Within each window, a local extremum search algorithm is used to identify current peaks and valleys, and their amplitudes and occurrence times are recorded. The current standard deviation within the window is calculated, and a dynamic threshold (e.g., mean ±2 standard deviation) is set to flag abnormal fluctuations exceeding the threshold. A three-layer wavelet decomposition of the current signal is performed using a Daubechies wavelet basis to separate high-frequency transient components (detail coefficients) from low-frequency steady-state components. The raw current data, peak / valley value marks, and three-phase imbalance indicators are encapsulated into JSON format data packets, attached with millisecond timestamps, and stored in the power load data.

[0037] In a shopping mall's power distribution system, smart meters upload A / B / C three-phase current data every second (for example, the peak current of phase A is 150A and the valley current is 80B). The calculated three-phase imbalance is 18%. The data timestamp is strictly aligned with the microseismic sensor data to ensure the accuracy of subsequent correlation analysis.

[0038] 104. Extracting a current mutation feature segment within a corresponding time window in the power load data according to a start timestamp of the abnormal vibration event in the stress wave propagation path data; In this step, an abnormal vibration event is a vibration signal in the microseismic data whose amplitude exceeds a preset threshold (such as 0.05g), which may be caused by elevator start-up and stop, equipment failure or external impact.

[0039] The current mutation feature segment is a current data segment corresponding to the time window of the abnormal vibration event, and contains the phase angle offset (the phase lag or lead angle of the current waveform relative to the standard sine wave).

[0040] In this application example, a dynamic time warping algorithm is used to align stress wave data (high-frequency sampling) and electric meter data (low-frequency sampling), eliminating time axis offsets caused by sampling frequency differences. Based on the building's maximum span and the propagation speed of stress waves in the structure (e.g., approximately 4000 m / s in concrete), the time required for stress wave propagation is calculated, and the time window is expanded to twice the propagation time to ensure complete event coverage. Within the time window, a cumulative sum algorithm is used to detect the start and end points of the current signal's sudden change. The sudden change interval is determined by comparing the cumulative deviation with a preset threshold. The current signal in the sudden change segment undergoes analytical signal transformation to extract the instantaneous phase angle. The phase angle change from the start to the end of the sudden change is calculated as a characteristic parameter of the load disturbance. Power spectrum analysis is performed on the stress wave signal of the abnormal vibration event to extract frequency domain energy distribution characteristics (e.g., energy value at the 100 Hz frequency point). The current sudden change characteristics (phase angle offset, energy entropy) are aligned with the stress wave frequency domain energy and three-phase imbalance within the time window. The temporal correlation coefficient between the two is calculated, and characteristic segments with strong correlation are selected as valid input. A structured dataset containing the time window, current mutation parameters (phase angle offset, peak amplitude), stress wave energy distribution, and three-phase imbalance is generated and pushed to the deep learning model in real time via the stream processing engine, triggering dynamic updates of the 3D model. The final output is the current mutation feature segment.

[0041] After an abnormal vibration event (amplitude 0.08g, timestamp 14:30:05.500) was triggered in an office building, current data for the corresponding time window was extracted. It was found that the phase angle of the B phase current shifted by +8°, indicating that the sudden start of the elevator caused an imbalance in the three-phase load and was highly correlated with the structural vibration event.

[0042] 105. Input the static load parameters of the building structure, the frequency domain energy distribution of the stress wave corresponding to the abnormal vibration event, and the phase angle offset of the current mutation characteristic segment into a pre-built deep learning model, so as to generate a three-dimensional model update instruction by coupling the association rules between the structural mechanical response and the electrical load disturbance through the deep learning model; In this step, the stress wave frequency-domain energy distribution is determined by converting the time-domain vibration signal into a frequency-domain energy spectrum using a fast Fourier transform (FFT). This quantifies the energy contribution of different frequency components (e.g., 70% of low-frequency energy is in the 0-100 Hz range). The association rules between structural mechanical response and electrical load perturbations describe the dynamic coupling between the mechanical behavior of a building structure under external or internal loads and the changes in the electrical load of the power distribution system. By analyzing the correlation between the two in time, space, and frequency domains, we can explore their inherent causal relationships or patterns of synergistic variation, providing a cross-disciplinary data fusion basis for building safety monitoring and 3D modeling.

[0043] Association rule coupling is a deep learning model that extracts spatial features through convolutional neural networks (CNNs), combines them with long short-term memory networks (LSTMs) to capture temporal dependencies, and learns the nonlinear relationship between mechanical responses and electrical disturbances.

[0044] In this application example, the model is first trained using historical data. The deep learning module is trained with the optimization goal of minimizing the error between the predicted deformation and the actual monitored value. Next, static load parameters (such as an elastic modulus of 32.5 GPa) are encoded as vectors and concatenated with the frequency domain energy distribution (such as 30% energy at 100 Hz) and the phase angle offset (+8°) to form multidimensional input features. Finally, the model outputs instructions for updating the 3D model (such as node displacements).

[0045] Input the static load parameters (elastic modulus 32.5 GPa), frequency domain energy distribution (100 Hz energy accounts for 30%), and phase angle offset (+8°) for a specific event. Also included are the direction and magnitude of the load acting on the node (e.g., X / Y wind loads, equipment vibration), material properties (elastic modulus, etc.), and geometric dimensions. Use the finite element method to divide the mesh, define node constraints, and assign load directional components. Based on the material and geometric parameters, calculate the local and global stiffness matrices of the node in the X / Y directions. Simultaneously solve the force-displacement equilibrium equations and numerically determine the preliminary deformation in the X / Y directions. Introduce material nonlinearity (e.g., plastic deformation) to adjust the results, compare the results with measured data to verify the error, and output the final deformation (e.g., X+3mm / Y-1mm). Generate a 3D model update command.

[0046] 106. Based on the three-dimensional model update instruction, dynamically adjust the grid vertex coordinates associated with the load-bearing walls and frame beam-column nodes of the target building in the initial three-dimensional city model, and overlay and display the spatial influence range boundary of the current mutation feature segment.

[0047] In this step, the mesh vertex coordinates are adjusted to modify the spatial coordinates of the beam-column nodes in the 3D BIM or CAD model based on the model instructions to reflect the actual deformation.

[0048] The spatial impact range boundary is interpolated to visualize the impact area of the current mutation, for example, using a heat map to show the structural area affected by the electrical disturbance.

[0049] In this application example, in Revit or AutoCAD, the grid vertex coordinates in the initial 3D city model are adjusted according to instructions (for example, the node coordinates are updated from (1000, 2000, 3000) to (1000, 2003, 3000)). A heat map is generated using Python's Matplotlib library or GIS software, overlaying the spatial impact range boundaries of the current mutation characteristic segment (for example, a red area with a radius of 10m).

[0050] In a 3D model of a hospital, the frame column node coordinates are updated according to instructions (Y direction +3mm), and the area affected by the current mutation is superimposed (the red area covers the operating room floor), intuitively indicating the structural impact of electrical load disturbances on sensitive areas and assisting engineers in formulating reinforcement plans.

[0051] In summary, this solution integrates building static load parameters, stress wave propagation data from microseismic sensor arrays, and electricity load data from smart meters. It leverages a deep learning model to dynamically couple structural mechanical responses with electrical disturbances, generating 3D model update instructions. This dynamically adjusts the building model's grid coordinates and visualizes the impact of electrical disturbances. This enables real-time monitoring and precise assessment of building structural deformation under complex loads, providing comprehensive technical support for structural health diagnosis, risk warning, and maintenance decision-making. For example, this solution enabled a commercial complex to proactively detect local beam deformation caused by sudden load changes in air conditioning units, enabling timely reinforcement to avert potential safety hazards.

[0052] In order to solve the problems of misjudgment and delayed response of abnormal events caused by single physical domain modeling in complex working conditions in traditional methods, in some embodiments, the static load parameters of the building structure, the stress wave frequency domain energy distribution corresponding to the abnormal vibration event, and the phase angle offset of the current mutation characteristic segment are input into a pre-built deep learning model. The deep learning model is used to couple the association rules of the mechanical response characteristics in the stress wave propagation path data and the load disturbance characteristics in the power load data to generate a model update instruction, including: 201. Perform equivalent load mapping on the input static load parameters of the building structure, the stress wave frequency domain energy distribution corresponding to the abnormal vibration event, and the phase angle offset of the current mutation characteristic segment, generate an energy density vector and a phase angle offset sequence, and output a static load parameter matrix; In step 201, equivalent load mapping converts non-mechanical parameters (such as current phase angle offset) into equivalent static load parameters (N / m²), establishing a mechanical equivalence across physical domains. The energy density vector represents the energy distribution vector of the stress wave in each frequency subband (e.g., 0.1Hz-10Hz is divided into 10 subbands). The phase angle offset sequence represents the temporal variation of the phase angle during a sudden current event (unit: degrees). Static load parameters, including beam and column cross-sectional dimensions, concrete elastic modulus, and steel reinforcement ratio, are extracted from the building BIM model or design documents to generate an initial static load parameter matrix (matrix dimension: parameter type × spatial location). Spectral analysis is performed on the stress wave signals collected by the microseismic sensors to extract the energy contribution of different frequency bands (e.g., low frequency 0-100Hz, medium frequency 100-500Hz, and high frequency >500Hz). These energy contributions are aggregated by spatial location (e.g., nodes, walls) into energy density vectors (vector dimension: number of frequency bands × number of locations). Phase angle changes during current mutations are sampled segment by segment (e.g., one sampling point per millisecond) to generate time series data, which is then normalized to eliminate dimensional differences. Based on the energy-to-load conversion coefficient (calibrated by the material's mechanical properties), the energy density vector is converted to an equivalent dynamic load. The phase angle offset sequence is mapped to a load perturbation directional weight (e.g., a positive phase angle offset corresponds to an increased load in the X-direction). The static load parameter matrix, energy density vector, and phase angle weight sequence are integrated to generate an energy density vector and phase angle offset sequence, and the static load parameter matrix is output.

[0053] 202. Based on the static load parameter matrix, the stress wave energy density vector and the phase angle offset sequence are time-frequency aligned, and the structural mechanics field distribution diagram and the time-frequency joint characteristic matrix are output based on the time-frequency alignment results; In step 202, time-frequency alignment interpolates the stress wave energy density (1000 Hz) and phase angle offset sequences (50 Hz) at different sampling rates onto a unified time axis. The time-frequency joint feature matrix contains the joint feature matrix of the time-domain phase offset and the frequency-domain energy distribution (dimensions: time × frequency × space).

[0054] In this embodiment, a dynamic time warping (DTW) algorithm is used to align the time axes of stress waves (high-frequency sampling) and current data (low-frequency sampling) to compensate for sampling rate differences. Frequency-domain interpolation is performed on the stiffness distribution in the static load parameter matrix to ensure a one-to-one correspondence with the stress wave frequency bands. Based on the equivalent load data, finite element simulation is used to calculate the stress-strain distribution at each spatial location, outputting a three-dimensional mechanical field diagram (e.g., color mapping to represent deformation and heat maps to represent energy concentration areas).

[0055] The phase angle offset sequence is sliced into time windows and superimposed with the stress wave frequency domain energy in the corresponding time window to construct a time-frequency two-dimensional matrix, and the structural mechanics field distribution map and the time-frequency joint characteristic matrix are output.

[0056] 203. Convert the structural mechanics field distribution map into a high-dimensional gridded tensor, superimpose the energy information in the time-frequency joint feature matrix, establish a mapping relationship between the three-dimensional space grid and the frequency domain energy, and output a multi-dimensional feature tensor; In step 203, the high-dimensional gridded tensor discretizes the structural mechanics field distribution map into a three-dimensional voxel grid (voxel size 0.1m×0.1m×0.1m). Frequency-domain energy mapping maps the energy information in the joint time-frequency matrix to the three-dimensional grid nodes, forming an energy-space coupling field.

[0057] In an embodiment of the present application, the mechanical field distribution map is discretized according to a spatial grid (such as a 1m×1m×1m cube), and each grid vertex stores parameters such as local deformation variables and stress values to form a three-dimensional grid tensor (dimensions: X×Y×Z×parameters). The energy information in the time-frequency joint feature matrix is superimposed: each grid is associated with the main frequency band energy value and phase angle weight of the spatial region in which it is located, and the tensor dimension is expanded (such as X×Y×Z×parameter×frequency band). High-frequency energy areas (such as stress wave energy concentration areas) are marked as "sensitive grids", and low-frequency energy areas are marked as "stable grids", and the directionality of energy distribution is adjusted based on the phase angle weight. A multi-dimensional feature tensor is output for deep learning model input.

[0058] 204. Input the multidimensional feature tensor into a pre-built deep learning model, extract the association rules between the mechanical response characteristics in the stress wave propagation path data and the load disturbance characteristics in the power load data through an adaptive convolution kernel, couple them with the deep learning model, and generate a model update instruction based on the coupling result.

[0059] In step 204, the adaptive convolution kernel dynamically adjusts the convolution kernel weights based on the input features, prioritizing the capture of strong mechanical-electrical correlations. Association rules are coupled to establish a mathematical correlation model between the stress wave propagation path (e.g., wave velocity decay rate) and the load disturbance (e.g., phase mutation frequency).

[0060] In this embodiment, multi-scale convolution kernels are used (e.g., 3×3×3 grid convolution to extract local features, 7×7×7 convolution to capture global correlations), dynamically adjusting the receptive field of the convolution kernels to adapt to the coupled relationship between energy and spatial deformation in different frequency bands. During the model training phase, an attention mechanism is used to identify key grids (e.g., high-energy sensitive areas) and key time windows (e.g., phase angle mutation periods). During the inference phase, the model automatically matches similar patterns in historical data (e.g., energy surges in a certain frequency band accompanied by X-direction deformation) and outputs an association rule confidence score. The association rule is compared with a preset threshold (e.g., confidence > 0.8). If the criteria are met, an update command is triggered: the corresponding grid coordinates are adjusted based on the deformation amount (e.g., X+3mm / Y-1mm), and the boundary affected by the current mutation (e.g., the red area represents the area of maximum deformation caused by load disturbance) is superimposed based on energy density and phase angle weights. Model update commands are generated based on the coupling results.

[0061] In a structural health monitoring scenario for a super-high-rise building (400m in height), a current phase change of 15° (Δθ) translates to an equivalent load (F_equ) of 150kN / m². The energy density vector indicates that the 2Hz frequency band accounts for 45%. Time-frequency alignment generates a mechanical field distribution map, showing a stress concentration of 120MPa at the 80th floor node. The time-frequency matrix dimensions are [3600×10×800×5]. The gridded tensor contains 8,000,000 voxels, and the energy in the 2Hz frequency band mapped to the stress concentration region reaches a value of 0.7. Adaptive convolution captures a strong correlation (R²=0.89) between the wave velocity decay rate (-2% / s) and the phase change frequency (0.5Hz) in the stress concentration region, triggering a model weight update.

[0062] In summary, this solution achieves cross-domain coupled modeling of mechanical response and electrical load disturbances through equivalent load mapping, time-frequency joint feature fusion, and adaptive convolution rule extraction. In super-high-rise building monitoring, the model's detection rate for hidden damage has increased to 95% (compared to 78% with traditional methods), the false alarm rate has been reduced to 3%, and the parameter update response time is less than 1 second, providing multi-physics collaborative analysis capabilities for real-time health assessment of complex engineering structures.

[0063] In order to solve the problem of insufficient association rule modeling caused by independent processing of mechanical and electrical characteristics in traditional deep learning models under complex working conditions, based on the multi-dimensional feature tensor (including the mechanical response characteristics of the stress wave propagation path, the load disturbance characteristics of the power load and the spatiotemporal energy distribution), the adaptive convolution kernel group in the pre-built deep learning model is used to extract mechanical characteristics such as stress wave velocity attenuation and energy diffusion direction in the three-dimensional spatial dimension, and at the same time, the electrical characteristics such as phase mutation frequency and harmonic distortion rate of the load disturbance are extracted in the time-frequency dimension; the cross-attention mechanism is used to dynamically weight the correlation strength of mechanical and electrical characteristics to construct cross-domain association rules (such as the nonlinear mapping relationship between wave velocity attenuation rate and phase mutation); through residual distribution monitoring and dynamic weight allocation strategy, the model convolution kernel weight update instructions and feature fusion rule optimization instructions are generated to realize closed-loop iterative optimization of multi-physics field coupling characteristics. In some embodiments, a multidimensional feature tensor is input into a pre-built deep learning model, an adaptive convolution kernel is used to extract association rules between the mechanical response characteristics in the stress wave propagation path data and the load disturbance characteristics in the power load data, and the association rules are coupled with the deep learning model. Based on the coupling results, a model update instruction is generated, including: 301. Input a multidimensional feature tensor, segment the continuous time series signal using a sliding window segmentation method, and output a standardized multidimensional spatiotemporal feature tensor based on the segmentation results; In step 301, the sliding window segmentation method segments the continuous time series signal into fixed time windows (e.g., 10 seconds) to ensure time series continuity and eliminate boundary effects. The standardized multidimensional spatiotemporal feature tensor is a four-dimensional data matrix (dimensions: time window × spatial location × physical field × feature channel).

[0064] In this embodiment, based on the stress wave propagation velocity and the duration of the load mutation, the window length is set to 200 milliseconds (covering a complete vibration event) and the step size is set to 50 milliseconds (to prevent truncation of key events). Mirror padding is used for incomplete windows at the beginning and end, copying adjacent data to complete the time series. A sliding window is performed along the time dimension, extracting multidimensional data blocks within each window (dimensions: spatial grid × parameter × frequency band × window time point). Dynamic time warping (DTW) is used to align data from different sensors within the window (e.g., microseismic sensors and electricity meters) to eliminate timing jitter caused by sampling rate differences. Maximum-minimum normalization is performed on the energy values in each frequency band to compress the data to the [0, 1] range. Mechanical parameters (e.g., deformation) are normalized using the Z-score. Structural mechanical parameters (e.g., stress values) and electrical load parameters (e.g., phase angle offset) are integrated according to spatial grid locations to form a unified multidimensional spatiotemporal feature tensor. The standardized multidimensional spatiotemporal feature tensor is then output.

[0065] 302. Based on the multidimensional spatiotemporal feature tensor, a multidimensional spatiotemporal channel attention mechanism is constructed to quantify the coupling relationship between mechanical response characteristics and load disturbance characteristics, and an association rule matrix is generated based on the coupling relationship.

[0066] In step 302, a multi-dimensional spatiotemporal inter-channel attention mechanism uses a multi-head attention mechanism to calculate cross-channel association weights between mechanical response features (e.g., stress wave energy) and load disturbance features (e.g., current harmonic distortion). The association rule matrix represents a three-dimensional matrix (dimensions: spatial position × physical field pair × weight value) that reflects the coupling strength between mechanical and electrical features.

[0067] In an embodiment of the present application, the channels of the spatiotemporal feature tensor are divided into a "structural mechanics group" (stress, deformation) and an "electrical load group" (phase angle, energy density). The self-attention score is calculated for each group of channels to identify key features within the group (such as deformation in high-frequency energy concentration areas). Through cross-group interactive attention, the association weights between the mechanics group and the electrical load group are calculated (for example, a certain phase angle mutation period corresponds to a high stress value of a certain spatial grid). Based on the attention weights, a spatiotemporal association matrix is constructed (rows: mechanical feature channels, columns: load feature channels). The matrix element values represent the coupling strength between the two within a specific time window. Threshold filtering is introduced (such as weight > 0.6) to retain only high-confidence association pairs to form a sparse association rule matrix. An association rule matrix is generated based on the coupling relationship.

[0068] 303. Couple the association rule matrix with the weight matrix of the deep learning model, achieve bidirectional coupling between the rules and the model through a layer-by-layer parameter fusion mechanism, and output the coupled joint parameter tensor; In step 303, the layer-by-layer parameter fusion mechanism injects the association rule matrix into the weight matrix layer by layer according to the deep learning model hierarchy (e.g., convolutional layer, fully connected layer), achieving dynamic coupling. The joint parameter tensor fuses the four-dimensional parameter set (dimensions: layer × channel × rule × weight value) of the original model weights and the association rule weights.

[0069] In the embodiment of the present application, in each layer of the model, the weight of the association rule matrix is mixed with the original model weight in proportion. For example, the kernel weight of the convolution layer is updated to Wnew=αWmodel+(1 α)Wrule, where α is dynamically adjusted by the rule confidence. A gating unit is designed to dynamically select the contribution ratio of rule weights or model weights based on the current input's spatiotemporal features. During the model's forward propagation, the output of each layer retains both the original features and the rule-enhanced features, forming a joint parameter tensor (dimensions: spatial grid × time × fused features). Cross-layer residual connections are used to fuse shallow physical rules with deep abstract features to prevent gradient vanishing. The output is a joint parameter tensor (dimensions: spatial grid × time × fused feature level), which contains the rule-guided enhanced features.

[0070] According to the characteristic distribution of the joint parameter tensor, residual connection and nonlinear activation function are used to spatially decouple the deformation instructions, and model update instructions are generated based on the spatial decoupling results.

[0071] In step 304, spatial decoupling separates the global trend and local perturbation components in the joint parameter tensor through residual connections. The model update instructions include a set of instructions for the weight adjustment amplitude, convolution kernel update priority, and learning rate adjustment strategy.

[0072] In an embodiment of the present application, multi-layer residual convolution is performed on the joint parameter tensor, and spatial features at different scales (such as local deformation details and global load distribution) are preserved through skip connections. A gated linear unit (GLU) is used as a nonlinear activation function to suppress noise characteristics and enhance the propagation of effective signals. The fused parameter tensor is split into a "structural deformation component" and a "load disturbance component," corresponding to the mesh vertex displacement and the electrical influence range, respectively. Spatial interpolation (such as bilinear interpolation) is performed on the deformation component to smoothly map the displacement of the discrete mesh vertices to the continuous three-dimensional model surface. Based on the decoupling results, a JSON-formatted instruction is generated, containing the mesh vertex ID to be adjusted and the X / Y / Z displacement (such as "Grid_205: X+3mm, Y-1mm"). The load disturbance influence range is superimposed: Based on the phase angle offset weight, the impact area of the current mutation event is marked in the three-dimensional model (for example, a red semi-transparent area indicates a high-risk deformation area). A model update instruction is generated.

[0073] In the offshore wind turbine structural health monitoring scenario, the tower is 120 meters tall. The monitoring data includes stress waves (1000 Hz) and electrical loads (50 Hz). A sliding window captures 10-second data blocks and normalizes them to generate a spatiotemporal feature tensor of dimensions [200 × 60 × 60 × 8]. Attention mechanism calculations show that the mechanical-electrical coupling weight at the tower base reaches 0.92, while the weight at the tower top reaches 0.35. After parameter fusion, the weight of the fifth convolution kernel is adjusted by ±15%, resulting in a combined parameter tensor of dimensions [8 × 512 × 8 × 512]. After spatial decoupling, the gradient amplitude of the local disturbance component at the tower base reaches 0.6, triggering a weight adjustment command (+25%, high priority).

[0074] This solution achieves dynamic collaborative optimization of multimodal data and deep learning models through sliding window segmentation, cross-physics field attention modeling, and parameter fusion and decoupling mechanisms. In wind tower monitoring scenarios, the model's detection rate for hidden damage increased from 82% to 96%, the false alarm rate decreased to 2%, and the weight update response time was shortened to 50ms, providing high-precision, low-latency intelligent decision-making capabilities for real-time health management of complex engineering structures.

[0075] In order to solve the problem of insufficient association rule modeling caused by the independent processing of physical field features in traditional methods, an inter-channel attention mechanism is constructed based on a multi-dimensional spatiotemporal feature tensor (including the spatial distribution of mechanical response features, the temporal changes of load disturbance features, and frequency domain energy information): the cross-channel correlation of the mechanical feature group (wave velocity attenuation rate of the stress wave propagation path, energy diffusion direction) and the load feature group (phase angle mutation frequency, harmonic distortion rate) is extracted through a multi-head attention model, and the cross-attention score of each physical field feature pair is calculated; based on the score distribution, high coupling strength feature pairs (such as the correlation coefficient between wave velocity attenuation rate and phase mutation > 0.8) are selected, and a matrix reflecting the mechanical-electrical cross-domain association rules is generated through spatial position weighting and frequency domain energy fusion to quantify the dynamic coupling relationship between different physical field features. In some embodiments, the inter-channel attention mechanism is constructed based on the multi-dimensional spatiotemporal feature tensor to quantify the coupling relationship between the mechanical response features and the load disturbance features, and the association rule matrix is generated based on the coupling relationship, including: 401. Based on the multi-dimensional spatiotemporal feature tensor, construct an inter-channel attention mechanism, and generate an inter-channel attention weight matrix by calculating the correlation weights between different channels; In step 401, the channels of the multidimensional spatiotemporal feature tensor are divided into logically related groups. The structural mechanics group includes channels such as stress wave energy, deformation, and modal frequency; the electrical load group includes channels such as phase angle offset, current harmonics, and three-phase imbalance. Feature dimensionality reduction performs global average pooling on each channel group, compressing the spatial dimensions while retaining key channel statistical features (such as mean and maximum values). Similarity weights are calculated between channels within a group. For example, the dot product of the feature vectors is used to measure the correlation between different channels within the same group (such as the correlation between deformation and stress wave energy). Cross-attention calculations are performed on the channels of the mechanics and load groups to capture cross-domain feature correlations (such as the synchronous changes in phase angle offset and beam node strain). The attention scores are normalized using the Softmax function to ensure that the sum of the weights is 1. The within-group and cross-group attention weights are concatenated into a global inter-channel attention weight matrix (dimensions: number of mechanics channels × number of load channels × time window). This generates the inter-channel attention weight matrix.

[0076] In this embodiment, the multidimensional spatiotemporal feature tensor is divided into a mechanical channel group (stress wave energy, wave velocity decay rate) and a load channel group (phase angle offset, harmonic distortion rate) based on the physical field type. For each channel pair (e.g., wave velocity decay rate - phase offset), cosine similarity is calculated, and the similarity scores are normalized using the Softmax function to generate an inter-channel attention weight matrix.

[0077] 402. Quantify the coupling relationship between the mechanical response characteristics and the load disturbance characteristics using the inter-channel attention weight matrix, obtain the contribution of each channel to the coupling relationship, and generate a quantitative result of the coupling relationship; In step 402, the coupling relationship is quantified as the contribution (0-1) of each channel to the cross-domain coupling relationship, calculated by weighting the attention weights. The contribution is a score of the importance of the channel pair in the overall coupling relationship. For example, a high contribution (>0.8) indicates a strong correlation.

[0078] In an embodiment of the present application, the attention weight matrix in each time window is weighted and summed according to channel pairs (such as "stress wave energy-phase angle offset") to generate a contribution score (for example: the contribution of stress wave energy to phase angle offset in a certain time window is 0.75). A sliding average filter is introduced to eliminate the influence of instantaneous noise on the contribution calculation (such as a window length of 50 milliseconds). The windows are divided along the time axis (such as a slice every 10 milliseconds), and the distribution pattern of the contribution scores in each slice is counted (such as the delayed enhancement of the mechanical response corresponding to high-frequency load disturbances). The contribution scores are clustered according to spatial grid positions (such as floors, beam-column nodes) to generate a spatial heat map of the coupling relationship (such as the contribution of a certain area of the grid is significantly higher than other positions). The contribution scores and spatiotemporal distribution characteristics are integrated into structured data to generate quantitative results of the coupling relationship.

[0079] Based on the quantitative results of the coupling relationship, an association rule matrix is constructed. By describing the association strength between the mechanical response characteristics and the load disturbance characteristics, an association rule matrix is generated.

[0080] In step 403, a fixed association pattern based on historical data statistics (such as "when the three-phase imbalance is greater than 5%, the node deformation increases by 2mm"). The association strength is dynamically adjusted according to the real-time contribution score (such as a contribution of 0.8 corresponds to a deformation adjustment coefficient of 0.1). The rows of the matrix represent mechanical characteristic channels (such as stress wave energy and deformation), and the columns represent load characteristic channels (such as phase angle offset and harmonic distortion). The matrix element values are the association strength levels (such as 0-1, 1 indicates strong association), and an independent association rule matrix is generated for each time window to support dynamic rule updates. The association rule matrix is embedded in the weight update process of the deep learning model, and the association rule matrix (dimensions: number of mechanical channels × number of load channels × time window × rule level) is output to drive the generation of model update instructions.

[0081] In this embodiment, the contribution is multiplied by the spatial distribution (e.g., the weighted proportion of high-stress regions) to generate the initial correlation strength. Channel pairs with correlation strengths > 0.7 are retained, and low-correlation noise (e.g., channel pairs with strengths < 0.3) are removed. The valid channel pairs are integrated to output the association rule matrix.

[0082] In a large-scale bridge structural health monitoring scenario, the monitoring data includes stress waves (500 Hz) and electrical loads (60 Hz). The multidimensional spatiotemporal feature tensor has dimensions of [time × 100 × 50 × 8], with four channels each for the mechanical channel group (wave velocity, energy) and the load channel group (phase, harmonics). The cosine similarity score for the velocity-phase pair is 0.91, and the weight after softmax normalization is 0.89. After weighted summation, the contribution of the velocity-phase pair in the pier area is 0.92, and 0.75 in the main beam area. The calculated association strength for the pier area is 0.92 × 0.95 (percentage of high-risk areas) = 0.87, generating an association rule matrix (wave velocity-phase pair strength 0.87).

[0083] In summary, this solution, through an inter-channel attention mechanism and quantitative contribution modeling, enables cross-domain association rule extraction of mechanical response and load disturbance characteristics. In bridge monitoring scenarios, the detection rate of high-association-strength rules (>0.8) increases to 90%, while the false alarm rate decreases to 5%, providing a quantifiable and interpretable correlation basis for the collaborative analysis of structural damage and electrical disturbances.

[0084] In order to solve the problem of difficulty in modeling the correlation between mechanical vibration events and electrical load data due to sampling rate differences or noise interference, based on the starting timestamp of the abnormal vibration event in the stress wave propagation path data, the sampling time axis of the power load data is aligned through the timestamp synchronization method, and a sliding window is used to intercept the current time series data that matches the vibration event time window (such as 5 seconds before the event to 10 seconds after the event); combined with differential threshold detection and waveform morphology analysis, the current mutation feature segment (such as phase angle jump, amplitude surge) is extracted, and the characteristic vector representing the mutation intensity is generated by time domain integration and frequency domain energy ratio calculation to achieve accurate spatiotemporal correlation between mechanical vibration events and electrical load disturbances. In some embodiments, based on the starting timestamp of the abnormal vibration event in the stress wave propagation path data, the current mutation feature segment within the corresponding time window in the power load data is extracted, including: 501. Calculate, based on the start timestamp of the abnormal vibration event in the stress wave propagation path data, a transmission delay between the start timestamp of the abnormal vibration event and the current signal receiving end, and generate a current signal trigger timestamp sequence based on the transmission delay; In step 501, the transmission delay is the time delay (in milliseconds) it takes for the stress wave to propagate from the vibration source to the current signal receiving end. This delay is calculated from the wave velocity and the propagation path length. The current signal triggers a timestamp sequence aligned with the delay, marking the start of the corresponding vibration event in the current signal.

[0085] In this embodiment, a high-precision clock module (e.g., one that supports the IEEE 1588 PTP protocol) is deployed in the microseismic sensor array and smart meter to ensure microsecond-level synchronization of their timestamps and eliminate system clock skew. If network transmission delays exist (e.g., meter data is forwarded via a gateway), the round-trip time (RTT) measurement and timestamp compensation algorithm are used to estimate the transmission delay from the acquisition end to the processing center (e.g., an average delay of 5ms). This value is deducted from the timestamp. A short-term energy detection algorithm is applied to the microseismic sensor data. When the signal energy exceeds a background noise threshold (e.g., three times the standard deviation), it is marked as the start timestamp of an abnormal vibration event. Based on the building structure topology, the theoretical propagation time of the vibration wave from the event location to the nearest microseismic sensor is calculated. The total transmission delay is calculated by combining hardware synchronization errors, transmission path delays, and signal processing time (e.g., a 1ms delay in the filtering algorithm). The vibration event start timestamp is inversely compensated for the transmission delay to obtain the current signal trigger reference point. For consecutive vibration events, a trigger timestamp sequence is generated in chronological order. This calibrated current signal trigger timestamp sequence serves as the reference point for extracting the characteristic segment.

[0086] 502. Taking the current signal triggering timestamp sequence as a reference point, dynamically calculate the forward extension window length and the backward retention window length, and intercept the current mutation feature segment containing the complete vibration feature before and after the reference point based on the window length.

[0087] In step 502, the forward window length is extended to the time range (e.g., 200ms) from the trigger timestamp forward to capture the steady-state current characteristics before the vibration event. The backward window length is maintained to the time range (e.g., 500ms) from the trigger timestamp backward to capture the entire process of the current mutation caused by the vibration event.

[0088] In this embodiment of the present application, based on historical data statistics, the lead time (e.g., 50ms) of a typical load mutation event (e.g., motor startup) is determined to capture the steady-state characteristics before the current mutation. Based on the frequency-domain energy distribution of the current vibration event (low-frequency events require a longer forward window), the window is expanded by integer multiples of the energy dominant frequency (e.g., 100Hz corresponds to a 10ms period) (e.g., 3 times the period → 30ms). Using the decay time constant of the stress wave signal (e.g., an RC circuit model), the time required for the vibration energy to drop to a steady-state level is estimated (e.g., 20ms), which serves as the backward window length. To ensure the phase angle offset of the current mutation segment is complete (e.g., a complete cycle waveform), the window is expanded by the length of an entire cycle of the power supply frequency (50Hz / 60Hz) (e.g., 50Hz corresponds to a 20ms window). The current signal within the window is checked to see if it contains a complete mutation waveform (e.g., rising edge, peak, falling edge). If the edge is truncated (e.g., the signal at the end of the window does not return to zero), it is interpolated and completed using an exponential decay curve, outputting a current mutation segment dataset containing complete vibration-related characteristics. The time window length dynamically adapts to the event characteristics.

[0089] Offshore wind turbine tower vibration monitoring. The vibration event was triggered by unbalanced blade loads. The stress wave propagation path length was 30m, and the wave velocity was 3000m / s. The propagation delay was calculated as: 30m / 3000m / s = 10ms. The vibration event start timestamp T = 2023ms, and the current signal trigger timestamp T_trigger = 2013ms. A trigger sequence of [2013ms, 4055ms, 6120ms] was generated (corresponding to three vibration events). The event type was determined to be shock-type, with a forward window of 200ms and a backward window of 500ms. The current signal window for the second event was captured: 4055ms - 200ms = 3855ms to 4055ms + 500ms = 4555ms. A phase angle jump (Δθ = 15°) and an amplitude surge (+200A) were detected within the captured segment. The characteristic segment length was 700ms.

[0090] In summary, this solution achieves high-precision spatiotemporal alignment of mechanical vibration events and current mutation signatures through transmission delay compensation and dynamic windowing. In wind turbine scenarios, the time synchronization error is reduced from ±20ms with traditional methods to ±2ms, and the completeness of the mutation signature segment (including the complete rising and falling edges) is improved from 70% to 98%, providing a reliable data foundation for the collaborative analysis of multi-physics field anomaly events.

[0091] In order to solve the problem of misjudgment of load disturbance mode caused by noise interference or waveform distortion in the current mutation feature segment, after intercepting the current mutation feature segment containing the complete vibration feature, the key parameters such as phase angle offset, amplitude mutation slope and harmonic distortion rate in the feature segment are extracted by time domain-frequency domain joint analysis method, and the dynamic feature vector representing the load disturbance mode is constructed by combining the time series stationarity test and waveform morphology matching algorithm; the sliding window dynamic weighting mechanism is used to fuse the disturbance features of multiple time scales, and the high confidence abnormal events are screened by adaptive threshold segmentation to generate a load disturbance feature library and abnormal event labels, providing standardized input data for multi-physics field coupling analysis. In some embodiments, after intercepting the current mutation feature segment containing the complete vibration feature before and after the reference point based on the window length, it also includes: 601. Perform multi-resolution time-frequency decomposition on the intercepted current mutation feature segment to separate a narrowband high-frequency transient component feature subset that is strongly correlated with the abnormal vibration event; In step 601, multi-resolution time-frequency decomposition uses analysis methods with different time-frequency resolutions (such as wavelet packet decomposition) to decompose the current signal into multiple frequency band subsets. Narrowband high-frequency transient components are transient signal components with a narrow frequency range (e.g., 1kHz to 5kHz) and a short duration (<100ms), and are strongly correlated with abnormal mechanical vibration events.

[0092] In the embodiment of the present application, wavelet packet decomposition is used, and a wavelet basis is selected as the mother wavelet, taking into account the balance between time-frequency resolution. The number of decomposition layers is set to 6, and the current signal is divided into 64 frequency bands (2^6), covering the range of 0Hz to 500Hz (assuming a sampling rate of 1kHz). According to historical data statistics, the high-frequency transient energy associated with abnormal vibration events is concentrated in the 100-300Hz frequency band. The target sub-band is filtered by the frequency band energy threshold (such as exceeding 3 times the baseline energy). The signal of the filtered sub-band is reconstructed to obtain a narrowband high-frequency transient component time series. Adaptive threshold denoising is applied to the reconstructed signal: the threshold is dynamically adjusted according to the noise level of each sub-band, and transient pulses above the threshold are retained. The amplitude envelope of the high-frequency component is extracted through envelope detection, highlighting the mutation characteristics synchronized with the vibration event. A subset of narrowband high-frequency transient component features that are strongly correlated with the abnormal vibration event is separated.

[0093] 602. Perform time-warping matching on the narrowband high-frequency transient component feature subset using a sliding time window, and filter out the current mutation feature segment within the corresponding time window in the power load data based on the time-warping matching result.

[0094] In step 602, time alignment and matching are performed using a dynamic time warping (DTW) algorithm to align the length and phase differences of different time series. A sliding time window, which slides with a fixed step size (e.g., 10ms), is used for local time series matching and feature segment screening.

[0095] In this embodiment, a window length of 10ms and a step size of 1ms are set based on the average duration of the high-frequency component (e.g., 5ms) to ensure coverage of the complete transient process. If an unattenuated transient signal is detected at the edge of the window, the window is automatically extended until the signal returns to zero. Dynamic time warping aligns the timing of the high-frequency transient component and the stress wave signal: The morphological similarity of the two waveforms (e.g., rising edge slope and peak position) is calculated to generate an optimal alignment path. After alignment, the time window in the current signal that is strongly synchronized with the onset of the stress wave event is marked. Based on the cumulative distance of the aligned path (e.g., the smaller the distance, the higher the match), a threshold (e.g., a score > 0.8) is set to select high-confidence associated segments. The matching window is centered and extended forward and backward by 50ms to capture the complete current mutation segment (including precursor, mutation, and decay). The extracted segment is checked to ensure that it contains a complete phase angle offset (e.g., at least one complete power frequency cycle). If not, it is completed by interpolation from adjacent windows. The current mutation feature segment within the corresponding time window in the load data is output.

[0096] Abnormal vibration monitoring of wind turbine gearboxes was performed with a current sampling rate of 10kHz. Current data was captured after aligning the vibration event trigger timestamps. Wavelet packet decomposition was performed to identify a sub-band between 3.5kHz and 4.5kHz, showing a 4.2-fold energy surge (baseline standard deviation 0.5A², current value 2.1A²). The feature subset matrix dimensions were [500×3×1] (500ms duration, three frequency bands). The DTW template matching threshold was set to 0.08, and a sliding window was used to detect three consecutive matching intervals (time windows 20ms–70ms, 85ms–135ms, and 150ms–200ms). The merged output current mutation feature segment lasted from 20ms to 200ms, encompassing the complete high-frequency transient waveform.

[0097] In summary, this solution uses multi-resolution time-frequency decomposition to accurately separate high-frequency transient components, and combines sliding window dynamic matching to achieve high-precision extraction of current mutation feature segments. In the wind turbine scenario, the high-frequency noise suppression rate is increased to 90%, and the time alignment error is less than 2ms, providing reliable data support for cross-domain correlation analysis of mechanical vibration and electrical disturbances.

[0098] To address the problem of insufficient representation of hidden risks caused by the static nature and single-physics perspective of traditional three-dimensional models, based on the structural parameter changes of the load-bearing walls and frame beam-column nodes in the model update instructions, the mesh vertex coordinate adjustment values are dynamically calculated using a finite element inverse solution algorithm. Combined with the spatial propagation attenuation model of the current mutation characteristic segment (such as the exponential decay law of electromagnetic field intensity with distance), the impact range of the current mutation event is mapped to the spatial grid of the three-dimensional city model. Boundary interpolation and color gradient rendering techniques are used to overlay and display the impact range boundaries, and the geometric deformation parameters of the load-bearing structure and the electrical disturbance risk hotspots are simultaneously updated, realizing dynamic reconstruction of the three-dimensional model and multi-dimensional risk visualization under the action of mechanical-electrical coupling. In some embodiments, based on the model update instructions, the mesh vertex coordinates associated with the load-bearing walls and frame beam-column nodes of the target building are dynamically adjusted in the initial three-dimensional city model, and the spatial impact range boundaries of the current mutation characteristic segment are overlaid and displayed, including: 701. Based on the model update instruction, extract the geometric topological relationship between the load-bearing walls and the frame beam-column nodes in the initial three-dimensional city model, traverse the mesh vertex data of the geometric topological relationship and establish a vertex index table, and generate a vertex mapping relationship table based on the vertex index table; In step 701, geometric topological relationships describe the spatial connectivity (e.g., adjacency and containment) between load-bearing walls and beam-column nodes, as well as the distribution of mesh vertices. A vertex index table records the mapping between mesh vertex unique identifiers (IDs) and initial coordinates, enabling rapid vertex location. The vertex mapping table extends the index table to include attributes such as the component type (e.g., load-bearing wall vertices are labeled Type_A) and the layer level (e.g., layer 5).

[0099] In this embodiment, the connection relationship between load-bearing walls and beam-column nodes is extracted from the BIM data or mesh file of the initial 3D model. A unique ID is assigned to each vertex, and its spatial coordinates, component type (beam, column, wall), and adjacent vertex IDs are recorded. A local-to-global coordinate mapping is established based on a vertex index table. The local coordinates of each vertex are based on the local coordinate system of the component to which it belongs (e.g., the beam endpoint is the origin), and the global coordinates are based on the overall building coordinate system. Fields are reserved in the mapping table to store the X / Y / Z displacement values (e.g., +3mm, -1mm) in the model update instructions. The vertex mapping table is traversed, and the dynamic offset vector (from the update instruction) is superimposed dimensionally with the original vertex coordinates to generate new coordinates. Coordinated displacement is performed on associated vertices (e.g., multiple vertices at the same beam-column node). If a vertex is displaced, its adjacent vertices are adjusted synchronously based on the connection stiffness coefficient (synchronous displacement for rigid connections, partial displacement for hinged connections). An updated set of vertex coordinates is output, containing the displaced global coordinates and a record of the coordinated adjustments.

[0100] 702. Based on the vertex mapping table of the target building structure, extract the dynamic offset vector and superimpose it point by point with the local coordinate values of the vertices in the vertex mapping table, update the vertex coordinate values, and output an updated vertex coordinate set based on the vertex coordinate values; In step 702, the dynamic offset vector is the vertex coordinate adjustment amount provided by the model update instruction (such as +0.3m in the X direction and -0.1m in the Y direction). The local coordinate value of the vertex is the coordinate of the vertex in the local coordinate system of the component to which it belongs (relative to the component origin).

[0101] In an embodiment of the present application, based on the vertex mapping table of the target building structure, the system first parses the local coordinate values and dynamic offset vectors of the vertices recorded in the table, where the local coordinate value represents the relative position of the vertex in the local coordinate system of the component to which it belongs, and the dynamic offset vector comes from the displacement calculated in the model update instruction based on the mechanical response and load perturbation association rule). During the coordinate update process, a dimension-by-dimensional superposition operation of the local coordinates and the offset vector is performed for each vertex. For vertices with rigid connections, the system automatically triggers the collaborative displacement mechanism and synchronously adjusts the coordinates of adjacent vertices based on the constraint relationship between components to avoid structural topological contradictions due to local deformation. At the same time, a physical rule verification module is introduced, and the axis-aligned bounding box algorithm is used to detect the risk of collision between vertices. The displacement amplitude is limited in combination with the material stiffness coefficient to prevent exceeding the material deformation threshold. After the update is completed, the system encapsulates the corrected global coordinate value, displacement and verification status into a standardized vertex coordinate set, uses a binary protocol for efficient storage, records version logs and operation traceability information, and finally outputs the updated vertex coordinate set.

[0102] 703. According to the current mutation feature segment, extract a vertex subset affected by the offset from the updated vertex coordinate set, generate a geometric outline of the spatial influence range boundary using a convex hull algorithm, and output the spatial influence range boundary of the current mutation feature segment based on the geometric outline.

[0103] In step 703, the vertex subset affected by the offset has its coordinates updated by more than a threshold (eg, Δ>0.1 m). The convex hull algorithm calculates the minimum convex polygon boundary of the point set to define the geometric outline of the spatial impact range.

[0104] In an embodiment of the present application, vertices whose displacement exceeds a threshold (e.g., ≥1 mm) are filtered from the updated vertex coordinate set to form a subset of affected vertices. Further filtering is performed based on the component type to which the vertex belongs (e.g., only load-bearing wall nodes), excluding irrelevant vertices of non-structural components. The global coordinates of the affected vertices are projected onto the horizontal plane (XY plane) and the vertical plane (XZ / YZ plane), and the two-dimensional convex hull is calculated respectively. The vertices are sorted by polar angle, and non-convex points are removed point by point to generate the minimum enclosing polygon. For multi-story buildings, the convex hull is calculated layered in the vertical direction (Z axis) and superimposed to form a three-dimensional polyhedron boundary. B-spline curve fitting is performed on the convex hull contour to eliminate jagged edges and generate a smooth geometric boundary. A semi-transparent color block (e.g., red indicates a high-risk area) is superimposed on the three-dimensional model to mark the affected range, and the timestamp and phase angle offset of the current mutation event are associated to output the spatial impact range boundary of the current mutation feature segment.

[0105] A commercial complex suffered local structural deformation due to an electrical short circuit. The model needed to be dynamically updated and the affected area displayed. A total of 5,000 vertices from the 10-story load-bearing walls were extracted from the BIM model. A mapping table containing ID, coordinates, and type (Type_A) was constructed. The model update command required that vertices from ID_1001 to ID_1300 be offset by +0.3m in the X direction. After the update, the positions of 1,200 vertices in the coordinate set changed. 800 vertices with an offset greater than 0.2m were selected, and a 3D convex hull was calculated to generate a boundary outline containing 150 vertices. The rendering covered an area of 120 square meters.

[0106] In summary, this solution, through dynamic vertex offsetting and convex hull boundary generation, enables efficient, coordinated updates of structural deformation and the impact range of electrical disturbances in 3D city models. In a commercial complex scenario, model update response time is reduced to less than 5 seconds, and impact range visualization accuracy reaches 95%, providing real-time, intuitive, multi-dimensional data support for safety assessments and emergency decision-making.

[0107] Figure 2 The present invention provides a schematic diagram of a real-time three-dimensional city modeling and dynamic updating device based on deep learning, as shown in FIG. Figure 2 As shown, the device includes: An acquisition module 21 acquires the static load parameters of the building structure of the target building, wherein the static load parameters of the building structure include the cross-sectional dimensions of beams and columns, the elastic modulus of the materials, and the historical data of foundation settlement; Deployment module 22, deploying a microseismic sensor array at the nodes between the load-bearing walls and frame beams of the target building, wherein the microseismic sensor array collects stress wave propagation path data generated inside the building structure at a preset sampling frequency; The receiving module 23 synchronously receives power load data uploaded by a smart meter connected to the power distribution system of the target building, wherein the power load data includes hourly current fluctuation characteristics and three-phase imbalance index; An extraction module 24 extracts a current mutation feature segment within a corresponding time window in the power load data according to a start timestamp of the abnormal vibration event in the stress wave propagation path data; A generation module 25 inputs the static load parameters of the building structure, the frequency domain energy distribution of the stress wave corresponding to the abnormal vibration event, and the phase angle offset of the current mutation characteristic segment into a pre-built deep learning model, so as to generate a three-dimensional model update instruction by coupling the association rules between the structural mechanical response and the electrical load disturbance through the deep learning model; The overlay module 26 dynamically adjusts the grid vertex coordinates associated with the load-bearing walls and frame beam-column nodes of the target building in the initial three-dimensional city model based on the three-dimensional model update instruction, and overlays and displays the spatial influence range boundary of the current mutation feature segment.

[0108] Figure 2 The real-time 3D city modeling and dynamic updating device based on deep learning can perform Figure 1 The implementation principles and technical effects of the deep learning-based real-time 3D city modeling and dynamic updating method described in the illustrated embodiment will not be elaborated upon. The specific manner in which the various modules and units of the deep learning-based real-time 3D city modeling and dynamic updating device in the aforementioned embodiment perform their operations has been described in detail in the relevant embodiments of the method and will not be elaborated upon here.

[0109] In one possible design, Figure 2 The real-time 3D city modeling and dynamic updating device based on deep learning of the embodiment shown can be implemented as a computing device, such as Figure 3 As shown, the computing device may include a storage component 31 and a processing component 32; The storage component 31 stores one or more computer instructions, wherein the one or more computer instructions are called and executed by the processing component 32 .

[0110] The processing component 32 is used for the above Figure 1 The embodiment provides a real-time three-dimensional city modeling and dynamic updating method based on deep learning.

[0111] The processing component 32 may include one or more processors to execute computer instructions to complete all or part of the steps in the above method. Of course, the processing component may also be implemented as one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to perform the above method.

[0112] The storage component 31 is configured to store various types of data to support operations at the terminal. The storage component can be implemented by any type of volatile or non-volatile memory device, or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk, or optical disk.

[0113] Of course, a computing device may also include other components, such as input / output interfaces, display components, communication components, etc.

[0114] The input / output interface provides an interface between the processing component and the peripheral interface module, which can be an output device, an input device, etc.

[0115] The communication component is configured to facilitate, among other things, wired or wireless communications between the computing device and other devices.

[0116] Among them, the computing device can be a physical device or an elastic computing host provided by a cloud computing platform, etc. In this case, the computing device can refer to a cloud server, and the above-mentioned processing components, storage components, etc. can be basic server resources rented or purchased from the cloud computing platform.

[0117] The present application also provides a computer storage medium storing a computer program, wherein the computer program can achieve the above-mentioned Figure 1 The illustrated embodiment is a real-time three-dimensional city modeling and dynamic updating method based on deep learning.

[0118] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0119] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, i.e., they may be located in one location or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of the present embodiment. Persons of ordinary skill in the art will be able to understand and implement the present invention without inventive effort.

[0120] Through the description of the above embodiments, those skilled in the art will clearly understand that each embodiment can be implemented using software plus a necessary general-purpose hardware platform, or of course, hardware. Based on this understanding, the essence of the above technical solution, or the portion that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, or an optical disk, and includes a number of instructions for causing a computer device (such as a personal computer, server, or network device) to execute the methods described in each embodiment or certain portions of the embodiments.

[0121] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A real-time 3D city modeling and dynamic updating method based on deep learning, characterized in that: include: Obtain the static load parameters of the building structure of the target building; Deploy a microseismic sensor array at the nodes between the load-bearing walls and frame beams of the target building, wherein the microseismic sensor array collects data on the propagation path of stress waves generated inside the building structure at a preset sampling frequency; Synchronously receiving power load data uploaded by a smart meter connected to a power distribution system of the target building; Extracting a current mutation feature segment within a corresponding time window in the power load data according to a start timestamp of the abnormal vibration event in the stress wave propagation path data; Inputting the static load parameters of the building structure, the frequency-domain energy distribution of the stress wave corresponding to the abnormal vibration event, and the phase angle offset of the current mutation characteristic segment into a pre-built deep learning model, so as to generate a model update instruction by coupling the association rules of the mechanical response characteristics in the stress wave propagation path data and the load disturbance characteristics in the power load data through the deep learning model; Based on the model update instruction, the grid vertex coordinates associated with the load-bearing walls and frame beam-column nodes of the target building are dynamically adjusted in the initial three-dimensional city model, and the spatial influence range boundary of the current mutation feature segment is superimposed and displayed.

2. The method according to claim 1, characterized in that Inputting the static load parameters of the building structure, the frequency domain energy distribution of the stress wave corresponding to the abnormal vibration event, and the phase angle offset of the current mutation characteristic segment into a pre-built deep learning model, and generating a model update instruction by coupling the association rules of the mechanical response characteristics in the stress wave propagation path data and the load disturbance characteristics in the power load data through the deep learning model, including: Perform equivalent load mapping on the input static load parameters of the building structure, the stress wave frequency domain energy distribution corresponding to the abnormal vibration event, and the phase angle offset of the current mutation characteristic segment, generate an energy density vector and a phase angle offset sequence, and output a static load parameter matrix; Based on the static load parameter matrix, the stress wave energy density vector and the phase angle offset sequence are time-frequency aligned, and the structural mechanical field distribution diagram and the time-frequency joint characteristic matrix are output based on the time-frequency alignment results; Convert the structural mechanics field distribution map into a high-dimensional gridded tensor, superimpose the energy information in the time-frequency joint feature matrix, establish a mapping relationship between the three-dimensional space grid and the frequency domain energy, and output a multi-dimensional feature tensor; The multidimensional feature tensor is input into a pre-built deep learning model, and the association rules between the mechanical response characteristics in the stress wave propagation path data and the load disturbance characteristics in the power load data are extracted through an adaptive convolution kernel and coupled with the deep learning model, and a model update instruction is generated based on the coupling result.

3. The method according to claim 2, characterized in that Inputting the multidimensional feature tensor into a pre-built deep learning model, extracting association rules between the mechanical response features in the stress wave propagation path data and the load disturbance features in the power load data through an adaptive convolution kernel, coupling them with the deep learning model, and generating a model update instruction based on the coupling result, including: Input a multidimensional feature tensor, use the sliding window segmentation method to segment the continuous time series signal, and output a standardized multidimensional spatiotemporal feature tensor based on the segmentation results; Based on the multi-dimensional spatiotemporal feature tensor, a multi-dimensional spatiotemporal channel attention mechanism is constructed to quantify the coupling relationship between the mechanical response characteristics and the load disturbance characteristics, and an association rule matrix is generated based on the coupling relationship; The association rule matrix is coupled with the weight matrix of the deep learning model, and the bidirectional coupling between the rules and the model is achieved through a layer-by-layer parameter fusion mechanism, and the coupled joint parameter tensor is output; According to the characteristic distribution of the joint parameter tensor, residual connection and nonlinear activation function are used to spatially decouple the deformation instructions, and model update instructions are generated based on the spatial decoupling results.

4. The method according to claim 3, characterized in that The method constructs an inter-channel attention mechanism based on the multi-dimensional spatiotemporal feature tensor to quantify the coupling relationship between the mechanical response characteristics and the load disturbance characteristics, and generates an association rule matrix based on the coupling relationship, including: Based on the multi-dimensional spatiotemporal feature tensor, an inter-channel attention mechanism is constructed, and an inter-channel attention weight matrix is generated by calculating the correlation weights between different channels; The coupling relationship between the mechanical response characteristics and the load disturbance characteristics is quantified using the inter-channel attention weight matrix, and the contribution of each channel to the coupling relationship is obtained to generate the quantitative results of the coupling relationship. Based on the quantitative results of the coupling relationship, an association rule matrix is constructed, and the association rule matrix is generated by describing the association strength between the mechanical response characteristics and the load disturbance characteristics.

5. The method according to claim 1, wherein Extracting a current mutation feature segment within a corresponding time window in the power load data according to a start timestamp of the abnormal vibration event in the stress wave propagation path data includes: According to the start timestamp of the abnormal vibration event in the stress wave propagation path data, the transmission delay between the start timestamp of the abnormal vibration event and the current signal receiving end is calculated, and a current signal trigger timestamp sequence is generated based on the transmission delay; Taking the current signal trigger timestamp sequence as the reference point, the forward extension window length and the backward retention window length are dynamically calculated, and the current mutation feature segment containing the complete vibration feature is intercepted before and after the reference point based on the window length.

6. The method according to claim 5, characterized in that After intercepting the current mutation feature segment containing the complete vibration feature before and after the reference point based on the window length, the method further includes: Multi-resolution time-frequency decomposition is performed on the intercepted current mutation feature segment to separate the narrowband high-frequency transient component feature subset that is strongly correlated with the abnormal vibration event; A sliding time window is used to perform time-warping matching on a narrowband high-frequency transient component feature subset, and a current mutation feature segment within a corresponding time window in the power load data is screened out based on the time-warping matching result.

7. The method according to claim 1, characterized in that Based on the model update instruction, the mesh vertex coordinates associated with the load-bearing walls and frame beam-column nodes of the target building are dynamically adjusted in the initial three-dimensional city model, and the spatial influence range boundary of the current mutation characteristic segment is superimposed and displayed, including: Based on the model update instruction, extract the geometric topological relationship between the load-bearing walls and the frame beam-column nodes in the initial three-dimensional city model, traverse the mesh vertex data of the geometric topological relationship and establish a vertex index table, and generate a vertex mapping relationship table based on the vertex index table; Based on the vertex mapping table of the target building structure, the dynamic offset vector is extracted and superimposed point by point with the local coordinate values of the vertices in the vertex mapping table, the vertex coordinate values are updated, and an updated vertex coordinate set is output based on the vertex coordinate values; According to the current mutation feature segment, a vertex subset affected by the offset is extracted from the updated vertex coordinate set, a geometric outline of the spatial influence range boundary is generated by a convex hull algorithm, and the spatial influence range boundary of the current mutation feature segment is output based on the geometric outline.

8. A real-time 3D city modeling and dynamic updating system based on deep learning, characterized by: include: An acquisition module is used to obtain the static load parameters of the building structure of the target building; A deployment module deploys a microseismic sensor array at the nodes between the load-bearing walls and frame beams of the target building, wherein the microseismic sensor array collects data on the propagation path of stress waves generated inside the building structure at a preset sampling frequency; A receiving module, synchronously receiving power load data uploaded by a smart meter connected to the power distribution system of the target building; an extraction module, which extracts a current mutation characteristic segment within a corresponding time window in the power load data according to a start timestamp of the abnormal vibration event in the stress wave propagation path data; A generation module inputs the static load parameters of the building structure, the stress wave frequency domain energy distribution corresponding to the abnormal vibration event, and the phase angle offset of the current mutation characteristic segment into a pre-built deep learning model, so as to generate a three-dimensional model update instruction by coupling the association rules between the structural mechanical response and the electrical load disturbance through the deep learning model; The overlay module dynamically adjusts the grid vertex coordinates associated with the load-bearing walls and frame beam-column nodes of the target building in the initial three-dimensional city model based on the three-dimensional model update instruction, and overlays and displays the spatial influence range boundary of the current mutation feature segment.

9. A computing device, characterized in that It includes a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are used to be called and executed by the processing component to implement the real-time three-dimensional city modeling and dynamic updating method based on deep learning as described in any one of claims 1 to 7.

10. A computer storage medium, characterized in that A computer program is stored, and when the computer program is executed by a computer, the real-time three-dimensional city modeling and dynamic updating method based on deep learning as described in any one of claims 1 to 7 is implemented.

Images