A method and system for assessing the operational health status of power transmission towers
By combining multi-channel vibration signal processing and sparse environmental sensing parameters, and utilizing solar position calculation and structural self-shading ray tracing, a rod-level steady-state thermal balance equation is established. A physical constraint variable decoupling autoencoder network is used to remove temperature interference, thus solving the stiffness distortion problem caused by non-uniform temperature field in the dynamic monitoring of power transmission towers and realizing accurate identification and assessment of early damage.
Patent Information
- Application Number
- CN202611369113.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-04
- Publication Date
- 2026-11-17
AI Technical Summary
In existing monitoring of the dynamic vibration of power transmission towers, the nonlinear distortion of structural stiffness caused by the non-uniform temperature field in space masks the characteristics of weak physical damage. Furthermore, under limited temperature measurement hardware conditions, it is difficult to accurately isolate temperature interference, resulting in a high false alarm rate and a high missed alarm rate in the monitoring system.
By acquiring multi-channel vibration acceleration response signals and sparse environment sensing parameters, multivariate variational mode decomposition and random subspace identification are performed. Combined with solar position calculation and structural self-shading ray tracing, a rod-level steady-state thermal balance equation is established. A physical constraint variational decoupling autoencoder network is used to remove temperature interference, and a temperature-independent damage index is output for health assessment.
Under conditions of a very small number of environmental measurement points, the temperature field with component-level resolution is reconstructed, and the nonlinear modal drift caused by non-uniform temperature internal stress is eliminated. This improves the accuracy of early weak damage identification and the ability to resist environmental interference, reduces hardware and power consumption requirements, and enables reliable health status assessment.
Smart Images

Figure CN123116701A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power transmission line condition monitoring and structural health assessment technology, and in particular to a method and system for assessing the operational health status of power transmission towers. Background Technology
[0002] Transmission towers are the main supporting structures for overhead transmission lines, carrying both conductors and ground wires. They have large spatial spans, long service lives, and are widely distributed in coal mining subsidence areas, areas with high groundwater levels, seasonally frozen soil areas, and zones with frequent geological activity. Under long-term wind vibration, icing, foundation deformation, and alternating loads, towers may experience deterioration phenomena such as uneven foundation settlement, tower tilting, loss of preload on node bolts, and local buckling of main materials. Such deterioration often lacks visible appearance in its early stages, but once it reaches a macroscopic stage, it can cause sudden changes in conductor tension and insufficient electrical safety distances, potentially leading to line breaks or even tower collapse. Therefore, continuous online condition monitoring of towers without interrupting line operation is of significant engineering importance.
[0003] Early monitoring methods primarily fell under the category of statics, including displacement measurement based on visual images, tilt sensor over-limit alarms, and basic deformation observation based on remote sensing interferometry. These methods share the characteristic of using changes in macroscopic geometric quantities as criteria, only providing a response when significant observable deformation has occurred, making it difficult to intervene in early, minor damage. In contrast, vibration monitoring technology based on dynamic response is more sensitive to changes in the internal state of the structure. The principle is that any physical damage to the structure ultimately manifests as a change in local mass or stiffness matrix, and even a small change in the stiffness matrix inevitably causes drift in modal parameters such as natural frequencies and mode shapes. Therefore, by deploying accelerometers at key nodes of the tower to collect weak vibration responses under natural wind loads, and using signal processing and system identification methods such as variational modal decomposition and random subspace identification to extract modal parameters, and then comparing these parameters with baseline values under healthy conditions, the damage status of the structure can be inverted.
[0004] However, the diagnostic logic based on the ideal mechanical model encountered serious obstacles when applied to engineering sites. Temperature is the most common and severe disturbance load in the field environment. It not only changes the elastic modulus of steel but also alters the contact state of bolt nodes and the stiffness of boundary constraints through thermal expansion and contraction, introducing considerable internal temperature stresses into the structure. Long-term field measurements show that the natural frequency drift caused by temperature cycling within a single day often exceeds the frequency change corresponding to a 15% stiffness degradation in the structure. Existing publicly available technologies, when dealing with this disturbance, either limit temperature compensation to hardware correction of the temperature drift of the probe and transmitter itself, or rely on linear regression and covariate detrending processing based on the readings of a single ambient temperature sensor. Both approaches assume that the entire tower is in a uniform temperature field and that there is a simple linear relationship between stiffness and temperature. Transmission towers are actually spatial trusses composed of hundreds or even thousands of angle steels interlocked in three-dimensional space. Under unilateral sunlight, the members facing the sun rapidly heat up and expand, while the members facing away from the sun remain at a relatively low temperature due to the tower's own shading. At any given moment, the temperature difference between different members can reach more than 10°C. This non-uniform thermal expansion induces complex bending moments and axial internal forces within the statically indeterminate structure, causing a highly nonlinear distortion in the overall stiffness matrix. The resulting modal variations are sufficient to completely mask early bolt loosening or minor foundation settlement. Compensation methods based on the assumption of a uniform temperature field and linear relationship fail under these conditions, leading to a high false alarm rate and a high missed alarm rate in the actual operation of the monitoring system.
[0005] Attempting to directly obtain the temperature field distribution by installing temperature measuring elements on every critical member is impractical due to the construction costs of towers tens of meters high, the power consumption budget of nodes under solar power conditions, and the channel capacity of low-power wireless networks when hundreds of nodes are concurrently uploading. Therefore, quantifying and isolating the nonlinear modal drift caused by the spatially non-uniform temperature field on the tower stiffness matrix under hardware constraints of configuring only a very small number of environmental measuring points, and thus obtaining health characteristics that only reflect the physical damage to the structure itself, has become a pressing technical problem to be solved in this field. Summary of the Invention
[0006] The purpose of this invention is to provide a method and system for assessing the operational health status of power transmission towers, in order to solve the technical problem in existing power transmission tower dynamic vibration monitoring where nonlinear distortion of structural stiffness caused by non-uniform temperature fields in space masks weak physical damage characteristics, and where it is difficult to accurately isolate the temperature interference under limited temperature measurement hardware conditions.
[0007] To achieve the above objectives, the present invention provides a method for assessing the operational health status of power transmission towers, comprising the following steps: The system acquires multi-channel vibration acceleration response signals and sparse environment sensing parameters of the monitored transmission tower, including environmental reference temperature, wind speed, wind direction, and total solar irradiance on the horizontal plane. Multivariate variational mode decomposition is performed on the multi-channel vibration acceleration response signal to purify the vibration signal, and the natural frequencies and mode shapes are extracted by covariance-driven random subspace identification. After normalization, a hybrid mode feature vector is constructed. Based on the latitude and longitude coordinates of the transmission tower and the current sampling time, the solar spatial position vector is calculated. According to the azimuth angle of the tower installation, the pre-set three-dimensional simplified truss topology dictionary of the transmission tower is transformed to the local horizontal coordinate system and structural self-shading ray tracing is performed to determine the direct sunlight exposure coefficient of each member. Combined with sparse environment perception parameters, the total solar irradiance on the horizontal plane is separated into direct and diffuse values. The member-level steady-state thermal balance equation is established and solved nonlinearly. Then, after spatial interpolation residual compensation, the spatial non-uniform temperature distribution feature vector at each node of the entire tower is output. The hybrid modal feature vectors are input into a pre-trained physical constraint variational decoupled autoencoder network. The hybrid modal feature vectors are decoupled and mapped to mutually orthogonal environmental latent spaces and health latent spaces through a branch encoder, and the pure health latent variables representing the true state of the structure are output. In the training phase, the physical constraint variational decoupled autoencoder network uses the spatial non-uniform temperature distribution feature vector as a physical mentor signal, and uses cross-domain physical correlation constraints and spatial orthogonal independent constraints to drive the environmental latent space to absorb the modal nonlinear variation caused by temperature fluctuations. The Mahalanobis distance of the current pure health latent variable relative to the preset health baseline distribution is calculated as the temperature-independent damage index. The temperature-independent damage index is compared with the first-level warning threshold and the second-level alarm threshold adaptively calibrated based on the health baseline stage sample distribution. Combined with the continuous multi-window judgment logic, the operation health status assessment and hierarchical early warning linkage of the transmission tower are realized.
[0008] Through the above steps, a temperature field with component-level resolution can be reconstructed using only extremely sparse environmental measurement points. Furthermore, by leveraging a physically constrained decoupling network, nonlinear modal drift caused by non-uniform temperature internal stress can be removed from the feature source, thereby improving the accuracy of early weak damage identification and the ability to resist environmental interference.
[0009] As a preferred embodiment of the present invention, the multi-channel vibration acceleration response signal is acquired by multiple triaxial accelerometers arranged on the transmission tower. The measurement point positions of each triaxial accelerometer are determined based on the finite element model of the transmission tower system, using the effective independence method combined with the information entropy corresponding to the posterior covariance of the modal parameters as the termination criterion. When performing multivariate variational mode decomposition on the multi-channel vibration acceleration response signal, each channel shares the same set of center frequencies. The number of decomposed modes and the secondary penalty factor of the multivariate variational mode decomposition are jointly optimized using a particle swarm optimization algorithm. The fitness function uses the minimum envelope entropy of each modal component of the reference channel as the main term, and applies a penalty term to cases where the ratio of adjacent center frequencies is less than a set threshold for over-decomposition. This scheme reduces the number of measurement points while ensuring the recognizability of vibration modes. The decomposition method with shared center frequencies maintains the amplitude ratio and phase relationship between channels and suppresses modal aliasing and over-decomposition.
[0010] As a preferred technical solution of the present invention, when using covariance-driven random subspace identification to extract natural frequencies and mode shapes, a block Topplitz matrix is constructed based on the output covariance of each order. The principal subspace is truncated through singular value decomposition, and the discrete state matrix is solved based on shift invariance. Then, the natural frequencies, damping ratios, and mode shapes are calculated from the eigenvalues and eigenvectors of the discrete state matrix. Steady-state diagrams are constructed at different system orders, and spurious poles are eliminated based on a triple screening condition of relative frequency deviation, relative damping ratio deviation, and modal confidence criterion values. The extracted mode shapes are normalized according to unit modulus, and after unifying the sign polarity with the horizontal component along the line at the end of the tower top ground wire crossarm as a reference, they are expanded and cascaded with each natural frequency to form a mixed modal eigenvector. This scheme filters out spurious mathematical poles caused by noise and eliminates the uncertainty of mode shape amplitude scaling and sign, ensuring the comparability of features across time periods.
[0011] As a preferred technical solution of the present invention, when calculating the solar spatial position vector, the solar declination angle and hour angle are calculated based on the annual day sequence of the sampling time, the longitude of the central meridian of the time zone and the local longitude, and the true solar time. The solar altitude angle and solar azimuth angle in the local horizontal coordinate system are then calculated. The solar azimuth angle is calculated with due north as zero degrees and clockwise as the positive direction, and the unit vector of the light propagation direction is determined accordingly. When performing structural self-shading ray tracing, the node coordinates in the topology dictionary are converted to the local horizontal coordinate system according to the tower installation azimuth angle. Several surface sampling points are set along the axis of each member. Rays are projected from each surface sampling point along the reflected light direction. Non-intersecting members are quickly eliminated by axis-aligned bounding box hierarchical cross detection, and the candidate members are accurately cross-detected. If the ray intersects with other members, the sampling point is determined to be in self-shade. The proportion of unshaded sampling points on a single member is counted as the direct light exposure coefficient of that member. This scheme uses astronomical laws and spatial geometric relationships to calculate the occlusion state, so that the spatial resolution of the temperature field depends on the discretization accuracy of the topological dictionary and is decoupled from the number of physical measurement points.
[0012] As a preferred technical solution of the present invention, when separating the direct and diffuse total solar irradiance on the horizontal plane and establishing the steady-state thermal balance equation for the rod level, the total solar irradiance on the horizontal plane is separated into normal direct irradiance, sky diffuse irradiance, and ground reflected irradiance according to the clear sky index. The direct light exposure coefficient is applied only to the direct irradiance term to calculate the heat flux absorbed by the rod. A nonlinear thermal balance equation including convective heat transfer and radiative heat dissipation terms is established, wherein the convective heat transfer coefficient is extracted based on the angle between the wind vector and the rod axis, and the lateral sweep perpendicular to the rod axis has a value. The effective airflow component is included, along with the leeward / frontal correction coefficients obtained from the cross-section of the enclosing box wind direction. The Newton-Raphson iterative method is used to solve for the surface temperature of each member element by element. After the surface temperature of each member is weighted by area and allocated to each node, during the long-term operation phase, the measured residuals of the reference temperature measurement points arranged on the shaded side of the tower foot and the sunlit side of the tower top are used, combined with the semi-variogram parameters pre-fitted during the calibration period, to establish a Kriging interpolation equation system to estimate the temperature residuals of each node in the entire tower and compensate for the calculated temperature, generating a spatially non-uniform temperature distribution characteristic vector. This scheme applies the exposure coefficient only to the direct sunlight component, avoiding the elimination of scattered heat gain from the shaded side members, and compensates for the systematic deviations introduced by material parameters and local micrometeorological conditions through residual interpolation at a few reference measurement points.
[0013] As a preferred embodiment of the present invention, the physically constrained variational decoupled autoencoder network includes a shared backbone coding layer, independently configured environmental latent variable branches and health latent variable branches, and a reconstruction decoder. The composite loss function used during the training phase includes: a reconstruction loss term, used to measure the squared error between the input mixed-modality feature vector and the decoder reconstruction output; a divergence penalty term, used to constrain the environmental and health latent variables to follow a standard normal distribution prior; a cross-domain physical correlation constraint term, used to extract principal components whose cumulative variance contribution rate reaches a set threshold from the spatially non-uniform temperature distribution feature vector, and to apply a reverse penalty to the Pearson correlation coefficient between each dimension component of the environmental latent space and the corresponding temperature principal component in the batch dimension; and a spatial orthogonality independence constraint term, used as a penalty amount, to force the environmental and health latent spaces to be statistically orthogonal. This scheme achieves structural isolation between the two subspaces through branched connections and completes attribute allocation using the physical field as a mentor signal, suppressing residual leakage of temperature information into the health dimension.
[0014] As a preferred embodiment of the present invention, the method further includes online out-of-distribution detection and incremental model update steps: During online inference, the reconstruction error of the mixed modality feature vector at the current moment and the Mahalanobis distance of the environmental latent variables relative to the training set distribution are calculated. If either indicator exceeds the preset high quantile threshold of the training set, the current moment is marked as an out-of-distribution sample, and its evaluation result is not included in the early warning decision. When the proportion of out-of-distribution samples within a set time window exceeds a preset proportion threshold, incremental retraining of the model using newly added healthy time period data as samples is triggered, and the weights of the decoupled autoencoder network of physical constraint variation are updated simultaneously. This scheme can still maintain the reliability of the decoupling results when micro-meteorological conditions or structural states undergo slow migration or online input deviates from the training distribution.
[0015] As a preferred embodiment of the present invention, when calculating the temperature-independent damage index, the baseline mean vector and baseline covariance matrix are calculated using long-term pure healthy latent variable samples accumulated during the healthy operation phase. A trace-based ridge regularization term is applied when inverting the baseline covariance matrix, and its condition number is continuously monitored. During long-term operation, a sliding time window is used to periodically update the baseline mean vector and baseline covariance matrix using samples determined to be in a healthy state, and a limit constraint based on the baseline standard deviation is applied to the mean drift of each update. This scheme avoids numerical divergence caused by ill-conditioned inversion of the covariance matrix and prevents progressive, minor damage from being absorbed by the synchronous translation of the baseline.
[0016] As a preferred technical solution of the present invention, the adaptive determination method of the multi-level warning threshold is as follows: based on the temperature-independent damage index samples accumulated during the healthy baseline stage, the probability density distribution is fitted using kernel density estimation and combined with tail extremum extrapolation, and the first-level warning threshold and the second-level alarm threshold are adaptively calibrated using the corresponding high cumulative probability quantile values respectively; the continuous multi-window persistence determination logic includes: when the temperature-independent damage index is less than or equal to the first-level warning threshold, it is determined to be in healthy operation, and the first persistence counter and the second persistence counter are cleared; when the temperature-independent damage index is greater than the first-level warning threshold and less than or equal to the second-level alarm threshold, the first persistence counter is incremented, and when the increment reaches a preset first count value, A Level 1 warning is triggered, and the acquisition cycle of the vibration acceleration signal is shortened for encrypted acquisition. When the temperature-independent damage index exceeds the Level 2 alarm threshold, the first and second persistent counters accumulate synchronously. Encrypted acquisition is immediately executed upon the first exceedance. When the second persistent counter continuously accumulates to a preset second count value under encrypted acquisition, a Level 2 alarm is triggered. During encrypted acquisition, if the temperature-independent damage index falls below the Level 1 warning threshold, both persistent counters are reset to zero and the normal acquisition cycle is restored. If it falls between the Level 1 warning threshold and the Level 2 alarm threshold, the second persistent counter stops accumulating and is reset to zero after failing to exceed the Level 2 alarm threshold again within a set number of consecutive windows. This scheme ensures that the decision logic covers the entire range of index values, balances the ability to capture weak damage with the suppression of transient gust interference, and controls the duration of encrypted acquisition.
[0017] This invention also provides a power transmission tower operation health status assessment system, including a sensing fusion layer, an edge resolution layer, and a cloud assessment layer. The sensing fusion layer is located on the monitored power transmission tower body and includes a vibration acquisition unit, an environmental sensing unit, and a data aggregation and transmission unit. It is used to acquire multi-channel vibration acceleration response signals and sparse environmental sensing parameters, and then aggregate and transmit them after synchronization via a unified time slot. The edge resolution layer is installed on the tower base and includes a signal purification unit, a modal identification unit, a temperature field reconstruction unit, and a feature uploading unit. The signal purification unit purifies vibration signals through multivariate variational mode decomposition. The modal identification unit extracts natural frequencies and mode shapes through covariance-driven random subspace identification and constructs a mixed modal feature vector. The temperature field reconstruction unit calculates the instantaneous solar spatial position vector, performs structural self-shading ray tracing based on a pre-set three-dimensional simplified truss topology dictionary of the power transmission tower to determine the direct sunlight exposure coefficient of each member, and combines... The environmental perception parameters are used to solve the steady-state thermal balance equation at the rod level and perform spatial interpolation compensation to reconstruct the spatial non-uniform temperature distribution feature vector at each node of the entire tower. The feature upload unit is used to upload the mixed-mode feature vector and the spatial non-uniform temperature distribution feature vector. The cloud evaluation layer is deployed on a server or data center and includes a decoupled network unit, a health measurement unit, and an early warning linkage unit. The decoupled network unit is equipped with a physically constrained variational decoupled autoencoder network, which is used to train the spatial non-uniform temperature distribution feature vector after principal component dimensionality reduction as a mentor signal for cross-domain correlation and orthogonal constraint training. During online inference, the mixed-mode feature vector is decoupled to output a pure health latent variable. The health measurement unit is used to calculate the Mahalanobis distance of the pure health latent variable relative to the baseline distribution to obtain the temperature-independent damage index. The early warning linkage unit is used to determine the temperature-independent damage index through adaptive threshold and continuous multi-window persistence logic, and output graded early warning and scheduling linkage instructions.
[0018] Compared with the prior art, the present invention has the following beneficial effects.
[0019] Based on solar position calculation and structural self-occlusion ray tracing, combined with steady-state thermal balance solution at the member level and spatial interpolation correction of sparse measured residuals, the temperature distribution of all members and nodes in the tower can be calculated from geometric topology and celestial motion laws. Except for a small number of temporary temperature measurement points added during the commissioning calibration period for calibration of interpolation parameters and empirical constants of convective heat transfer, only 2 to 4 reference environmental measurement points need to be retained in the entire tower during the long-term operation phase. There is no need to lay dense temperature measurement arrays and supporting cables inside the tower, which reduces the investment in hardware materials, construction and subsequent maintenance. The spatial resolution of temperature distribution depends on the discrete accuracy of the topology dictionary rather than the number of physical measurement points. The reduction in the number of sensing nodes also reduces the power consumption of edge side acquisition and the concurrent pressure of wireless channels, which is conducive to extending the non-intervention operation time of field monitoring stations powered by solar energy and batteries.
[0020] The non-uniform temperature distribution feature vector obtained from the above reconstruction is used as the loss function of the physical prior embedded variational autoencoder. Through the combined effect of cross-domain correlation constraints and spatial orthogonality constraints, the latent variables output by the branch encoder are divided into environmental components that are highly correlated with the temperature field and health components that are statistically orthogonal to them. This achieves the separation of thermally induced modal drift and structural damage response at the feature level. Compared with the linear detrending method that relies on single-point temperature, this mechanism does not assume a linear relationship between temperature and stiffness. Therefore, even under conditions where the diurnal temperature difference reaches 30°C and there is strong unilateral sunlight, it can still suppress the environmental components in frequency and mode drift, allowing weak damage features such as early bolt loosening, local buckling of main materials, and minor uneven settlement of the foundation to be preserved. Combined with out-of-distribution detection based on the reconstruction error and the distribution of environmental latent variables, samples outside the model coverage are separately marked and do not participate in the decision, giving the evaluation conclusions a credibility indicator.
[0021] Within a clean, healthy hidden space after eliminating environmental interference, a temperature-independent damage index is constructed using Mahalanobis distance. Two-level thresholds are adaptively determined using kernel density estimation and tail extremum extrapolation, ensuring that the decision threshold does not rely on universally applicable empirical values across towers but is automatically calibrated based on the actual operating conditions of each tower. Combined with a continuous window persistence criterion and counter reset rules, this approach maintains low-damage sensitivity while suppressing erroneous actions caused by transient disturbances, and the decision logic covers the entire range of index values. The long-term sequence of this index reflects the gradual degradation process of structural performance. Combined with tiered early warning and the linkage with external inspection resources, it provides a quantitative basis for the maintenance department's shift from planned maintenance to condition-based maintenance. Attached Figure Description
[0022] Figure 1 This is a schematic diagram of the overall architecture of the power transmission tower operation health status assessment system provided in an embodiment of the present invention; Figure 2 This is a flowchart illustrating the method for assessing the operational health status of power transmission towers provided in an embodiment of the present invention.
[0023] Figure reference numerals: 100, Perception Fusion Layer; 110, Vibration Acquisition Unit; 120, Environmental Perception Unit; 130, Data Aggregation and Transmission Unit; 200, Edge Resolution Layer; 210, Signal Purification Unit; 220, Modal Identification Unit; 230, Temperature Field Reconstruction Unit; 240, Feature Upload Unit; 300, Cloud Evaluation Layer; 310, Decoupling Network Unit; 320, Health Measurement Unit; 330, Early Warning and Linkage Unit. Detailed Implementation
[0024] To make the objectives, technical solutions, and implementation processes of this invention clearer, the specific embodiments of this invention will be further described below with reference to the accompanying drawings. The embodiments described below are merely preferred implementations of this invention and are not intended to limit the scope of protection of this invention. Equivalent substitutions, parameter adjustments, and local modifications made by those skilled in the art to the following embodiments without departing from the technical concept of this invention should all fall within the scope of protection of this invention.
[0025] like Figure 1 As shown in the figure, the power transmission tower operation health status assessment system provided in this embodiment is divided into a perception fusion layer 100, an edge resolution layer 200, and a cloud assessment layer 300 from bottom to top according to computing power allocation and data flow. The three layers are connected to a wide area network via wireless communication links. The perception fusion layer 100 is deployed on the monitored power transmission tower body and is responsible for low-power acquisition of heterogeneous physical quantities; the edge resolution layer 200 is installed in the tower base control box and is responsible for local resolution of high-frequency waveforms and estimation of spatial physical quantities; the cloud assessment layer 300 is deployed on the power grid side server or data center and is responsible for feature decoupling and status determination. The essence of this layered approach is to sink the processing of the raw waveform, which has the largest amount of data, to the tower base to complete the process, so that only highly compressed feature vectors need to be uploaded via the public network, thereby maintaining the continuous operation of the system in the bandwidth-constrained field environment.
[0026] The perception fusion layer 100 includes a vibration acquisition unit 110, an environmental perception unit 120, and a data aggregation and transmission unit 130. The vibration acquisition unit 110 consists of several triaxial microelectromechanical accelerometers, and the positions of its measurement points are pre-optimized using the effective independence method combined with the information entropy criterion. During optimization, the preceding degrees of freedom of the candidate system are first calculated using the finite element model of the rod system established from the design drawings. Order mode matrix The candidate measurement point set includes the end of the ground wire crossarm at the top of the tower, the root of the crossarm of each conductor layer, the main material nodes at the cross-section of the tower body, and the upper nodes of the tower legs, totaling 30 to 50 points; the Fischer information matrix is calculated. and effective independent distribution vector The degrees of freedom with the smallest values are successively eliminated and recalculated; the information entropy corresponding to the posterior covariance of the modal parameters is used. As a termination criterion, elimination stops when the increase in information entropy caused by a certain elimination exceeds 5% of the initial value. The remaining measurement points at this point represent the optimized layout. For typical goblet-shaped or cat-head-shaped straight towers, the number generally converges to 4–8. The range of the triaxial microelectromechanical accelerometer is ±2g, and the broadband noise density is no higher than 100 μg / √Hz. Considering that such devices are dominated by flicker noise in the low-frequency band, devices with a noise inflection point frequency below 0.5 Hz and an equivalent input noise of no more than 300 μg in the 0.5–20 Hz bandwidth should also be selected, with a minimum resolvable acceleration amplitude better than 5 × 10⁻⁶. -4 m / s 2 It can cover the main modal frequency band of such tall truss structures from 0.5 to 20 Hz; sampling frequency Take 100 Hz, sampling time interval s.
[0027] The configuration of the environmental sensing unit 120 follows the principle of extreme sparseness, with only one temperature and humidity sensor with a temperature measurement error of no more than ±0.2℃ installed on the shaded side of the tower base and one on the sun-facing side of the tower top. A sensor with a measurement range of 0–1500 W / m is installed at the tower top. 2 The tower is equipped with a total radiation meter and one ultrasonic anemometer / wind direction sensor. During long-term operation, the total number of environmental measurement points across the entire tower is controlled to be no more than four, with a sampling interval of 1 minute. The total radiation meter outputs the total solar irradiance on the horizontal plane. The data aggregation and transmission unit 130 adopts a low-power wireless networking method with time-slot synchronization. Under the constraint of the timing signal, each sensing node controls the time synchronization error to within 1ms, and adds a unified timestamp to each frame of data before aggregating it to the edge resolution layer 200 to ensure strict alignment of the vibration response with the environmental scalar on the time axis. The sensing fusion layer 100 is powered by a solar panel and a battery. The vibration acquisition unit 110 operates according to a duty cycle, triggering once per hour in normal mode, with a single continuous high-frequency sampling of 10 minutes, and entering sleep mode during the rest of the time.
[0028] The edge resolution layer 200 is an industrial-grade gateway integrating a multi-core processor, a vector computing acceleration unit supporting single-instruction multiple-data floating-point operations, and a micro-time series database. Its processor clock speed is no less than 1.5 GHz. The reason for not configuring neural network inference acceleration components is that the multivariate variational mode decomposition, random subspace identification, bounding box ray intersection, and Newton iteration solutions handled by this layer are all floating-point intensive numerical operations rather than neural network inference. Such components are only needed in the variant implementation described later, where the decoupled network is deployed in a lower-level configuration. The edge resolution layer 200 has three types of pre-built key prior data. The first is a three-dimensional simplified truss topology dictionary corresponding to the monitored tower type. This dictionary can be directly imported from design drawings or as-built drawings, or generated by extracting axes after reconstructing point clouds from UAV oblique photogrammetry. Its simplification principle is to fully retain load-bearing components such as main members, diagonal members, transverse diaphragms, and crossbeams. Auxiliary members, ladders, and hardware with a length less than 0.5 m or a cross-sectional projection width less than 50 mm are merged or omitted. The total number of nodes after simplification is [not specified]. The shading area contributed by the omitted components is generally no more than 5%, and can be compensated by multiplying the shading rate calculation result by an empirical correction factor of 0.95 to 1.00. Secondly, the precise longitude of the tower location is crucial. ,latitude and the azimuth angle of the tower installation ,in Defined as the tower body coordinate system The azimuth angle measured clockwise from true north, in the tower coordinate system. Axis azimuth angle is , The axis is vertically upward. Calibration is performed by as-built surveying or a dual-antenna directional module, with a calibration error of no more than 1°. The transformation from tower coordinates to local horizontal coordinates is as follows: , , The 1° deviation of this angle causes the overall occlusion determination to rotate by 1°, thus it is a necessary technical feature for the implementation of this scheme. Thirdly, there are the epoch parameters required for astronomical calculations. The edge resolution layer 200 contains, in sequence, a signal purification unit 210, a modal identification unit 220, a temperature field reconstruction unit 230, and a feature uploading unit 240. The cloud evaluation layer 300 includes a decoupling network unit 310, a health measurement unit 320, and an early warning linkage unit 330.
[0029] like Figure 2 As shown, the power transmission tower operation health status assessment method based on the above system is executed progressively in four mutually coupled calculation stages within each acquisition cycle. In the acquisition stage, the vibration acquisition unit 110 outputs the multi-channel time series matrix of the acceleration response at each measuring point. ,in The total number of channels. Number of acceleration measurement points, number of sampling points per channel per scan The environmental sensing unit 120 synchronously outputs the reference temperature within this time window. Wind speed ,wind direction Total solar irradiance on horizontal plane The aforementioned heterogeneous data is aggregated by the data aggregation and transmission unit 130 and then enters the edge resolution layer 200.
[0030] Signal purification unit 210 addresses the non-stationarity and broadband noise of vibration signals under random wind load excitation by employing multivariate variational mode decomposition to transform signal separation into a constrained variational problem. Unlike independent channel-by-channel decomposition, the multivariate form forces all channels to share the same set of center frequencies, thus preserving the amplitude ratio and phase relationship between channels. This is crucial for subsequent mode shape identification. The constrained variational problem is expressed as follows:
[0031] In the formula For the first The first channel One modal component, For its center frequency, Let be the Dirac impulse function. The imaginary unit is used. A quadratic penalty factor is introduced. With Lagrange multipliers Transform it into an unconstrained augmented Lagrangian function.
[0032] The solution is obtained iteratively using the alternating direction multiplier method, and its frequency domain update formula is: The center frequency is updated jointly based on the first moment of the power spectrum of each channel. , according to Ascend, step size Under strong noise conditions, the value is taken as 0, typically between 0.01 and 0.10. The convergence criterion for iteration is that the sum of the relative changes of two adjacent components is less than 0.01. Therefore, the penalty factor The position of action is the bandwidth term coefficient of the augmented Lagrange function mentioned above, and its value directly determines the balance between the bandwidth of each component and the reconstruction fidelity.
[0033] Signal purification unit 210 decomposes the number of modes With penalty factor Particle swarm optimization is used for joint optimization. It should be noted that if the fitness is the sum of the envelope entropies of each component, the number of terms in the objective function increases with... If the entropy increases and all terms are non-negative, the optimization result will inevitably converge to the lower bound of the search interval, rendering the search meaningless. Therefore, in this embodiment, the minimum value of the envelope entropy of each component is taken as the fitness principal term, and a penalty is imposed on over-decomposition.
[0034] In the formula In response to the reference channel with the highest energy, For indicator functions, Take 10 times the typical magnitude of the main term; envelope entropy is calculated as follows: calculate, , This is the envelope sequence obtained by Hilbert transforming this component. During optimization, the particle swarm size is set to 20, the maximum number of iterations to 50, the inertia weight to 0.8, the learning factor to 2.0, and the search interval to be... , Considering that a single optimization requires approximately 1000 decompositions, its computational load is far greater than that of conventional analysis. The signal purification unit 210 only triggers optimization when the device is initially deployed, during seasonal changes, or when the drift of environmental statistics relative to the previous optimization exceeds a set threshold. For other acquisition cycles, the fixed optimal parameters are directly used. The optimization process is scheduled to be performed offline during idle acquisition periods. After decomposition, components are retained and recombined into purified signals based on the triple criteria of center frequency falling between 0.3 and 25 Hz, correlation coefficient with the original signal not less than 0.1, and energy proportion not less than 1%. The remaining components are discarded as noise.
[0035] Modality identification unit 220 receives the multi-channel purified signal output by signal purification unit 210 and uses covariance-driven random subspace identification. It first... Estimate the output covariance of each order and stack them into a block Toplitz matrix. Perform singular value decomposition on it. After truncating the principal subspace based on the singular value fall position, the observable matrix is obtained. The discrete state matrix is obtained by utilizing its shift invariance. .right Discrete poles are obtained by performing eigenvalue decomposition. ,according to Mapped to a continuous domain, and then from and The natural frequency and damping ratio are obtained, and the mode shape is given by the product of the output matrix and the eigenvectors. The above calculations are repeated for different system orders to generate steady-state diagrams. A triple screening condition is set: relative frequency deviation no greater than 1%, relative damping ratio deviation no greater than 5%, and modal confidence criterion value no less than 0.95. This eliminates spurious mathematical poles introduced by noise, and finally outputs the forward frequency. First-order stable natural frequency vector With mode matrix Its rows correspond to the three orthogonal directional degrees of freedom of each measuring point, and its columns correspond to the modal orders. Generally, 3 to 5 are chosen. To ensure comparability across time intervals, each mode shape is normalized to a unit mode, i.e. Furthermore, the horizontal component along the line at the end of the ground wire crossarm at the top of the tower is used as a reference to unify the sign, so that the component is always positive, thereby eliminating the inherent sign uncertainty of the mode shape. and The mixed-mode feature vector at that moment is formed by unfolding the vectors column-wise and concatenating them vertically.
[0036] Its dimension is typically 39–123. It should be noted that this vector is simultaneously entangled with information reflecting the true health state of the structure and the nonlinear stiffness drift caused by non-uniform thermal stress. If it is directly compared with the health benchmark, the result will inevitably be dominated by the latter.
[0037] The temperature field reconstruction unit 230 is used to quantize the aforementioned thermally induced drift into a physical field that can be monitored by the subsequent network. Its calculation begins with a high-precision solution of the sun's position, derived from the Julian day at the time of data acquisition. Julian Day on January 0th of that year according to To obtain the date sequence within the year, from standard time... according to Converted to true solar time, of which The longitude of the central meridian of the time zone, and the time difference. , , and Hourly In minutes, this formula represents the step where longitude a priori factors are involved in the solution. Then, according to... The solar declination angle is obtained, according to... Find the hour angle, and from...
[0038] Solve for the solar altitude angle With azimuth This is an agreement. With true north as zero degrees and clockwise as positive, when That is, take during the morning period ,when That is, take it in the afternoon. The inverse cosine independent variable is truncated to the interval before calculation. To avoid numerical overflow at critical moments. In the local horizontal coordinate system with east, north, and vertically upward as coordinate axes, the unit vector in the direction of the sun is... The unit vector of the direction of light propagation is taken in the opposite direction, that is This definition is strictly consistent with the above expression for azimuth angles calculated from true north, avoiding the interchange of the sunlit and shaded sides caused by the mixed use of north and south references.
[0039] After obtaining the instantaneous light direction, the temperature field reconstruction unit 230 reconstructs the node coordinates in the three-dimensional simplified truss topology dictionary according to the installation azimuth angle. Transform to the local horizontal coordinate system, and use Perform structural self-occlusion determination for the projection reference axis. Specifically, along each member... Approximately 20 surface sampling points are equidistantly arranged along the axial direction, and from each sampling point along... A ray is projected in the opposite direction, and hierarchical cross-detection using axis-aligned bounding boxes quickly eliminates most irrelevant members. Only candidate members whose bounding boxes intersect are precisely intersected. If the ray intersects any other member before exiting the tower's envelope, the sampling point is determined to be within the tower's self-shadow. The percentage of unobstructed sampling points on a single member is counted to obtain the direct light exposure coefficient for that member. A value of 0 indicates complete self-shadowing with no direct radiation, while a value of 1 indicates complete sun exposure. When the solar altitude angle... During the nighttime period, the vertical component of the light vector points towards the zenith, rendering ray tracing meaningless. In this case, all direct light terms are set to zero, and the occlusion determination is skipped. The solution is then obtained solely based on the nighttime radiation model described later. Since this determination is entirely based on geometric topology and celestial motion, its spatial resolution depends only on the discretization accuracy of the topological dictionary and is independent of the number of physical measurement points. This is the key to obtaining the temperature gradient across the entire tower under sparse configuration.
[0040] The temperature field reconstruction unit 230 then converts the measured total irradiance of the horizontal plane into the absorbed heat flux of each rod surface. Considering that cosine projection correction is physically only applicable to the direct component, it is first calculated according to the clear sky index. Perform direct-dispersion separation, where W / m 2 The solar constant, Earth-Sun distance correction factor; scattering ratio Values are determined based on the empirical relationships of segments. hour , hour , hour Normal direct irradiance is taken as For the axial unit vector, The rod has a direct projection factor of . Then, the shape conversion factor for the ratio of projected area to total heat exchange surface area is also taken into account. For circular cross-section members, the heat flux is taken as 0.30–0.35; for angle steel members, the heat flux is taken as 0.35–0.50 based on the normal projection of the two legs. Therefore, the heat flux absorbed per unit surface area is:
[0041] In the formula The solar radiation absorptivity of the galvanized steel surface is taken as 0.5 to 0.6. and These are the perspective factors for the sky and the ground, respectively, with a value of 0.5 for rods far from the ground. Let the ground albedo be 0.2. The above formula will then be used to calculate the exposure coefficient. Applying the direct sunlight term only, the scattering and ground reflection terms are still retained for rods in self-shade, thus avoiding the problem of the temperature on the shaded side being systematically underestimated, which in turn exaggerates the temperature difference between the sunlit and shaded sides.
[0042] Based on this, the temperature field reconstruction unit 230 establishes a steady-state thermal equilibrium equation using the rods as the calculation units, thereby eliminating the non-uniqueness of parameters caused by the different orientations and inclination angles of multiple rods at the nodes:
[0043] All temperatures are expressed in thermodynamic form. The value was calculated by adding 273.15 to the measured Celsius temperature. The thermal emissivity of galvanized steel is taken as 0.2 to 0.3. The Stefan-Boltzmann constant is used; the effective sky temperature is taken as... Sky emission rate For clear skies, use 0.70–0.80; for cloudy skies, use 0.90–1.00. The convective heat transfer coefficient is determined based on the sweep component, i.e.
[0044] In the formula Let be the angle between the wind vector and the axis of the member. Extract the effective flow component perpendicular to the axis of the rod, so that the heat transfer is strongest when swept horizontally and weakest when swept vertically, which is consistent with the heat transfer law of flow around a column. In m / s, For the natural convection term, take 3–6 W / (m³). 2 ·K), Both can be calibrated by regression from wind tunnel tests or measured temperature rise data during commissioning, or calculated using the classical forced convection correlation based on the characteristic dimensions of the angle steel; windward and leeward correction coefficients. The calculation for tower wake shielding reuses the same bounding box intersection module, only changing the projection reference axis from... The wind direction is replaced with the opposite vector. Members within the tower's wind shadow are assigned a value of 0.6–0.8, while members directly exposed to the wind are assigned 1.0. Due to the fourth-power nature of the radiation term, this equation exhibits strong nonlinearity. A Newton-Raphson iteration is used to solve it element by element. The iteration format is as follows: ,in Initial value is taken The convergence criterion is that the absolute value of two consecutive temperature increments is less than 0.01K, and convergence is usually achieved in 3 to 5 iterations. The above element-by-element solution ignores axial heat conduction between adjacent components. Its applicability is when the length-to-diameter ratio of the member is greater than 20 and the temperature of the member's mid-section is of interest. In this case, the influence of end-point heat conduction is limited to a region several times the cross-sectional size near the node. When the tower type uses large-section steel pipes or when the local temperature of the node area is of interest, it is advisable to use a transient solution that considers the coupling of heat capacity and thermal conduction, as described in the modified implementation method below. The member temperature is weighted by surface area and reduced to the node, i.e. ,in To converge at the node A collection of rods, For the surface area of the rod, The weights assigned to the members at both ends are both set to 0.5; the node exposure coefficient... According to the same weighting rule synthesis.
[0045] Inevitably, residuals exist between the model's calculated values and the actual values due to deviations in material parameters and local micro-meteorological disturbances. The temperature field reconstruction unit 230 compensates for this using spatial interpolation. It's important to clarify that the weighting equations for Kriging interpolation can be established as long as the semivariogram model is known; the number of sample points is not an obstacle. What truly cannot be accomplished with just two long-term measurement points is the estimation of the three parameters of the semivariogram. Therefore, in this embodiment, 10-20 temporary temperature measurement points are installed during the calibration period after the tower is put into operation, covering the sunny and shady sides and different heights. After continuous data collection for 1-2 weeks, the residual samples are used to fit the nugget value, sill value, and range of the spherical semivariogram. The parameters are then solidified into the edge analysis layer 200, and the temporary measurement points are removed. During long-term operation, only 2-4 benchmark measurement points are retained on the shady side of the tower base and the sunny side of the tower top, using their residuals... Together with the fixed semivariogram, a set of Kriging equations is established to obtain the residual estimates for each node. The final node temperature is taken as If calibration measurements cannot be conducted due to engineering limitations, an interpolation method independent of spatial structure fitting will be used, i.e., based on the exposure coefficient of each node. This step can be fully implemented by performing a bilinear weighted allocation of the residuals between the two benchmark measurement points based on relative height, or by using inverse distance weighted interpolation. This outputs a high-dimensional non-uniform temperature distribution feature vector at that moment. .
[0046] Feature uploading unit 240 uploads mixed modal feature vectors within the same time window. eigenvectors of non-uniform temperature distribution The data is encapsulated into a single feature message and uploaded to the cloud evaluation layer 300. This processing reduces the amount of data transmitted over the public network in a single acquisition cycle from megabytes to kilobytes. In terms of real-time performance, with the aforementioned hardware configuration, multivariate variational mode decomposition takes approximately 20 seconds, random subspace identification approximately 10 seconds, ray tracing and occlusion determination approximately 2 seconds, and unit-by-unit Newton iteration and interpolation compensation total no more than 1 second, for a total process time of approximately 40 seconds. This provides ample margin compared to a standard 60-minute acquisition cycle and a 15-minute encrypted acquisition cycle. Parameter optimization, limited to low-frequency offline triggering, is not included in the aforementioned real-time budget.
[0047] The decoupled network unit 310 in the cloud evaluation layer 300 carries a physically constrained variable decoupled autoencoder network. The encoder of this network consists of a shared backbone and two unconnected branches. The backbone compresses the input mixed-modality feature vector sequentially through 256-dimensional and 128-dimensional fully connected layers. Then, two independent branches, each containing a 64-dimensional hidden layer, output the means of environmental and health latent variables, respectively. With log variance The two dimensions All are set to 8; latent variables are set to... Reparameterized sampling, where To follow a standard normal distribution random vectors, This indicates element-wise multiplication; the decoder uses concatenated vectors. The input is used to reconstruct the original features through 128-dimensional and 256-dimensional fully connected layers. The branching encoder achieves structural isolation between the two subspaces in terms of network connectivity, and the activation functions of each layer are modified linear units. The environmental latent space component... Its physical mission is limited to characterizing the structural nonlinear stiffness variation induced by changes in spatially nonuniform temperature gradient distribution, and the healthy latent space component. This is used to characterize the pure characteristics of the tower body material mechanical properties and node connection state after eliminating all thermal stress interference. The actual allocation of this physical property is accomplished by two constraint terms in the following loss function.
[0048] The decoupling network unit 310 employs a composite loss function that includes four constraints:
[0049] The reconstruction loss is taken as the sum of squares of elements. , The batch size is used to ensure that the latent variables fully represent the original modal information; divergence penalty The latent variables are constrained to follow a standard Gaussian prior to ensure the smoothness and continuity of the latent manifold and the feasibility of subsequent statistical measures; cross-domain physical strong correlation constraints are also included. This is the core of the solution, which is the output of the temperature field reconstruction unit 230. As a strong mentor signal, the principal component projection matrix is first fitted and solidified using the training set to reduce its dimensionality, retaining only the top components whose cumulative variance contribution rate reaches 95%. Principal components, Typically, the value is 3 to 5 and no greater than 5. To avoid introducing reconstruction noise from higher-order principal components into the environmental latent space, and thus according to
[0050] Calculate, where For the Pearson correlation coefficient calculated on the batch dimension, the batch size should be no less than 64 to ensure the stability of the estimate. The principal components of the temperature field after dimensionality reduction; this term applies to The behavior that deviates from the true physical temperature field distribution is penalized, thereby driving the network to actively converge and incorporate the nonlinear distortion information related to temperature fluctuations. The remaining dimensions, not involved in the relevant constraints, are only constrained by the orthogonality and divergence terms. Spatial orthogonality and independence constraints are taken as... Using the Frobenius norm of the cross-covariance of the batch-centered latent variable matrix as a penalty, the two subspaces are forced to be statistically orthogonal, preventing residual leakage of temperature information into the health dimension. The weighting coefficients are... In the early stages of training, the model is linearly annealed from 0 to 1. and The value is tuned within the range of 0.1 to 10, and in this embodiment, it is set to 1.0. The training uses an adaptive moment estimation optimizer with a learning rate of 0.001, a batch size of 128, a maximum number of training rounds of 500, and early stopping.
[0051] The training of the decoupled network unit 310 was completed offline. The training samples were taken from long-term data during the initial structural health confirmation period of the tower's commissioning, with a time span of no less than 12 months to cover the complete four seasons and diurnal temperature cycle. Each sample was generated from data collected at the same time. and Pairing is used to construct the vector. During online inference, the mixed-modal feature vector at the current time step only needs to be fed into the encoder for one forward propagation, and the mean of the healthy branch is taken as the output to obtain the pure healthy vector at that time step. .at this time Although not used as direct input to the network, its physical constraints are already embedded in the network weights during the training phase. However, the effectiveness of this embedded relationship depends on the online data's temperature and modality mapping not significantly deviating from the training distribution. Therefore, the decoupled network unit 310 simultaneously performs out-of-distribution detection, i.e., online calculation of the reconstruction error. as well as If either the Mahalanobis distance relative to the training set distribution exceeds the 99.9th percentile of the training set, that moment is marked as an out-of-distribution sample, and its evaluation result will not participate in the early warning decision but will be transferred to the review queue. When the proportion of out-of-distribution samples exceeds 5% within 7 days, it is determined that the operating conditions have exceeded the model coverage, triggering incremental retraining with newly added health data as samples and simultaneously issuing new network weights, so that the decoupling capability can be maintained when micro-meteorological conditions or structural states slowly shift.
[0052] The health measurement unit 320 performs unsupervised state evaluation within a pure latent space. It first selects a set of long-period mappings representing generally accepted healthy states, with a sample size of at least 1000, and calculates their mean vector. With covariance matrix Considering that the covariance matrix may be ill-conditioned when the sliding window sample size is small or the latent variable dimension degenerates, ridge regularization is applied before inversion. Regularity coefficient Pick And continuously monitor its condition number; once it exceeds This triggers an alarm and suspends baseline updates. During real-time monitoring, the Mahalanobis distance of the current health vector relative to the baseline distribution is calculated using the following formula and defined as the temperature-independent damage index. :
[0053] Compared to Euclidean distance, Mahalanobis distance takes into account the correlation and variance scale differences between the dimensions of the latent space, and can provide higher resolution sensitivity for multidimensional feature migrations that only cause a slight decrease in stiffness, such as buckling of a single bar, while eliminating the influence of dimensions.
[0054] The early warning linkage unit 330 does not use a universally applicable empirical threshold across towers, but instead adaptively calibrates two levels of thresholds based on the baseline data of this tower. It uses kernel density estimation to perform probability density fitting on the exponential samples accumulated during the baseline phase, according to... calculate, Take the Gaussian kernel, The baseline sample size is no less than 1000, and the bandwidth is calculated using an empirical formula. Sure, The baseline sample standard deviation; Level 1 warning threshold. Take the quantile value corresponding to the cumulative probability of 99.7% as the level 2 alarm threshold. Taking the 99.99th percentile value, when the baseline sample size is insufficient to support a direct estimate of this quantile, for values exceeding [a certain threshold]... The tail is extrapolated from the generalized Pareto distribution to obtain the extreme values, thus enabling both threshold levels to be adaptively determined by the data rather than manually specified multiples. The decision logic covers the entire value range: when When the system is deemed to be in healthy operation, the data is archived for long-term plastic rheological curve plotting and baseline updates, while the first persistence counter is also activated. With the second persistence counter All cleared to zero; when At that time, the first persistent counter The accumulation reaches 2, triggering a Level 1 warning. This situation typically corresponds to a discrete loss of local bolt tightening force or slight slippage of the foundation. In this case, the sampling interval of the vibration acquisition unit 110 is shortened from 60 minutes to 15 minutes, the single sampling duration is shortened from 10 minutes to 5 minutes, and a key verification reminder is pushed to the operation and maintenance side. At that time, the first persistent counter With the second persistence counter Simultaneously, the data is accumulated, and from the first instance of exceeding the limit, encrypted data collection is immediately executed according to the first-level warning, with a second persistent counter under encrypted data collection. After accumulating to 6 (approximately 1.5 hours), the alarm level is upgraded to Level 2. The counter reset rule is: when any window falls back to... The following is the first persistence counter With the second persistence counter All data are reset to zero and the normal collection interval is restored; if the data falls back to zero... and Between, the second persistence counter Stop accumulating but do not immediately clear to zero, only after three consecutive windows have not been crossed. Reset to zero, first persistent counter The data collection continues to accumulate and is maintained using encrypted acquisition. The aforementioned accumulation and reset rules avoid undefined states between the first window and the confirmation window, suppress single-point jumps caused by transient events such as extreme gusts or lightning electromagnetic pulses, and ensure that the confirmation delay for secondary alarms is within 2 hours thanks to encrypted acquisition. Considering that the 99.7th percentile in a healthy state corresponds to approximately 1 / 333 of the single-window limit exceedance probability, by adopting the criterion of triggering a primary warning only after two consecutive windows, the equivalent false alarm interval can be extended to the annual level. Simultaneously, encrypted acquisition increases the duty cycle from approximately 1 / 6 to 1 / 3. To match the energy budget of solar power and batteries, the duration of the encrypted mode does not exceed 24 hours, and when the battery's state of charge falls below 30%, it automatically reverts to a 30-minute interval and reports an energy-limited flag. Level 2 alarms are pushed to the power grid geographic information dispatch platform and the UAV mission system in a standardized message. The message content includes tower number, latitude and longitude coordinates, timestamp, index value, over-limit level and recent index sequence summary. It is transmitted asynchronously through the message queue interface and requires acknowledgment.
[0055] Considering the inevitable slow aging and plastic flow of tower materials during long-term service, the health measurement unit 320 measures the tower's health using a 90-day sliding window, once every 7 days. and Periodic updates are performed, and only samples deemed healthy are included in the update. The mean shift for each update is limited to 0.1 times the baseline standard deviation to prevent true progressive damage from being absorbed by the synchronous shift of the baseline and thus missed. After the baseline update, the early warning linkage unit 330 synchronously recalculates and updates the kernel density estimate using the index samples within the same sliding window. and If the recalculated threshold changes by more than 20% compared to the previous period, the update will be paused and an alarm will be issued, prompting a review of data quality or model validity. The resulting long-term index sequence constitutes the health degradation trajectory of the tower, providing data support for the transition from planned maintenance to condition-based maintenance.
[0056] The technical solution of this invention allows for various equivalent modifications during implementation. The triaxial microelectromechanical accelerometer in the vibration acquisition unit 110 can be replaced with a piezoelectric accelerometer or a force-balanced accelerometer; the optimized arrangement of measurement points can also be replaced with the modal kinetic energy method or the information entropy method based on Bayesian model updates. The networking method of the data aggregation and transmission unit 130 can be either a time-slot synchronized low-power local area network or a narrowband Internet of Things or mobile cellular network for backhaul. The particle swarm optimization in the signal purification unit 210 can be replaced by other swarm intelligence optimization methods such as genetic algorithms or gray wolf optimization; the fitness function can also be replaced with sample entropy or kurtosis indices. The covariance-driven random subspace identification in the modal identification unit 220 can be replaced with data-driven random subspace identification, where a block Hankel matrix is constructed and projected through orthogonal triangular decomposition before singular value decomposition; alternatively, it can be replaced with a frequency domain decomposition method or a feature system implementation algorithm. The axis-aligned bounding box hierarchical detection in the temperature field reconstruction unit 230 can be replaced with a directed bounding box, hierarchical bounding box tree, or spatial binary tree acceleration structure. Its residual correction stage can be replaced with inverse distance weighted interpolation or radial basis function interpolation. When the cross-section of the tower components is large, the thermal inertia cannot be ignored, or the heat conduction along the axis between components needs to be considered, the steady-state thermal balance equation can be extended to a first-order hysteresis form that includes the heat capacity term, or replaced by transient heat conduction finite element solution. In this case, the temperature state at several previous moments needs to be used as the initial value, at the cost of a corresponding increase in the computing power required on the edge side. The encoder and decoder of the decoupling network unit 310 can be changed from a fully connected network to a one-dimensional convolutional network or a long short-term memory network to further utilize the temporal correlation of features. The Mahalanobis distance in the health measurement unit 320 can be replaced with the squared prediction error statistic of kernel principal component analysis. The kernel density estimation in the early warning linkage unit 330 can be replaced with Gaussian mixture model fitting. Furthermore, when computing power permits, the online inference portion of the decoupled network unit 310 can also be deployed on the edge resolution layer 200. In this case, the edge resolution layer 200 needs to be equipped with corresponding neural network inference acceleration components, while the cloud evaluation layer 300 only undertakes offline training of the model, version distribution, and horizontal comparison of multi-tower data. In cases where higher accuracy of the temperature field is required, a small number of fiber optic grating temperature measurement points can be added on the sunny and shady sides to enrich the residual samples. The above adjustments do not change the technical essence of this invention, which achieves orthogonal decoupling of features by reconstructing the latent space constrained by the temperature field.
[0057] The above description is only a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural transformations made based on the content of the present invention specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.
Claims
1. A method for assessing the operational health status of power transmission towers, characterized in that, Includes the following steps: The system acquires multi-channel vibration acceleration response signals and sparse environment sensing parameters of the monitored transmission tower, including environmental reference temperature, wind speed, wind direction, and total solar irradiance on the horizontal plane. The multi-channel vibration acceleration response signal is subjected to multivariate variational mode decomposition to purify the vibration signal, and the natural frequencies and mode shapes are extracted by covariance-driven random subspace identification. After normalization, a hybrid mode feature vector is constructed. The solar spatial position vector is calculated based on the latitude and longitude coordinates of the location of the transmission tower and the current sampling time. The pre-set three-dimensional simplified truss topology dictionary of the transmission tower is transformed to the local horizontal coordinate system according to the azimuth angle of the tower installation and structural self-shading ray tracing is performed to determine the direct sunlight exposure coefficient of each member. By combining the sparse environment sensing parameters, the total solar irradiance on the horizontal plane is separated into direct and diffuse values. A rod-level steady-state thermal balance equation is established and solved nonlinearly. Then, after spatial interpolation residual compensation, the spatial non-uniform temperature distribution feature vector at each node of the entire tower is output. The hybrid modal feature vector is input into a pre-trained physical constraint variational decoupled autoencoder network. The hybrid modal feature vector is decoupled and mapped to mutually orthogonal environmental latent spaces and health latent spaces through a branch encoder, and a pure health latent variable representing the true state of the structure is output. In the training phase, the physical constraint variational decoupled autoencoder network uses the spatial non-uniform temperature distribution feature vector as a physical mentor signal, and uses cross-domain physical correlation constraints and spatial orthogonal independent constraints to drive the environmental latent space to absorb the modal nonlinear variation caused by temperature fluctuations. The Mahalanobis distance of the current pure health latent variable relative to the preset health baseline distribution is calculated as the temperature-independent damage index. The temperature-independent damage index is compared with the first-level warning threshold and the second-level alarm threshold adaptively calibrated based on the health baseline stage sample distribution. Combined with the continuous multi-window persistent judgment logic, the operation health status assessment and hierarchical early warning linkage of the transmission tower are realized.
2. The method for assessing the operational health status of transmission towers according to claim 1, characterized in that: The multi-channel vibration acceleration response signal is acquired by multiple triaxial accelerometers arranged on the power transmission tower. The measurement point position of each triaxial accelerometer is determined based on the finite element model of the power transmission tower system, and the effective independence method is combined with the information entropy corresponding to the posterior covariance of the modal parameters as the termination criterion for optimization. When performing multivariate variational mode decomposition on the multi-channel vibration acceleration response signal, each channel shares the same set of center frequencies. The number of decomposed modes and the secondary penalty factor of the multivariate variational mode decomposition are jointly optimized using a particle swarm optimization algorithm. The fitness function takes the minimum envelope entropy of each modal component of the reference channel as the main term and applies a penalty term to the over-decomposition case where the ratio of adjacent center frequencies is less than a set threshold.
3. The method for assessing the operational health status of transmission towers according to claim 1, characterized in that: When using covariance-driven random subspace identification to extract natural frequencies and mode shapes, a block Toplitz matrix is constructed based on the output covariance of each order. The principal subspace is truncated through singular value decomposition, and the discrete state matrix is solved based on shift invariance. Then, the natural frequencies, damping ratios, and mode shapes are calculated from the eigenvalues and eigenvectors of the discrete state matrix. Steady-state diagrams are constructed at different system orders, and spurious poles are eliminated based on a triple screening condition of relative frequency deviation, relative damping ratio deviation, and modal confidence criterion. The extracted mode shapes are normalized according to the unit modulus, and the horizontal component along the line at the end of the ground wire crossarm at the top of the tower is used as a reference to unify the symbol polarity. Then, they are expanded and connected in series with the natural frequencies of each order to form the hybrid mode feature vector.
4. The method for assessing the operational health status of transmission towers according to claim 1, characterized in that: When calculating the solar spatial position vector, the time difference and true solar time are calculated based on the annual day sequence at the sampling time, the longitude of the central meridian of the time zone, and the local longitude. Then, the solar declination angle and hour angle are calculated to obtain the solar altitude angle and solar azimuth angle in the local horizontal coordinate system. The solar azimuth angle is calculated with due north as zero degrees and clockwise as the positive direction. The unit vector of the light propagation direction is determined based on the solar altitude angle and solar azimuth angle. When the solar altitude angle is not greater than zero, it is determined that it is nighttime, the direct radiation term is set to zero, and self-occlusion ray tracing is skipped. When performing structural self-shading ray tracing, the node coordinates in the three-dimensional simplified truss topology dictionary are converted to the local horizontal coordinate system based on the tower installation azimuth angle. Several surface sampling points are set along the axis of each member. Rays are projected from each surface sampling point along the direction of the reflected light. Non-intersecting members are eliminated by axis-aligned bounding box hierarchical cross detection. The candidate members are then precisely cross-checked. If the ray intersects with other members, the surface sampling point is determined to be in self-shade. The percentage of unshaded surface sampling points on a single member is counted as the direct light exposure coefficient of that member.
5. The method for assessing the operational health status of transmission towers according to claim 1, characterized in that: When performing direct and diffuse separation of the total solar irradiance on the horizontal plane and establishing the steady-state thermal balance equation for the rod level, the total solar irradiance on the horizontal plane is separated into normal direct irradiance, sky diffuse irradiance, and ground reflected irradiance according to the clear sky index. The direct light exposure coefficient is applied only to the direct irradiance term to calculate the heat flux absorbed by the rod. A nonlinear thermal balance equation including convective heat transfer term and radiative heat dissipation term is established, and each temperature quantity is substituted with thermodynamic temperature. The convective heat transfer coefficient is extracted based on the angle between the wind vector and the rod axis, and the effective inflow component perpendicular to the rod axis is extracted. The leeward correction coefficient obtained by cross-intersection based on the wind direction of the bounding box is also included. The surface temperature of each rod is solved element by element using the Newton-Raphson iterative method. After the surface temperature of each member is weighted by area and allocated to each node, during the long-term operation phase, the measured residuals of the reference temperature measurement points arranged on the shaded side of the tower foot and the sun-facing side of the tower top are used to establish a Kriging interpolation equation system in combination with the semi-variogram parameters pre-fitted during the calibration period. The temperature residuals of each node in the entire tower are estimated and the calculated temperature is compensated to generate the spatial non-uniform temperature distribution feature vector.
6. The method for assessing the operational health status of transmission towers according to claim 1, characterized in that: The physical constraint variational decoupled autoencoder network includes a shared backbone coding layer, independently configured environment latent variable branches and health latent variable branches, and a reconstruction decoder; The composite loss function used in the training phase includes: The reconstruction loss term measures the squared error between the input mixed-modality feature vector and the decoder reconstruction output. The divergence penalty term is used to constrain the prior that the environmental and health latent variables follow a standard normal distribution. Cross-domain physical related constraints are used to extract principal components whose cumulative variance contribution rate reaches a set threshold from the feature vector of the non-uniform temperature distribution in space, and to apply reverse penalties to the Pearson correlation coefficients of each dimension component of the environmental latent space and the corresponding temperature principal components in the batch dimension. The spatial orthogonality and independence constraint term is used to force the environmental latent space and the health latent space to be statistically orthogonal, using the norm of the cross-covariance of the in-batch centered environmental latent variable matrix and the health latent variable matrix as a penalty.
7. The method for assessing the operational health status of transmission towers according to claim 1, characterized in that, The method also includes online out-of-distribution detection and incremental model update steps: During online inference, the reconstruction error of the mixed modality feature vector at the current moment and the Mahalanobis distance of the environmental latent variables relative to the training set distribution are calculated. If any indicator exceeds the preset high quantile threshold of the training set, the current moment is marked as an out-of-distribution sample and its evaluation result is not included in the early warning decision. When the proportion of out-of-distribution samples within a set time window exceeds a preset threshold, incremental retraining of the model using newly added healthy time period data as samples is triggered, and the weights of the physical constraint variation decoupled autoencoder network are updated simultaneously.
8. The method for assessing the operational health status of transmission towers according to claim 1, characterized in that: When calculating the temperature-independent damage index, the mean vector of the healthy baseline and the baseline covariance matrix are calculated using long-term pure healthy latent variable samples accumulated during the healthy operation phase. When inverting the baseline covariance matrix, a trace-based ridge regularization term is applied, and its condition number is continuously monitored. During long-term operation, a sliding time window is used to periodically update the baseline mean vector and the baseline covariance matrix with samples determined to be in a healthy state, and a limiting constraint based on the baseline standard deviation is applied to the mean drift of a single update.
9. The method for assessing the operational health status of transmission towers according to claim 1, characterized in that: The adaptive calibration method for the first-level warning threshold and the second-level alarm threshold is as follows: based on the temperature-independent damage index samples accumulated during the healthy baseline stage, the probability density distribution is fitted by kernel density estimation and combined with tail extremum extrapolation, and the first-level warning threshold and the second-level alarm threshold are adaptively calibrated with the corresponding high cumulative probability quantile values respectively. The continuous multi-window persistence determination logic includes: When the temperature-independent damage index is less than or equal to the first-level warning threshold, it is determined to be in healthy operation, and the first and second persistence counters are cleared to zero. When the temperature-independent damage index is greater than the first-level warning threshold and less than or equal to the second-level alarm threshold, the first continuous counter increments. When the increment reaches a preset first count value, a first-level warning is triggered, and the acquisition cycle of the vibration acceleration response signal is shortened for encrypted acquisition. When the temperature-independent damage index is greater than the secondary alarm threshold, the first continuous counter and the second continuous counter are synchronously incremented. Encrypted data acquisition is immediately executed from the first time the limit is exceeded. When the second continuous counter continuously increments to a preset second count value under encrypted data acquisition, a secondary alarm is triggered. During the encrypted data acquisition process, if the temperature-independent damage index falls below the first-level warning threshold, the first and second persistent counters are reset to zero and the normal acquisition cycle is restored; if it falls between the first-level warning threshold and the second-level alarm threshold, the second persistent counter stops accumulating and is reset to zero after it has not exceeded the second-level alarm threshold within a set number of consecutive windows.
10. A system for assessing the operational health status of power transmission towers, characterized in that, The method for assessing the operational health status of transmission towers according to any one of claims 1 to 9 comprises a perception fusion layer, an edge resolution layer, and a cloud assessment layer, wherein: The sensing fusion layer is arranged on the monitored transmission tower body and includes a vibration acquisition unit, an environmental sensing unit, and a data aggregation and transmission unit. It is used to acquire multi-channel vibration acceleration response signals and sparse environmental sensing parameters, and then aggregate and transmit them after synchronization in a unified time slot. The edge resolution layer, installed at the tower base, includes a signal purification unit, a modal identification unit, a temperature field reconstruction unit, and a feature uploading unit. The signal purification unit purifies vibration signals through multivariate variational mode decomposition. The modal identification unit extracts natural frequencies and mode shapes using covariance-driven random subspace identification and constructs a hybrid modal feature vector. The temperature field reconstruction unit calculates the instantaneous solar spatial position vector, performs structural self-shading ray tracing based on a pre-set three-dimensional simplified truss topology dictionary for the transmission tower to determine the direct sunlight exposure coefficient of each member, and solves the member-level steady-state thermal balance equation and spatial interpolation residual compensation using the sparse environment perception parameters to reconstruct the spatial non-uniform temperature distribution feature vector at each node of the entire tower. The feature uploading unit uploads the hybrid modal feature vector and the spatial non-uniform temperature distribution feature vector to the cloud evaluation layer. The cloud-based assessment layer, deployed on a server or data center, includes a decoupled network unit, a health measurement unit, and an early warning linkage unit. The decoupled network unit is equipped with a physically constrained variable decoupled autoencoder network, used to train cross-domain correlation constraints and orthogonal independent constraints using the spatially non-uniform temperature distribution feature vector after principal component dimensionality reduction as the mentor signal. During online inference, it decouples the mixed-modality feature vector to output a pure health latent variable. The health measurement unit calculates the Mahalanobis distance of the pure health latent variable relative to the health baseline distribution to obtain a temperature-independent damage index. The early warning linkage unit determines the temperature-independent damage index using adaptively calibrated multi-level thresholds and continuous multi-window persistence logic, outputting tiered early warning and scheduling linkage instructions.
