Artificial intelligence-based power equipment three-dimensional model automatic generation and early warning method

Abstract

This invention discloses an artificial intelligence-based method for automatic generation and early warning of 3D models of power equipment, relating to the field of power equipment condition monitoring technology. Specifically, it includes: acquiring multimodal time-series data of the target power equipment; constructing a physical information-driven time-series 3D generation network; generating a sequence of time-series 3D models with physical attribute labels; introducing the heat conduction equation, elasticity equation, and electric field distribution equation as residual terms into the loss function through a coupled thermo-mechanical-electric field physical information neural network, ensuring the 3D model satisfies physical coupling relationships; extracting geometric deformation sequences and physical field change sequences as dynamic micro-variation features based on the time-series 3D model sequence; inputting the dynamic micro-variation features into an early warning analysis network to output early warning results; and achieving deep coupling between the early warning network and the generation network through a cross-task gradient sharing mechanism, eliminating temporal disconnect and realizing high-fidelity generation and accurate early warning of 3D models and physical fields of power equipment.

Description

Technical Field

[0001] This invention belongs to the field of power equipment condition monitoring technology, specifically relating to an artificial intelligence-based method for automatically generating and providing early warning of three-dimensional models of power equipment. Background Technology

[0002] As a core component of the power grid, the accurate perception and timely early warning of the operating status of power equipment are of great significance to ensuring the safe and stable operation of the power grid. Traditional power equipment condition monitoring mainly relies on manual inspections and discrete data collected by fixed sensors, which has problems such as monitoring blind spots, data silos, and lack of spatial correlation, making it difficult to achieve comprehensive perception and accurate assessment of equipment status.

[0003] In recent years, with the development of 3D scanning technology and computer vision, some studies have begun to attempt to reconstruct 3D geometric models of power equipment using lidar or visible light images for equipment deformation monitoring and defect identification. However, existing 3D reconstruction methods mainly focus on the appearance restoration of the geometric structure, ignoring the distribution and coupling effects of multiple physical fields such as temperature field, stress field, and electric field during equipment operation. This results in the generated 3D models lacking physical property information, making it difficult to support in-depth fault mechanism analysis.

[0004] In terms of equipment early warning, existing technologies mostly rely on single-source sensor data for threshold judgment or trend prediction, failing to effectively integrate the spatiotemporal correlation information of 3D geometric deformation and multi-physics field changes. Furthermore, 3D reconstruction and early warning tasks are typically designed as independent modules, lacking deep coupling at the feature level. This results in the generated 3D model failing to prioritize the retention of key features sensitive to the early warning task, leading to temporal disconnect and information loss.

[0005] Therefore, there is an urgent need for an artificial intelligence-based method for automatically generating and providing early warning of 3D models of power equipment to solve the above problems. Summary of the Invention

[0006] The purpose of this invention is to provide an automatic generation and early warning method for 3D models of power equipment based on artificial intelligence, which solves the technical problems in the prior art such as the lack of physical attribute information in 3D models, the failure to integrate the spatiotemporal correlation between geometry and physical field in early warning analysis, and the temporal disconnect between 3D reconstruction and early warning tasks.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: An AI-based method for automatic generation and early warning of 3D models of power equipment includes: Step 1: Acquire multimodal time-series data of the target power equipment. The multimodal time-series data includes lidar point cloud data, visible light image data, and sensor data reflecting the physical field distribution, which are continuously collected at a preset sampling frequency. Step 2: Construct a physical information-driven temporal 3D generation network. Input multimodal temporal data into the temporal 3D generation network to generate a sequence of temporal 3D models with physical attribute labels. A physical information neural network coupled with thermal, mechanical, and electric fields is constructed. The heat conduction equation, elasticity equation, and electric field distribution equation are introduced as coupling residual terms into the loss function of the temporal 3D generation network, so that the temperature field distribution, stress field distribution, and geometric structure in the generated 3D model satisfy the preset physical coupling relationship. Step 3: Based on the time-series three-dimensional model sequence, extract dynamic micro-variation features, which include geometric deformation sequence and physical field change sequence; Step 4: Input the dynamic micro-variation features into the early warning analysis network and output the early warning results. The early warning analysis network and the temporal 3D generation network are deeply coupled through a cross-task gradient sharing mechanism. The cross-task gradient sharing mechanism includes: during the backpropagation process of the early warning analysis network, the feature learning gradient of the early warning task is used to update the low-level parameters related to the extraction of dynamic micro-variation features in the temporal 3D generation network. This allows the temporal 3D generation network to prioritize the retention of key geometric features and key physical field features that are sensitive to subsequent early warning tasks when generating 3D models, thus eliminating the temporal disconnect between 3D generation and early warning analysis.

[0008] Furthermore, the multimodal time-series data of the target power equipment is obtained, specifically through the following method: The system collects three-dimensional geometric data of the target power equipment, records the geospatial coordinates through differential GPS, and simultaneously collects physical field sensing data such as temperature field and corona discharge intensity. The three-dimensional geometric data and physical field data are voxelized and mapped to generate a multimodal data cube with geometric-physical association. The data is continuously collected at a preset sampling frequency to form a time series, and missing or noisy data is interpolated and filtered.

[0009] Furthermore, during the generation of the temporal 3D model sequence with physical attribute labels, the following constraint operations are performed: Based on the standard design parameters of power equipment, the spatial distance between charged and non-charged bodies in the generated model is dynamically calculated. The difference between this distance and the preset safety threshold is used as the first constraint loss term and backpropagated to the temporal three-dimensional generation network to force the spatial distance between the two to be no less than the preset safety threshold.

[0010] Furthermore, a physical information neural network coupling heat, force, and electric fields is constructed, specifically using the following method: A multi-output physical information neural network architecture is constructed, using spatiotemporal coordinates as input and temperature field distribution, displacement field distribution, and electric potential field distribution as output. In the loss function of the multi-output physical information neural network, residual terms of the heat conduction equation, elasticity equation, and electric field distribution equation are constructed respectively. Then, coupled residual terms including thermo-mechanical coupling terms and thermo-electric coupling terms are constructed. After weighted summation, a joint loss function is formed. The network parameters of the multi-output physical information neural network are optimized through backpropagation.

[0011] Furthermore, a coupled residual term including thermo-mechanical coupling terms and thermo-electric coupling terms is constructed, specifically as follows: Based on the constitutive relationship between the temperature field and the displacement field, the deviation between the thermal strain tensor and the displacement field gradient is calculated as a thermo-mechanical coupling residual term. A temperature-dependent conductivity function is constructed, and the difference between the actual electric potential field and the theoretical electric potential field based on this function is taken as the first thermo-electric coupling residual term. The deviation between the Joule thermal power density and the heat source term of the heat conduction equation is taken as the second thermo-electric coupling residual term. The above three residual terms are weighted and combined to form the overall coupling residual term.

[0012] Furthermore, based on the temporal 3D model sequence, dynamic micro-variation features are extracted. The specific method is as follows: Non-rigid registration is performed on adjacent time-series 3D model models to calculate the displacement vector field of each spatial point on the surface. Spatiotemporal filtering separates structural deformation and random disturbance components, and the geometric deformation sequence is formed by arranging them in time and calculating the deformation rate and cumulative deformation. The temperature field, stress field, and electric potential field distribution at each time are extracted, the difference field between adjacent time moments is calculated, and spatial clustering is used to identify abnormal regions where the rate of change of the physical field exceeds the limit. The physical field change sequence is formed by arranging them in time. The two types of sequences are spatiotemporally aligned to construct a geometric-physical correlation matrix, and the correlation features and time-series lead-lag relationship between the two are extracted as dynamic micro-variation features output.

[0013] Furthermore, the dynamic micro-change characteristics are input into the early warning analysis network, and the early warning results are output. The specific method is as follows: A warning analysis network with three branches—geometric deformation, physical field change, and spatiotemporal correlation—is constructed: the geometric deformation branch takes the geometric deformation sequence as input, extracts the temporal features of deformation rate and cumulative deformation, compares them with preset thresholds, generates warnings and marks the locations when the limits are exceeded; the physical field change branch takes the physical field change sequence as input, predicts the distribution of each physical field in time series, compares it with the corresponding safety thresholds, generates warnings and marks the locations when the limits are exceeded; the spatiotemporal correlation branch takes the spatiotemporal correlation features of the two branches as input and conducts correlation warning analysis.

[0014] Furthermore, the spatiotemporal correlation branch inputs the spatiotemporal correlation features of the two to conduct correlation early warning analysis. The specific method is as follows: By extracting temporal causal patterns of geometric deformation and physical field changes through temporal convolutional networks, and generating cooperative evolution early warning signals when a preset abnormal causal pattern is identified; The comprehensive warning level is determined by weighted fusion of geometric deformation warning signals, physical field anomaly warning signals, and co-evolution warning signals based on confidence level and spatiotemporal overlap. When the comprehensive warning level exceeds a preset threshold, the warning result, which includes warning type, location, timestamp, and level, is output.

[0015] Furthermore, the cross-task gradient sharing mechanism specifically includes: In the backpropagation phase of the early warning analysis network, the loss gradient of the early warning task with respect to the dynamic micro-variable features is calculated, and this gradient is projected onto the feature extraction layer of the temporal 3D generative network. The bottom convolutional layer and spatiotemporal feature fusion layer responsible for geometric structure feature extraction and physical field feature extraction in the temporal 3D generative network are selected as the parameter layers to be updated; The loss gradient of the early warning task is propagated backward along the feature propagation path to the bottom parameter layer, and the bottom parameters are updated by the gradient descent algorithm.

[0016] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. This invention constructs a physical information-driven temporal 3D generation network and introduces the heat conduction equation, elasticity equation, and electric field distribution equation as coupling residual terms into the loss function. This enables high-fidelity generation of temporal 3D model sequences with physical attribute labels such as temperature field, stress field, and electric potential field. The 3D model not only contains geometric structure information but also strictly satisfies the coupling relationship of multiple physical fields such as heat, force, and electricity. This solves the problem that the existing 3D model lacks physical attributes and cannot support in-depth fault mechanism analysis. 2. This invention extracts geometric deformation sequences and physical field change sequences based on time-series three-dimensional model sequences, constructs a geometric-physical correlation matrix, and extracts correlation features and time-series lead-lag relationships, thereby achieving accurate capture of early weak signs of equipment. This solves the problem that existing early warning technologies fail to effectively integrate spatiotemporal correlation information of geometric deformation and multiple physical fields, resulting in low early warning accuracy. 3. This invention, through a cross-task gradient sharing mechanism, simultaneously updates the low-level parameters related to dynamic micro-variation feature extraction in the temporal 3D generation network with the feature learning gradient of the early warning task during the backpropagation process of the early warning analysis network. This enables the generation network to prioritize the retention of key geometric features and key physical field features that are sensitive to the early warning task during 3D modeling, achieving deep collaboration at the feature level between 3D generation and early warning analysis, eliminating the temporal disconnect between the two, and significantly improving the overall performance of equipment status monitoring and early warning. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 The figure shows the steps of the method for automatic generation and early warning of three-dimensional models of power equipment based on artificial intelligence according to the present invention; Figure 2 The flowchart illustrating the construction of the temporal three-dimensional generative network and the thermo-mechanical-electric field coupling constraint of the present invention is shown. Figure 3 The flowchart of the deep coupling of temporal 3D generation and early warning analysis based on gradient sharing in this invention is shown. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Example 1, such as Figure 1 The method for automatic generation and early warning of 3D models of power equipment based on artificial intelligence, as shown, specifically includes the following steps: Step 1: Acquire multimodal time-series data of the target power equipment. The multimodal time-series data includes lidar point cloud data, visible light image data, and sensor data reflecting the physical field distribution, which are continuously collected at a preset sampling frequency. Multiple sensors are deployed for fusion data acquisition, including a lidar scanner, a high-resolution visible light camera, a differential GPS receiver, and a physical field sensor array. The physical field sensor array includes at least a distributed temperature sensor and a corona discharge intensity sensor. Depending on the type of equipment and monitoring requirements, electric field sensors, strain sensors, etc., may also be added.

[0021] During the data acquisition process, a lidar scanner is used to collect three-dimensional geometric data of the target power equipment, obtaining dense point cloud information of the equipment surface and key components. Simultaneously, a differential GPS receiver is used to record the geospatial coordinates of each spatial point in the point cloud data, achieving precise registration between the three-dimensional geometric data and the absolute spatial position. At the same time, a physical field sensor array is used to simultaneously collect temperature field distribution data and corona discharge intensity data on the equipment surface, ensuring strict temporal and spatial alignment between the geometric data and the physical field data.

[0022] After data acquisition, the 3D geometric data (including point cloud and GPS coordinates) and physical field sensing data (temperature field, corona discharge intensity, etc.) are voxelized and mapped. Specifically, the space where the target power equipment is located is divided into a regular 3D voxel grid at a preset resolution (e.g., 1cm×1cm×1cm). Each voxel unit serves as a data storage unit, storing multi-source information about that spatial location. Specifically, each voxel unit contains the following information: Spatial coordinate information: The three-dimensional coordinates (x, y, z) of the voxel center point are derived from the spatial position of the lidar point cloud data after differential GPS positioning; Geometric information: Whether the voxel belongs to the device body, surface normal vector, point cloud density, and other geometric features; Physical field information: the temperature value, corona discharge intensity value, and other synchronously acquired physical field sensing data (such as electric potential, stress σ, etc.) corresponding to this voxel. Timestamp information: The data collection time t corresponding to this data cube.

[0023] The above information forms a structured representation in the voxel grid. Each voxel unit is equivalent to a multidimensional feature vector containing geometric and physical properties. All voxel units together constitute a complete three-dimensional data cube, namely a multimodal data cube.

[0024] The above acquisition and voxel mapping process is repeated according to the preset sampling frequency (set according to the time-varying characteristics of the target power equipment's state changes and monitoring requirements, such as once every 30 seconds or adaptively adjusted according to the equipment's operating status). Multimodal data cubes at multiple time points are continuously acquired to construct a time series. For missing data in the time series due to sensor failure, communication interruption, etc., linear interpolation or spatiotemporal interpolation methods based on adjacent time point data are used to complete the data. For noisy data, Gaussian filtering or median filtering is used for smoothing to filter out random interference during the acquisition process.

[0025] Finally, the multimodal time-series data, after interpolation and filtering, is output as input to the subsequent time-series 3D generation network, providing a high-quality data foundation for constructing time-series 3D model sequences with physical attribute labels.

[0026] Step 2: Construct a physical information-driven temporal 3D generation network. Input multimodal temporal data into the temporal 3D generation network to generate a sequence of temporal 3D models with physical attribute labels. A physical information neural network coupled with thermal, mechanical, and electric fields is constructed. The heat conduction equation, elasticity equation, and electric field distribution equation are introduced as coupling residual terms into the loss function of the temporal 3D generation network, so that the temperature field distribution, stress field distribution, and geometric structure in the generated 3D model satisfy the preset physical coupling relationship. like Figure 2 As shown, a physically-driven temporal 3D generation network is constructed. This network employs an encoder-decoder architecture. The encoder consists of a cascaded 3D convolutional neural network and a long short-term memory network, used to extract spatial geometric features and temporal evolution features from multimodal temporal data. The decoder uses a 3D deconvolutional network to progressively recover the 3D geometric structure and simultaneously output physical attribute labels for each voxel. These labels include temperature values, stress components, and potential values, forming a 3D model with physical attribute labels. The network input is the multimodal temporal data cube sequence generated in step one, and the output is the temporal 3D model sequence at the corresponding time point.

[0027] To ensure that the temperature field distribution and stress field distribution (stress is calculated from displacement through constitutive relations) in the generated 3D model satisfy a preset physical coupling relationship with the geometric structure, a physical information neural network coupling the thermal, mechanical, and electric fields is constructed. The specific method is as follows: Construct a multi-output physical information neural network architecture, using spatiotemporal coordinates The temperature field distribution, displacement field distribution, and electric potential field distribution are used as inputs and as outputs. The following residual terms are constructed in the loss function of the multi-output physical information neural network: The residual terms of the heat conduction equation are constructed based on Fourier's law of heat conduction. Core parameters such as thermal conductivity and specific heat capacity are obtained from equipment design / material manuals, and fitting corrections are made for temperature-sensitive materials. The heat source term integrates Joule heat, equipment loss heat, and environmental heat transfer additional terms, which are obtained from electric field calculation results, power grid monitoring system, and field sensor data, respectively. The boundary conditions are mainly based on convective heat transfer boundaries, combined with adiabatic and isothermal boundaries. The parameters refer to engineering specifications and are corrected according to wind speed and field temperature. The residual terms of the elasticity equations are constructed based on the governing equations of elasticity (Navier-Cauchy equations). The dielectric constant and conductivity are obtained from material handbooks / industry standards. The conductivity is fitted and corrected according to temperature. The parameters of the corona discharge region are calibrated separately. The free charge density is set to 0 during normal operation. The discharge region is calibrated by corona sensing data. The boundary conditions combine constant potential (charged body, grounded component), natural boundary (insulating component), and surface current boundary (discharge region). The potential values ​​are obtained from the power grid voltage monitoring system. The residual terms of the electric field distribution equation are constructed based on the Poisson equation. Constitutive parameters such as elastic modulus, Poisson's ratio, and coefficient of thermal expansion are obtained from national standards / design drawings. Corrections are made for temperature-sensitive materials, and contact stiffness coefficients are introduced at connection points. Loads are divided into static loads (equipment self-weight, fastening force) and dynamic loads (wind loads, vibration loads), which are calculated and obtained from equipment nameplates, design documents, and on-site sensor data, respectively. Constraints are set according to the equipment installation method, including full displacement, rotational free linear displacement constraints, flexible constraints, elastic constraints, etc., and some constraint values ​​are calibrated from lidar point cloud data. Spatiotemporal coordinate samples are extracted from the multimodal time-series data cube sequence. Based on the three residual terms mentioned above, coupled residual terms, including thermal-mechanical coupling terms and thermal-electric coupling terms, are further constructed. The specific construction method is as follows: Based on the constitutive relationship between the temperature field and the displacement field, the deviation between the thermal strain tensor and the displacement field gradient is calculated as a thermo-mechanical coupling residual term. Specifically, based on thermoelastic theory, the thermal strain tensor is expressed as the product of the coefficient of thermal expansion, the temperature change, and the unit tensor. The specific value of the coefficient of thermal expansion is preset according to the equipment design drawings or material standards. The temperature change represents the change of the temperature field relative to the reference temperature at the current moment (the reference temperature is usually the initial operating temperature of the equipment, such as an ambient temperature of 20℃ or the reference temperature under normal operating conditions). The unit tensor refers to the second-order unit tensor. The temperature change of the scalar is mapped to the form of a second-order tensor using the unit tensor, so that it can be directly compared with the displacement field gradient (geometric deformation tensor). The deviation between the thermal strain tensor and the displacement field gradient is calculated to form a thermo-mechanical coupling residual term. A temperature-dependent conductivity function is constructed, and the difference between the actual electric potential field and the theoretical electric potential field based on this function is taken as the first thermo-electric coupling residual term. The deviation between the Joule heat power density and the heat source term in the heat conduction equation is taken as the second thermo-electric coupling residual term. Specifically, the first thermo-electric coupling residual term is constructed based on the current continuity equation, and the second thermo-electric coupling residual term is constructed based on the deviation between the Joule heat and the heat source term in the heat conduction equation.

[0028] The residual terms of the heat conduction equation, elasticity equation, electric field distribution equation, thermo-mechanical coupling residual term, first thermo-electric coupling residual term, and second thermo-electric coupling residual term are weighted and combined to form a whole coupling residual term. Then, a joint loss function is constructed, and the network parameters of the multi-output physical information neural network are optimized through backpropagation.

[0029] The joint loss function is used as a physical constraint term in the temporal 3D generation network and introduced into the total loss function of the temporal 3D generation network, so that the temperature field distribution, stress field distribution and geometric structure in the generated 3D model strictly satisfy the preset thermo-mechanical-electric coupling physical relationship.

[0030] In the process of generating a time-series 3D model sequence with physical attribute labels, to further enhance the safety compliance of the generated model, the following constraint operations are performed based on the standard design parameters of power equipment: First, charged and uncharged regions are identified from the generated temporal 3D model. Charged regions are determined based on the electric potential field distribution in the model and the preset conductive material properties, while uncharged regions include insulating components, grounding components, and structural support components. The minimum spatial distance between the surfaces of charged and uncharged components in the generated model is dynamically calculated. .

[0031] The calculated minimum spatial distance With preset safety threshold (Based on the minimum electrical clearance requirements between live and non-live parts specified in the power equipment design standards) a comparison is made to construct the first constraint loss term. This loss term represents the penalty for insufficient safe distance between charged and uncharged objects. This constraint loss term is backpropagated to the underlying parameters of the temporal 3D generation network. Network optimization forces the spatial distance between the two objects to be no less than a preset safety threshold, ensuring that the models at each time step in the generated temporal 3D model sequence meet the electrical safety distance requirements.

[0032] The total loss function of the temporal 3D generative network consists of a weighted sum of the geometric reconstruction loss, the joint loss function of the physical information neural network, and the safety distance constraint loss term. By jointly optimizing the above loss terms, the network generates a physical property field that conforms to the physical laws of thermo-mechanical-electric coupling while generating the 3D geometric structure, and ensures that the model meets the safety specifications for power equipment.

[0033] Finally, the temporal 3D generation network outputs a sequence of temporal 3D models with physical attribute labels, providing model inputs containing both accurate geometric information and high-fidelity physical field information for the subsequent step three dynamic micro-variation feature extraction.

[0034] Step 3: Based on the time-series three-dimensional model sequence, extract dynamic micro-variation features, which include geometric deformation sequence and physical field change sequence; The temporal 3D model sequence with physical attribute labels generated in step two is denoted as... The three-dimensional model corresponding to each time t It also includes the geometric structure information of the power grid equipment, as well as the distribution information of physical properties such as temperature field, stress field, and electric potential field. By comparing and analyzing the models at adjacent time points, dynamic micro-variation features that can reflect the evolution of the equipment state are extracted. These dynamic micro-variation features mainly include two categories: geometric deformation sequence and physical field change sequence. The geometric deformation sequence is extracted, and the displacement vector field of each spatial point on the model surface is calculated by non-rigid registration of the model at adjacent time points in the temporal 3D model sequence. The specific implementation method is as follows: Selecting three-dimensional models at adjacent times t and t+1 and The registration operation is performed using a non-rigid iterative nearest-point algorithm. This algorithm introduces a smoothing constraint on the basis of the traditional ICP algorithm, allowing the model to undergo local non-rigid deformation. By minimizing the weighted sum of the distance error between point clouds and the smoothness of the deformation field, the algorithm obtains the result from... arrive displacement vector field ,in These are the spatial coordinates of points on the surface of the model within the temporal 3D model sequence generated in step two. The structural deformation and random disturbance components contained within are separated using a spatiotemporal filtering method: In the spatial dimension, the displacement field is smoothed by Gaussian filtering or bilateral filtering to filter out noise and random disturbance components caused by local measurement errors; in the temporal dimension, the displacement field sequence at multiple consecutive moments is low-pass filtered to further suppress high-frequency noise and retain the structural deformation component that reflects the long-term evolution trend of the equipment. The filtered displacement vector fields are arranged in chronological order to form a geometric deformation sequence. For each spatial point p, two derived features are calculated: one is the deformation rate. (Δt is the time interval between adjacent moments), and secondly, the cumulative deformation. The geometric deformation sequence and its derived features are output together and used as the geometric feature input for subsequent early warning analysis. The physical field change sequence is extracted. First, the distribution of various physical fields is extracted from the model at each time step, and the difference field between adjacent time steps is calculated. Then, spatial clustering is used to identify abnormal regions where the rate of change of the physical field exceeds the limit. The specific implementation method is as follows: From the three-dimensional model at each time t In the middle, the temperature field distribution is extracted respectively. Stress field distribution and electric potential field distribution And calculate the difference field of each physical field between adjacent times t and t+1, that is, the temperature field difference. Stress field difference electric potential difference ; After normalizing each difference field, spatial clustering analysis is performed using a density-based spatial clustering algorithm such as DBSCAN, with preset thresholds for the rate of change of each physical field. , , (Based on the statistical distribution of the rate of change of the physical field under normal operating conditions of the equipment, the mean plus 3 times the standard deviation is taken as the initial threshold, and the value is determined by inversion optimization in combination with historical failure cases.) Points in the difference field whose rate of change exceeds the corresponding threshold are marked as candidate anomaly points. Then, the spatially adjacent candidate anomaly points are aggregated into anomaly regions by clustering algorithm, and the location, range and rate of change intensity of the anomaly region are recorded. The abnormal regions identified at each time point are arranged in chronological order to form a sequence of physical field changes. In this sequence It contains key information such as the location, extent, rate of change, and intensity of the anomalous region for each physical field at time t.

[0035] The geometric deformation sequence and the physical field change sequence are spatiotemporally aligned, a geometric-physical correlation matrix is ​​constructed, and finally, the correlation features and temporal lead-lag relationships between the two are extracted. The specific implementation method is as follows: The geometric deformation sequence and the physical field change sequence are unified under the same spatiotemporal coordinate system. For the geometric deformation sequence with spatial point p as the unit and the physical field change sequence with clustered anomalous regions as the unit, the geometric deformation time series data near the reference point is extracted with the geometric center of each anomalous region as the reference point to form spatiotemporally aligned geometric-physical pairing data. Construct the geometric-physical correlation matrix C(t), whose matrix elements The correlation coefficient represents the spatial relationship between the i-th geometric deformation feature (deformation rate, cumulative deformation, etc.) and the j-th physical field change feature (temperature change rate, stress change rate, etc.). This coefficient can be obtained using the Pearson correlation coefficient formula, specifically as follows: For each time t, a set of spatial sampling points is obtained, denoted as p1, p2, ..., pN, and each point corresponds to a set of geometric deformation characteristic values. and a set of physical field change eigenvalues Correlation coefficient That is, the calculation is performed based on the feature value sequence of these N spatial sampling points, and the specific calculation formula is as follows: ; in This represents the i-th geometric deformation eigenvalue of the k-th spatial point at time t. This represents the characteristic value of the change in the physical field at the k-th spatial point at time t. This represents the spatial mean of the i-th geometric deformation feature. This represents the spatial mean of the variation characteristics of the j-th physical field. This indicates that geometric deformation and physical field change are positively correlated spatially (regions with large geometric deformation also have large physical field changes), and vice versa. Indicates a negative correlation. This indicates that there is no spatial correlation.

[0036] Extracting temporal correlation features: First, correlation features, calculating the synchronous correlation coefficient between geometric deformation and physical field changes in the time dimension to reflect whether the two change in tandem at the same time; Second, temporal lead-lag relationship, using temporal cross-correlation analysis to calculate the correlation coefficient between the two sequences at different time offsets, identifying temporal patterns where physical field changes precede geometric deformation or geometric deformation precedes physical field changes, thereby determining the causes of equipment state evolution; The geometric deformation sequence, physical field change sequence, geometric-physical correlation matrix, correlation features, and time series lead-lag relationship are integrated and packaged, and output as a whole dynamic micro-variation feature to the early warning analysis network in step four.

[0037] Step 4: Input the dynamic micro-variation features into the early warning analysis network and output the early warning results. The early warning analysis network and the temporal 3D generation network are deeply coupled through a cross-task gradient sharing mechanism. The cross-task gradient sharing mechanism includes: during the backpropagation process of the early warning analysis network, the feature learning gradient of the early warning task is used to update the low-level parameters related to the extraction of dynamic micro-variation features in the temporal 3D generation network. This allows the temporal 3D generation network to prioritize the retention of key geometric features and key physical field features that are sensitive to subsequent early warning tasks when generating 3D models, thus eliminating the temporal disconnect between 3D generation and early warning analysis.

[0038] like Figure 3 As shown, dynamic micro-variation features are input into the early warning analysis network. The early warning analysis network adopts a three-branch parallel architecture, corresponding to the analysis of three types of features: geometric deformation, physical field change, and spatiotemporal correlation. After each branch performs independent calculations, a comprehensive early warning is output through fusion judgment. The specific design is as follows: Geometric deformation branch, input: geometric deformation sequence, including the deformation rate of each spatial point. and cumulative deformation Temporal convolutional networks are used to extract temporal features of deformation rate and cumulative deformation, including deformation acceleration features, deformation trend features (linear trend, exponential trend, etc.) and local fluctuation features. The extracted time-series features are compared with a preset threshold: When the deformation rate exceeds the preset rate threshold (based on the fatigue limit and creep characteristics of the equipment material, the allowable strain rate of the material is calculated by finite element simulation and multiplied by a safety factor), a deformation rate warning is generated and the location of the excess space is marked. When the cumulative deformation exceeds the preset deformation threshold (based on the equipment structural design tolerance and allowable deformation, and set with reference to the deformation limit in the technical specifications or industry standards provided by the equipment manufacturer), a cumulative deformation warning is generated and the location of the excess space is marked. Output geometric deformation early warning signal It includes the warning type (rate warning / cumulative warning), warning location coordinates, and warning timestamp.

[0039] The physical field change branch takes the following input: a sequence of physical field changes, including information on temperature, stress, and electric potential changes in various abnormal regions. A long short-term memory network is used to predict the evolution trends of various physical fields in time series, obtaining the future distribution of the physical fields. The measured values ​​of the physical fields at the current moment and the predicted values ​​for the future are compared with the corresponding safety thresholds. If the temperature field exceeds the preset temperature threshold (set according to the heat resistance level of the insulation material of the power equipment and its corresponding maximum allowable operating temperature), a temperature anomaly warning is generated. If the stress field exceeds the preset stress threshold (set based on the yield strength and allowable stress of the equipment structural materials, referring to the stress limits in the mechanical design manual and equipment design drawings, and considering the safety factor), a stress over-limit warning is generated. If the electric potential field distortion exceeds the preset distortion threshold (calculated through electromagnetic field simulation of the electric potential distribution characteristics under normal operating conditions, taking a deviation from the theoretical distribution exceeding 15%~20% as the distortion judgment threshold, and combined with the corona discharge initiation voltage calibration), an electric field distortion warning is generated. Output physical field anomaly warning signal It includes the warning type (temperature / stress / potential), warning location coordinates, warning timestamp, and intensity of abnormal change rate.

[0040] The spatiotemporal correlation branch takes as input spatiotemporal correlation features, including geometric-physical correlation matrices, correlation features, and temporal lead-lag relationships. It extracts temporal causal patterns of geometric deformation and physical field changes through a temporal convolutional network. The Granger causality test is used to identify the causal direction between the two. A temporal pattern matching method is employed to identify preset anomalous causal patterns (including thermally induced deformation patterns, stress-induced cracking patterns, electric field distortion deformation patterns, and co-deterioration patterns). Upon identification of any anomalous causal pattern, a co-evolution warning signal is generated. Record the pattern type and associated spatial location; Output: Co-evolutionary early warning signal, including the type of abnormal causal pattern and the spatial location of the associated region; Will , , The three types of early warning signals are fused, and the comprehensive early warning level is determined by weighting confidence level and spatiotemporal overlap. Confidence level weights are assigned to each early warning signal based on its historical accuracy and current saliency. , i represents the i-th class, such as For i=1, calculate the overlap 'o' of different warning signals in spatial location and time window. The higher the overlap, the higher the comprehensive warning level. The overall warning level is ,in Basic warning level, This is the overlap adjustment coefficient. The specific values ​​need to be dynamically adjusted based on the type of power equipment (such as transformers, circuit breakers, transmission lines, etc.), operating conditions, and safety regulations. The preset hyperparameters are set based on the operating characteristics of the power equipment and historical fault data; when When the preset comprehensive warning threshold is exceeded (based on historical operation data and fault cases, the critical value that achieves the optimal balance between warning accuracy and false alarm rate is selected as the final threshold through ROC curve analysis), the final warning result is output, including warning type, warning location, warning timestamp, and warning level (e.g., level 1-5, with higher levels indicating greater urgency).

[0041] During the backpropagation phase of the early warning analysis network, the loss gradient of the early warning task with respect to the dynamically changing characteristics is calculated. , For early warning network loss function, To address the dynamic, micro-variable features, the gradient is backpropagated to the feature extraction layer of the temporal 3D generative network. The gradient is then projected onto the orthogonal complement space of the gradient space of the temporal 3D generative network using a gradient projection method, thus avoiding gradient conflicts that could lead to training instability.

[0042] The underlying parameter layers in the temporal 3D generative network related to dynamic micro-variation feature extraction are selected as the parameter layers to be updated, including: Geometric structure feature extraction layer (the first three layers of a 3D convolutional network, etc.); Physical field feature extraction layer (bottom convolutional layer that extracts features of temperature field, stress field, and electric potential field). Spatiotemporal feature fusion layer (spatiotemporal attention layer or feature splicing layer).

[0043] The shared underlying parameters are updated simultaneously using the gradient descent algorithm, and the update formula is as follows: ; in To share underlying parameters, For learning rate, The total loss of the temporal 3D generative network (including geometric reconstruction loss, physical constraint loss, and safety distance constraint loss) is denoted as . The gradient fusion weight coefficient is selected through comparative experiments to achieve the best overall performance of the generation task and the early warning task on the validation set. The value is usually in the range of 0.1 to 0.5. It is used to balance the supervision intensity of the generation task and the early warning task. This mechanism enables the temporal 3D generation network to prioritize the retention of key features that are sensitive to the early warning task (such as features of weak areas of equipment and features of easily abnormal connection parts), so as to achieve deep collaboration at the feature level between the generation and early warning tasks.

[0044] The comprehensive early warning level determination result and the early warning signals from each branch are packaged in a preset format and output to the user interface or monitoring system, which can trigger the following subsequent operations: Highlight abnormal areas on the 3D model; Push early warning notifications to the terminals of maintenance personnel; Automatically record warning events to the equipment health management database; Trigger high-precision acquisition mode to perform encrypted monitoring of abnormal areas. This completes the full-process implementation of the AI-based method for automatically generating and providing early warning of 3D models of power equipment.

[0045] The above formulas are all dimensionless calculations, and the preset parameters in the formulas should be set by those skilled in the art according to the actual situation.

[0046] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0047] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for automatic generation and early warning of 3D models of power equipment based on artificial intelligence, characterized in that, include: Step 1: Acquire multimodal time-series data of the target power equipment. The multimodal time-series data includes lidar point cloud data, visible light image data, and sensor data reflecting the physical field distribution, which are continuously collected at a preset sampling frequency. Step 2: Construct a physical information-driven temporal 3D generation network. Input multimodal temporal data into the temporal 3D generation network to generate a sequence of temporal 3D models with physical attribute labels. A physical information neural network coupled with thermal, mechanical, and electric fields is constructed. The heat conduction equation, elasticity equation, and electric field distribution equation are introduced as coupling residual terms into the loss function of the temporal 3D generation network, so that the temperature field distribution, stress field distribution, and geometric structure in the generated 3D model satisfy the preset physical coupling relationship. Step 3: Based on the time-series three-dimensional model sequence, extract dynamic micro-variation features, which include geometric deformation sequence and physical field change sequence; Step 4: Input the dynamic micro-variation features into the early warning analysis network and output the early warning results. The early warning analysis network and the temporal 3D generation network are deeply coupled through a cross-task gradient sharing mechanism. The cross-task gradient sharing mechanism includes: during the backpropagation process of the early warning analysis network, the feature learning gradient of the early warning task is used to update the low-level parameters related to the extraction of dynamic micro-variation features in the temporal 3D generation network. This allows the temporal 3D generation network to prioritize the retention of key geometric features and key physical field features that are sensitive to subsequent early warning tasks when generating 3D models, thus eliminating the temporal disconnect between 3D generation and early warning analysis.

2. The method for automatic generation and early warning of 3D models of power equipment based on artificial intelligence according to claim 1, characterized in that, The specific method for obtaining multimodal time-series data of the target power equipment is as follows: The system collects three-dimensional geometric data of the target power equipment, records the geospatial coordinates through differential GPS, and simultaneously collects physical field sensing data such as temperature field and corona discharge intensity. The three-dimensional geometric data and physical field data are voxelized and mapped to generate a multimodal data cube with geometric-physical association. The data is continuously collected at a preset sampling frequency to form a time series, and missing or noisy data is interpolated and filtered.

3. The method for automatic generation and early warning of 3D models of power equipment based on artificial intelligence according to claim 1, characterized in that, During the generation of a temporal 3D model sequence with physical attribute labels, the following constraint operations are performed: Based on the standard design parameters of power equipment, the spatial distance between charged and non-charged bodies in the generated model is dynamically calculated. The difference between this distance and the preset safety threshold is used as the first constraint loss term and backpropagated to the temporal three-dimensional generation network to force the spatial distance between the two to be no less than the preset safety threshold.

4. The method for automatic generation and early warning of three-dimensional models of power equipment based on artificial intelligence according to claim 1, characterized in that, The specific method for constructing a physical information neural network that couples thermal, mechanical, and electric fields is as follows: A multi-output physical information neural network architecture is constructed, using spatiotemporal coordinates as input and temperature field distribution, displacement field distribution, and electric potential field distribution as output. In the loss function of the multi-output physical information neural network, residual terms of the heat conduction equation, elasticity equation, and electric field distribution equation are constructed respectively. Then, coupled residual terms including thermo-mechanical coupling terms and thermo-electric coupling terms are constructed. After weighted summation, a joint loss function is formed. The network parameters of the multi-output physical information neural network are optimized through backpropagation.

5. The method for automatic generation and early warning of three-dimensional models of power equipment based on artificial intelligence according to claim 4, characterized in that, The coupling residuals, including thermo-mechanical coupling terms and thermo-electric coupling terms, are constructed using the following method: Based on the constitutive relationship between the temperature field and the displacement field, the deviation between the thermal strain tensor and the displacement field gradient is calculated as a thermo-mechanical coupling residual term. A temperature-dependent conductivity function is constructed, and the difference between the actual electric potential field and the theoretical electric potential field based on this function is taken as the first thermo-electric coupling residual term. The deviation between the Joule thermal power density and the heat source term of the heat conduction equation is taken as the second thermo-electric coupling residual term. The above three residual terms are weighted and combined to form the overall coupling residual term.

6. The method for automatic generation and early warning of 3D models of power equipment based on artificial intelligence according to claim 1, characterized in that, Based on a time-series 3D model sequence, dynamic micro-variation features are extracted. The specific method is as follows: Non-rigid registration is performed on adjacent time-series 3D model models to calculate the displacement vector field of each spatial point on the surface. Spatiotemporal filtering separates structural deformation and random disturbance components, and the geometric deformation sequence is formed by arranging them in time and calculating the deformation rate and cumulative deformation. The temperature field, stress field, and electric potential field distribution at each time are extracted, the difference field between adjacent time moments is calculated, and spatial clustering is used to identify abnormal regions where the rate of change of the physical field exceeds the limit. The physical field change sequence is formed by arranging them in time. The two types of sequences are spatiotemporally aligned to construct a geometric-physical correlation matrix, and the correlation features and time-series lead-lag relationship between the two are extracted as dynamic micro-variation features output.

7. The method for automatic generation and early warning of three-dimensional models of power equipment based on artificial intelligence according to claim 1, characterized in that, The dynamic micro-change characteristics are input into the early warning analysis network, and the early warning results are output. The specific method is as follows: A warning analysis network with three branches—geometric deformation, physical field change, and spatiotemporal correlation—is constructed: the geometric deformation branch takes the geometric deformation sequence as input, extracts the temporal features of deformation rate and cumulative deformation, compares them with preset thresholds, generates warnings and marks the locations when the limits are exceeded; the physical field change branch takes the physical field change sequence as input, predicts the distribution of each physical field in time series, compares it with the corresponding safety thresholds, generates warnings and marks the locations when the limits are exceeded; the spatiotemporal correlation branch takes the spatiotemporal correlation features of the two branches as input and conducts correlation warning analysis.

8. The method for automatic generation and early warning of three-dimensional models of power equipment based on artificial intelligence according to claim 7, characterized in that, The spatiotemporal correlation branch takes the spatiotemporal correlation features of the two as input and performs correlation early warning analysis. The specific method is as follows: By extracting temporal causal patterns of geometric deformation and physical field changes through temporal convolutional networks, and generating cooperative evolution early warning signals when a preset abnormal causal pattern is identified; The comprehensive warning level is determined by weighted fusion of geometric deformation warning signals, physical field anomaly warning signals, and co-evolution warning signals based on confidence level and spatiotemporal overlap. When the comprehensive warning level exceeds a preset threshold, the warning result, which includes warning type, location, timestamp, and level, is output.

9. The method for automatic generation and early warning of three-dimensional models of power equipment based on artificial intelligence according to claim 1, characterized in that, Cross-task gradient sharing mechanisms specifically include: In the backpropagation phase of the early warning analysis network, the loss gradient of the early warning task with respect to the dynamic micro-variable features is calculated, and this gradient is projected onto the feature extraction layer of the temporal 3D generative network. The bottom convolutional layer and spatiotemporal feature fusion layer responsible for geometric structure feature extraction and physical field feature extraction in the temporal 3D generative network are selected as the parameter layers to be updated; The loss gradient of the early warning task is propagated backward along the feature propagation path to the bottom parameter layer, and the bottom parameters are updated by the gradient descent algorithm.

Images