A method and system for monitoring the inclination state of a tower

Abstract

The application relates to the technical field of state monitoring, and discloses a tower and pole tilting state monitoring method and system. The method comprises the following steps: performing multi-field digital coupling modeling on a power tower and pole to obtain a digital twin model and calculate an electrical load-structure response sensitivity matrix; collecting a multi-modal sensing feature data set comprising tower and pole tilting angle time series data, insulator string displacement, tower body micro-vibration spectrum and foundation settlement gradient; performing time lag correlation analysis on electrical events and structure responses to obtain an electrical-structure mapping matrix; performing tensor decomposition processing through a multi-modal tilting feature fusion network to obtain a tower and pole tilting modal fingerprint; and performing abnormal separation processing on a tower and pole foundation area, a connecting area and an upper structure area to output structural tilting and non-structural tilting determination results. The application realizes accurate mapping relationship between electrical load and structure response, effectively reduces the false alarm rate, and improves the accuracy and timeliness of early warning.

Description

Technical Field

[0001] The present invention relates to the technical field of condition monitoring, and in particular to a method and system for monitoring the tilt condition of a tower. Background Art

[0002] Traditional methods for monitoring power tower tilt rely primarily on single tilt sensors or manual inspections for condition assessment. These methods are relatively simple in terms of technical implementation, directly measuring changes in the tower's tilt angle to determine its stability. Existing monitoring systems typically employ independent sensor placement schemes and lack systematic multi-physics coupling analysis. Consequently, understanding of tower structural response remains limited to monitoring changes in a single physical quantity.

[0003] Existing technologies suffer from significant deficiencies, primarily due to a lack of consideration of the unique physical factors of power systems, such as the current-carrying state of conductors, unbalanced currents between phases, and the impact of load variations on the tower structure. This results in an inability to establish a correlation between electrical events and structural responses, rendering early warning analysis lacking in systematicity and comprehensiveness. More seriously, the core challenge facing existing technologies is the inability to effectively distinguish between dangerous structural tilt and safe non-structural tilt. This failure to fully consider the effects of multi-physics coupling and the collaborative analysis of multiple sensor data can easily lead to false alarms or missed alarms under complex operating conditions.

[0004] Based on the above shortcomings, further analysis found that the existing technology still has deeper problems: first, there is a lack of accurate methods for establishing the mapping relationship between electrical load and structural response, which makes it impossible to quantify the specific impact of current changes on the tower structure; second, there is a lack of an effective fusion mechanism for multimodal sensing features, which makes it difficult to extract highly discriminative structural status features from multi-dimensional information such as inclination, vibration, offset, and settlement; finally, there is a lack of anomaly separation and processing methods based on structural risk propagation paths, which makes it impossible to achieve refined risk assessment of different areas of the tower and accurate determination of the tilt type. Summary of the Invention

[0005] The present invention provides a method and system for monitoring the tilt status of a tower, which realizes an accurate mapping relationship between electrical load and structural response, effectively reduces the false alarm rate, and improves the accuracy and timeliness of early warning.

[0006] In a first aspect, the present invention provides a method for monitoring a tower tilt state, the method comprising:

[0007] Conduct multi-field digital coupling modeling of power towers to obtain a digital twin model and calculate the electrical load-structure response sensitivity matrix;

[0008] Based on the digital twin model, a multimodal sensing feature data set including tower inclination time series data, insulator string offset, tower body micro-vibration spectrum and foundation settlement gradient is collected;

[0009] performing a time-lag correlation analysis on electrical events and structural responses based on the multimodal sensing feature data set and the electrical load-structural response sensitivity matrix to obtain an electrical-structural mapping matrix;

[0010] Inputting the electrical-structural mapping matrix into a multimodal tilt feature fusion network for tensor decomposition processing to obtain a tower tilt modal fingerprint;

[0011] According to the tower tilt modal fingerprint, abnormal separation processing is performed on the tower foundation area, connection area and superstructure area, and the structural tilt and non-structural tilt judgment results are output.

[0012] Optionally, in a first implementation of the first aspect of the present invention, performing multi-field digital coupling modeling on the power tower to obtain a digital twin model and calculate the electrical load-structure response sensitivity matrix includes:

[0013] The height, cross arm width, tower segment size, connection node coordinates, conductor connection point locations, and foundation burial depth of the power tower are collected to obtain a tower geometric parameter set.

[0014] Inputting the tower geometric parameter set into a three-dimensional finite element modeling system for meshing and material property assignment to obtain a digital skeleton model of the tower;

[0015] Based on the digital skeleton model of the tower, a multi-physics field coupling calculation is performed on the conductor current-carrying temperature field and the tower structure displacement field to obtain a digital twin model;

[0016] The structural response sensitivity caused by current changes is quantified according to the digital twin model to obtain an electrical load-structural response sensitivity matrix.

[0017] Optionally, in a second implementation of the first aspect of the present invention, performing multi-physics field coupling calculations on the conductor current-carrying temperature field and the tower structure displacement field based on the tower digital skeleton model to obtain a digital twin model includes:

[0018] Based on the digital skeleton model of the tower, a thermodynamic model is performed on the current parameters of the conductor to obtain a conductor current-carrying temperature field, which is used to quantify the impact of current changes on conductor temperature distribution;

[0019] Based on the digital skeleton model of the tower, a finite element analysis is performed on the deformation characteristics of the tower structure under the action of external loads to obtain a displacement field of the tower structure. The displacement field of the tower structure is used to describe the displacement-stress mapping relationship of each target node of the tower;

[0020] Based on the digital skeleton model of the tower, the soil bearing capacity, hydrological conditions and seismic subsidence coefficient of the geological environment in which the tower foundation is located are parameterized to obtain a geological parameter tensor;

[0021] The conductor current-carrying temperature field, the tower structure displacement field and the geological parameter tensor are subjected to multi-physics field coupling calculation to construct a digital twin model.

[0022] Optionally, in a third implementation of the first aspect of the present invention, the collecting of a multimodal sensing feature dataset including tower inclination time series data, insulator string offset, tower micro-vibration spectrum, and foundation settlement gradient based on the digital twin model includes:

[0023] Calculating tower structure sensitivity distribution information based on the digital twin model, and determining optimal monitoring point locations based on the tower structure sensitivity distribution information;

[0024] Collecting the original sensor signal matrix of the tilt angle time series data, the insulator string offset, the tower body micro-vibration spectrum and the foundation settlement gradient on the power tower according to the optimal monitoring point position;

[0025] Comparing and calibrating the original sensor signal matrix with the theoretical response value predicted by the digital twin model to obtain a calibrated preprocessed signal set;

[0026] Multi-scale feature extraction is performed on the calibrated preprocessed signal set and fused with the feature template generated by the digital twin model to obtain a multimodal sensing feature data set.

[0027] Optionally, in a fourth implementation of the first aspect of the present invention, performing a time-lag correlation analysis on electrical events and structural responses based on the multimodal sensing feature dataset and the electrical load-structure response sensitivity matrix to obtain an electrical-structure mapping matrix includes:

[0028] Matching the multimodal sensing feature data set with the electrical load-structural response sensitivity matrix to obtain a structural response feature set;

[0029] Correlating the phase current, inter-phase imbalance, power factor change rate, and system harmonic content in the power tower operation data with the electrical load-structure response sensitivity matrix to obtain an electrical event feature set;

[0030] A time-delay convolution operation and a correlation degree calculation are performed based on the structural response feature set and the electrical event feature set to obtain an electrical-structural mapping matrix, wherein the electrical-structural mapping matrix includes a tilt correlation vector, an offset correlation vector, a vibration correlation vector, and a settlement correlation vector.

[0031] Optionally, in a fifth implementation of the first aspect of the present invention, inputting the electrical-structural mapping matrix into a multimodal tilt feature fusion network for tensor decomposition processing to obtain a tower tilt modal fingerprint includes:

[0032] Inputting the electrical-structural mapping matrix into the four-channel parallel feature extraction layer in the multimodal tilt feature fusion network for processing, thereby obtaining a tilt angle feature vector, an offset feature vector, a vibration feature vector, and a settlement feature vector;

[0033] Performing tensor construction and decomposition on the tilt angle eigenvector, the offset eigenvector, the vibration eigenvector, and the settlement eigenvector to obtain a potential feature tensor;

[0034] Inputting the latent feature tensor into the cross-modal attention fusion layer in the multimodal tilted feature fusion network for feature weighted summation to obtain a fused feature vector;

[0035] The fused feature vector is sent to the deep autoencoder in the multimodal tilt feature fusion network for compression representation to obtain the tower tilt modal fingerprint.

[0036] Optionally, in a sixth implementation of the first aspect of the present invention, the step of sending the fused feature vector to a deep autoencoder in the multimodal tilt feature fusion network for compression representation to obtain a tower tilt modal fingerprint includes:

[0037] Inputting the fused feature vector into the encoder of the deep autoencoder for dimensionality reduction to obtain a first latent space representation;

[0038] Inputting the first latent space representation into the decoder of the deep autoencoder for reconstruction to obtain a reconstructed feature vector;

[0039] Calculating a mean square error between the reconstructed feature vector and the fused feature vector, and optimizing network parameters of the encoder and the decoder by a gradient descent method based on the mean square error to obtain a second latent space representation;

[0040] The contrastive learning loss function is used to perform feature enhancement processing on the second latent space representation to obtain the tower tilt modal fingerprint.

[0041] Optionally, in a seventh implementation of the first aspect of the present invention, performing abnormal separation processing on the tower base area, the connection area, and the superstructure area according to the tower tilt modal fingerprint and outputting the structural tilt and non-structural tilt determination results includes:

[0042] Performing abnormal separation on the tower tilt modal fingerprint to obtain abnormal characteristics of the tower structure;

[0043] Based on the abnormal characteristics of the tower structure, abnormal scores are calculated for the tower foundation area, connection area and superstructure area respectively to obtain a structural abnormality location vector;

[0044] The structural anomaly location vector is used to model the structural risk propagation path to obtain a risk propagation model, wherein the risk propagation model uses the sensor layout locations as nodes and determines the topological connection based on the actual structural relationship of the tower;

[0045] The structural reliability index is calculated based on the risk propagation model and the displacement-stress mapping relationship in the digital twin model to obtain the judgment results of structural tilt and non-structural tilt. The judgment results of structural tilt and non-structural tilt are used to distinguish between temporary tilt caused by transient events in the power system and structural tilt caused by foundation erosion and loose bolts.

[0046] In a second aspect, the present invention provides a tower tilt state monitoring system, the tower tilt state monitoring system comprising:

[0047] Modeling module, used to perform multi-field digital coupling modeling of power towers, obtain digital twin models and calculate electrical load-structure response sensitivity matrices;

[0048] An acquisition module is used to acquire a multimodal sensing feature data set including tower inclination time series data, insulator string offset, tower body micro-vibration spectrum and foundation settlement gradient based on the digital twin model;

[0049] an analysis module, configured to perform a time-lag correlation analysis on electrical events and structural responses based on the multimodal sensing feature data set and the electrical load-structural response sensitivity matrix to obtain an electrical-structural mapping matrix;

[0050] A processing module is used to input the electrical-structural mapping matrix into a multimodal tilt feature fusion network for tensor decomposition processing to obtain a tower tilt modal fingerprint;

[0051] The output module is used to perform abnormal separation processing on the tower foundation area, the connection area and the upper structure area according to the tower tilt modal fingerprint, and output the structural tilt and non-structural tilt judgment results.

[0052] The technical solution provided by this invention achieves a precise mapping between electrical load and structural response by constructing a multi-field coupled digital twin model of power towers. This overcomes the limitation of traditional monitoring methods that only consider mechanical factors and effectively correlates the influencing mechanisms between conductor current-carrying status and tower tilt. A multimodal sensor deployment scheme guided by the digital twin model enables comprehensive acquisition and preprocessing of tower tilt time-series data, insulator string offset, tower microvibration spectrum, and foundation settlement gradient, improving the quality and relevance of monitoring data. Asynchronous electrical-structural event correlation analysis establishes a time-delayed correlation between electrical events and structural responses. A multimodal tilt feature fusion network performs tensor decomposition processing, extracting a highly discriminative tower tilt modal fingerprint. This fingerprint can accurately distinguish temporary tilt caused by power system-specific transient events such as lightning strikes and short-circuit current surges from structural tilt caused by foundation erosion and loose bolts. Anomaly separation based on a structural risk propagation path diagram enables refined risk assessment of the tower foundation, connection, and superstructure areas, effectively reducing false alarm rates and improving the accuracy and timeliness of early warnings. A complete tower tilt status monitoring and early warning decision-making closed-loop system has been established. By deeply integrating the unique physical factors and structural engineering characteristics of the power system, it provides substantial guarantees for the safe operation of the power grid and reduces the risk of tower collapse accidents. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0054] Figure 1 A schematic diagram of an embodiment of a method for monitoring a tower tilt state according to an embodiment of the present invention;

[0055] Figure 2 FIG. 1 is a schematic diagram of an embodiment of a tower tilt state monitoring system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0056] An embodiment of the present invention provides a method and system for monitoring the tilt condition of a tower. The terms "first," "second," "third," "fourth," and so on (if any) in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the numbers used in this way are interchangeable where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to these processes, methods, products, or apparatus.

[0057] For ease of understanding, the specific process of the embodiment of the present invention is described below. Figure 1 , an embodiment of a tower tilt state monitoring method in an embodiment of the present invention includes:

[0058] Step S101: Perform multi-field digital coupling modeling on the power tower to obtain a digital twin model and calculate the electrical load-structure response sensitivity matrix;

[0059] It is understandable that the execution subject of the present invention may be a tower tilt state monitoring system, or a terminal or a server, which is not limited here. The embodiment of the present invention is described by taking a server as the execution subject as an example.

[0060] Specifically, a systematic collection of geometric structure and operating environment parameters for power towers is conducted, including key features such as the overall tower height, crossarm width and distribution, tower segment dimensions, 3D coordinates of connection nodes, conductor suspension point locations, and foundation depth. This constitutes a geometric parameter set that describes the tower's physical geometry. This geometric parameter set is imported into a 3D finite element modeling system, and the tower is divided into multiple finite element units using a meshing strategy. Mesh density is increased at key locations such as crossarm-to-tower connection points, conductor suspension points, and foundation transition areas to improve local simulation accuracy. Material property parameters such as elastic modulus, Poisson's ratio, density, thermal conductivity, thermal expansion coefficient, and damping ratio are assigned to each unit, enabling the unit to respond to both mechanical forces and thermal field changes.

[0061] After the digital skeleton is constructed, the multi-physics coupling phase begins. The temperature field generated by the conductors being energized is coupled in real time with the tower's structural response field. This simulates the process by which heat from the conductors is transferred to the tower under different current-carrying conditions, causing local thermal expansion or stress deformation. The deformation and stress distribution of the tower under the combined thermal and mechanical influences are then calculated in a time-varying manner. Coupled modeling simultaneously accounts for geological heterogeneity and local stiffness differences at structural nodes, enabling the digital twin model to dynamically reflect the interaction between the structure and its environment.

[0062] Based on the digital skeleton model of the tower, a thermodynamic modeling of the conductor current load state is performed. Considering factors such as current intensity, conductor resistivity, surface heat transfer conditions, and ambient wind speed and temperature, a set of conductor thermal transfer equations is established to obtain a time-varying temperature distribution field. Simultaneously, a finite element analysis of the tower under external loads such as conductor tension, wind load, gravity, and thermal expansion is performed to solve for internal stress distribution and node displacement responses, resulting in a structural displacement field describing the spatial displacement and stress state of key nodes. A parametric modeling of the tower foundation's geological environment is performed, taking into account the nonlinear bearing characteristics of the soil, hydrological permeability conditions, and seismic subsidence sensitivity. A geological parameter tensor is constructed, including indicators such as soil modulus, water saturation, compressibility, and seismic response factor. By setting boundary conditions and coupling interfaces, a multi-physics coupling calculation is performed to combine the conductor current-carrying temperature field, the tower structural displacement field, and the geological parameter tensor. Heat conduction equations, mechanical equilibrium equations, and foundation-structure coupling constraints are jointly solved to construct a digital twin model of the tower.

[0063] Step S102: Based on the digital twin model, a multimodal sensing feature data set including tower inclination time series data, insulator string offset, tower body micro-vibration spectrum, and foundation settlement gradient is collected;

[0064] Specifically, based on the digital twin model, multiple representative load combinations are applied in the simulation space to determine the stress, displacement, and temperature gradient responses at key nodes. The structural response rate is then differentiated to construct a sensitivity distribution function for input variables such as current fluctuations, wind load gradients, and foundation settlement, quantifying the sensitivity of various structural components to external excitations. Based on this sensitivity distribution, a multi-objective optimization algorithm is used to determine the optimal monitoring point locations, ensuring that sensor installation feasibility and communication path accessibility are met. This ensures that the inclination, image, vibration, and displacement sensors can best reflect the true state of the structure. A multimodal sensor array is deployed on the tower structure, including inclination sensors at the tower head and crossarm nodes, image or laser displacement measurement equipment near the insulator strings, triaxial high-frequency accelerometers at key tower connections, and high-precision displacement meters or electronic levels at the four corners of the foundation or at the base of the piles. Each sensor outputs a matrix of raw sensor signals at varying sampling frequencies and accuracies, including inclination, insulator deflection, vibration spectrum, and foundation settlement gradient in the time domain, forming the raw set of multidimensional observational data for the structural state.

[0065] The theoretical response value predicted by the digital twin model is introduced as a reference, and the original signal is corrected for errors and offset calibration is performed through residual analysis and dynamic calibration algorithms. Based on the response comparison within the sliding window, the measured data and the simulation model are aligned using least squares registration, sample normalized difference calibration, or Bayesian residual optimization methods to form a calibrated preprocessed signal set. Multi-scale feature extraction is performed on the calibration signal set to extract stability indicators, fluctuation characteristics, frequency domain peaks, statistical distribution parameters, and coupling correlation characteristics at different time scales, frequency bandwidths, and spatial distribution ranges. For example, statistical quantities such as trend line slope, variance, and zero-crossing rate are extracted from the inclination data; feature point motion trajectory extraction and three-dimensional reconstruction are performed on the offset image of the insulator string; short-time Fourier transform or Hilbert-Huang transform is used to obtain the main vibration mode and instantaneous frequency characteristics of the vibration signal; spatial gradient fitting and settlement surface reconstruction are performed on the foundation settlement signal. The multi-dimensional features are fused with the standard feature templates extracted during the simulation evolution of the digital twin model. The fused high-dimensional representation vector is constructed through feature-level weighted matching, deep feature space mapping or principal component correlation analysis to form a multimodal sensing feature dataset with a unified coding format.

[0066] Step S103: performing a time-lag correlation analysis on the electrical event and the structural response based on the multimodal sensing feature data set and the electrical load-structural response sensitivity matrix to obtain an electrical-structural mapping matrix;

[0067] Specifically, the multimodal sensing feature dataset is matched to the electrical load-structure response sensitivity matrix. The multimodal sensing feature dataset contains a dynamic evolution sequence of structural states, including inclination changes, insulator string offsets, tower vibration spectra, and foundation settlement gradients. The electrical load-structure response sensitivity matrix provides a quantitative description of the response of each electrical perturbation to a specific structural node. Through timestamp alignment and sensitivity matrix guidance, the physical locations of the sensing data are logically mapped to the structural regions affected by electrical parameter changes. This yields a set of mapping indicators for the true structural response under various electrical driving factors, forming a structural response feature set.

[0068] The actual operational data stream of the power system is introduced, including the evolution patterns of three-phase current time series, interphase imbalance, power factor change rate, and harmonic analysis results (total harmonic distortion and higher-order harmonic components) in the time and frequency domains. These variables characterize the dynamic behavior of the electrical system and are converted into physical action quantities that cause disturbances to the structural system through coupling with the sensitivity matrix. These action quantities are then mapped to the structural response sensitivity matrix to obtain a set of electrical event signatures representing the characteristics of the electrical disturbance. This process is implemented through vector inner product operations, weighted mapping coefficients, or tensor projection, ensuring that each electrical quantity change has a clear response indicator in the structural influence domain.

[0069] Time-delayed convolution and correlation calculation are performed on the two feature sets. Time-delayed convolution considers the asynchronous occurrence of electrical disturbances and structural responses. By designing a sliding window function and kernel function, the cross-influence of signals within different lag times τ is modeled. The convolution operation weights the lag times using exponential, Gaussian, or chi-square kernels, highlighting periods of high correlation and suppressing uncorrelated disturbances. After performing bidirectional convolution between each pair of electrical variables and structural responses, the overall correlation is measured using statistical methods such as mutual information entropy, dynamic time warping distance, and correlation coefficient. This results in an explicit vector representing the path of the electrical variable's influence on the structural state. These results are organized into an electrical-structural mapping matrix containing four components: the inclination correlation vector characterizes the causal relationship between current or unbalance changes and tower posture changes; the offset correlation vector reveals the impact of short-circuit current surges or harmonic components on the stress state of the insulator string; the vibration correlation vector reflects the intensity of the excitation of local vibration modes of the tower body by system switching or electromagnetic harmonics; and the settlement correlation vector correlates the effects of long-period load offsets or uneven thermal stress distribution on the foundation response.

[0070] Step S104: inputting the electrical-structural mapping matrix into the multimodal tilt feature fusion network for tensor decomposition processing to obtain the tower tilt modal fingerprint;

[0071] Specifically, the electrical-structural mapping matrix is ​​input into the four-channel parallel feature extraction layer of the multimodal tilt feature fusion network. This mapping matrix contains the correlation characteristics of four types of structural responses driven by specific electrical events: tilt angle change, insulator string offset, tower microvibration, and foundation settlement. The parallel feature extraction layer designs different processing channels for each of the four types of input data: the tilt angle channel introduces a long-short-term memory network to capture evolutionary trends and mutation patterns; the offset channel uses a one-dimensional convolutional network to analyze the micro-displacement characteristics of the insulator string; the vibration channel incorporates a spectral attention mechanism to perform frequency-domain weighted analysis of multi-frequency vibration signals; and the settlement channel introduces a graph convolutional network to construct the topological correlation of the basic displacement space. After processing each channel, the tilt angle feature vector, offset feature vector, vibration feature vector, and settlement feature vector are output, respectively, to form the basic modal representation of the structural tilt.

[0072] The four types of eigenvectors are uniformly mapped to a high-order tensor space, constructing a multidimensional feature tensor that maintains information complementarity in the channel dimension and overall structural logic in the sample and time dimensions. The original tensor is rank-compressed through Tucker or CP decomposition to extract the potential synergistic relationships between channels and the dominant factors of structural deformation, forming a potential feature tensor with lower dimensionality but stronger expressiveness.

[0073] The latent feature tensor is fed into the cross-modal attention fusion layer, which dynamically calculates the contribution of each feature type to the overall tilt trend by constructing an attention weight matrix. Based on a soft attention mechanism, the tilt, offset, vibration, and settlement feature tensor components are fused through weighted summation, giving higher weights to key features with significant response patterns or spatiotemporal coupling. This generates a fused feature vector representing the overall tilt state of the current structure.

[0074] The fused feature vector is fed into a deep autoencoder for latent representation learning. The encoder uses a multi-layer nonlinear network mapping to reduce the fused vector layer by layer into a low-dimensional latent representation space, forming the first latent space representation. The decoder then attempts to restore the original fused feature vector by increasing the dimensionality of the latent variables layer by layer, outputting a reconstructed feature vector. The mean squared error between the reconstructed and original fused vectors is calculated as the loss function, and a gradient descent algorithm is used to jointly optimize the encoder and decoder network weights. This minimizes the reconstruction error and improves the latent space representation capability, resulting in the second latent space representation.

[0075] Finally, a contrastive learning loss function was introduced. By constructing pairs of positive samples (samples of the same tilt type but different time periods) and negative samples (samples of different tilt types), similarity constraints were imposed in the latent space. The cosine similarity between samples of the same type was calculated to bring them closer together in the latent space, while samples of different types were separated. This resulted in a representation distribution with a distinct clustering structure. Ultimately, the tower tilt modal fingerprint was obtained—a compressed, enhanced, and classifiable feature encoding representing the multimodal response state of the tower structure under electrical induction.

[0076] Step S105: performing abnormal separation processing on the tower foundation area, the connection area and the upper structure area according to the tower tilt modal fingerprint, and outputting the determination results of structural tilt and non-structural tilt.

[0077] Specifically, the tower tilt modal fingerprint is subjected to anomaly separation. Using a pre-trained feature decoding network or an anomaly measurement mechanism based on projection distance, significant features representing the current structure's deviation from the normal operating mode are extracted to obtain the tower structural anomaly features. The anomaly feature vector is mapped to the three key sub-regions of the tower structure: the base region, the connection region, and the superstructure region. A positional calculation is performed using the regional attribution matrix and the spatial label of each component in the modal fingerprint. Local anomaly activation values ​​within each region are counted. Combined with the sensor density and sensitivity weighted distribution in that region, a regional-level anomaly score is calculated to construct a structural anomaly location vector that describes the health status of each structural component in the tower's current state.

[0078] Using the structural anomaly location vector as the core input and combining the physical connections between actual sensor locations and tower components, a structural risk propagation path model is constructed. This model employs a graph network structure, with sensor nodes as graph vertices and connections between tower components as edges. Each edge is assigned a propagation weight, representing the probability or impact of an anomaly spreading from one node to adjacent nodes. Using a graph neural network message passing strategy, a neighboring node anomaly state aggregation function is introduced for each node, simulating the propagation links of structural damage or instability within the tower and the evolution of anomaly states to adjacent structural units.

[0079] After the risk map is constructed, the displacement-stress mapping relationship within the digital twin model is introduced to map anomaly scores to structural stress response indicators in the physical domain, thereby constructing a structural reliability index function. This function comprehensively assesses the future failure risk of each key unit based on factors such as the current stress level, residual load-bearing capacity, and predicted displacement trend. By integrating and normalizing the reliability index across multiple nodes of the entire structure and combining it with regional risk propagation results, the final determination of structural and non-structural tilt is obtained.

[0080] Structural tilt is manifested as a continuous decline in reliability, with the risk extending upward from the foundation area or connection nodes, accompanied by slow evolution and persistence, and is caused by foundation erosion, concrete settlement, bolt loosening or metal fatigue. Non-structural tilt is stimulated by electrical disturbances (short-circuit shock, harmonic interference, switching operation, etc.) in a short period of time. After the disturbance ends, the structural state returns to normal, manifested as a high-frequency short-time response, and the reliability index fluctuates and then tends to stabilize.

[0081] In a specific embodiment, the process of executing step S101 may specifically include the following steps:

[0082] The height, cross arm width, tower segment size, connection node coordinates, conductor connection point locations, and foundation burial depth of the power tower are collected to obtain a tower geometric parameter set.

[0083] Input the tower geometric parameter set into the 3D finite element modeling system for meshing and material property assignment to obtain a digital skeleton model of the tower;

[0084] Based on the digital skeleton model of the tower, multi-physics field coupling calculations are performed on the conductor current-carrying temperature field and the tower structure displacement field to obtain a digital twin model.

[0085] The structural response sensitivity caused by current changes is quantified based on the digital twin model to obtain the electrical load-structure response sensitivity matrix.

[0086] Specifically, raw data is collected around key geometric parameters of the tower structure and its operating environment. Basic geometric features of the power tower are collected, including the total tower height, the number and unit width of crossarms, the dimensions of the tower's longitudinal segments, and their spatial distribution ratios. Spatial parameters of structural connection features are also collected, including the three-dimensional coordinates of all connection nodes on the tower, including angle steel intersections, support truss nodes, and primary load-bearing contact points with the foundation or crossarms. For conductor routing, the location distribution of each conductor connection point is measured, including its height on the tower structure, horizontal offset from the main axis, and relative distance from the crossarms and insulator strings. Furthermore, the foundation depth, a foundation constraint that affects the stability of the entire tower, is calibrated according to construction drawings and field survey data, taking into account the non-uniformity of foundation stiffness caused by inconsistent foundation depths. Ultimately, a structured tower geometric parameter set is compiled. This tower geometric parameter set is then imported into a finite element modeling system equipped with 3D modeling and structural analysis capabilities. Standard structural modeling processes are then implemented, and the tower is automatically or semi-automatically meshed based on the structural type and mechanical requirements. The partitioning strategy is differentiated based on the force complexity of different locations. For example, high-density meshes are used in key areas such as the connection between the tower base and the tower body, crossarm corners, and conductor hanging points to improve local calculation accuracy, while the mesh density is moderately reduced in uniform components such as the middle of the tower to improve calculation efficiency. Appropriate material property parameters are assigned to each finite element, including physical parameters such as the elastic modulus, Poisson's ratio, density, thermal conductivity, thermal expansion coefficient, and damping ratio of the structural steel, to ensure that the structural model accurately represents the true behavior of the material when responding to loads. On this basis, a digital skeleton model of the tower is constructed in combination with boundary conditions such as foundation constraints, tower top degree of freedom restrictions, and conductor tension input methods. Based on the digital skeleton model of the tower, multi-physics field coupling calculations are performed on the conductor current-carrying temperature field and the tower structure displacement field. The heat conduction modeling process of the conductor's current-carrying temperature rise is analyzed. Based on the conductor's actual current value, resistance parameters, surface radiation capacity, ambient wind speed, and air temperature, the temperature field distribution generated by the energization is calculated. This temperature field is applied as a thermal boundary condition to the tower nodes connected to the conductors. Localized thermal stress concentrations occur in areas such as the conductor attachment points, crossarm ends, and insulator supports. Simultaneously, the structural displacements of the tower structure itself, caused by wind loads, deadweight, conductor tension, and electromagnetic forces, must be calculated. Based on the classical elastic-plastic mechanics equations, the static deformation and stress states of the nodes are solved using the finite element method, thereby constructing the tower structure's response behavior under actual electrical operating conditions. To ensure the accuracy of the interaction between the thermal and mechanical fields, a thermal strain term is established between the temperature change and the thermal expansion of the material. The secondary force caused by thermal expansion is fed back into the mechanical solution model and an iterative coupled solution is performed within the time step to form a dynamically evolving thermal-structural multi-field response model.The digital twin model quantifies the structural response sensitivity to current changes. By making small perturbations to the conductor current parameters and calculating the displacement changes at key nodes before and after the perturbation, the derivatives—the incremental node displacements caused by a unit current change—are further calculated to construct an electrical load-structure response sensitivity matrix. Each row of this matrix represents the response channel of a structural node, and each column corresponds to a current input channel. The numerical values ​​reflect the node's response sensitivity to current perturbations. In practice, matrix compression is performed based on structural symmetry, or Laplace regularization is introduced to avoid overfitting of boundary responses. To ensure that this matrix has practical application value, it is fitted and calibrated with the dynamic response decoupling results of the digital twin model to enable it to reflect the optimal prediction capability of electrical events on structural behavior. The final output is a sensitivity matrix.

[0087] In a specific embodiment, the execution step of performing multi-physics field coupling calculation on the conductor current-carrying temperature field and the tower structure displacement field based on the tower digital skeleton model to obtain the digital twin model may specifically include the following steps:

[0088] Based on the digital skeleton model of the tower, the conductor current parameters are thermodynamically modeled to obtain the conductor current temperature field. The conductor current temperature field is used to quantify the impact of current changes on the conductor temperature distribution.

[0089] Based on the digital skeleton model of the tower, a finite element analysis is performed on the deformation characteristics of the tower structure under external loads to obtain the tower structure displacement field. The tower structure displacement field is used to describe the displacement-stress mapping relationship of each target node of the tower.

[0090] Based on the digital skeleton model of the tower, the soil bearing capacity, hydrological conditions and seismic subsidence coefficient of the geological environment where the tower foundation is located are parameterized to obtain the geological parameter tensor;

[0091] The conductor current-carrying temperature field, tower structure displacement field and geological parameter tensor are calculated through multi-physics field coupling to construct a digital twin model.

[0092] Specifically, based on a digital skeleton model of the tower, a thermodynamic modeling of the thermal field evolution behavior of conductors under current flow was performed. Conductors, as crucial components for power transmission, experience Joule heating during current-carrying. The heat generation power density is determined by the product of current and resistance. This heat is dissipated into the surrounding air through convection and radiation on the conductor surface. To simulate this process, a heat conduction differential equation was introduced, and the conductor cross-section was divided into finite element sub-elements. The heat balance equation was solved for each element through time-step iteration. The specific heat conduction model considered the effects of axial temperature rise, radial heat diffusion, and the gradient between the surface convective heat transfer coefficient and the ambient air temperature. Thermophysical properties such as the thermal conductivity, resistivity, current density, and specific heat capacity of the conductor insulation material were incorporated as parameters. By setting boundary conditions such as the convective heat transfer coefficient under natural wind speed, the effect of ground radiation, and an air pressure correction factor, the spatially distributed conductor temperature field, which varies with time, was obtained. This temperature field reflects the dynamic heat distribution caused by changes in current intensity. It also conducts heat energy to the tower structure at the attachment point, indirectly contributing to the local deformation response of the tower and being the first heat input source field formed in the coupled model. Simultaneously, a finite element analysis of the tower structure's stress and deformation behavior under combined loads was performed based on the same digital skeleton model. This analysis comprehensively considers multiple load types, including wind loads, conductor tension, tower deadweight, electromagnetic forces, and thermally induced expansion stresses. These loads are applied to the corresponding parts of the tower in the model as nodal forces or surface forces. During the solution process, a unit stiffness matrix is ​​constructed based on the geometric information and material parameters of the tower components. Linear or nonlinear static solution methods from structural mechanics are used to obtain nodal displacement and stress solutions under boundary constraints. The displacement solution characterizes the overall deformation trend of the tower under the influence of multiple load sources, including top horizontal drift, crossarm sagging, and tower bending. The stress solution reflects the stress concentration within the components and can be used for subsequent failure diagnosis and fatigue life assessment. The solution vectors at all key nodes can be further organized into continuous functions, forming a structural displacement field in three-dimensional space. Their gradient fields are used to describe the strain distribution and stiffness trends among local components. To ensure that the model has boundary constraints and a mechanical response basis consistent with field conditions, a geological parameter tensor reflecting the foundation's support capacity and soil properties is introduced. Based on regional gridding, this tensor assigns mechanical properties to soil layers at different depths beneath the tower foundation, including geomechanical parameters such as soil bearing capacity, compression modulus, water saturation, permeability, and seismic subsidence coefficient. Calibration is performed through inversion using field exploration data or existing geological data. Each component of this tensor not only reflects the deformation capacity of the soil layer under different stress paths but also provides a physical basis for modeling complex behaviors such as foundation settlement, slip, lateral displacement, and uneven subsidence. Furthermore, by coupling the foundation with the mesh nodes at the bottom of the structure, the geological tensor influences the boundary displacement field of the tower foundation, altering the overall stiffness distribution and dynamic response pattern of the structure.A multi-physics coupling calculation is performed on the conductor current-carrying temperature field, the tower structural displacement field, and the geological parameter tensor. A coupling channel is established between the thermal and structural force fields and the geological boundary, and simultaneous iterative updates of the multi-field variables are achieved within the solution framework. In the numerical solution, the thermal expansion deformation caused by the thermal field is applied to the structural mechanics equations in the form of thermal strain, forming a thermal-structural coupling. Simultaneously, the boundary response of the structural displacement field at the tower base influences the tower rebound and settlement feedback through the foundation stiffness matrix in the geological tensor, forming a structural-geological coupling. Furthermore, the changes in the ground stress state caused by the foundation temperature gradient can also be fed back into the thermal-stress field, forming a triple-coupled system. This coupling process is implemented using the multi-physics module or external coupling iterative algorithm in the finite element software. Solving the problem in either weak or strong coupling forms, a high-fidelity digital twin model is constructed under the three-dimensional dynamic effects of realistic electrical loads, structural responses, and geological boundaries.

[0093] In a specific embodiment, the process of executing step S102 may specifically include the following steps:

[0094] Calculate the sensitivity distribution information of the tower structure based on the digital twin model, and determine the optimal monitoring point location based on the sensitivity distribution information of the tower structure;

[0095] According to the optimal monitoring point location, the original sensor signal matrix of the tilt angle time series data, insulator string offset, tower body micro-vibration spectrum and foundation settlement gradient is collected on the power tower;

[0096] The original sensor signal matrix is ​​compared and calibrated with the theoretical response value predicted by the digital twin model to obtain the calibrated preprocessed signal set;

[0097] Multi-scale feature extraction is performed on the calibrated preprocessed signal set and fused with the feature template generated by the digital twin model to obtain a multimodal sensing feature dataset.

[0098] Specifically, based on a digital twin model, the tower structure's geometric characteristics, material properties, electrical load boundary conditions, thermal-mechanical coupling, and geological parameters are integrated into the model to construct a complete stress-response system. Under multiple loading conditions, such as simulating conductor current inputs of varying magnitudes, wind load variations, and foundation settlement disturbances, the displacement, stress, and strain responses of key nodes throughout the tower structure are calculated. A unit perturbation is applied to each node or region, and the rate of change of the local response is calculated to determine the magnitude of the change in the global response under the unit perturbation. Using a gradient calculation method based on perturbation analysis, a structural response sensitivity map covering the entire tower structure is obtained in a spatial coordinate system. This map reflects the intensity of response to external perturbation inputs at different locations. High values ​​in the range indicate that subtle state changes, i.e., locations with the most significant electrical-structural coupling effects, are more easily captured. After obtaining this structural sensitivity distribution information, an optimal monitoring point deployment plan is designed, taking into account multiple constraints, such as sensor type characteristics, construction feasibility, power supply, and communication conditions. This solution prioritizes points in highly sensitive areas and, incorporating geometric structural features such as tower symmetry, intersections of key components, and the locations of the main modal couplings in the dynamic model, utilizes multi-objective optimization algorithms such as genetic optimization or particle swarm optimization to determine the optimal set of monitoring points for collecting inclination, offset, vibration, and settlement signals. This ensures that a limited number of sensors can capture the maximum amount of structural response information, improving overall monitoring efficiency and data quality. After on-site deployment based on the optimal monitoring point locations, the system continuously collects raw signal data from the structure during operation. Tilt data is sampled by high-resolution MEMS tilt sensors installed at the tower head and crossarm endpoints, outputting a high-frequency, low-noise sequence of angle changes. Insulator string offset is detected by a binocular vision system, which identifies the motion trajectory of characteristic points installed on the insulators and reconstructs their three-dimensional spatial coordinates to form a time series offset. Tower micro-vibration signals are acquired by highly sensitive triaxial accelerometers located at the main rod connection nodes. The signal frequency ranges from 0.1 Hz to several hundred Hz, reflecting transient excitation responses and high-order modal frequency energy. Foundation settlement gradient data is collected by an array of electronic levels placed at the four corners of the tower base, recording the evolution of the settlement surface in both time and space. These four types of information constitute the raw sensor signal matrix representing the multi-source operating state of the structure. Data calibration of the raw sensor signals is performed based on the theoretical response curve derived from the digital twin model. The sensor signal matrix is ​​matched with the ideal response value predicted by the model. Sliding window residual analysis, wavelet filtering denoising, multi-scale cross-correlation analysis, and range correction are used to construct an error evaluation function based on time step differences and response gradient changes. Adaptive adjustment is performed through the gradient descent method to minimize the error to obtain the calibrated preprocessed signal set.During this process, to ensure global consistency in the calibration, structural mechanics constraints are introduced as regularization terms, such as maximum allowable offset thresholds, physical acceleration bounds, or upper limits on displacement growth rates. This enhances the consistency between the signal and physical behavior, ensuring that the data conforms to both the actual acquisition conditions and the physical logic of the model predictions. After signal calibration, the calibration signal set undergoes multi-scale feature extraction. By jointly analyzing information in the time and frequency domains, temporal, frequency, statistical, and coupled representations of the structural response are constructed. In the time domain, sequence statistics such as the sliding mean, extreme point distribution, coefficient of variation, and periodicity are extracted. In the frequency domain, short-time Fourier transforms and Hilbert-Huang transforms are applied to the vibration signals to obtain features such as dominant frequency, frequency band energy, spectral peak shift, and modal coupling. In the statistical domain, metrics such as kurtosis, skewness, entropy, and zero-crossing rate are introduced to characterize the signal morphology. In the multivariate coupling domain, autocorrelation and cross-correlation functions are used to identify time lags and phase coupling relationships between multi-channel signals. These features are ultimately constructed into a set of high-dimensional feature vectors, representing the key manifestations of each modal response in the current structural state. To achieve unified modeling and logical fusion of multimodal information, the above-mentioned feature vectors are fused and compared with the standard structural response feature templates generated by the digital twin model during the training phase. This fusion process includes vector alignment, scale normalization, principal component matching, and multi-layer feature superposition. Deep feature fusion algorithms such as self-attention mechanisms or tensor decomposition modules are used to achieve dynamic fusion of multimodal features in spatial, temporal, and physical dimensions, generating the final multimodal sensing feature dataset.

[0099] In a specific embodiment, the process of executing step S103 may specifically include the following steps:

[0100] The multimodal sensing feature data set is matched with the electrical load-structural response sensitivity matrix to obtain the structural response feature set;

[0101] The phase current, inter-phase imbalance, power factor change rate and system harmonic content in the power tower operation data are correlated with the electrical load-structure response sensitivity matrix to obtain the electrical event feature set;

[0102] Based on the structural response feature set and the electrical event feature set, time-delay convolution operation and correlation calculation are performed to obtain the electrical-structural mapping matrix, which includes the inclination correlation vector, offset correlation vector, vibration correlation vector and settlement correlation vector.

[0103] Specifically, the multimodal sensing feature dataset is a fusion of inclination angle variation, insulator string offset, tower vibration spectrum, and foundation settlement data collected at sensor locations. Calibration and multi-scale feature extraction processes form a unified high-dimensional feature matrix. Each type of signal feature encodes the corresponding response mode and spatial location label. Therefore, to achieve a corresponding match with the electrical load-structure response sensitivity matrix, a mapping index from signal channels to structural nodes is established based on the principles of node-level coordinate consistency and modal response channel alignment. A channel identifier in the sensing feature is associated with a node number in the sensitivity matrix. Based on the mapping result, a weighted expression of the sensitivity value multiplied by the eigenvalue is extracted, thereby achieving physical amplification or suppression of the structural response characteristics. Furthermore, using each mode (e.g., inclination angle, offset, vibration, settlement) as the classification dimension, the response feature vectors of the corresponding channel are combined according to time slices or sampling periods to obtain a structural response feature set that covers both spatial sensitivity and temporal evolution. While obtaining the structural response feature set, data from the power system's operating side is processed and feature extracted to construct an electrical event feature set with temporal logic and frequency domain distribution. This data source includes the phase currents on the power supply channels of power towers. The resulting phase current trend curves reflect load fluctuations, transient shocks, and long-term stability issues during operation. Furthermore, by defining interphase imbalance, time series data on the degree of symmetry violation of the phase current distribution is obtained. Combined with the power factor change rate, this describes the dynamic variation of power quality within a short time window. Frequency domain feature extraction technology based on discrete Fourier transforms is also introduced to perform harmonic analysis on the current signal, obtaining the amplitude and phase angle of high-order harmonic components. The total harmonic distortion rate is then used to quantify the system's harmonic pollution level. These electrical variables are encoded in the time and frequency domains, and feature normalization and scale alignment are performed to construct the electrical event feature set. The structural response feature set and the electrical event feature set are then placed on a unified time axis for time-delay convolution and correlation calculation. Considering that the occurrence of electrical events does not immediately generate a response on the structural side, a strided convolution function is introduced through a sliding time window mechanism, that is, a set of convolution integral operations are performed on the electrical event vector E(t) and the structural response vector S(t). The time delay variable τ introduced is a dynamically adjustable parameter, which is set as a multi-level time-delay structure in the calculation to form The product integral model of S(t) is used to improve the rationality and physical consistency of the convolution weights. Specific kernel functions, such as exponential decay kernels and Gaussian kernels, are designed to emphasize the most likely response lag interval and suppress the effect of edge interference data in the convolution. Through the kernel-weighted convolution operation, a set of time-delay coupled response functions between the electrical and structural components are formed to characterize the specific impact trend of a certain type of electrical event on the structure under the condition of τ time lag. After the convolution is completed, in order to quantify the global correlation between the two types of features, metric functions such as mutual information entropy calculation or Pearson correlation coefficient and dynamic time warping are introduced. From the perspective of information theory or distance function, a full-scale comparison of the structural response and the electrical event is performed to obtain an electrical-structural mapping matrix classified by modal dimension. This matrix constructs four sub-vectors according to the four major structural response dimensions of inclination, offset, vibration, and settlement. Each sub-vector records the impact intensity, lag time, impact duration, and maximum response amplitude of different electrical event variables on the corresponding modal response, and is recorded as the inclination correlation vector, offset correlation vector, vibration correlation vector, and settlement correlation vector, respectively.

[0104] In a specific embodiment, the process of executing step S104 may specifically include the following steps:

[0105] The electrical-structural mapping matrix is ​​input into the four-channel parallel feature extraction layer in the multimodal tilt feature fusion network for processing to obtain the tilt angle feature vector, offset feature vector, vibration feature vector and settlement feature vector;

[0106] Perform tensor construction and decomposition on the inclination eigenvector, displacement eigenvector, vibration eigenvector and settlement eigenvector to obtain the potential feature tensor;

[0107] The latent feature tensor is input into the cross-modal attention fusion layer in the multimodal tilted feature fusion network for feature weighted summation to obtain the fused feature vector;

[0108] The fused feature vector is fed into the deep autoencoder in the multimodal tilt feature fusion network for compression representation to obtain the tower tilt modal fingerprint.

[0109] Specifically, the electrical-structural mapping matrix serves as input. This mapping matrix represents the causal relationship between electrical perturbations and structural responses in four key modal dimensions in a time-aligned manner: the tilt, offset, vibration, and settlement correlation vectors. These vectors structurally share the same time series length, but their distributions reflect the response characteristics, time lags, and peak differences of their respective modes. To maximize the extraction of the implicit physical laws and key perturbation features within each mode, this matrix is ​​fed into a four-channel parallel feature extraction layer within the multimodal tilt feature fusion network. In the four-channel structure, each type of modal data is fed into a dedicated feature extraction channel. The inclination channel adopts a time series modeling structure and selects the long short-term memory network (LSTM) as the processing unit. Its ability to retain historical states is used to process the inclination change sequence caused by electrical disturbances, thereby extracting the inclination feature vector containing trend information and mutation response. The offset channel adopts a one-dimensional convolutional neural network (Conv1D) structure. The dynamic change pattern of the insulator string offset is abstracted and extracted through convolution kernels of different scales. It is suitable for capturing continuous offset, pulse rebound and non-periodic disturbance patterns. The vibration channel is designed as a spectrum attention network. The frequency domain weight mechanism is used to enhance the local frequency mutation caused by electrical harmonic components or system switching events, so that the vibration feature vector has high responsiveness and directionality in the spectrum space. The settlement channel adopts a graph convolutional network (GCN) to establish a node-adjacency matrix relationship on the spatial topology structure formed by the basic measuring points. The feature aggregation mechanism of the graph neural network in the local space propagation process is used to extract the evolutionary relationship between the settlement gradient and the spatial distribution characteristics. Through this four-channel processing, the tilt angle eigenvector, offset eigenvector, vibration eigenvector, and settlement eigenvector are obtained, respectively. After the four eigenvectors are extracted, a unified representation is established in the structured space, and a high-order multimodal tensor structure is constructed. The four modal vectors are concatenated and expanded in the modal dimension, forming a third-order tensor structure between the time, channel, and modal dimensions. This structured representation not only preserves the temporal evolution of the features within each mode but also establishes a semantic expression of the coupling between modalities in the tensor dimension. To reduce the dimensionality of redundant features and explore the core shared subspace between modalities, tensor decomposition techniques such as Tucker decomposition or CP decomposition are used to represent the high-order tensor as the product of a set of low-rank kernel tensors and an orthogonal modal basis. This compresses its representation complexity and extracts a latent feature tensor that expresses the coupling and variation characteristics between modalities. This tensor is essentially the interactive feature result of the four modal eigenvectors in low-dimensional space, preserving the intra-modal independence and inter-modal correlation in three-dimensional space. The latent feature tensor is input into the cross-modal attention fusion layer in the network. The core of this layer is to construct the attention weight matrix between modalities. By calculating the correlation and mutual dependence between the features between modalities, the weight influence of each modality in the final fusion expression is dynamically adjusted.A multi-head attention mechanism is used to input different modal features in the potential feature tensor as query vector Q, key vector K and value vector V, and then the softmax(. / )·V structure generates a modally weighted representation, so that the contribution to the final fused vector depends not only on feature amplitude but also on inter-modal coupling, frequency similarity, and temporal response consistency. Especially in the context of complex multi-source disturbances, it can automatically enhance the most representative modes and suppress less correlated modes, outputting a highly expressive fused feature vector. The fused feature vector is then input into the network's deep autoencoder structure for compressed representation. The autoencoder, consisting of an encoder and a decoder, uses a nonlinear network to map the high-dimensional fused features into a low-dimensional latent space Z, which is a compressed representation of the tilt state. During the encoding phase, the fused vector undergoes multiple layers of dimensionality reduction and nonlinear mapping, resulting in a low-dimensional embedding representation that preserves the dominant features and response characteristics of all modal information. Subsequently, during the decoding phase, the network attempts to reconstruct the original fused vector from the Z space. The network parameters are optimized by minimizing the reconstruction error (e.g., mean squared error (MSE)) to ensure that the Z space retains the full information representation capability. After training convergence, the Z vector output from the encoder is the tower tilt modal fingerprint representing the structural response mode of the current tower under a specific electrical disturbance.

[0110] In a specific embodiment, the execution step of sending the fused feature vector to the deep autoencoder in the multimodal tilt feature fusion network for compression representation to obtain the tower tilt modal fingerprint may specifically include the following steps:

[0111] The fused feature vector is input into the encoder of the deep autoencoder for dimensionality reduction to obtain the first latent space representation;

[0112] Inputting the first latent space representation into the decoder of the deep autoencoder for reconstruction to obtain a reconstructed feature vector;

[0113] The mean square error (MSE) of the reconstructed feature vector and the fused feature vector is calculated, and the network parameters of the encoder and decoder are optimized by gradient descent based on the MSE to obtain the second latent space representation.

[0114] The contrastive learning loss function is used to enhance the features of the second latent space representation to obtain the tower tilt modal fingerprint.

[0115] Specifically, the fused feature vector is fed into the encoder of a deep autoencoder for dimensionality reduction. The encoder compresses the high-dimensional, complex fused feature vector into a low-dimensional embedding space through multiple layers of nonlinear transformations, obtaining low-dimensional features that can compressively represent the original state distribution. The encoder structure consists of multiple hidden layers. Each layer performs a linear transformation followed by a nonlinear mapping using an activation function, such as ReLU or LeakyReLU, to ensure the expressiveness of high-order features and gradient stability. Through layer-by-layer dimensionality reduction, a compressed representation vector is obtained, namely the first latent space representation. To test the fidelity of this latent space representation—that is, to determine whether it can fully restore the original state information—the first latent space representation is fed into the decoder subnetwork for feature reconstruction. The decoder, symmetrical in structure to the encoder, performs the reverse expansion of the compressed features, gradually restoring the feature dimensions through multiple dimensionality-raising layers, and attempting to reconstruct the fused feature vector of the original input. The entire reconstruction process retains the nonlinear mapping strategy used during encoding, restoring the original information through gradual dimensionality-raising combined with activation functions, and outputs a reconstructed vector with the same dimensionality as the input fused feature vector. Due to the inevitable information compression loss during the encoding and decoding processes, the reconstructed vector deviates from the original fused feature vector. This requires precise comparison and error calculation to assess the effectiveness of the latent space. During the error evaluation phase, the difference between the reconstructed and fused feature vectors in each dimension is calculated, squared, and averaged to form a standard mean squared error loss function, which serves as a measure of the encoder and decoder capabilities at the current neural network state. To optimize network parameters and ensure that the reconstruction result is as close as possible to the original input, network weight parameters are backpropagated and updated based on the mean squared error. Using a gradient descent algorithm or its variants, such as the Adam optimizer, the weights and bias parameters of each layer are adjusted according to the gradient of the error. This gradually reduces the reconstruction error during the training iterations and improves the accuracy of the latent space representation. As the network iterative training converges, the latent space representation output by the encoder not only exhibits good compression capabilities but also fully represents the characteristics of the original structural state, with strong discriminability and embedding consistency. This trained and optimized latent representation is referred to as the second latent space representation. To improve the clustering ability and discrimination boundary of the latent representation, a contrastive learning mechanism is introduced to enhance the latent space. This process takes the second latent space representation as input and uses structural state labels or temporal proximity logic to construct positive and negative pairs of samples. Positive pairs are embedding vectors belonging to the same tilt state or originating from the same physical perturbation type, while negative pairs are derived from structural response samples in different states and perturbation types. A typical contrastive learning loss function, such as NT-Xent or InfoNCE, is then used to calculate the distance between latent vectors using cosine similarity. This approach brings positive pairs closer together in the embedding space and negative pairs further apart, thereby forming a latent space structure that is separable between classes and compact within classes.This optimization process utilizes a batch training approach. In each training batch, multiple sets of positive and negative pairs are constructed from the latent vector set. A fixed temperature coefficient is used to control the scale of the similarity function, defining a contrastive loss function. By minimizing this loss function, the model continuously optimizes the distribution of the structural response fingerprint in the latent space, clustering similar tilt states together while naturally separating different states. Ultimately, clustered regions with clear structure and physical interpretation are formed in the latent space. After training is complete, each compressed and contrast-enhanced vector in the latent space constitutes the tower tilt modal fingerprint.

[0116] In a specific embodiment, the process of executing step S105 may specifically include the following steps:

[0117] Perform abnormal separation on the tower tilt modal fingerprint to obtain the abnormal characteristics of the tower structure;

[0118] Based on the abnormal characteristics of the tower structure, the abnormality scores of the tower foundation area, connection area and superstructure area are calculated respectively to obtain the structural abnormality location vector;

[0119] The structural anomaly location vector is used to model the structural risk propagation path and obtain a risk propagation model. The risk propagation model uses the sensor layout locations as nodes and the actual structural relationship of the tower to determine the topological connection.

[0120] The structural reliability index is calculated based on the risk propagation model and the displacement-stress mapping relationship in the digital twin model to obtain the judgment results of structural tilt and non-structural tilt. The judgment results of structural tilt and non-structural tilt are used to distinguish temporary tilt caused by transient events in the power system from structural tilt caused by foundation erosion and bolt loosening.

[0121] Specifically, the tower tilt modal fingerprint is subjected to anomaly separation. This modal fingerprint is mapped to a historical modal feature database or cluster center set. Departure from the normal state range is determined by calculating metrics such as the Euclidean distance, cosine similarity, or Mahalanobis distance between it and the normal working state cluster center. When the degree of deviation exceeds a set judgment threshold, the fingerprint sample is considered abnormal. The degree of abnormal deviation of each modal component in the fingerprint is simultaneously decomposed to obtain the local anomaly contribution of the inclination component, vibration component, offset component, and settlement component. These anomaly data are then reorganized based on the modal distribution and spatial position relationship to form a tower structural anomaly feature vector with physical positioning significance. This vector represents the abnormal pattern, amplitude, and direction of the current structure in the multimodal response and serves as the basic data for zoning diagnosis. Based on the structural anomaly feature vector and the structural geometric partitioning logic defined in the digital twin model, the entire tower is divided into three functional structural sections: the foundation area, the connection area, and the superstructure area. Based on the correspondence between each modal feature and the area to which the node belongs, the anomaly components are aggregated into their respective structural sub-areas. For example, foundation settlement and vertical acceleration anomalies are primarily attributed to the foundation region, while tower vibration modal shifts and tower head inclination mutations are assigned to the connection region and the superstructure region, respectively. A regional-level anomaly score is then calculated for each structural region using aggregated modal anomaly indicators. This score is expressed quantitatively using methods such as distribution deviation, signal energy mutation value, or feature sparsity change rate. This score is then normalized using statistical thresholds based on historical normal conditions to form a structural anomaly location vector consisting of three numerical components. Each component represents the current anomaly level in the corresponding region and serves as an important input for spatial localization to determine the location of structural damage, damage expansion trend, and dynamic risk level. The structural anomaly location vector is then used to model the structural risk propagation path. Sensor locations are represented as nodes in a graph, with each node representing a structural response unit with monitoring capabilities. Topological connecting edges are defined based on the tower's geometric connection, force transmission path, and material continuity. The presence of an edge indicates a direct physical connection or force transmission channel between two nodes, while edge weights are determined based on connection stiffness, displacement synergy, and vibration coupling strength. Combined with the risk activation value of each node in the anomaly location vector, multi-layer iterative propagation is performed through graph neural networks or message passing algorithms. In each round of propagation, each node propagates its current anomaly state through adjacent nodes, performs weighted aggregation and transfer, and dynamically updates its state vector in combination with edge weights. After several propagation layers, a structural risk propagation model that includes local anomaly amplification, linked response of adjacent nodes, and the impact of the entire tower stability can be formed. This model can reflect whether the current anomaly is contagious, and can also evaluate its impact radius on the structure and the possible induced chain damage path.The risk propagation model is combined with the displacement-stress mapping relationship defined in the digital twin model to map the abnormal state of each node during the propagation process to the corresponding displacement response, which is then converted into a local stress field change. The residual reliability value at the current moment is then calculated for each node based on material strength, fatigue limit, and yield threshold. A weighted average or minimum value strategy is then taken across all nodes of the entire tower structure to obtain a macrostructural reliability index for the entire tower. If the reliability value drops sharply within a short period of time and then shows a steady downward trend, and the risk propagation path shows upward propagation from the foundation or connection area, the anomaly is determined to be structural tilt, i.e., a stability shift caused by structural deterioration factors such as foundation settlement, loose connections, and component damage. If the reliability index changes little, and the risk propagation is limited to short-term transitions in certain modal signals, and the anomaly quickly returns to normal after the external disturbance disappears, the anomaly is attributed to non-structural tilt induced by transient electrical events such as short-circuit current surges, harmonic excitation, and conductor thermal expansion.

[0122] The above describes the method for monitoring the tilt state of a tower in an embodiment of the present invention. The following describes the system for monitoring the tilt state of a tower in an embodiment of the present invention. Figure 2 In one embodiment of the present invention, a system for monitoring the tilt state of a tower includes:

[0123] Modeling module 201, used to perform multi-field digital coupling modeling on power towers, obtain a digital twin model and calculate the electrical load-structure response sensitivity matrix;

[0124] The acquisition module 202 is used to acquire a multimodal sensing feature data set including tower inclination time series data, insulator string offset, tower micro-vibration spectrum, and foundation settlement gradient based on the digital twin model;

[0125] An analysis module 203 is configured to perform a time-lag correlation analysis on electrical events and structural responses based on a multimodal sensing feature data set and an electrical load-structural response sensitivity matrix to obtain an electrical-structural mapping matrix;

[0126] The processing module 204 is used to input the electrical-structural mapping matrix into the multi-modal tilt feature fusion network for tensor decomposition processing to obtain the tower tilt modal fingerprint;

[0127] The output module 205 is used to perform abnormal separation processing on the tower foundation area, the connection area and the upper structure area according to the tower tilt modal fingerprint, and output the determination results of structural tilt and non-structural tilt.

[0128] Through the collaborative efforts of these components, a multi-field coupled digital twin model of power towers was constructed, enabling a precise mapping of electrical loads and structural responses. This model transcends the limitations of traditional monitoring methods that only consider mechanical factors and effectively correlates the influencing mechanisms between conductor current-carrying status and tower tilt. A multimodal sensor deployment scheme guided by the digital twin model enables comprehensive acquisition and preprocessing of tower tilt time-series data, insulator string offsets, tower microvibration spectra, and foundation settlement gradients, improving the quality and relevance of monitoring data. Asynchronous electrical-structural event correlation analysis establishes a time-delayed correlation between electrical events and structural responses. A multimodal tilt feature fusion network performs tensor decomposition processing, extracting a highly discriminative modal fingerprint of tower tilt. This fingerprint can accurately distinguish temporary tilt caused by power system-specific transient events such as lightning strikes and short-circuit current surges from structural tilt caused by foundation erosion and loose bolts. Anomaly separation based on a structural risk propagation path diagram enables refined risk assessment of the tower foundation, connection areas, and superstructure, effectively reducing false alarm rates and improving the accuracy and timeliness of early warnings. A complete tower tilt status monitoring and early warning decision-making closed-loop system has been established. By deeply integrating the unique physical factors and structural engineering characteristics of the power system, it provides substantial guarantees for the safe operation of the power grid and reduces the risk of tower collapse accidents.

[0129] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0130] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a tower tilt status monitoring device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0131] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions described in the above embodiments can still be modified, or some of the technical features thereof can be replaced by equivalents. 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 invention.

Claims

1. A method for monitoring the tilt state of a tower, characterized in that: include: Conduct multi-field digital coupling modeling of power towers to obtain a digital twin model and calculate the electrical load-structure response sensitivity matrix; Based on the digital twin model, a multimodal sensing feature data set including tower inclination time series data, insulator string offset, tower body micro-vibration spectrum and foundation settlement gradient is collected; performing a time-lag correlation analysis on electrical events and structural responses based on the multimodal sensing feature data set and the electrical load-structural response sensitivity matrix to obtain an electrical-structural mapping matrix; Inputting the electrical-structural mapping matrix into a multimodal tilt feature fusion network for tensor decomposition processing to obtain a tower tilt modal fingerprint; Performing abnormal separation processing on the tower foundation area, the connection area and the upper structure area according to the tower tilt modal fingerprint, and outputting the structural tilt and non-structural tilt determination results, including: performing abnormal separation on the tower tilt modal fingerprint to obtain the tower structural abnormality characteristics; Based on the abnormal characteristics of the tower structure, the abnormality scores are calculated for the tower foundation area, connection area and upper structure area respectively to obtain the structural abnormality location vector; the structural abnormality location vector is used to model the structural risk propagation path to obtain a risk propagation model, which uses the sensor layout location as the node and the actual structural relationship of the tower to determine the topological connection; based on the risk propagation model and the displacement-stress mapping relationship in the digital twin model, the structural reliability index is calculated to obtain the structural tilt and non-structural tilt judgment results, which are used to distinguish between temporary tilt caused by transient events in the power system and structural tilt caused by foundation erosion and loose bolts.

2. The tower tilt state monitoring method according to claim 1, characterized in that: The multi-field digital coupling modeling of the power tower is performed to obtain a digital twin model and calculate the electrical load-structure response sensitivity matrix, including: The height, cross arm width, tower segment size, connection node coordinates, conductor connection point locations, and foundation burial depth of the power tower are collected to obtain a tower geometric parameter set. Inputting the tower geometric parameter set into a three-dimensional finite element modeling system for meshing and material property assignment to obtain a digital skeleton model of the tower; Based on the digital skeleton model of the tower, a multi-physics field coupling calculation is performed on the conductor current-carrying temperature field and the tower structure displacement field to obtain a digital twin model; The structural response sensitivity caused by current changes is quantified according to the digital twin model to obtain an electrical load-structural response sensitivity matrix.

3. The tower tilt state monitoring method according to claim 2, characterized in that: The multi-physics field coupling calculation of the conductor current-carrying temperature field and the tower structure displacement field is performed based on the tower digital skeleton model to obtain a digital twin model, including: Based on the digital skeleton model of the tower, a thermodynamic model is performed on the current parameters of the conductor to obtain a conductor current-carrying temperature field, which is used to quantify the impact of current changes on conductor temperature distribution; Based on the digital skeleton model of the tower, a finite element analysis is performed on the deformation characteristics of the tower structure under the action of external loads to obtain a displacement field of the tower structure. The displacement field of the tower structure is used to describe the displacement-stress mapping relationship of each target node of the tower; Based on the digital skeleton model of the tower, the soil bearing capacity, hydrological conditions and seismic subsidence coefficient of the geological environment in which the tower foundation is located are parameterized to obtain a geological parameter tensor; The conductor current-carrying temperature field, the tower structure displacement field and the geological parameter tensor are subjected to multi-physics field coupling calculation to construct a digital twin model.

4. The tower tilt state monitoring method according to claim 1, characterized in that: The multimodal sensing feature data set collected based on the digital twin model includes tower inclination time series data, insulator string offset, tower micro-vibration spectrum and foundation settlement gradient, including: Calculating tower structure sensitivity distribution information based on the digital twin model, and determining optimal monitoring point locations based on the tower structure sensitivity distribution information; Collecting the original sensor signal matrix of the tilt angle time series data, the insulator string offset, the tower body micro-vibration spectrum and the foundation settlement gradient on the power tower according to the optimal monitoring point position; Comparing and calibrating the original sensor signal matrix with the theoretical response value predicted by the digital twin model to obtain a calibrated preprocessed signal set; Multi-scale feature extraction is performed on the calibrated preprocessed signal set and fused with the feature template generated by the digital twin model to obtain a multimodal sensing feature data set.

5. The tower tilt state monitoring method according to claim 1, characterized in that: The step of performing a time-lag correlation analysis on electrical events and structural responses based on the multimodal sensing feature data set and the electrical load-structure response sensitivity matrix to obtain an electrical-structure mapping matrix includes: Matching the multimodal sensing feature data set with the electrical load-structural response sensitivity matrix to obtain a structural response feature set; Correlating the phase current, inter-phase imbalance, power factor change rate, and system harmonic content in the power tower operation data with the electrical load-structure response sensitivity matrix to obtain an electrical event feature set; A time-delay convolution operation and a correlation degree calculation are performed based on the structural response feature set and the electrical event feature set to obtain an electrical-structural mapping matrix, wherein the electrical-structural mapping matrix includes a tilt correlation vector, an offset correlation vector, a vibration correlation vector, and a settlement correlation vector.

6. The tower tilt state monitoring method according to claim 5, characterized in that: The step of inputting the electrical-structural mapping matrix into a multimodal tilt feature fusion network for tensor decomposition processing to obtain a tower tilt modal fingerprint includes: Inputting the electrical-structural mapping matrix into the four-channel parallel feature extraction layer in the multimodal tilt feature fusion network for processing, thereby obtaining a tilt angle feature vector, an offset feature vector, a vibration feature vector, and a settlement feature vector; Performing tensor construction and decomposition on the tilt angle eigenvector, the offset eigenvector, the vibration eigenvector, and the settlement eigenvector to obtain a potential feature tensor; Inputting the latent feature tensor into the cross-modal attention fusion layer in the multimodal tilted feature fusion network for feature weighted summation to obtain a fused feature vector; The fused feature vector is sent to the deep autoencoder in the multimodal tilt feature fusion network for compression representation to obtain the tower tilt modal fingerprint.

7. The tower tilt state monitoring method according to claim 6, characterized in that: The step of sending the fused feature vector to a deep autoencoder in the multimodal tilt feature fusion network for compression representation to obtain a tower tilt modal fingerprint includes: Inputting the fused feature vector into the encoder of the deep autoencoder for dimensionality reduction to obtain a first latent space representation; Inputting the first latent space representation into the decoder of the deep autoencoder for reconstruction to obtain a reconstructed feature vector; Calculating a mean square error between the reconstructed feature vector and the fused feature vector, and optimizing network parameters of the encoder and the decoder by a gradient descent method based on the mean square error to obtain a second latent space representation; The contrastive learning loss function is used to perform feature enhancement processing on the second latent space representation to obtain the tower tilt modal fingerprint.

8. A tower tilt status monitoring system, characterized in that: For implementing the tower tilt state monitoring method according to any one of claims 1 to 7, the tower tilt state monitoring system comprises: Modeling module, used to perform multi-field digital coupling modeling of power towers, obtain digital twin models and calculate electrical load-structure response sensitivity matrices; An acquisition module is used to acquire a multimodal sensing feature data set including tower inclination time series data, insulator string offset, tower body micro-vibration spectrum and foundation settlement gradient based on the digital twin model; an analysis module, configured to perform a time-lag correlation analysis on electrical events and structural responses based on the multimodal sensing feature data set and the electrical load-structural response sensitivity matrix to obtain an electrical-structural mapping matrix; A processing module is used to input the electrical-structural mapping matrix into a multimodal tilt feature fusion network for tensor decomposition processing to obtain a tower tilt modal fingerprint; An output module is used to perform abnormal separation processing on the tower foundation area, connection area and upper structure area according to the tower tilt modal fingerprint, and output the structural tilt and non-structural tilt judgment results, including: performing abnormal separation on the tower tilt modal fingerprint to obtain the tower structural abnormality characteristics; based on the tower structural abnormality characteristics, calculating the abnormality scores for the tower foundation area, connection area and upper structure area respectively to obtain the structural abnormality location vector; using the structural abnormality location vector to model the structural risk propagation path to obtain a risk propagation model, wherein the risk propagation model uses the sensor layout position as the node and the actual structural relationship of the tower to determine the topological connection; based on the risk propagation model and the displacement-stress mapping relationship in the digital twin model, a structural reliability index is calculated to obtain the structural tilt and non-structural tilt judgment results, and the structural tilt and non-structural tilt judgment results are used to distinguish between temporary tilt caused by transient events in the power system and structural tilt caused by foundation erosion and loose bolts.

Images